1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
25 #include <linux/page_counter.h>
26 #include <linux/memcontrol.h>
27 #include <linux/cgroup.h>
28 #include <linux/pagewalk.h>
29 #include <linux/sched/mm.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/hugetlb.h>
32 #include <linux/pagemap.h>
33 #include <linux/vm_event_item.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
59 #include <linux/tracehook.h>
60 #include <linux/psi.h>
61 #include <linux/seq_buf.h>
67 #include <linux/uaccess.h>
69 #include <trace/events/vmscan.h>
71 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
72 EXPORT_SYMBOL(memory_cgrp_subsys);
74 struct mem_cgroup *root_mem_cgroup __read_mostly;
76 /* Active memory cgroup to use from an interrupt context */
77 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
79 /* Socket memory accounting disabled? */
80 static bool cgroup_memory_nosocket;
82 /* Kernel memory accounting disabled? */
83 static bool cgroup_memory_nokmem;
85 /* Whether the swap controller is active */
86 #ifdef CONFIG_MEMCG_SWAP
87 bool cgroup_memory_noswap __read_mostly;
89 #define cgroup_memory_noswap 1
92 #ifdef CONFIG_CGROUP_WRITEBACK
93 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
96 /* Whether legacy memory+swap accounting is active */
97 static bool do_memsw_account(void)
99 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
102 #define THRESHOLDS_EVENTS_TARGET 128
103 #define SOFTLIMIT_EVENTS_TARGET 1024
106 * Cgroups above their limits are maintained in a RB-Tree, independent of
107 * their hierarchy representation
110 struct mem_cgroup_tree_per_node {
111 struct rb_root rb_root;
112 struct rb_node *rb_rightmost;
116 struct mem_cgroup_tree {
117 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
120 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
123 struct mem_cgroup_eventfd_list {
124 struct list_head list;
125 struct eventfd_ctx *eventfd;
129 * cgroup_event represents events which userspace want to receive.
131 struct mem_cgroup_event {
133 * memcg which the event belongs to.
135 struct mem_cgroup *memcg;
137 * eventfd to signal userspace about the event.
139 struct eventfd_ctx *eventfd;
141 * Each of these stored in a list by the cgroup.
143 struct list_head list;
145 * register_event() callback will be used to add new userspace
146 * waiter for changes related to this event. Use eventfd_signal()
147 * on eventfd to send notification to userspace.
149 int (*register_event)(struct mem_cgroup *memcg,
150 struct eventfd_ctx *eventfd, const char *args);
152 * unregister_event() callback will be called when userspace closes
153 * the eventfd or on cgroup removing. This callback must be set,
154 * if you want provide notification functionality.
156 void (*unregister_event)(struct mem_cgroup *memcg,
157 struct eventfd_ctx *eventfd);
159 * All fields below needed to unregister event when
160 * userspace closes eventfd.
163 wait_queue_head_t *wqh;
164 wait_queue_entry_t wait;
165 struct work_struct remove;
168 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
169 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
171 /* Stuffs for move charges at task migration. */
173 * Types of charges to be moved.
175 #define MOVE_ANON 0x1U
176 #define MOVE_FILE 0x2U
177 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
179 /* "mc" and its members are protected by cgroup_mutex */
180 static struct move_charge_struct {
181 spinlock_t lock; /* for from, to */
182 struct mm_struct *mm;
183 struct mem_cgroup *from;
184 struct mem_cgroup *to;
186 unsigned long precharge;
187 unsigned long moved_charge;
188 unsigned long moved_swap;
189 struct task_struct *moving_task; /* a task moving charges */
190 wait_queue_head_t waitq; /* a waitq for other context */
192 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
193 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
197 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
198 * limit reclaim to prevent infinite loops, if they ever occur.
200 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
201 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
203 /* for encoding cft->private value on file */
212 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
213 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
214 #define MEMFILE_ATTR(val) ((val) & 0xffff)
215 /* Used for OOM nofiier */
216 #define OOM_CONTROL (0)
219 * Iteration constructs for visiting all cgroups (under a tree). If
220 * loops are exited prematurely (break), mem_cgroup_iter_break() must
221 * be used for reference counting.
223 #define for_each_mem_cgroup_tree(iter, root) \
224 for (iter = mem_cgroup_iter(root, NULL, NULL); \
226 iter = mem_cgroup_iter(root, iter, NULL))
228 #define for_each_mem_cgroup(iter) \
229 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
231 iter = mem_cgroup_iter(NULL, iter, NULL))
233 static inline bool should_force_charge(void)
235 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
236 (current->flags & PF_EXITING);
239 /* Some nice accessors for the vmpressure. */
240 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
243 memcg = root_mem_cgroup;
244 return &memcg->vmpressure;
247 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
249 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
252 #ifdef CONFIG_MEMCG_KMEM
253 extern spinlock_t css_set_lock;
255 static void obj_cgroup_release(struct percpu_ref *ref)
257 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
258 struct mem_cgroup *memcg;
259 unsigned int nr_bytes;
260 unsigned int nr_pages;
264 * At this point all allocated objects are freed, and
265 * objcg->nr_charged_bytes can't have an arbitrary byte value.
266 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
268 * The following sequence can lead to it:
269 * 1) CPU0: objcg == stock->cached_objcg
270 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
271 * PAGE_SIZE bytes are charged
272 * 3) CPU1: a process from another memcg is allocating something,
273 * the stock if flushed,
274 * objcg->nr_charged_bytes = PAGE_SIZE - 92
275 * 5) CPU0: we do release this object,
276 * 92 bytes are added to stock->nr_bytes
277 * 6) CPU0: stock is flushed,
278 * 92 bytes are added to objcg->nr_charged_bytes
280 * In the result, nr_charged_bytes == PAGE_SIZE.
281 * This page will be uncharged in obj_cgroup_release().
283 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
284 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
285 nr_pages = nr_bytes >> PAGE_SHIFT;
287 spin_lock_irqsave(&css_set_lock, flags);
288 memcg = obj_cgroup_memcg(objcg);
290 __memcg_kmem_uncharge(memcg, nr_pages);
291 list_del(&objcg->list);
292 mem_cgroup_put(memcg);
293 spin_unlock_irqrestore(&css_set_lock, flags);
295 percpu_ref_exit(ref);
296 kfree_rcu(objcg, rcu);
299 static struct obj_cgroup *obj_cgroup_alloc(void)
301 struct obj_cgroup *objcg;
304 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
308 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
314 INIT_LIST_HEAD(&objcg->list);
318 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
319 struct mem_cgroup *parent)
321 struct obj_cgroup *objcg, *iter;
323 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
325 spin_lock_irq(&css_set_lock);
327 /* Move active objcg to the parent's list */
328 xchg(&objcg->memcg, parent);
329 css_get(&parent->css);
330 list_add(&objcg->list, &parent->objcg_list);
332 /* Move already reparented objcgs to the parent's list */
333 list_for_each_entry(iter, &memcg->objcg_list, list) {
334 css_get(&parent->css);
335 xchg(&iter->memcg, parent);
336 css_put(&memcg->css);
338 list_splice(&memcg->objcg_list, &parent->objcg_list);
340 spin_unlock_irq(&css_set_lock);
342 percpu_ref_kill(&objcg->refcnt);
346 * This will be used as a shrinker list's index.
347 * The main reason for not using cgroup id for this:
348 * this works better in sparse environments, where we have a lot of memcgs,
349 * but only a few kmem-limited. Or also, if we have, for instance, 200
350 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
351 * 200 entry array for that.
353 * The current size of the caches array is stored in memcg_nr_cache_ids. It
354 * will double each time we have to increase it.
356 static DEFINE_IDA(memcg_cache_ida);
357 int memcg_nr_cache_ids;
359 /* Protects memcg_nr_cache_ids */
360 static DECLARE_RWSEM(memcg_cache_ids_sem);
362 void memcg_get_cache_ids(void)
364 down_read(&memcg_cache_ids_sem);
367 void memcg_put_cache_ids(void)
369 up_read(&memcg_cache_ids_sem);
373 * MIN_SIZE is different than 1, because we would like to avoid going through
374 * the alloc/free process all the time. In a small machine, 4 kmem-limited
375 * cgroups is a reasonable guess. In the future, it could be a parameter or
376 * tunable, but that is strictly not necessary.
378 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
379 * this constant directly from cgroup, but it is understandable that this is
380 * better kept as an internal representation in cgroup.c. In any case, the
381 * cgrp_id space is not getting any smaller, and we don't have to necessarily
382 * increase ours as well if it increases.
384 #define MEMCG_CACHES_MIN_SIZE 4
385 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
388 * A lot of the calls to the cache allocation functions are expected to be
389 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
390 * conditional to this static branch, we'll have to allow modules that does
391 * kmem_cache_alloc and the such to see this symbol as well
393 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
394 EXPORT_SYMBOL(memcg_kmem_enabled_key);
397 static int memcg_shrinker_map_size;
398 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
400 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
402 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
405 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
406 int size, int old_size)
408 struct memcg_shrinker_map *new, *old;
411 lockdep_assert_held(&memcg_shrinker_map_mutex);
414 old = rcu_dereference_protected(
415 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
416 /* Not yet online memcg */
420 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
424 /* Set all old bits, clear all new bits */
425 memset(new->map, (int)0xff, old_size);
426 memset((void *)new->map + old_size, 0, size - old_size);
428 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
429 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
435 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
437 struct mem_cgroup_per_node *pn;
438 struct memcg_shrinker_map *map;
441 if (mem_cgroup_is_root(memcg))
445 pn = mem_cgroup_nodeinfo(memcg, nid);
446 map = rcu_dereference_protected(pn->shrinker_map, true);
449 rcu_assign_pointer(pn->shrinker_map, NULL);
453 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
455 struct memcg_shrinker_map *map;
456 int nid, size, ret = 0;
458 if (mem_cgroup_is_root(memcg))
461 mutex_lock(&memcg_shrinker_map_mutex);
462 size = memcg_shrinker_map_size;
464 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
466 memcg_free_shrinker_maps(memcg);
470 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
472 mutex_unlock(&memcg_shrinker_map_mutex);
477 int memcg_expand_shrinker_maps(int new_id)
479 int size, old_size, ret = 0;
480 struct mem_cgroup *memcg;
482 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
483 old_size = memcg_shrinker_map_size;
484 if (size <= old_size)
487 mutex_lock(&memcg_shrinker_map_mutex);
488 if (!root_mem_cgroup)
491 for_each_mem_cgroup(memcg) {
492 if (mem_cgroup_is_root(memcg))
494 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
496 mem_cgroup_iter_break(NULL, memcg);
502 memcg_shrinker_map_size = size;
503 mutex_unlock(&memcg_shrinker_map_mutex);
507 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
509 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
510 struct memcg_shrinker_map *map;
513 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
514 /* Pairs with smp mb in shrink_slab() */
515 smp_mb__before_atomic();
516 set_bit(shrinker_id, map->map);
522 * mem_cgroup_css_from_page - css of the memcg associated with a page
523 * @page: page of interest
525 * If memcg is bound to the default hierarchy, css of the memcg associated
526 * with @page is returned. The returned css remains associated with @page
527 * until it is released.
529 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
532 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
534 struct mem_cgroup *memcg;
536 memcg = page->mem_cgroup;
538 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
539 memcg = root_mem_cgroup;
545 * page_cgroup_ino - return inode number of the memcg a page is charged to
548 * Look up the closest online ancestor of the memory cgroup @page is charged to
549 * and return its inode number or 0 if @page is not charged to any cgroup. It
550 * is safe to call this function without holding a reference to @page.
552 * Note, this function is inherently racy, because there is nothing to prevent
553 * the cgroup inode from getting torn down and potentially reallocated a moment
554 * after page_cgroup_ino() returns, so it only should be used by callers that
555 * do not care (such as procfs interfaces).
557 ino_t page_cgroup_ino(struct page *page)
559 struct mem_cgroup *memcg;
560 unsigned long ino = 0;
563 memcg = page->mem_cgroup;
566 * The lowest bit set means that memcg isn't a valid
567 * memcg pointer, but a obj_cgroups pointer.
568 * In this case the page is shared and doesn't belong
569 * to any specific memory cgroup.
571 if ((unsigned long) memcg & 0x1UL)
574 while (memcg && !(memcg->css.flags & CSS_ONLINE))
575 memcg = parent_mem_cgroup(memcg);
577 ino = cgroup_ino(memcg->css.cgroup);
582 static struct mem_cgroup_per_node *
583 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
585 int nid = page_to_nid(page);
587 return memcg->nodeinfo[nid];
590 static struct mem_cgroup_tree_per_node *
591 soft_limit_tree_node(int nid)
593 return soft_limit_tree.rb_tree_per_node[nid];
596 static struct mem_cgroup_tree_per_node *
597 soft_limit_tree_from_page(struct page *page)
599 int nid = page_to_nid(page);
601 return soft_limit_tree.rb_tree_per_node[nid];
604 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
605 struct mem_cgroup_tree_per_node *mctz,
606 unsigned long new_usage_in_excess)
608 struct rb_node **p = &mctz->rb_root.rb_node;
609 struct rb_node *parent = NULL;
610 struct mem_cgroup_per_node *mz_node;
611 bool rightmost = true;
616 mz->usage_in_excess = new_usage_in_excess;
617 if (!mz->usage_in_excess)
621 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
623 if (mz->usage_in_excess < mz_node->usage_in_excess) {
632 mctz->rb_rightmost = &mz->tree_node;
634 rb_link_node(&mz->tree_node, parent, p);
635 rb_insert_color(&mz->tree_node, &mctz->rb_root);
639 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
640 struct mem_cgroup_tree_per_node *mctz)
645 if (&mz->tree_node == mctz->rb_rightmost)
646 mctz->rb_rightmost = rb_prev(&mz->tree_node);
648 rb_erase(&mz->tree_node, &mctz->rb_root);
652 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
653 struct mem_cgroup_tree_per_node *mctz)
657 spin_lock_irqsave(&mctz->lock, flags);
658 __mem_cgroup_remove_exceeded(mz, mctz);
659 spin_unlock_irqrestore(&mctz->lock, flags);
662 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
664 unsigned long nr_pages = page_counter_read(&memcg->memory);
665 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
666 unsigned long excess = 0;
668 if (nr_pages > soft_limit)
669 excess = nr_pages - soft_limit;
674 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
676 unsigned long excess;
677 struct mem_cgroup_per_node *mz;
678 struct mem_cgroup_tree_per_node *mctz;
680 mctz = soft_limit_tree_from_page(page);
684 * Necessary to update all ancestors when hierarchy is used.
685 * because their event counter is not touched.
687 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
688 mz = mem_cgroup_page_nodeinfo(memcg, page);
689 excess = soft_limit_excess(memcg);
691 * We have to update the tree if mz is on RB-tree or
692 * mem is over its softlimit.
694 if (excess || mz->on_tree) {
697 spin_lock_irqsave(&mctz->lock, flags);
698 /* if on-tree, remove it */
700 __mem_cgroup_remove_exceeded(mz, mctz);
702 * Insert again. mz->usage_in_excess will be updated.
703 * If excess is 0, no tree ops.
705 __mem_cgroup_insert_exceeded(mz, mctz, excess);
706 spin_unlock_irqrestore(&mctz->lock, flags);
711 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
713 struct mem_cgroup_tree_per_node *mctz;
714 struct mem_cgroup_per_node *mz;
718 mz = mem_cgroup_nodeinfo(memcg, nid);
719 mctz = soft_limit_tree_node(nid);
721 mem_cgroup_remove_exceeded(mz, mctz);
725 static struct mem_cgroup_per_node *
726 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
728 struct mem_cgroup_per_node *mz;
732 if (!mctz->rb_rightmost)
733 goto done; /* Nothing to reclaim from */
735 mz = rb_entry(mctz->rb_rightmost,
736 struct mem_cgroup_per_node, tree_node);
738 * Remove the node now but someone else can add it back,
739 * we will to add it back at the end of reclaim to its correct
740 * position in the tree.
742 __mem_cgroup_remove_exceeded(mz, mctz);
743 if (!soft_limit_excess(mz->memcg) ||
744 !css_tryget(&mz->memcg->css))
750 static struct mem_cgroup_per_node *
751 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
753 struct mem_cgroup_per_node *mz;
755 spin_lock_irq(&mctz->lock);
756 mz = __mem_cgroup_largest_soft_limit_node(mctz);
757 spin_unlock_irq(&mctz->lock);
762 * __mod_memcg_state - update cgroup memory statistics
763 * @memcg: the memory cgroup
764 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
765 * @val: delta to add to the counter, can be negative
767 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
769 long x, threshold = MEMCG_CHARGE_BATCH;
771 if (mem_cgroup_disabled())
774 if (memcg_stat_item_in_bytes(idx))
775 threshold <<= PAGE_SHIFT;
777 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
778 if (unlikely(abs(x) > threshold)) {
779 struct mem_cgroup *mi;
782 * Batch local counters to keep them in sync with
783 * the hierarchical ones.
785 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
786 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
787 atomic_long_add(x, &mi->vmstats[idx]);
790 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
793 static struct mem_cgroup_per_node *
794 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
796 struct mem_cgroup *parent;
798 parent = parent_mem_cgroup(pn->memcg);
801 return mem_cgroup_nodeinfo(parent, nid);
804 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
807 struct mem_cgroup_per_node *pn;
808 struct mem_cgroup *memcg;
809 long x, threshold = MEMCG_CHARGE_BATCH;
811 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
815 __mod_memcg_state(memcg, idx, val);
818 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
820 if (vmstat_item_in_bytes(idx))
821 threshold <<= PAGE_SHIFT;
823 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
824 if (unlikely(abs(x) > threshold)) {
825 pg_data_t *pgdat = lruvec_pgdat(lruvec);
826 struct mem_cgroup_per_node *pi;
828 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
829 atomic_long_add(x, &pi->lruvec_stat[idx]);
832 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
836 * __mod_lruvec_state - update lruvec memory statistics
837 * @lruvec: the lruvec
838 * @idx: the stat item
839 * @val: delta to add to the counter, can be negative
841 * The lruvec is the intersection of the NUMA node and a cgroup. This
842 * function updates the all three counters that are affected by a
843 * change of state at this level: per-node, per-cgroup, per-lruvec.
845 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
849 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
851 /* Update memcg and lruvec */
852 if (!mem_cgroup_disabled())
853 __mod_memcg_lruvec_state(lruvec, idx, val);
856 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
859 struct page *head = compound_head(page); /* rmap on tail pages */
860 pg_data_t *pgdat = page_pgdat(page);
861 struct lruvec *lruvec;
863 /* Untracked pages have no memcg, no lruvec. Update only the node */
864 if (!head->mem_cgroup) {
865 __mod_node_page_state(pgdat, idx, val);
869 lruvec = mem_cgroup_lruvec(head->mem_cgroup, pgdat);
870 __mod_lruvec_state(lruvec, idx, val);
873 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
875 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
876 struct mem_cgroup *memcg;
877 struct lruvec *lruvec;
880 memcg = mem_cgroup_from_obj(p);
883 * Untracked pages have no memcg, no lruvec. Update only the
884 * node. If we reparent the slab objects to the root memcg,
885 * when we free the slab object, we need to update the per-memcg
886 * vmstats to keep it correct for the root memcg.
889 __mod_node_page_state(pgdat, idx, val);
891 lruvec = mem_cgroup_lruvec(memcg, pgdat);
892 __mod_lruvec_state(lruvec, idx, val);
898 * __count_memcg_events - account VM events in a cgroup
899 * @memcg: the memory cgroup
900 * @idx: the event item
901 * @count: the number of events that occured
903 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
908 if (mem_cgroup_disabled())
911 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
912 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
913 struct mem_cgroup *mi;
916 * Batch local counters to keep them in sync with
917 * the hierarchical ones.
919 __this_cpu_add(memcg->vmstats_local->events[idx], x);
920 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
921 atomic_long_add(x, &mi->vmevents[idx]);
924 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
927 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
929 return atomic_long_read(&memcg->vmevents[event]);
932 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
937 for_each_possible_cpu(cpu)
938 x += per_cpu(memcg->vmstats_local->events[event], cpu);
942 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
946 /* pagein of a big page is an event. So, ignore page size */
948 __count_memcg_events(memcg, PGPGIN, 1);
950 __count_memcg_events(memcg, PGPGOUT, 1);
951 nr_pages = -nr_pages; /* for event */
954 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
957 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
958 enum mem_cgroup_events_target target)
960 unsigned long val, next;
962 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
963 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
964 /* from time_after() in jiffies.h */
965 if ((long)(next - val) < 0) {
967 case MEM_CGROUP_TARGET_THRESH:
968 next = val + THRESHOLDS_EVENTS_TARGET;
970 case MEM_CGROUP_TARGET_SOFTLIMIT:
971 next = val + SOFTLIMIT_EVENTS_TARGET;
976 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
983 * Check events in order.
986 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
988 /* threshold event is triggered in finer grain than soft limit */
989 if (unlikely(mem_cgroup_event_ratelimit(memcg,
990 MEM_CGROUP_TARGET_THRESH))) {
993 do_softlimit = mem_cgroup_event_ratelimit(memcg,
994 MEM_CGROUP_TARGET_SOFTLIMIT);
995 mem_cgroup_threshold(memcg);
996 if (unlikely(do_softlimit))
997 mem_cgroup_update_tree(memcg, page);
1001 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1004 * mm_update_next_owner() may clear mm->owner to NULL
1005 * if it races with swapoff, page migration, etc.
1006 * So this can be called with p == NULL.
1011 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1013 EXPORT_SYMBOL(mem_cgroup_from_task);
1016 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1017 * @mm: mm from which memcg should be extracted. It can be NULL.
1019 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1020 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1023 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1025 struct mem_cgroup *memcg;
1027 if (mem_cgroup_disabled())
1033 * Page cache insertions can happen withou an
1034 * actual mm context, e.g. during disk probing
1035 * on boot, loopback IO, acct() writes etc.
1038 memcg = root_mem_cgroup;
1040 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1041 if (unlikely(!memcg))
1042 memcg = root_mem_cgroup;
1044 } while (!css_tryget(&memcg->css));
1048 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1051 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1052 * @page: page from which memcg should be extracted.
1054 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1055 * root_mem_cgroup is returned.
1057 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
1059 struct mem_cgroup *memcg = page->mem_cgroup;
1061 if (mem_cgroup_disabled())
1065 /* Page should not get uncharged and freed memcg under us. */
1066 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
1067 memcg = root_mem_cgroup;
1071 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1073 static __always_inline struct mem_cgroup *active_memcg(void)
1076 return this_cpu_read(int_active_memcg);
1078 return current->active_memcg;
1081 static __always_inline struct mem_cgroup *get_active_memcg(void)
1083 struct mem_cgroup *memcg;
1086 memcg = active_memcg();
1088 /* current->active_memcg must hold a ref. */
1089 if (WARN_ON_ONCE(!css_tryget(&memcg->css)))
1090 memcg = root_mem_cgroup;
1092 memcg = current->active_memcg;
1099 static __always_inline bool memcg_kmem_bypass(void)
1101 /* Allow remote memcg charging from any context. */
1102 if (unlikely(active_memcg()))
1105 /* Memcg to charge can't be determined. */
1106 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
1113 * If active memcg is set, do not fallback to current->mm->memcg.
1115 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1117 if (memcg_kmem_bypass())
1120 if (unlikely(active_memcg()))
1121 return get_active_memcg();
1123 return get_mem_cgroup_from_mm(current->mm);
1127 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1128 * @root: hierarchy root
1129 * @prev: previously returned memcg, NULL on first invocation
1130 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1132 * Returns references to children of the hierarchy below @root, or
1133 * @root itself, or %NULL after a full round-trip.
1135 * Caller must pass the return value in @prev on subsequent
1136 * invocations for reference counting, or use mem_cgroup_iter_break()
1137 * to cancel a hierarchy walk before the round-trip is complete.
1139 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1140 * in the hierarchy among all concurrent reclaimers operating on the
1143 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1144 struct mem_cgroup *prev,
1145 struct mem_cgroup_reclaim_cookie *reclaim)
1147 struct mem_cgroup_reclaim_iter *iter;
1148 struct cgroup_subsys_state *css = NULL;
1149 struct mem_cgroup *memcg = NULL;
1150 struct mem_cgroup *pos = NULL;
1152 if (mem_cgroup_disabled())
1156 root = root_mem_cgroup;
1158 if (prev && !reclaim)
1164 struct mem_cgroup_per_node *mz;
1166 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1169 if (prev && reclaim->generation != iter->generation)
1173 pos = READ_ONCE(iter->position);
1174 if (!pos || css_tryget(&pos->css))
1177 * css reference reached zero, so iter->position will
1178 * be cleared by ->css_released. However, we should not
1179 * rely on this happening soon, because ->css_released
1180 * is called from a work queue, and by busy-waiting we
1181 * might block it. So we clear iter->position right
1184 (void)cmpxchg(&iter->position, pos, NULL);
1192 css = css_next_descendant_pre(css, &root->css);
1195 * Reclaimers share the hierarchy walk, and a
1196 * new one might jump in right at the end of
1197 * the hierarchy - make sure they see at least
1198 * one group and restart from the beginning.
1206 * Verify the css and acquire a reference. The root
1207 * is provided by the caller, so we know it's alive
1208 * and kicking, and don't take an extra reference.
1210 memcg = mem_cgroup_from_css(css);
1212 if (css == &root->css)
1215 if (css_tryget(css))
1223 * The position could have already been updated by a competing
1224 * thread, so check that the value hasn't changed since we read
1225 * it to avoid reclaiming from the same cgroup twice.
1227 (void)cmpxchg(&iter->position, pos, memcg);
1235 reclaim->generation = iter->generation;
1240 if (prev && prev != root)
1241 css_put(&prev->css);
1247 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1248 * @root: hierarchy root
1249 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1251 void mem_cgroup_iter_break(struct mem_cgroup *root,
1252 struct mem_cgroup *prev)
1255 root = root_mem_cgroup;
1256 if (prev && prev != root)
1257 css_put(&prev->css);
1260 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1261 struct mem_cgroup *dead_memcg)
1263 struct mem_cgroup_reclaim_iter *iter;
1264 struct mem_cgroup_per_node *mz;
1267 for_each_node(nid) {
1268 mz = mem_cgroup_nodeinfo(from, nid);
1270 cmpxchg(&iter->position, dead_memcg, NULL);
1274 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1276 struct mem_cgroup *memcg = dead_memcg;
1277 struct mem_cgroup *last;
1280 __invalidate_reclaim_iterators(memcg, dead_memcg);
1282 } while ((memcg = parent_mem_cgroup(memcg)));
1285 * When cgruop1 non-hierarchy mode is used,
1286 * parent_mem_cgroup() does not walk all the way up to the
1287 * cgroup root (root_mem_cgroup). So we have to handle
1288 * dead_memcg from cgroup root separately.
1290 if (last != root_mem_cgroup)
1291 __invalidate_reclaim_iterators(root_mem_cgroup,
1296 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1297 * @memcg: hierarchy root
1298 * @fn: function to call for each task
1299 * @arg: argument passed to @fn
1301 * This function iterates over tasks attached to @memcg or to any of its
1302 * descendants and calls @fn for each task. If @fn returns a non-zero
1303 * value, the function breaks the iteration loop and returns the value.
1304 * Otherwise, it will iterate over all tasks and return 0.
1306 * This function must not be called for the root memory cgroup.
1308 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1309 int (*fn)(struct task_struct *, void *), void *arg)
1311 struct mem_cgroup *iter;
1314 BUG_ON(memcg == root_mem_cgroup);
1316 for_each_mem_cgroup_tree(iter, memcg) {
1317 struct css_task_iter it;
1318 struct task_struct *task;
1320 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1321 while (!ret && (task = css_task_iter_next(&it)))
1322 ret = fn(task, arg);
1323 css_task_iter_end(&it);
1325 mem_cgroup_iter_break(memcg, iter);
1333 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1335 * @pgdat: pgdat of the page
1337 * This function relies on page's memcg being stable - see the
1338 * access rules in commit_charge().
1340 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1342 struct mem_cgroup_per_node *mz;
1343 struct mem_cgroup *memcg;
1344 struct lruvec *lruvec;
1346 if (mem_cgroup_disabled()) {
1347 lruvec = &pgdat->__lruvec;
1351 memcg = page->mem_cgroup;
1353 * Swapcache readahead pages are added to the LRU - and
1354 * possibly migrated - before they are charged.
1357 memcg = root_mem_cgroup;
1359 mz = mem_cgroup_page_nodeinfo(memcg, page);
1360 lruvec = &mz->lruvec;
1363 * Since a node can be onlined after the mem_cgroup was created,
1364 * we have to be prepared to initialize lruvec->zone here;
1365 * and if offlined then reonlined, we need to reinitialize it.
1367 if (unlikely(lruvec->pgdat != pgdat))
1368 lruvec->pgdat = pgdat;
1373 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1374 * @lruvec: mem_cgroup per zone lru vector
1375 * @lru: index of lru list the page is sitting on
1376 * @zid: zone id of the accounted pages
1377 * @nr_pages: positive when adding or negative when removing
1379 * This function must be called under lru_lock, just before a page is added
1380 * to or just after a page is removed from an lru list (that ordering being
1381 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1383 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1384 int zid, int nr_pages)
1386 struct mem_cgroup_per_node *mz;
1387 unsigned long *lru_size;
1390 if (mem_cgroup_disabled())
1393 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1394 lru_size = &mz->lru_zone_size[zid][lru];
1397 *lru_size += nr_pages;
1400 if (WARN_ONCE(size < 0,
1401 "%s(%p, %d, %d): lru_size %ld\n",
1402 __func__, lruvec, lru, nr_pages, size)) {
1408 *lru_size += nr_pages;
1412 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1413 * @memcg: the memory cgroup
1415 * Returns the maximum amount of memory @mem can be charged with, in
1418 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1420 unsigned long margin = 0;
1421 unsigned long count;
1422 unsigned long limit;
1424 count = page_counter_read(&memcg->memory);
1425 limit = READ_ONCE(memcg->memory.max);
1427 margin = limit - count;
1429 if (do_memsw_account()) {
1430 count = page_counter_read(&memcg->memsw);
1431 limit = READ_ONCE(memcg->memsw.max);
1433 margin = min(margin, limit - count);
1442 * A routine for checking "mem" is under move_account() or not.
1444 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1445 * moving cgroups. This is for waiting at high-memory pressure
1448 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1450 struct mem_cgroup *from;
1451 struct mem_cgroup *to;
1454 * Unlike task_move routines, we access mc.to, mc.from not under
1455 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1457 spin_lock(&mc.lock);
1463 ret = mem_cgroup_is_descendant(from, memcg) ||
1464 mem_cgroup_is_descendant(to, memcg);
1466 spin_unlock(&mc.lock);
1470 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1472 if (mc.moving_task && current != mc.moving_task) {
1473 if (mem_cgroup_under_move(memcg)) {
1475 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1476 /* moving charge context might have finished. */
1479 finish_wait(&mc.waitq, &wait);
1486 struct memory_stat {
1492 static struct memory_stat memory_stats[] = {
1493 { "anon", PAGE_SIZE, NR_ANON_MAPPED },
1494 { "file", PAGE_SIZE, NR_FILE_PAGES },
1495 { "kernel_stack", 1024, NR_KERNEL_STACK_KB },
1496 { "percpu", 1, MEMCG_PERCPU_B },
1497 { "sock", PAGE_SIZE, MEMCG_SOCK },
1498 { "shmem", PAGE_SIZE, NR_SHMEM },
1499 { "file_mapped", PAGE_SIZE, NR_FILE_MAPPED },
1500 { "file_dirty", PAGE_SIZE, NR_FILE_DIRTY },
1501 { "file_writeback", PAGE_SIZE, NR_WRITEBACK },
1502 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1504 * The ratio will be initialized in memory_stats_init(). Because
1505 * on some architectures, the macro of HPAGE_PMD_SIZE is not
1506 * constant(e.g. powerpc).
1508 { "anon_thp", 0, NR_ANON_THPS },
1509 { "file_thp", 0, NR_FILE_THPS },
1510 { "shmem_thp", 0, NR_SHMEM_THPS },
1512 { "inactive_anon", PAGE_SIZE, NR_INACTIVE_ANON },
1513 { "active_anon", PAGE_SIZE, NR_ACTIVE_ANON },
1514 { "inactive_file", PAGE_SIZE, NR_INACTIVE_FILE },
1515 { "active_file", PAGE_SIZE, NR_ACTIVE_FILE },
1516 { "unevictable", PAGE_SIZE, NR_UNEVICTABLE },
1519 * Note: The slab_reclaimable and slab_unreclaimable must be
1520 * together and slab_reclaimable must be in front.
1522 { "slab_reclaimable", 1, NR_SLAB_RECLAIMABLE_B },
1523 { "slab_unreclaimable", 1, NR_SLAB_UNRECLAIMABLE_B },
1525 /* The memory events */
1526 { "workingset_refault_anon", 1, WORKINGSET_REFAULT_ANON },
1527 { "workingset_refault_file", 1, WORKINGSET_REFAULT_FILE },
1528 { "workingset_activate_anon", 1, WORKINGSET_ACTIVATE_ANON },
1529 { "workingset_activate_file", 1, WORKINGSET_ACTIVATE_FILE },
1530 { "workingset_restore_anon", 1, WORKINGSET_RESTORE_ANON },
1531 { "workingset_restore_file", 1, WORKINGSET_RESTORE_FILE },
1532 { "workingset_nodereclaim", 1, WORKINGSET_NODERECLAIM },
1535 static int __init memory_stats_init(void)
1539 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1540 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1541 if (memory_stats[i].idx == NR_ANON_THPS ||
1542 memory_stats[i].idx == NR_FILE_THPS ||
1543 memory_stats[i].idx == NR_SHMEM_THPS)
1544 memory_stats[i].ratio = HPAGE_PMD_SIZE;
1546 VM_BUG_ON(!memory_stats[i].ratio);
1547 VM_BUG_ON(memory_stats[i].idx >= MEMCG_NR_STAT);
1552 pure_initcall(memory_stats_init);
1554 static char *memory_stat_format(struct mem_cgroup *memcg)
1559 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1564 * Provide statistics on the state of the memory subsystem as
1565 * well as cumulative event counters that show past behavior.
1567 * This list is ordered following a combination of these gradients:
1568 * 1) generic big picture -> specifics and details
1569 * 2) reflecting userspace activity -> reflecting kernel heuristics
1571 * Current memory state:
1574 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1577 size = memcg_page_state(memcg, memory_stats[i].idx);
1578 size *= memory_stats[i].ratio;
1579 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1581 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1582 size = memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B) +
1583 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B);
1584 seq_buf_printf(&s, "slab %llu\n", size);
1588 /* Accumulated memory events */
1590 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1591 memcg_events(memcg, PGFAULT));
1592 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1593 memcg_events(memcg, PGMAJFAULT));
1594 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1595 memcg_events(memcg, PGREFILL));
1596 seq_buf_printf(&s, "pgscan %lu\n",
1597 memcg_events(memcg, PGSCAN_KSWAPD) +
1598 memcg_events(memcg, PGSCAN_DIRECT));
1599 seq_buf_printf(&s, "pgsteal %lu\n",
1600 memcg_events(memcg, PGSTEAL_KSWAPD) +
1601 memcg_events(memcg, PGSTEAL_DIRECT));
1602 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1603 memcg_events(memcg, PGACTIVATE));
1604 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1605 memcg_events(memcg, PGDEACTIVATE));
1606 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1607 memcg_events(memcg, PGLAZYFREE));
1608 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1609 memcg_events(memcg, PGLAZYFREED));
1611 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1612 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1613 memcg_events(memcg, THP_FAULT_ALLOC));
1614 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1615 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1616 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1618 /* The above should easily fit into one page */
1619 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1624 #define K(x) ((x) << (PAGE_SHIFT-10))
1626 * mem_cgroup_print_oom_context: Print OOM information relevant to
1627 * memory controller.
1628 * @memcg: The memory cgroup that went over limit
1629 * @p: Task that is going to be killed
1631 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1634 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1639 pr_cont(",oom_memcg=");
1640 pr_cont_cgroup_path(memcg->css.cgroup);
1642 pr_cont(",global_oom");
1644 pr_cont(",task_memcg=");
1645 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1651 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1652 * memory controller.
1653 * @memcg: The memory cgroup that went over limit
1655 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1659 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1660 K((u64)page_counter_read(&memcg->memory)),
1661 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1662 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1663 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1664 K((u64)page_counter_read(&memcg->swap)),
1665 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1667 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1668 K((u64)page_counter_read(&memcg->memsw)),
1669 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1670 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1671 K((u64)page_counter_read(&memcg->kmem)),
1672 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1675 pr_info("Memory cgroup stats for ");
1676 pr_cont_cgroup_path(memcg->css.cgroup);
1678 buf = memory_stat_format(memcg);
1686 * Return the memory (and swap, if configured) limit for a memcg.
1688 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1690 unsigned long max = READ_ONCE(memcg->memory.max);
1692 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1693 if (mem_cgroup_swappiness(memcg))
1694 max += min(READ_ONCE(memcg->swap.max),
1695 (unsigned long)total_swap_pages);
1697 if (mem_cgroup_swappiness(memcg)) {
1698 /* Calculate swap excess capacity from memsw limit */
1699 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1701 max += min(swap, (unsigned long)total_swap_pages);
1707 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1709 return page_counter_read(&memcg->memory);
1712 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1715 struct oom_control oc = {
1719 .gfp_mask = gfp_mask,
1724 if (mutex_lock_killable(&oom_lock))
1727 if (mem_cgroup_margin(memcg) >= (1 << order))
1731 * A few threads which were not waiting at mutex_lock_killable() can
1732 * fail to bail out. Therefore, check again after holding oom_lock.
1734 ret = should_force_charge() || out_of_memory(&oc);
1737 mutex_unlock(&oom_lock);
1741 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1744 unsigned long *total_scanned)
1746 struct mem_cgroup *victim = NULL;
1749 unsigned long excess;
1750 unsigned long nr_scanned;
1751 struct mem_cgroup_reclaim_cookie reclaim = {
1755 excess = soft_limit_excess(root_memcg);
1758 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1763 * If we have not been able to reclaim
1764 * anything, it might because there are
1765 * no reclaimable pages under this hierarchy
1770 * We want to do more targeted reclaim.
1771 * excess >> 2 is not to excessive so as to
1772 * reclaim too much, nor too less that we keep
1773 * coming back to reclaim from this cgroup
1775 if (total >= (excess >> 2) ||
1776 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1781 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1782 pgdat, &nr_scanned);
1783 *total_scanned += nr_scanned;
1784 if (!soft_limit_excess(root_memcg))
1787 mem_cgroup_iter_break(root_memcg, victim);
1791 #ifdef CONFIG_LOCKDEP
1792 static struct lockdep_map memcg_oom_lock_dep_map = {
1793 .name = "memcg_oom_lock",
1797 static DEFINE_SPINLOCK(memcg_oom_lock);
1800 * Check OOM-Killer is already running under our hierarchy.
1801 * If someone is running, return false.
1803 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1805 struct mem_cgroup *iter, *failed = NULL;
1807 spin_lock(&memcg_oom_lock);
1809 for_each_mem_cgroup_tree(iter, memcg) {
1810 if (iter->oom_lock) {
1812 * this subtree of our hierarchy is already locked
1813 * so we cannot give a lock.
1816 mem_cgroup_iter_break(memcg, iter);
1819 iter->oom_lock = true;
1824 * OK, we failed to lock the whole subtree so we have
1825 * to clean up what we set up to the failing subtree
1827 for_each_mem_cgroup_tree(iter, memcg) {
1828 if (iter == failed) {
1829 mem_cgroup_iter_break(memcg, iter);
1832 iter->oom_lock = false;
1835 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1837 spin_unlock(&memcg_oom_lock);
1842 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1844 struct mem_cgroup *iter;
1846 spin_lock(&memcg_oom_lock);
1847 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1848 for_each_mem_cgroup_tree(iter, memcg)
1849 iter->oom_lock = false;
1850 spin_unlock(&memcg_oom_lock);
1853 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1855 struct mem_cgroup *iter;
1857 spin_lock(&memcg_oom_lock);
1858 for_each_mem_cgroup_tree(iter, memcg)
1860 spin_unlock(&memcg_oom_lock);
1863 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1865 struct mem_cgroup *iter;
1868 * Be careful about under_oom underflows becase a child memcg
1869 * could have been added after mem_cgroup_mark_under_oom.
1871 spin_lock(&memcg_oom_lock);
1872 for_each_mem_cgroup_tree(iter, memcg)
1873 if (iter->under_oom > 0)
1875 spin_unlock(&memcg_oom_lock);
1878 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1880 struct oom_wait_info {
1881 struct mem_cgroup *memcg;
1882 wait_queue_entry_t wait;
1885 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1886 unsigned mode, int sync, void *arg)
1888 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1889 struct mem_cgroup *oom_wait_memcg;
1890 struct oom_wait_info *oom_wait_info;
1892 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1893 oom_wait_memcg = oom_wait_info->memcg;
1895 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1896 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1898 return autoremove_wake_function(wait, mode, sync, arg);
1901 static void memcg_oom_recover(struct mem_cgroup *memcg)
1904 * For the following lockless ->under_oom test, the only required
1905 * guarantee is that it must see the state asserted by an OOM when
1906 * this function is called as a result of userland actions
1907 * triggered by the notification of the OOM. This is trivially
1908 * achieved by invoking mem_cgroup_mark_under_oom() before
1909 * triggering notification.
1911 if (memcg && memcg->under_oom)
1912 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1922 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1924 enum oom_status ret;
1927 if (order > PAGE_ALLOC_COSTLY_ORDER)
1930 memcg_memory_event(memcg, MEMCG_OOM);
1933 * We are in the middle of the charge context here, so we
1934 * don't want to block when potentially sitting on a callstack
1935 * that holds all kinds of filesystem and mm locks.
1937 * cgroup1 allows disabling the OOM killer and waiting for outside
1938 * handling until the charge can succeed; remember the context and put
1939 * the task to sleep at the end of the page fault when all locks are
1942 * On the other hand, in-kernel OOM killer allows for an async victim
1943 * memory reclaim (oom_reaper) and that means that we are not solely
1944 * relying on the oom victim to make a forward progress and we can
1945 * invoke the oom killer here.
1947 * Please note that mem_cgroup_out_of_memory might fail to find a
1948 * victim and then we have to bail out from the charge path.
1950 if (memcg->oom_kill_disable) {
1951 if (!current->in_user_fault)
1953 css_get(&memcg->css);
1954 current->memcg_in_oom = memcg;
1955 current->memcg_oom_gfp_mask = mask;
1956 current->memcg_oom_order = order;
1961 mem_cgroup_mark_under_oom(memcg);
1963 locked = mem_cgroup_oom_trylock(memcg);
1966 mem_cgroup_oom_notify(memcg);
1968 mem_cgroup_unmark_under_oom(memcg);
1969 if (mem_cgroup_out_of_memory(memcg, mask, order))
1975 mem_cgroup_oom_unlock(memcg);
1981 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1982 * @handle: actually kill/wait or just clean up the OOM state
1984 * This has to be called at the end of a page fault if the memcg OOM
1985 * handler was enabled.
1987 * Memcg supports userspace OOM handling where failed allocations must
1988 * sleep on a waitqueue until the userspace task resolves the
1989 * situation. Sleeping directly in the charge context with all kinds
1990 * of locks held is not a good idea, instead we remember an OOM state
1991 * in the task and mem_cgroup_oom_synchronize() has to be called at
1992 * the end of the page fault to complete the OOM handling.
1994 * Returns %true if an ongoing memcg OOM situation was detected and
1995 * completed, %false otherwise.
1997 bool mem_cgroup_oom_synchronize(bool handle)
1999 struct mem_cgroup *memcg = current->memcg_in_oom;
2000 struct oom_wait_info owait;
2003 /* OOM is global, do not handle */
2010 owait.memcg = memcg;
2011 owait.wait.flags = 0;
2012 owait.wait.func = memcg_oom_wake_function;
2013 owait.wait.private = current;
2014 INIT_LIST_HEAD(&owait.wait.entry);
2016 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2017 mem_cgroup_mark_under_oom(memcg);
2019 locked = mem_cgroup_oom_trylock(memcg);
2022 mem_cgroup_oom_notify(memcg);
2024 if (locked && !memcg->oom_kill_disable) {
2025 mem_cgroup_unmark_under_oom(memcg);
2026 finish_wait(&memcg_oom_waitq, &owait.wait);
2027 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
2028 current->memcg_oom_order);
2031 mem_cgroup_unmark_under_oom(memcg);
2032 finish_wait(&memcg_oom_waitq, &owait.wait);
2036 mem_cgroup_oom_unlock(memcg);
2038 * There is no guarantee that an OOM-lock contender
2039 * sees the wakeups triggered by the OOM kill
2040 * uncharges. Wake any sleepers explicitely.
2042 memcg_oom_recover(memcg);
2045 current->memcg_in_oom = NULL;
2046 css_put(&memcg->css);
2051 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2052 * @victim: task to be killed by the OOM killer
2053 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2055 * Returns a pointer to a memory cgroup, which has to be cleaned up
2056 * by killing all belonging OOM-killable tasks.
2058 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2060 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2061 struct mem_cgroup *oom_domain)
2063 struct mem_cgroup *oom_group = NULL;
2064 struct mem_cgroup *memcg;
2066 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2070 oom_domain = root_mem_cgroup;
2074 memcg = mem_cgroup_from_task(victim);
2075 if (memcg == root_mem_cgroup)
2079 * If the victim task has been asynchronously moved to a different
2080 * memory cgroup, we might end up killing tasks outside oom_domain.
2081 * In this case it's better to ignore memory.group.oom.
2083 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2087 * Traverse the memory cgroup hierarchy from the victim task's
2088 * cgroup up to the OOMing cgroup (or root) to find the
2089 * highest-level memory cgroup with oom.group set.
2091 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2092 if (memcg->oom_group)
2095 if (memcg == oom_domain)
2100 css_get(&oom_group->css);
2107 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2109 pr_info("Tasks in ");
2110 pr_cont_cgroup_path(memcg->css.cgroup);
2111 pr_cont(" are going to be killed due to memory.oom.group set\n");
2115 * lock_page_memcg - lock a page->mem_cgroup binding
2118 * This function protects unlocked LRU pages from being moved to
2121 * It ensures lifetime of the returned memcg. Caller is responsible
2122 * for the lifetime of the page; __unlock_page_memcg() is available
2123 * when @page might get freed inside the locked section.
2125 struct mem_cgroup *lock_page_memcg(struct page *page)
2127 struct page *head = compound_head(page); /* rmap on tail pages */
2128 struct mem_cgroup *memcg;
2129 unsigned long flags;
2132 * The RCU lock is held throughout the transaction. The fast
2133 * path can get away without acquiring the memcg->move_lock
2134 * because page moving starts with an RCU grace period.
2136 * The RCU lock also protects the memcg from being freed when
2137 * the page state that is going to change is the only thing
2138 * preventing the page itself from being freed. E.g. writeback
2139 * doesn't hold a page reference and relies on PG_writeback to
2140 * keep off truncation, migration and so forth.
2144 if (mem_cgroup_disabled())
2147 memcg = head->mem_cgroup;
2148 if (unlikely(!memcg))
2151 if (atomic_read(&memcg->moving_account) <= 0)
2154 spin_lock_irqsave(&memcg->move_lock, flags);
2155 if (memcg != head->mem_cgroup) {
2156 spin_unlock_irqrestore(&memcg->move_lock, flags);
2161 * When charge migration first begins, we can have locked and
2162 * unlocked page stat updates happening concurrently. Track
2163 * the task who has the lock for unlock_page_memcg().
2165 memcg->move_lock_task = current;
2166 memcg->move_lock_flags = flags;
2170 EXPORT_SYMBOL(lock_page_memcg);
2173 * __unlock_page_memcg - unlock and unpin a memcg
2176 * Unlock and unpin a memcg returned by lock_page_memcg().
2178 void __unlock_page_memcg(struct mem_cgroup *memcg)
2180 if (memcg && memcg->move_lock_task == current) {
2181 unsigned long flags = memcg->move_lock_flags;
2183 memcg->move_lock_task = NULL;
2184 memcg->move_lock_flags = 0;
2186 spin_unlock_irqrestore(&memcg->move_lock, flags);
2193 * unlock_page_memcg - unlock a page->mem_cgroup binding
2196 void unlock_page_memcg(struct page *page)
2198 struct page *head = compound_head(page);
2200 __unlock_page_memcg(head->mem_cgroup);
2202 EXPORT_SYMBOL(unlock_page_memcg);
2204 struct memcg_stock_pcp {
2205 struct mem_cgroup *cached; /* this never be root cgroup */
2206 unsigned int nr_pages;
2208 #ifdef CONFIG_MEMCG_KMEM
2209 struct obj_cgroup *cached_objcg;
2210 unsigned int nr_bytes;
2213 struct work_struct work;
2214 unsigned long flags;
2215 #define FLUSHING_CACHED_CHARGE 0
2217 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2218 static DEFINE_MUTEX(percpu_charge_mutex);
2220 #ifdef CONFIG_MEMCG_KMEM
2221 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2222 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2223 struct mem_cgroup *root_memcg);
2226 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2229 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2230 struct mem_cgroup *root_memcg)
2237 * consume_stock: Try to consume stocked charge on this cpu.
2238 * @memcg: memcg to consume from.
2239 * @nr_pages: how many pages to charge.
2241 * The charges will only happen if @memcg matches the current cpu's memcg
2242 * stock, and at least @nr_pages are available in that stock. Failure to
2243 * service an allocation will refill the stock.
2245 * returns true if successful, false otherwise.
2247 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2249 struct memcg_stock_pcp *stock;
2250 unsigned long flags;
2253 if (nr_pages > MEMCG_CHARGE_BATCH)
2256 local_irq_save(flags);
2258 stock = this_cpu_ptr(&memcg_stock);
2259 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2260 stock->nr_pages -= nr_pages;
2264 local_irq_restore(flags);
2270 * Returns stocks cached in percpu and reset cached information.
2272 static void drain_stock(struct memcg_stock_pcp *stock)
2274 struct mem_cgroup *old = stock->cached;
2279 if (stock->nr_pages) {
2280 page_counter_uncharge(&old->memory, stock->nr_pages);
2281 if (do_memsw_account())
2282 page_counter_uncharge(&old->memsw, stock->nr_pages);
2283 stock->nr_pages = 0;
2287 stock->cached = NULL;
2290 static void drain_local_stock(struct work_struct *dummy)
2292 struct memcg_stock_pcp *stock;
2293 unsigned long flags;
2296 * The only protection from memory hotplug vs. drain_stock races is
2297 * that we always operate on local CPU stock here with IRQ disabled
2299 local_irq_save(flags);
2301 stock = this_cpu_ptr(&memcg_stock);
2302 drain_obj_stock(stock);
2304 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2306 local_irq_restore(flags);
2310 * Cache charges(val) to local per_cpu area.
2311 * This will be consumed by consume_stock() function, later.
2313 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2315 struct memcg_stock_pcp *stock;
2316 unsigned long flags;
2318 local_irq_save(flags);
2320 stock = this_cpu_ptr(&memcg_stock);
2321 if (stock->cached != memcg) { /* reset if necessary */
2323 css_get(&memcg->css);
2324 stock->cached = memcg;
2326 stock->nr_pages += nr_pages;
2328 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2331 local_irq_restore(flags);
2335 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2336 * of the hierarchy under it.
2338 static void drain_all_stock(struct mem_cgroup *root_memcg)
2342 /* If someone's already draining, avoid adding running more workers. */
2343 if (!mutex_trylock(&percpu_charge_mutex))
2346 * Notify other cpus that system-wide "drain" is running
2347 * We do not care about races with the cpu hotplug because cpu down
2348 * as well as workers from this path always operate on the local
2349 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2352 for_each_online_cpu(cpu) {
2353 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2354 struct mem_cgroup *memcg;
2358 memcg = stock->cached;
2359 if (memcg && stock->nr_pages &&
2360 mem_cgroup_is_descendant(memcg, root_memcg))
2362 if (obj_stock_flush_required(stock, root_memcg))
2367 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2369 drain_local_stock(&stock->work);
2371 schedule_work_on(cpu, &stock->work);
2375 mutex_unlock(&percpu_charge_mutex);
2378 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2380 struct memcg_stock_pcp *stock;
2381 struct mem_cgroup *memcg, *mi;
2383 stock = &per_cpu(memcg_stock, cpu);
2386 for_each_mem_cgroup(memcg) {
2389 for (i = 0; i < MEMCG_NR_STAT; i++) {
2393 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2395 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2396 atomic_long_add(x, &memcg->vmstats[i]);
2398 if (i >= NR_VM_NODE_STAT_ITEMS)
2401 for_each_node(nid) {
2402 struct mem_cgroup_per_node *pn;
2404 pn = mem_cgroup_nodeinfo(memcg, nid);
2405 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2408 atomic_long_add(x, &pn->lruvec_stat[i]);
2409 } while ((pn = parent_nodeinfo(pn, nid)));
2413 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2416 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2418 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2419 atomic_long_add(x, &memcg->vmevents[i]);
2426 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2427 unsigned int nr_pages,
2430 unsigned long nr_reclaimed = 0;
2433 unsigned long pflags;
2435 if (page_counter_read(&memcg->memory) <=
2436 READ_ONCE(memcg->memory.high))
2439 memcg_memory_event(memcg, MEMCG_HIGH);
2441 psi_memstall_enter(&pflags);
2442 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2444 psi_memstall_leave(&pflags);
2445 } while ((memcg = parent_mem_cgroup(memcg)) &&
2446 !mem_cgroup_is_root(memcg));
2448 return nr_reclaimed;
2451 static void high_work_func(struct work_struct *work)
2453 struct mem_cgroup *memcg;
2455 memcg = container_of(work, struct mem_cgroup, high_work);
2456 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2460 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2461 * enough to still cause a significant slowdown in most cases, while still
2462 * allowing diagnostics and tracing to proceed without becoming stuck.
2464 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2467 * When calculating the delay, we use these either side of the exponentiation to
2468 * maintain precision and scale to a reasonable number of jiffies (see the table
2471 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2472 * overage ratio to a delay.
2473 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2474 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2475 * to produce a reasonable delay curve.
2477 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2478 * reasonable delay curve compared to precision-adjusted overage, not
2479 * penalising heavily at first, but still making sure that growth beyond the
2480 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2481 * example, with a high of 100 megabytes:
2483 * +-------+------------------------+
2484 * | usage | time to allocate in ms |
2485 * +-------+------------------------+
2507 * +-------+------------------------+
2509 #define MEMCG_DELAY_PRECISION_SHIFT 20
2510 #define MEMCG_DELAY_SCALING_SHIFT 14
2512 static u64 calculate_overage(unsigned long usage, unsigned long high)
2520 * Prevent division by 0 in overage calculation by acting as if
2521 * it was a threshold of 1 page
2523 high = max(high, 1UL);
2525 overage = usage - high;
2526 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2527 return div64_u64(overage, high);
2530 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2532 u64 overage, max_overage = 0;
2535 overage = calculate_overage(page_counter_read(&memcg->memory),
2536 READ_ONCE(memcg->memory.high));
2537 max_overage = max(overage, max_overage);
2538 } while ((memcg = parent_mem_cgroup(memcg)) &&
2539 !mem_cgroup_is_root(memcg));
2544 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2546 u64 overage, max_overage = 0;
2549 overage = calculate_overage(page_counter_read(&memcg->swap),
2550 READ_ONCE(memcg->swap.high));
2552 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2553 max_overage = max(overage, max_overage);
2554 } while ((memcg = parent_mem_cgroup(memcg)) &&
2555 !mem_cgroup_is_root(memcg));
2561 * Get the number of jiffies that we should penalise a mischievous cgroup which
2562 * is exceeding its memory.high by checking both it and its ancestors.
2564 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2565 unsigned int nr_pages,
2568 unsigned long penalty_jiffies;
2574 * We use overage compared to memory.high to calculate the number of
2575 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2576 * fairly lenient on small overages, and increasingly harsh when the
2577 * memcg in question makes it clear that it has no intention of stopping
2578 * its crazy behaviour, so we exponentially increase the delay based on
2581 penalty_jiffies = max_overage * max_overage * HZ;
2582 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2583 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2586 * Factor in the task's own contribution to the overage, such that four
2587 * N-sized allocations are throttled approximately the same as one
2588 * 4N-sized allocation.
2590 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2591 * larger the current charge patch is than that.
2593 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2597 * Scheduled by try_charge() to be executed from the userland return path
2598 * and reclaims memory over the high limit.
2600 void mem_cgroup_handle_over_high(void)
2602 unsigned long penalty_jiffies;
2603 unsigned long pflags;
2604 unsigned long nr_reclaimed;
2605 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2606 int nr_retries = MAX_RECLAIM_RETRIES;
2607 struct mem_cgroup *memcg;
2608 bool in_retry = false;
2610 if (likely(!nr_pages))
2613 memcg = get_mem_cgroup_from_mm(current->mm);
2614 current->memcg_nr_pages_over_high = 0;
2618 * The allocating task should reclaim at least the batch size, but for
2619 * subsequent retries we only want to do what's necessary to prevent oom
2620 * or breaching resource isolation.
2622 * This is distinct from memory.max or page allocator behaviour because
2623 * memory.high is currently batched, whereas memory.max and the page
2624 * allocator run every time an allocation is made.
2626 nr_reclaimed = reclaim_high(memcg,
2627 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2631 * memory.high is breached and reclaim is unable to keep up. Throttle
2632 * allocators proactively to slow down excessive growth.
2634 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2635 mem_find_max_overage(memcg));
2637 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2638 swap_find_max_overage(memcg));
2641 * Clamp the max delay per usermode return so as to still keep the
2642 * application moving forwards and also permit diagnostics, albeit
2645 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2648 * Don't sleep if the amount of jiffies this memcg owes us is so low
2649 * that it's not even worth doing, in an attempt to be nice to those who
2650 * go only a small amount over their memory.high value and maybe haven't
2651 * been aggressively reclaimed enough yet.
2653 if (penalty_jiffies <= HZ / 100)
2657 * If reclaim is making forward progress but we're still over
2658 * memory.high, we want to encourage that rather than doing allocator
2661 if (nr_reclaimed || nr_retries--) {
2667 * If we exit early, we're guaranteed to die (since
2668 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2669 * need to account for any ill-begotten jiffies to pay them off later.
2671 psi_memstall_enter(&pflags);
2672 schedule_timeout_killable(penalty_jiffies);
2673 psi_memstall_leave(&pflags);
2676 css_put(&memcg->css);
2679 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2680 unsigned int nr_pages)
2682 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2683 int nr_retries = MAX_RECLAIM_RETRIES;
2684 struct mem_cgroup *mem_over_limit;
2685 struct page_counter *counter;
2686 enum oom_status oom_status;
2687 unsigned long nr_reclaimed;
2688 bool may_swap = true;
2689 bool drained = false;
2690 unsigned long pflags;
2692 if (mem_cgroup_is_root(memcg))
2695 if (consume_stock(memcg, nr_pages))
2698 if (!do_memsw_account() ||
2699 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2700 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2702 if (do_memsw_account())
2703 page_counter_uncharge(&memcg->memsw, batch);
2704 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2706 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2710 if (batch > nr_pages) {
2716 * Memcg doesn't have a dedicated reserve for atomic
2717 * allocations. But like the global atomic pool, we need to
2718 * put the burden of reclaim on regular allocation requests
2719 * and let these go through as privileged allocations.
2721 if (gfp_mask & __GFP_ATOMIC)
2725 * Unlike in global OOM situations, memcg is not in a physical
2726 * memory shortage. Allow dying and OOM-killed tasks to
2727 * bypass the last charges so that they can exit quickly and
2728 * free their memory.
2730 if (unlikely(should_force_charge()))
2734 * Prevent unbounded recursion when reclaim operations need to
2735 * allocate memory. This might exceed the limits temporarily,
2736 * but we prefer facilitating memory reclaim and getting back
2737 * under the limit over triggering OOM kills in these cases.
2739 if (unlikely(current->flags & PF_MEMALLOC))
2742 if (unlikely(task_in_memcg_oom(current)))
2745 if (!gfpflags_allow_blocking(gfp_mask))
2748 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2750 psi_memstall_enter(&pflags);
2751 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2752 gfp_mask, may_swap);
2753 psi_memstall_leave(&pflags);
2755 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2759 drain_all_stock(mem_over_limit);
2764 if (gfp_mask & __GFP_NORETRY)
2767 * Even though the limit is exceeded at this point, reclaim
2768 * may have been able to free some pages. Retry the charge
2769 * before killing the task.
2771 * Only for regular pages, though: huge pages are rather
2772 * unlikely to succeed so close to the limit, and we fall back
2773 * to regular pages anyway in case of failure.
2775 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2778 * At task move, charge accounts can be doubly counted. So, it's
2779 * better to wait until the end of task_move if something is going on.
2781 if (mem_cgroup_wait_acct_move(mem_over_limit))
2787 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2790 if (gfp_mask & __GFP_NOFAIL)
2793 if (fatal_signal_pending(current))
2797 * keep retrying as long as the memcg oom killer is able to make
2798 * a forward progress or bypass the charge if the oom killer
2799 * couldn't make any progress.
2801 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2802 get_order(nr_pages * PAGE_SIZE));
2803 switch (oom_status) {
2805 nr_retries = MAX_RECLAIM_RETRIES;
2813 if (!(gfp_mask & __GFP_NOFAIL))
2817 * The allocation either can't fail or will lead to more memory
2818 * being freed very soon. Allow memory usage go over the limit
2819 * temporarily by force charging it.
2821 page_counter_charge(&memcg->memory, nr_pages);
2822 if (do_memsw_account())
2823 page_counter_charge(&memcg->memsw, nr_pages);
2828 if (batch > nr_pages)
2829 refill_stock(memcg, batch - nr_pages);
2832 * If the hierarchy is above the normal consumption range, schedule
2833 * reclaim on returning to userland. We can perform reclaim here
2834 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2835 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2836 * not recorded as it most likely matches current's and won't
2837 * change in the meantime. As high limit is checked again before
2838 * reclaim, the cost of mismatch is negligible.
2841 bool mem_high, swap_high;
2843 mem_high = page_counter_read(&memcg->memory) >
2844 READ_ONCE(memcg->memory.high);
2845 swap_high = page_counter_read(&memcg->swap) >
2846 READ_ONCE(memcg->swap.high);
2848 /* Don't bother a random interrupted task */
2849 if (in_interrupt()) {
2851 schedule_work(&memcg->high_work);
2857 if (mem_high || swap_high) {
2859 * The allocating tasks in this cgroup will need to do
2860 * reclaim or be throttled to prevent further growth
2861 * of the memory or swap footprints.
2863 * Target some best-effort fairness between the tasks,
2864 * and distribute reclaim work and delay penalties
2865 * based on how much each task is actually allocating.
2867 current->memcg_nr_pages_over_high += batch;
2868 set_notify_resume(current);
2871 } while ((memcg = parent_mem_cgroup(memcg)));
2876 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2877 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2879 if (mem_cgroup_is_root(memcg))
2882 page_counter_uncharge(&memcg->memory, nr_pages);
2883 if (do_memsw_account())
2884 page_counter_uncharge(&memcg->memsw, nr_pages);
2888 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2890 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2892 * Any of the following ensures page's memcg stability:
2896 * - lock_page_memcg()
2897 * - exclusive reference
2899 page->mem_cgroup = memcg;
2902 #ifdef CONFIG_MEMCG_KMEM
2903 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2906 unsigned int objects = objs_per_slab_page(s, page);
2909 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2914 if (cmpxchg(&page->obj_cgroups, NULL,
2915 (struct obj_cgroup **) ((unsigned long)vec | 0x1UL)))
2918 kmemleak_not_leak(vec);
2924 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2926 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2927 * cgroup_mutex, etc.
2929 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2933 if (mem_cgroup_disabled())
2936 page = virt_to_head_page(p);
2939 * If page->mem_cgroup is set, it's either a simple mem_cgroup pointer
2940 * or a pointer to obj_cgroup vector. In the latter case the lowest
2941 * bit of the pointer is set.
2942 * The page->mem_cgroup pointer can be asynchronously changed
2943 * from NULL to (obj_cgroup_vec | 0x1UL), but can't be changed
2944 * from a valid memcg pointer to objcg vector or back.
2946 if (!page->mem_cgroup)
2950 * Slab objects are accounted individually, not per-page.
2951 * Memcg membership data for each individual object is saved in
2952 * the page->obj_cgroups.
2954 if (page_has_obj_cgroups(page)) {
2955 struct obj_cgroup *objcg;
2958 off = obj_to_index(page->slab_cache, page, p);
2959 objcg = page_obj_cgroups(page)[off];
2961 return obj_cgroup_memcg(objcg);
2966 /* All other pages use page->mem_cgroup */
2967 return page->mem_cgroup;
2970 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2972 struct obj_cgroup *objcg = NULL;
2973 struct mem_cgroup *memcg;
2975 if (memcg_kmem_bypass())
2979 if (unlikely(active_memcg()))
2980 memcg = active_memcg();
2982 memcg = mem_cgroup_from_task(current);
2984 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2985 objcg = rcu_dereference(memcg->objcg);
2986 if (objcg && obj_cgroup_tryget(objcg))
2995 static int memcg_alloc_cache_id(void)
3000 id = ida_simple_get(&memcg_cache_ida,
3001 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3005 if (id < memcg_nr_cache_ids)
3009 * There's no space for the new id in memcg_caches arrays,
3010 * so we have to grow them.
3012 down_write(&memcg_cache_ids_sem);
3014 size = 2 * (id + 1);
3015 if (size < MEMCG_CACHES_MIN_SIZE)
3016 size = MEMCG_CACHES_MIN_SIZE;
3017 else if (size > MEMCG_CACHES_MAX_SIZE)
3018 size = MEMCG_CACHES_MAX_SIZE;
3020 err = memcg_update_all_list_lrus(size);
3022 memcg_nr_cache_ids = size;
3024 up_write(&memcg_cache_ids_sem);
3027 ida_simple_remove(&memcg_cache_ida, id);
3033 static void memcg_free_cache_id(int id)
3035 ida_simple_remove(&memcg_cache_ida, id);
3039 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3040 * @memcg: memory cgroup to charge
3041 * @gfp: reclaim mode
3042 * @nr_pages: number of pages to charge
3044 * Returns 0 on success, an error code on failure.
3046 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
3047 unsigned int nr_pages)
3049 struct page_counter *counter;
3052 ret = try_charge(memcg, gfp, nr_pages);
3056 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3057 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3060 * Enforce __GFP_NOFAIL allocation because callers are not
3061 * prepared to see failures and likely do not have any failure
3064 if (gfp & __GFP_NOFAIL) {
3065 page_counter_charge(&memcg->kmem, nr_pages);
3068 cancel_charge(memcg, nr_pages);
3075 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3076 * @memcg: memcg to uncharge
3077 * @nr_pages: number of pages to uncharge
3079 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3081 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3082 page_counter_uncharge(&memcg->kmem, nr_pages);
3084 page_counter_uncharge(&memcg->memory, nr_pages);
3085 if (do_memsw_account())
3086 page_counter_uncharge(&memcg->memsw, nr_pages);
3090 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3091 * @page: page to charge
3092 * @gfp: reclaim mode
3093 * @order: allocation order
3095 * Returns 0 on success, an error code on failure.
3097 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3099 struct mem_cgroup *memcg;
3102 memcg = get_mem_cgroup_from_current();
3103 if (memcg && !mem_cgroup_is_root(memcg)) {
3104 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3106 page->mem_cgroup = memcg;
3107 __SetPageKmemcg(page);
3110 css_put(&memcg->css);
3116 * __memcg_kmem_uncharge_page: uncharge a kmem page
3117 * @page: page to uncharge
3118 * @order: allocation order
3120 void __memcg_kmem_uncharge_page(struct page *page, int order)
3122 struct mem_cgroup *memcg = page->mem_cgroup;
3123 unsigned int nr_pages = 1 << order;
3128 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3129 __memcg_kmem_uncharge(memcg, nr_pages);
3130 page->mem_cgroup = NULL;
3131 css_put(&memcg->css);
3133 /* slab pages do not have PageKmemcg flag set */
3134 if (PageKmemcg(page))
3135 __ClearPageKmemcg(page);
3138 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3140 struct memcg_stock_pcp *stock;
3141 unsigned long flags;
3144 local_irq_save(flags);
3146 stock = this_cpu_ptr(&memcg_stock);
3147 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3148 stock->nr_bytes -= nr_bytes;
3152 local_irq_restore(flags);
3157 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3159 struct obj_cgroup *old = stock->cached_objcg;
3164 if (stock->nr_bytes) {
3165 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3166 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3170 __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
3175 * The leftover is flushed to the centralized per-memcg value.
3176 * On the next attempt to refill obj stock it will be moved
3177 * to a per-cpu stock (probably, on an other CPU), see
3178 * refill_obj_stock().
3180 * How often it's flushed is a trade-off between the memory
3181 * limit enforcement accuracy and potential CPU contention,
3182 * so it might be changed in the future.
3184 atomic_add(nr_bytes, &old->nr_charged_bytes);
3185 stock->nr_bytes = 0;
3188 obj_cgroup_put(old);
3189 stock->cached_objcg = NULL;
3192 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3193 struct mem_cgroup *root_memcg)
3195 struct mem_cgroup *memcg;
3197 if (stock->cached_objcg) {
3198 memcg = obj_cgroup_memcg(stock->cached_objcg);
3199 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3206 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3208 struct memcg_stock_pcp *stock;
3209 unsigned long flags;
3211 local_irq_save(flags);
3213 stock = this_cpu_ptr(&memcg_stock);
3214 if (stock->cached_objcg != objcg) { /* reset if necessary */
3215 drain_obj_stock(stock);
3216 obj_cgroup_get(objcg);
3217 stock->cached_objcg = objcg;
3218 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3220 stock->nr_bytes += nr_bytes;
3222 if (stock->nr_bytes > PAGE_SIZE)
3223 drain_obj_stock(stock);
3225 local_irq_restore(flags);
3228 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3230 struct mem_cgroup *memcg;
3231 unsigned int nr_pages, nr_bytes;
3234 if (consume_obj_stock(objcg, size))
3238 * In theory, memcg->nr_charged_bytes can have enough
3239 * pre-charged bytes to satisfy the allocation. However,
3240 * flushing memcg->nr_charged_bytes requires two atomic
3241 * operations, and memcg->nr_charged_bytes can't be big,
3242 * so it's better to ignore it and try grab some new pages.
3243 * memcg->nr_charged_bytes will be flushed in
3244 * refill_obj_stock(), called from this function or
3245 * independently later.
3249 memcg = obj_cgroup_memcg(objcg);
3250 if (unlikely(!css_tryget(&memcg->css)))
3254 nr_pages = size >> PAGE_SHIFT;
3255 nr_bytes = size & (PAGE_SIZE - 1);
3260 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3261 if (!ret && nr_bytes)
3262 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3264 css_put(&memcg->css);
3268 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3270 refill_obj_stock(objcg, size);
3273 #endif /* CONFIG_MEMCG_KMEM */
3275 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3278 * Because tail pages are not marked as "used", set it. We're under
3279 * pgdat->lru_lock and migration entries setup in all page mappings.
3281 void mem_cgroup_split_huge_fixup(struct page *head)
3283 struct mem_cgroup *memcg = head->mem_cgroup;
3286 if (mem_cgroup_disabled())
3289 for (i = 1; i < HPAGE_PMD_NR; i++) {
3290 css_get(&memcg->css);
3291 head[i].mem_cgroup = memcg;
3294 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3296 #ifdef CONFIG_MEMCG_SWAP
3298 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3299 * @entry: swap entry to be moved
3300 * @from: mem_cgroup which the entry is moved from
3301 * @to: mem_cgroup which the entry is moved to
3303 * It succeeds only when the swap_cgroup's record for this entry is the same
3304 * as the mem_cgroup's id of @from.
3306 * Returns 0 on success, -EINVAL on failure.
3308 * The caller must have charged to @to, IOW, called page_counter_charge() about
3309 * both res and memsw, and called css_get().
3311 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3312 struct mem_cgroup *from, struct mem_cgroup *to)
3314 unsigned short old_id, new_id;
3316 old_id = mem_cgroup_id(from);
3317 new_id = mem_cgroup_id(to);
3319 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3320 mod_memcg_state(from, MEMCG_SWAP, -1);
3321 mod_memcg_state(to, MEMCG_SWAP, 1);
3327 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3328 struct mem_cgroup *from, struct mem_cgroup *to)
3334 static DEFINE_MUTEX(memcg_max_mutex);
3336 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3337 unsigned long max, bool memsw)
3339 bool enlarge = false;
3340 bool drained = false;
3342 bool limits_invariant;
3343 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3346 if (signal_pending(current)) {
3351 mutex_lock(&memcg_max_mutex);
3353 * Make sure that the new limit (memsw or memory limit) doesn't
3354 * break our basic invariant rule memory.max <= memsw.max.
3356 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3357 max <= memcg->memsw.max;
3358 if (!limits_invariant) {
3359 mutex_unlock(&memcg_max_mutex);
3363 if (max > counter->max)
3365 ret = page_counter_set_max(counter, max);
3366 mutex_unlock(&memcg_max_mutex);
3372 drain_all_stock(memcg);
3377 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3378 GFP_KERNEL, !memsw)) {
3384 if (!ret && enlarge)
3385 memcg_oom_recover(memcg);
3390 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3392 unsigned long *total_scanned)
3394 unsigned long nr_reclaimed = 0;
3395 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3396 unsigned long reclaimed;
3398 struct mem_cgroup_tree_per_node *mctz;
3399 unsigned long excess;
3400 unsigned long nr_scanned;
3405 mctz = soft_limit_tree_node(pgdat->node_id);
3408 * Do not even bother to check the largest node if the root
3409 * is empty. Do it lockless to prevent lock bouncing. Races
3410 * are acceptable as soft limit is best effort anyway.
3412 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3416 * This loop can run a while, specially if mem_cgroup's continuously
3417 * keep exceeding their soft limit and putting the system under
3424 mz = mem_cgroup_largest_soft_limit_node(mctz);
3429 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3430 gfp_mask, &nr_scanned);
3431 nr_reclaimed += reclaimed;
3432 *total_scanned += nr_scanned;
3433 spin_lock_irq(&mctz->lock);
3434 __mem_cgroup_remove_exceeded(mz, mctz);
3437 * If we failed to reclaim anything from this memory cgroup
3438 * it is time to move on to the next cgroup
3442 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3444 excess = soft_limit_excess(mz->memcg);
3446 * One school of thought says that we should not add
3447 * back the node to the tree if reclaim returns 0.
3448 * But our reclaim could return 0, simply because due
3449 * to priority we are exposing a smaller subset of
3450 * memory to reclaim from. Consider this as a longer
3453 /* If excess == 0, no tree ops */
3454 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3455 spin_unlock_irq(&mctz->lock);
3456 css_put(&mz->memcg->css);
3459 * Could not reclaim anything and there are no more
3460 * mem cgroups to try or we seem to be looping without
3461 * reclaiming anything.
3463 if (!nr_reclaimed &&
3465 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3467 } while (!nr_reclaimed);
3469 css_put(&next_mz->memcg->css);
3470 return nr_reclaimed;
3474 * Reclaims as many pages from the given memcg as possible.
3476 * Caller is responsible for holding css reference for memcg.
3478 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3480 int nr_retries = MAX_RECLAIM_RETRIES;
3482 /* we call try-to-free pages for make this cgroup empty */
3483 lru_add_drain_all();
3485 drain_all_stock(memcg);
3487 /* try to free all pages in this cgroup */
3488 while (nr_retries && page_counter_read(&memcg->memory)) {
3491 if (signal_pending(current))
3494 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3498 /* maybe some writeback is necessary */
3499 congestion_wait(BLK_RW_ASYNC, HZ/10);
3507 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3508 char *buf, size_t nbytes,
3511 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3513 if (mem_cgroup_is_root(memcg))
3515 return mem_cgroup_force_empty(memcg) ?: nbytes;
3518 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3524 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3525 struct cftype *cft, u64 val)
3530 pr_warn_once("Non-hierarchical mode is deprecated. "
3531 "Please report your usecase to linux-mm@kvack.org if you "
3532 "depend on this functionality.\n");
3537 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3541 if (mem_cgroup_is_root(memcg)) {
3542 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3543 memcg_page_state(memcg, NR_ANON_MAPPED);
3545 val += memcg_page_state(memcg, MEMCG_SWAP);
3548 val = page_counter_read(&memcg->memory);
3550 val = page_counter_read(&memcg->memsw);
3563 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3566 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3567 struct page_counter *counter;
3569 switch (MEMFILE_TYPE(cft->private)) {
3571 counter = &memcg->memory;
3574 counter = &memcg->memsw;
3577 counter = &memcg->kmem;
3580 counter = &memcg->tcpmem;
3586 switch (MEMFILE_ATTR(cft->private)) {
3588 if (counter == &memcg->memory)
3589 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3590 if (counter == &memcg->memsw)
3591 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3592 return (u64)page_counter_read(counter) * PAGE_SIZE;
3594 return (u64)counter->max * PAGE_SIZE;
3596 return (u64)counter->watermark * PAGE_SIZE;
3598 return counter->failcnt;
3599 case RES_SOFT_LIMIT:
3600 return (u64)memcg->soft_limit * PAGE_SIZE;
3606 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3608 unsigned long stat[MEMCG_NR_STAT] = {0};
3609 struct mem_cgroup *mi;
3612 for_each_online_cpu(cpu)
3613 for (i = 0; i < MEMCG_NR_STAT; i++)
3614 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3616 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3617 for (i = 0; i < MEMCG_NR_STAT; i++)
3618 atomic_long_add(stat[i], &mi->vmstats[i]);
3620 for_each_node(node) {
3621 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3622 struct mem_cgroup_per_node *pi;
3624 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3627 for_each_online_cpu(cpu)
3628 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3630 pn->lruvec_stat_cpu->count[i], cpu);
3632 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3633 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3634 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3638 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3640 unsigned long events[NR_VM_EVENT_ITEMS];
3641 struct mem_cgroup *mi;
3644 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3647 for_each_online_cpu(cpu)
3648 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3649 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3652 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3653 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3654 atomic_long_add(events[i], &mi->vmevents[i]);
3657 #ifdef CONFIG_MEMCG_KMEM
3658 static int memcg_online_kmem(struct mem_cgroup *memcg)
3660 struct obj_cgroup *objcg;
3663 if (cgroup_memory_nokmem)
3666 BUG_ON(memcg->kmemcg_id >= 0);
3667 BUG_ON(memcg->kmem_state);
3669 memcg_id = memcg_alloc_cache_id();
3673 objcg = obj_cgroup_alloc();
3675 memcg_free_cache_id(memcg_id);
3678 objcg->memcg = memcg;
3679 rcu_assign_pointer(memcg->objcg, objcg);
3681 static_branch_enable(&memcg_kmem_enabled_key);
3683 memcg->kmemcg_id = memcg_id;
3684 memcg->kmem_state = KMEM_ONLINE;
3689 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3691 struct cgroup_subsys_state *css;
3692 struct mem_cgroup *parent, *child;
3695 if (memcg->kmem_state != KMEM_ONLINE)
3698 memcg->kmem_state = KMEM_ALLOCATED;
3700 parent = parent_mem_cgroup(memcg);
3702 parent = root_mem_cgroup;
3704 memcg_reparent_objcgs(memcg, parent);
3706 kmemcg_id = memcg->kmemcg_id;
3707 BUG_ON(kmemcg_id < 0);
3710 * Change kmemcg_id of this cgroup and all its descendants to the
3711 * parent's id, and then move all entries from this cgroup's list_lrus
3712 * to ones of the parent. After we have finished, all list_lrus
3713 * corresponding to this cgroup are guaranteed to remain empty. The
3714 * ordering is imposed by list_lru_node->lock taken by
3715 * memcg_drain_all_list_lrus().
3717 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3718 css_for_each_descendant_pre(css, &memcg->css) {
3719 child = mem_cgroup_from_css(css);
3720 BUG_ON(child->kmemcg_id != kmemcg_id);
3721 child->kmemcg_id = parent->kmemcg_id;
3725 memcg_drain_all_list_lrus(kmemcg_id, parent);
3727 memcg_free_cache_id(kmemcg_id);
3730 static void memcg_free_kmem(struct mem_cgroup *memcg)
3732 /* css_alloc() failed, offlining didn't happen */
3733 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3734 memcg_offline_kmem(memcg);
3737 static int memcg_online_kmem(struct mem_cgroup *memcg)
3741 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3744 static void memcg_free_kmem(struct mem_cgroup *memcg)
3747 #endif /* CONFIG_MEMCG_KMEM */
3749 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3754 mutex_lock(&memcg_max_mutex);
3755 ret = page_counter_set_max(&memcg->kmem, max);
3756 mutex_unlock(&memcg_max_mutex);
3760 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3764 mutex_lock(&memcg_max_mutex);
3766 ret = page_counter_set_max(&memcg->tcpmem, max);
3770 if (!memcg->tcpmem_active) {
3772 * The active flag needs to be written after the static_key
3773 * update. This is what guarantees that the socket activation
3774 * function is the last one to run. See mem_cgroup_sk_alloc()
3775 * for details, and note that we don't mark any socket as
3776 * belonging to this memcg until that flag is up.
3778 * We need to do this, because static_keys will span multiple
3779 * sites, but we can't control their order. If we mark a socket
3780 * as accounted, but the accounting functions are not patched in
3781 * yet, we'll lose accounting.
3783 * We never race with the readers in mem_cgroup_sk_alloc(),
3784 * because when this value change, the code to process it is not
3787 static_branch_inc(&memcg_sockets_enabled_key);
3788 memcg->tcpmem_active = true;
3791 mutex_unlock(&memcg_max_mutex);
3796 * The user of this function is...
3799 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3800 char *buf, size_t nbytes, loff_t off)
3802 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3803 unsigned long nr_pages;
3806 buf = strstrip(buf);
3807 ret = page_counter_memparse(buf, "-1", &nr_pages);
3811 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3813 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3817 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3819 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3822 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3825 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3826 "Please report your usecase to linux-mm@kvack.org if you "
3827 "depend on this functionality.\n");
3828 ret = memcg_update_kmem_max(memcg, nr_pages);
3831 ret = memcg_update_tcp_max(memcg, nr_pages);
3835 case RES_SOFT_LIMIT:
3836 memcg->soft_limit = nr_pages;
3840 return ret ?: nbytes;
3843 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3844 size_t nbytes, loff_t off)
3846 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3847 struct page_counter *counter;
3849 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3851 counter = &memcg->memory;
3854 counter = &memcg->memsw;
3857 counter = &memcg->kmem;
3860 counter = &memcg->tcpmem;
3866 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3868 page_counter_reset_watermark(counter);
3871 counter->failcnt = 0;
3880 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3883 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3887 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3888 struct cftype *cft, u64 val)
3890 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3892 if (val & ~MOVE_MASK)
3896 * No kind of locking is needed in here, because ->can_attach() will
3897 * check this value once in the beginning of the process, and then carry
3898 * on with stale data. This means that changes to this value will only
3899 * affect task migrations starting after the change.
3901 memcg->move_charge_at_immigrate = val;
3905 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3906 struct cftype *cft, u64 val)
3914 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3915 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3916 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3918 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3919 int nid, unsigned int lru_mask, bool tree)
3921 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3922 unsigned long nr = 0;
3925 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3928 if (!(BIT(lru) & lru_mask))
3931 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3933 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3938 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3939 unsigned int lru_mask,
3942 unsigned long nr = 0;
3946 if (!(BIT(lru) & lru_mask))
3949 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3951 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3956 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3960 unsigned int lru_mask;
3963 static const struct numa_stat stats[] = {
3964 { "total", LRU_ALL },
3965 { "file", LRU_ALL_FILE },
3966 { "anon", LRU_ALL_ANON },
3967 { "unevictable", BIT(LRU_UNEVICTABLE) },
3969 const struct numa_stat *stat;
3971 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3973 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3974 seq_printf(m, "%s=%lu", stat->name,
3975 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3977 for_each_node_state(nid, N_MEMORY)
3978 seq_printf(m, " N%d=%lu", nid,
3979 mem_cgroup_node_nr_lru_pages(memcg, nid,
3980 stat->lru_mask, false));
3984 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3986 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3987 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3989 for_each_node_state(nid, N_MEMORY)
3990 seq_printf(m, " N%d=%lu", nid,
3991 mem_cgroup_node_nr_lru_pages(memcg, nid,
3992 stat->lru_mask, true));
3998 #endif /* CONFIG_NUMA */
4000 static const unsigned int memcg1_stats[] = {
4003 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4013 static const char *const memcg1_stat_names[] = {
4016 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4026 /* Universal VM events cgroup1 shows, original sort order */
4027 static const unsigned int memcg1_events[] = {
4034 static int memcg_stat_show(struct seq_file *m, void *v)
4036 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4037 unsigned long memory, memsw;
4038 struct mem_cgroup *mi;
4041 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4043 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4046 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4048 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4049 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4050 if (memcg1_stats[i] == NR_ANON_THPS)
4053 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4056 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4057 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4058 memcg_events_local(memcg, memcg1_events[i]));
4060 for (i = 0; i < NR_LRU_LISTS; i++)
4061 seq_printf(m, "%s %lu\n", lru_list_name(i),
4062 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4065 /* Hierarchical information */
4066 memory = memsw = PAGE_COUNTER_MAX;
4067 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4068 memory = min(memory, READ_ONCE(mi->memory.max));
4069 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4071 seq_printf(m, "hierarchical_memory_limit %llu\n",
4072 (u64)memory * PAGE_SIZE);
4073 if (do_memsw_account())
4074 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4075 (u64)memsw * PAGE_SIZE);
4077 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4080 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4082 nr = memcg_page_state(memcg, memcg1_stats[i]);
4083 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4084 if (memcg1_stats[i] == NR_ANON_THPS)
4087 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4088 (u64)nr * PAGE_SIZE);
4091 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4092 seq_printf(m, "total_%s %llu\n",
4093 vm_event_name(memcg1_events[i]),
4094 (u64)memcg_events(memcg, memcg1_events[i]));
4096 for (i = 0; i < NR_LRU_LISTS; i++)
4097 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4098 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4101 #ifdef CONFIG_DEBUG_VM
4104 struct mem_cgroup_per_node *mz;
4105 unsigned long anon_cost = 0;
4106 unsigned long file_cost = 0;
4108 for_each_online_pgdat(pgdat) {
4109 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4111 anon_cost += mz->lruvec.anon_cost;
4112 file_cost += mz->lruvec.file_cost;
4114 seq_printf(m, "anon_cost %lu\n", anon_cost);
4115 seq_printf(m, "file_cost %lu\n", file_cost);
4122 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4125 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4127 return mem_cgroup_swappiness(memcg);
4130 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4131 struct cftype *cft, u64 val)
4133 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4139 memcg->swappiness = val;
4141 vm_swappiness = val;
4146 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4148 struct mem_cgroup_threshold_ary *t;
4149 unsigned long usage;
4154 t = rcu_dereference(memcg->thresholds.primary);
4156 t = rcu_dereference(memcg->memsw_thresholds.primary);
4161 usage = mem_cgroup_usage(memcg, swap);
4164 * current_threshold points to threshold just below or equal to usage.
4165 * If it's not true, a threshold was crossed after last
4166 * call of __mem_cgroup_threshold().
4168 i = t->current_threshold;
4171 * Iterate backward over array of thresholds starting from
4172 * current_threshold and check if a threshold is crossed.
4173 * If none of thresholds below usage is crossed, we read
4174 * only one element of the array here.
4176 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4177 eventfd_signal(t->entries[i].eventfd, 1);
4179 /* i = current_threshold + 1 */
4183 * Iterate forward over array of thresholds starting from
4184 * current_threshold+1 and check if a threshold is crossed.
4185 * If none of thresholds above usage is crossed, we read
4186 * only one element of the array here.
4188 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4189 eventfd_signal(t->entries[i].eventfd, 1);
4191 /* Update current_threshold */
4192 t->current_threshold = i - 1;
4197 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4200 __mem_cgroup_threshold(memcg, false);
4201 if (do_memsw_account())
4202 __mem_cgroup_threshold(memcg, true);
4204 memcg = parent_mem_cgroup(memcg);
4208 static int compare_thresholds(const void *a, const void *b)
4210 const struct mem_cgroup_threshold *_a = a;
4211 const struct mem_cgroup_threshold *_b = b;
4213 if (_a->threshold > _b->threshold)
4216 if (_a->threshold < _b->threshold)
4222 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4224 struct mem_cgroup_eventfd_list *ev;
4226 spin_lock(&memcg_oom_lock);
4228 list_for_each_entry(ev, &memcg->oom_notify, list)
4229 eventfd_signal(ev->eventfd, 1);
4231 spin_unlock(&memcg_oom_lock);
4235 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4237 struct mem_cgroup *iter;
4239 for_each_mem_cgroup_tree(iter, memcg)
4240 mem_cgroup_oom_notify_cb(iter);
4243 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4244 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4246 struct mem_cgroup_thresholds *thresholds;
4247 struct mem_cgroup_threshold_ary *new;
4248 unsigned long threshold;
4249 unsigned long usage;
4252 ret = page_counter_memparse(args, "-1", &threshold);
4256 mutex_lock(&memcg->thresholds_lock);
4259 thresholds = &memcg->thresholds;
4260 usage = mem_cgroup_usage(memcg, false);
4261 } else if (type == _MEMSWAP) {
4262 thresholds = &memcg->memsw_thresholds;
4263 usage = mem_cgroup_usage(memcg, true);
4267 /* Check if a threshold crossed before adding a new one */
4268 if (thresholds->primary)
4269 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4271 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4273 /* Allocate memory for new array of thresholds */
4274 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4281 /* Copy thresholds (if any) to new array */
4282 if (thresholds->primary)
4283 memcpy(new->entries, thresholds->primary->entries,
4284 flex_array_size(new, entries, size - 1));
4286 /* Add new threshold */
4287 new->entries[size - 1].eventfd = eventfd;
4288 new->entries[size - 1].threshold = threshold;
4290 /* Sort thresholds. Registering of new threshold isn't time-critical */
4291 sort(new->entries, size, sizeof(*new->entries),
4292 compare_thresholds, NULL);
4294 /* Find current threshold */
4295 new->current_threshold = -1;
4296 for (i = 0; i < size; i++) {
4297 if (new->entries[i].threshold <= usage) {
4299 * new->current_threshold will not be used until
4300 * rcu_assign_pointer(), so it's safe to increment
4303 ++new->current_threshold;
4308 /* Free old spare buffer and save old primary buffer as spare */
4309 kfree(thresholds->spare);
4310 thresholds->spare = thresholds->primary;
4312 rcu_assign_pointer(thresholds->primary, new);
4314 /* To be sure that nobody uses thresholds */
4318 mutex_unlock(&memcg->thresholds_lock);
4323 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4324 struct eventfd_ctx *eventfd, const char *args)
4326 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4329 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4330 struct eventfd_ctx *eventfd, const char *args)
4332 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4335 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4336 struct eventfd_ctx *eventfd, enum res_type type)
4338 struct mem_cgroup_thresholds *thresholds;
4339 struct mem_cgroup_threshold_ary *new;
4340 unsigned long usage;
4341 int i, j, size, entries;
4343 mutex_lock(&memcg->thresholds_lock);
4346 thresholds = &memcg->thresholds;
4347 usage = mem_cgroup_usage(memcg, false);
4348 } else if (type == _MEMSWAP) {
4349 thresholds = &memcg->memsw_thresholds;
4350 usage = mem_cgroup_usage(memcg, true);
4354 if (!thresholds->primary)
4357 /* Check if a threshold crossed before removing */
4358 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4360 /* Calculate new number of threshold */
4362 for (i = 0; i < thresholds->primary->size; i++) {
4363 if (thresholds->primary->entries[i].eventfd != eventfd)
4369 new = thresholds->spare;
4371 /* If no items related to eventfd have been cleared, nothing to do */
4375 /* Set thresholds array to NULL if we don't have thresholds */
4384 /* Copy thresholds and find current threshold */
4385 new->current_threshold = -1;
4386 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4387 if (thresholds->primary->entries[i].eventfd == eventfd)
4390 new->entries[j] = thresholds->primary->entries[i];
4391 if (new->entries[j].threshold <= usage) {
4393 * new->current_threshold will not be used
4394 * until rcu_assign_pointer(), so it's safe to increment
4397 ++new->current_threshold;
4403 /* Swap primary and spare array */
4404 thresholds->spare = thresholds->primary;
4406 rcu_assign_pointer(thresholds->primary, new);
4408 /* To be sure that nobody uses thresholds */
4411 /* If all events are unregistered, free the spare array */
4413 kfree(thresholds->spare);
4414 thresholds->spare = NULL;
4417 mutex_unlock(&memcg->thresholds_lock);
4420 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4421 struct eventfd_ctx *eventfd)
4423 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4426 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4427 struct eventfd_ctx *eventfd)
4429 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4432 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4433 struct eventfd_ctx *eventfd, const char *args)
4435 struct mem_cgroup_eventfd_list *event;
4437 event = kmalloc(sizeof(*event), GFP_KERNEL);
4441 spin_lock(&memcg_oom_lock);
4443 event->eventfd = eventfd;
4444 list_add(&event->list, &memcg->oom_notify);
4446 /* already in OOM ? */
4447 if (memcg->under_oom)
4448 eventfd_signal(eventfd, 1);
4449 spin_unlock(&memcg_oom_lock);
4454 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4455 struct eventfd_ctx *eventfd)
4457 struct mem_cgroup_eventfd_list *ev, *tmp;
4459 spin_lock(&memcg_oom_lock);
4461 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4462 if (ev->eventfd == eventfd) {
4463 list_del(&ev->list);
4468 spin_unlock(&memcg_oom_lock);
4471 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4473 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4475 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4476 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4477 seq_printf(sf, "oom_kill %lu\n",
4478 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4482 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4483 struct cftype *cft, u64 val)
4485 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4487 /* cannot set to root cgroup and only 0 and 1 are allowed */
4488 if (!css->parent || !((val == 0) || (val == 1)))
4491 memcg->oom_kill_disable = val;
4493 memcg_oom_recover(memcg);
4498 #ifdef CONFIG_CGROUP_WRITEBACK
4500 #include <trace/events/writeback.h>
4502 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4504 return wb_domain_init(&memcg->cgwb_domain, gfp);
4507 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4509 wb_domain_exit(&memcg->cgwb_domain);
4512 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4514 wb_domain_size_changed(&memcg->cgwb_domain);
4517 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4519 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4521 if (!memcg->css.parent)
4524 return &memcg->cgwb_domain;
4528 * idx can be of type enum memcg_stat_item or node_stat_item.
4529 * Keep in sync with memcg_exact_page().
4531 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4533 long x = atomic_long_read(&memcg->vmstats[idx]);
4536 for_each_online_cpu(cpu)
4537 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4544 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4545 * @wb: bdi_writeback in question
4546 * @pfilepages: out parameter for number of file pages
4547 * @pheadroom: out parameter for number of allocatable pages according to memcg
4548 * @pdirty: out parameter for number of dirty pages
4549 * @pwriteback: out parameter for number of pages under writeback
4551 * Determine the numbers of file, headroom, dirty, and writeback pages in
4552 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4553 * is a bit more involved.
4555 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4556 * headroom is calculated as the lowest headroom of itself and the
4557 * ancestors. Note that this doesn't consider the actual amount of
4558 * available memory in the system. The caller should further cap
4559 * *@pheadroom accordingly.
4561 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4562 unsigned long *pheadroom, unsigned long *pdirty,
4563 unsigned long *pwriteback)
4565 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4566 struct mem_cgroup *parent;
4568 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4570 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4571 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4572 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4573 *pheadroom = PAGE_COUNTER_MAX;
4575 while ((parent = parent_mem_cgroup(memcg))) {
4576 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4577 READ_ONCE(memcg->memory.high));
4578 unsigned long used = page_counter_read(&memcg->memory);
4580 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4586 * Foreign dirty flushing
4588 * There's an inherent mismatch between memcg and writeback. The former
4589 * trackes ownership per-page while the latter per-inode. This was a
4590 * deliberate design decision because honoring per-page ownership in the
4591 * writeback path is complicated, may lead to higher CPU and IO overheads
4592 * and deemed unnecessary given that write-sharing an inode across
4593 * different cgroups isn't a common use-case.
4595 * Combined with inode majority-writer ownership switching, this works well
4596 * enough in most cases but there are some pathological cases. For
4597 * example, let's say there are two cgroups A and B which keep writing to
4598 * different but confined parts of the same inode. B owns the inode and
4599 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4600 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4601 * triggering background writeback. A will be slowed down without a way to
4602 * make writeback of the dirty pages happen.
4604 * Conditions like the above can lead to a cgroup getting repatedly and
4605 * severely throttled after making some progress after each
4606 * dirty_expire_interval while the underyling IO device is almost
4609 * Solving this problem completely requires matching the ownership tracking
4610 * granularities between memcg and writeback in either direction. However,
4611 * the more egregious behaviors can be avoided by simply remembering the
4612 * most recent foreign dirtying events and initiating remote flushes on
4613 * them when local writeback isn't enough to keep the memory clean enough.
4615 * The following two functions implement such mechanism. When a foreign
4616 * page - a page whose memcg and writeback ownerships don't match - is
4617 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4618 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4619 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4620 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4621 * foreign bdi_writebacks which haven't expired. Both the numbers of
4622 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4623 * limited to MEMCG_CGWB_FRN_CNT.
4625 * The mechanism only remembers IDs and doesn't hold any object references.
4626 * As being wrong occasionally doesn't matter, updates and accesses to the
4627 * records are lockless and racy.
4629 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4630 struct bdi_writeback *wb)
4632 struct mem_cgroup *memcg = page->mem_cgroup;
4633 struct memcg_cgwb_frn *frn;
4634 u64 now = get_jiffies_64();
4635 u64 oldest_at = now;
4639 trace_track_foreign_dirty(page, wb);
4642 * Pick the slot to use. If there is already a slot for @wb, keep
4643 * using it. If not replace the oldest one which isn't being
4646 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4647 frn = &memcg->cgwb_frn[i];
4648 if (frn->bdi_id == wb->bdi->id &&
4649 frn->memcg_id == wb->memcg_css->id)
4651 if (time_before64(frn->at, oldest_at) &&
4652 atomic_read(&frn->done.cnt) == 1) {
4654 oldest_at = frn->at;
4658 if (i < MEMCG_CGWB_FRN_CNT) {
4660 * Re-using an existing one. Update timestamp lazily to
4661 * avoid making the cacheline hot. We want them to be
4662 * reasonably up-to-date and significantly shorter than
4663 * dirty_expire_interval as that's what expires the record.
4664 * Use the shorter of 1s and dirty_expire_interval / 8.
4666 unsigned long update_intv =
4667 min_t(unsigned long, HZ,
4668 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4670 if (time_before64(frn->at, now - update_intv))
4672 } else if (oldest >= 0) {
4673 /* replace the oldest free one */
4674 frn = &memcg->cgwb_frn[oldest];
4675 frn->bdi_id = wb->bdi->id;
4676 frn->memcg_id = wb->memcg_css->id;
4681 /* issue foreign writeback flushes for recorded foreign dirtying events */
4682 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4684 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4685 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4686 u64 now = jiffies_64;
4689 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4690 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4693 * If the record is older than dirty_expire_interval,
4694 * writeback on it has already started. No need to kick it
4695 * off again. Also, don't start a new one if there's
4696 * already one in flight.
4698 if (time_after64(frn->at, now - intv) &&
4699 atomic_read(&frn->done.cnt) == 1) {
4701 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4702 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4703 WB_REASON_FOREIGN_FLUSH,
4709 #else /* CONFIG_CGROUP_WRITEBACK */
4711 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4716 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4720 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4724 #endif /* CONFIG_CGROUP_WRITEBACK */
4727 * DO NOT USE IN NEW FILES.
4729 * "cgroup.event_control" implementation.
4731 * This is way over-engineered. It tries to support fully configurable
4732 * events for each user. Such level of flexibility is completely
4733 * unnecessary especially in the light of the planned unified hierarchy.
4735 * Please deprecate this and replace with something simpler if at all
4740 * Unregister event and free resources.
4742 * Gets called from workqueue.
4744 static void memcg_event_remove(struct work_struct *work)
4746 struct mem_cgroup_event *event =
4747 container_of(work, struct mem_cgroup_event, remove);
4748 struct mem_cgroup *memcg = event->memcg;
4750 remove_wait_queue(event->wqh, &event->wait);
4752 event->unregister_event(memcg, event->eventfd);
4754 /* Notify userspace the event is going away. */
4755 eventfd_signal(event->eventfd, 1);
4757 eventfd_ctx_put(event->eventfd);
4759 css_put(&memcg->css);
4763 * Gets called on EPOLLHUP on eventfd when user closes it.
4765 * Called with wqh->lock held and interrupts disabled.
4767 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4768 int sync, void *key)
4770 struct mem_cgroup_event *event =
4771 container_of(wait, struct mem_cgroup_event, wait);
4772 struct mem_cgroup *memcg = event->memcg;
4773 __poll_t flags = key_to_poll(key);
4775 if (flags & EPOLLHUP) {
4777 * If the event has been detached at cgroup removal, we
4778 * can simply return knowing the other side will cleanup
4781 * We can't race against event freeing since the other
4782 * side will require wqh->lock via remove_wait_queue(),
4785 spin_lock(&memcg->event_list_lock);
4786 if (!list_empty(&event->list)) {
4787 list_del_init(&event->list);
4789 * We are in atomic context, but cgroup_event_remove()
4790 * may sleep, so we have to call it in workqueue.
4792 schedule_work(&event->remove);
4794 spin_unlock(&memcg->event_list_lock);
4800 static void memcg_event_ptable_queue_proc(struct file *file,
4801 wait_queue_head_t *wqh, poll_table *pt)
4803 struct mem_cgroup_event *event =
4804 container_of(pt, struct mem_cgroup_event, pt);
4807 add_wait_queue(wqh, &event->wait);
4811 * DO NOT USE IN NEW FILES.
4813 * Parse input and register new cgroup event handler.
4815 * Input must be in format '<event_fd> <control_fd> <args>'.
4816 * Interpretation of args is defined by control file implementation.
4818 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4819 char *buf, size_t nbytes, loff_t off)
4821 struct cgroup_subsys_state *css = of_css(of);
4822 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4823 struct mem_cgroup_event *event;
4824 struct cgroup_subsys_state *cfile_css;
4825 unsigned int efd, cfd;
4832 buf = strstrip(buf);
4834 efd = simple_strtoul(buf, &endp, 10);
4839 cfd = simple_strtoul(buf, &endp, 10);
4840 if ((*endp != ' ') && (*endp != '\0'))
4844 event = kzalloc(sizeof(*event), GFP_KERNEL);
4848 event->memcg = memcg;
4849 INIT_LIST_HEAD(&event->list);
4850 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4851 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4852 INIT_WORK(&event->remove, memcg_event_remove);
4860 event->eventfd = eventfd_ctx_fileget(efile.file);
4861 if (IS_ERR(event->eventfd)) {
4862 ret = PTR_ERR(event->eventfd);
4869 goto out_put_eventfd;
4872 /* the process need read permission on control file */
4873 /* AV: shouldn't we check that it's been opened for read instead? */
4874 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4879 * Determine the event callbacks and set them in @event. This used
4880 * to be done via struct cftype but cgroup core no longer knows
4881 * about these events. The following is crude but the whole thing
4882 * is for compatibility anyway.
4884 * DO NOT ADD NEW FILES.
4886 name = cfile.file->f_path.dentry->d_name.name;
4888 if (!strcmp(name, "memory.usage_in_bytes")) {
4889 event->register_event = mem_cgroup_usage_register_event;
4890 event->unregister_event = mem_cgroup_usage_unregister_event;
4891 } else if (!strcmp(name, "memory.oom_control")) {
4892 event->register_event = mem_cgroup_oom_register_event;
4893 event->unregister_event = mem_cgroup_oom_unregister_event;
4894 } else if (!strcmp(name, "memory.pressure_level")) {
4895 event->register_event = vmpressure_register_event;
4896 event->unregister_event = vmpressure_unregister_event;
4897 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4898 event->register_event = memsw_cgroup_usage_register_event;
4899 event->unregister_event = memsw_cgroup_usage_unregister_event;
4906 * Verify @cfile should belong to @css. Also, remaining events are
4907 * automatically removed on cgroup destruction but the removal is
4908 * asynchronous, so take an extra ref on @css.
4910 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4911 &memory_cgrp_subsys);
4913 if (IS_ERR(cfile_css))
4915 if (cfile_css != css) {
4920 ret = event->register_event(memcg, event->eventfd, buf);
4924 vfs_poll(efile.file, &event->pt);
4926 spin_lock(&memcg->event_list_lock);
4927 list_add(&event->list, &memcg->event_list);
4928 spin_unlock(&memcg->event_list_lock);
4940 eventfd_ctx_put(event->eventfd);
4949 static struct cftype mem_cgroup_legacy_files[] = {
4951 .name = "usage_in_bytes",
4952 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4953 .read_u64 = mem_cgroup_read_u64,
4956 .name = "max_usage_in_bytes",
4957 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4958 .write = mem_cgroup_reset,
4959 .read_u64 = mem_cgroup_read_u64,
4962 .name = "limit_in_bytes",
4963 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4964 .write = mem_cgroup_write,
4965 .read_u64 = mem_cgroup_read_u64,
4968 .name = "soft_limit_in_bytes",
4969 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4970 .write = mem_cgroup_write,
4971 .read_u64 = mem_cgroup_read_u64,
4975 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4976 .write = mem_cgroup_reset,
4977 .read_u64 = mem_cgroup_read_u64,
4981 .seq_show = memcg_stat_show,
4984 .name = "force_empty",
4985 .write = mem_cgroup_force_empty_write,
4988 .name = "use_hierarchy",
4989 .write_u64 = mem_cgroup_hierarchy_write,
4990 .read_u64 = mem_cgroup_hierarchy_read,
4993 .name = "cgroup.event_control", /* XXX: for compat */
4994 .write = memcg_write_event_control,
4995 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4998 .name = "swappiness",
4999 .read_u64 = mem_cgroup_swappiness_read,
5000 .write_u64 = mem_cgroup_swappiness_write,
5003 .name = "move_charge_at_immigrate",
5004 .read_u64 = mem_cgroup_move_charge_read,
5005 .write_u64 = mem_cgroup_move_charge_write,
5008 .name = "oom_control",
5009 .seq_show = mem_cgroup_oom_control_read,
5010 .write_u64 = mem_cgroup_oom_control_write,
5011 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5014 .name = "pressure_level",
5018 .name = "numa_stat",
5019 .seq_show = memcg_numa_stat_show,
5023 .name = "kmem.limit_in_bytes",
5024 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5025 .write = mem_cgroup_write,
5026 .read_u64 = mem_cgroup_read_u64,
5029 .name = "kmem.usage_in_bytes",
5030 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5031 .read_u64 = mem_cgroup_read_u64,
5034 .name = "kmem.failcnt",
5035 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5036 .write = mem_cgroup_reset,
5037 .read_u64 = mem_cgroup_read_u64,
5040 .name = "kmem.max_usage_in_bytes",
5041 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5042 .write = mem_cgroup_reset,
5043 .read_u64 = mem_cgroup_read_u64,
5045 #if defined(CONFIG_MEMCG_KMEM) && \
5046 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5048 .name = "kmem.slabinfo",
5049 .seq_show = memcg_slab_show,
5053 .name = "kmem.tcp.limit_in_bytes",
5054 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5055 .write = mem_cgroup_write,
5056 .read_u64 = mem_cgroup_read_u64,
5059 .name = "kmem.tcp.usage_in_bytes",
5060 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5061 .read_u64 = mem_cgroup_read_u64,
5064 .name = "kmem.tcp.failcnt",
5065 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5066 .write = mem_cgroup_reset,
5067 .read_u64 = mem_cgroup_read_u64,
5070 .name = "kmem.tcp.max_usage_in_bytes",
5071 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5072 .write = mem_cgroup_reset,
5073 .read_u64 = mem_cgroup_read_u64,
5075 { }, /* terminate */
5079 * Private memory cgroup IDR
5081 * Swap-out records and page cache shadow entries need to store memcg
5082 * references in constrained space, so we maintain an ID space that is
5083 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5084 * memory-controlled cgroups to 64k.
5086 * However, there usually are many references to the offline CSS after
5087 * the cgroup has been destroyed, such as page cache or reclaimable
5088 * slab objects, that don't need to hang on to the ID. We want to keep
5089 * those dead CSS from occupying IDs, or we might quickly exhaust the
5090 * relatively small ID space and prevent the creation of new cgroups
5091 * even when there are much fewer than 64k cgroups - possibly none.
5093 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5094 * be freed and recycled when it's no longer needed, which is usually
5095 * when the CSS is offlined.
5097 * The only exception to that are records of swapped out tmpfs/shmem
5098 * pages that need to be attributed to live ancestors on swapin. But
5099 * those references are manageable from userspace.
5102 static DEFINE_IDR(mem_cgroup_idr);
5104 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5106 if (memcg->id.id > 0) {
5107 idr_remove(&mem_cgroup_idr, memcg->id.id);
5112 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5115 refcount_add(n, &memcg->id.ref);
5118 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5120 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5121 mem_cgroup_id_remove(memcg);
5123 /* Memcg ID pins CSS */
5124 css_put(&memcg->css);
5128 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5130 mem_cgroup_id_put_many(memcg, 1);
5134 * mem_cgroup_from_id - look up a memcg from a memcg id
5135 * @id: the memcg id to look up
5137 * Caller must hold rcu_read_lock().
5139 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5141 WARN_ON_ONCE(!rcu_read_lock_held());
5142 return idr_find(&mem_cgroup_idr, id);
5145 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5147 struct mem_cgroup_per_node *pn;
5150 * This routine is called against possible nodes.
5151 * But it's BUG to call kmalloc() against offline node.
5153 * TODO: this routine can waste much memory for nodes which will
5154 * never be onlined. It's better to use memory hotplug callback
5157 if (!node_state(node, N_NORMAL_MEMORY))
5159 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5163 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5164 GFP_KERNEL_ACCOUNT);
5165 if (!pn->lruvec_stat_local) {
5170 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct lruvec_stat,
5171 GFP_KERNEL_ACCOUNT);
5172 if (!pn->lruvec_stat_cpu) {
5173 free_percpu(pn->lruvec_stat_local);
5178 lruvec_init(&pn->lruvec);
5179 pn->usage_in_excess = 0;
5180 pn->on_tree = false;
5183 memcg->nodeinfo[node] = pn;
5187 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5189 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5194 free_percpu(pn->lruvec_stat_cpu);
5195 free_percpu(pn->lruvec_stat_local);
5199 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5204 free_mem_cgroup_per_node_info(memcg, node);
5205 free_percpu(memcg->vmstats_percpu);
5206 free_percpu(memcg->vmstats_local);
5210 static void mem_cgroup_free(struct mem_cgroup *memcg)
5212 memcg_wb_domain_exit(memcg);
5214 * Flush percpu vmstats and vmevents to guarantee the value correctness
5215 * on parent's and all ancestor levels.
5217 memcg_flush_percpu_vmstats(memcg);
5218 memcg_flush_percpu_vmevents(memcg);
5219 __mem_cgroup_free(memcg);
5222 static struct mem_cgroup *mem_cgroup_alloc(void)
5224 struct mem_cgroup *memcg;
5227 int __maybe_unused i;
5228 long error = -ENOMEM;
5230 size = sizeof(struct mem_cgroup);
5231 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5233 memcg = kzalloc(size, GFP_KERNEL);
5235 return ERR_PTR(error);
5237 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5238 1, MEM_CGROUP_ID_MAX,
5240 if (memcg->id.id < 0) {
5241 error = memcg->id.id;
5245 memcg->vmstats_local = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5246 GFP_KERNEL_ACCOUNT);
5247 if (!memcg->vmstats_local)
5250 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5251 GFP_KERNEL_ACCOUNT);
5252 if (!memcg->vmstats_percpu)
5256 if (alloc_mem_cgroup_per_node_info(memcg, node))
5259 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5262 INIT_WORK(&memcg->high_work, high_work_func);
5263 INIT_LIST_HEAD(&memcg->oom_notify);
5264 mutex_init(&memcg->thresholds_lock);
5265 spin_lock_init(&memcg->move_lock);
5266 vmpressure_init(&memcg->vmpressure);
5267 INIT_LIST_HEAD(&memcg->event_list);
5268 spin_lock_init(&memcg->event_list_lock);
5269 memcg->socket_pressure = jiffies;
5270 #ifdef CONFIG_MEMCG_KMEM
5271 memcg->kmemcg_id = -1;
5272 INIT_LIST_HEAD(&memcg->objcg_list);
5274 #ifdef CONFIG_CGROUP_WRITEBACK
5275 INIT_LIST_HEAD(&memcg->cgwb_list);
5276 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5277 memcg->cgwb_frn[i].done =
5278 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5280 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5281 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5282 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5283 memcg->deferred_split_queue.split_queue_len = 0;
5285 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5288 mem_cgroup_id_remove(memcg);
5289 __mem_cgroup_free(memcg);
5290 return ERR_PTR(error);
5293 static struct cgroup_subsys_state * __ref
5294 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5296 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5297 struct mem_cgroup *memcg, *old_memcg;
5298 long error = -ENOMEM;
5300 old_memcg = set_active_memcg(parent);
5301 memcg = mem_cgroup_alloc();
5302 set_active_memcg(old_memcg);
5304 return ERR_CAST(memcg);
5306 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5307 memcg->soft_limit = PAGE_COUNTER_MAX;
5308 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5310 memcg->swappiness = mem_cgroup_swappiness(parent);
5311 memcg->oom_kill_disable = parent->oom_kill_disable;
5313 page_counter_init(&memcg->memory, &parent->memory);
5314 page_counter_init(&memcg->swap, &parent->swap);
5315 page_counter_init(&memcg->kmem, &parent->kmem);
5316 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5318 page_counter_init(&memcg->memory, NULL);
5319 page_counter_init(&memcg->swap, NULL);
5320 page_counter_init(&memcg->kmem, NULL);
5321 page_counter_init(&memcg->tcpmem, NULL);
5323 root_mem_cgroup = memcg;
5327 /* The following stuff does not apply to the root */
5328 error = memcg_online_kmem(memcg);
5332 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5333 static_branch_inc(&memcg_sockets_enabled_key);
5337 mem_cgroup_id_remove(memcg);
5338 mem_cgroup_free(memcg);
5339 return ERR_PTR(error);
5342 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5344 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5347 * A memcg must be visible for memcg_expand_shrinker_maps()
5348 * by the time the maps are allocated. So, we allocate maps
5349 * here, when for_each_mem_cgroup() can't skip it.
5351 if (memcg_alloc_shrinker_maps(memcg)) {
5352 mem_cgroup_id_remove(memcg);
5356 /* Online state pins memcg ID, memcg ID pins CSS */
5357 refcount_set(&memcg->id.ref, 1);
5362 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5364 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5365 struct mem_cgroup_event *event, *tmp;
5368 * Unregister events and notify userspace.
5369 * Notify userspace about cgroup removing only after rmdir of cgroup
5370 * directory to avoid race between userspace and kernelspace.
5372 spin_lock(&memcg->event_list_lock);
5373 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5374 list_del_init(&event->list);
5375 schedule_work(&event->remove);
5377 spin_unlock(&memcg->event_list_lock);
5379 page_counter_set_min(&memcg->memory, 0);
5380 page_counter_set_low(&memcg->memory, 0);
5382 memcg_offline_kmem(memcg);
5383 wb_memcg_offline(memcg);
5385 drain_all_stock(memcg);
5387 mem_cgroup_id_put(memcg);
5390 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5392 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5394 invalidate_reclaim_iterators(memcg);
5397 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5399 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5400 int __maybe_unused i;
5402 #ifdef CONFIG_CGROUP_WRITEBACK
5403 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5404 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5406 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5407 static_branch_dec(&memcg_sockets_enabled_key);
5409 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5410 static_branch_dec(&memcg_sockets_enabled_key);
5412 vmpressure_cleanup(&memcg->vmpressure);
5413 cancel_work_sync(&memcg->high_work);
5414 mem_cgroup_remove_from_trees(memcg);
5415 memcg_free_shrinker_maps(memcg);
5416 memcg_free_kmem(memcg);
5417 mem_cgroup_free(memcg);
5421 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5422 * @css: the target css
5424 * Reset the states of the mem_cgroup associated with @css. This is
5425 * invoked when the userland requests disabling on the default hierarchy
5426 * but the memcg is pinned through dependency. The memcg should stop
5427 * applying policies and should revert to the vanilla state as it may be
5428 * made visible again.
5430 * The current implementation only resets the essential configurations.
5431 * This needs to be expanded to cover all the visible parts.
5433 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5435 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5437 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5438 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5439 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5440 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5441 page_counter_set_min(&memcg->memory, 0);
5442 page_counter_set_low(&memcg->memory, 0);
5443 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5444 memcg->soft_limit = PAGE_COUNTER_MAX;
5445 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5446 memcg_wb_domain_size_changed(memcg);
5450 /* Handlers for move charge at task migration. */
5451 static int mem_cgroup_do_precharge(unsigned long count)
5455 /* Try a single bulk charge without reclaim first, kswapd may wake */
5456 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5458 mc.precharge += count;
5462 /* Try charges one by one with reclaim, but do not retry */
5464 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5478 enum mc_target_type {
5485 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5486 unsigned long addr, pte_t ptent)
5488 struct page *page = vm_normal_page(vma, addr, ptent);
5490 if (!page || !page_mapped(page))
5492 if (PageAnon(page)) {
5493 if (!(mc.flags & MOVE_ANON))
5496 if (!(mc.flags & MOVE_FILE))
5499 if (!get_page_unless_zero(page))
5505 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5506 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5507 pte_t ptent, swp_entry_t *entry)
5509 struct page *page = NULL;
5510 swp_entry_t ent = pte_to_swp_entry(ptent);
5512 if (!(mc.flags & MOVE_ANON))
5516 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5517 * a device and because they are not accessible by CPU they are store
5518 * as special swap entry in the CPU page table.
5520 if (is_device_private_entry(ent)) {
5521 page = device_private_entry_to_page(ent);
5523 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5524 * a refcount of 1 when free (unlike normal page)
5526 if (!page_ref_add_unless(page, 1, 1))
5531 if (non_swap_entry(ent))
5535 * Because lookup_swap_cache() updates some statistics counter,
5536 * we call find_get_page() with swapper_space directly.
5538 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5539 entry->val = ent.val;
5544 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5545 pte_t ptent, swp_entry_t *entry)
5551 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5552 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5554 if (!vma->vm_file) /* anonymous vma */
5556 if (!(mc.flags & MOVE_FILE))
5559 /* page is moved even if it's not RSS of this task(page-faulted). */
5560 /* shmem/tmpfs may report page out on swap: account for that too. */
5561 return find_get_incore_page(vma->vm_file->f_mapping,
5562 linear_page_index(vma, addr));
5566 * mem_cgroup_move_account - move account of the page
5568 * @compound: charge the page as compound or small page
5569 * @from: mem_cgroup which the page is moved from.
5570 * @to: mem_cgroup which the page is moved to. @from != @to.
5572 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5574 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5577 static int mem_cgroup_move_account(struct page *page,
5579 struct mem_cgroup *from,
5580 struct mem_cgroup *to)
5582 struct lruvec *from_vec, *to_vec;
5583 struct pglist_data *pgdat;
5584 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5587 VM_BUG_ON(from == to);
5588 VM_BUG_ON_PAGE(PageLRU(page), page);
5589 VM_BUG_ON(compound && !PageTransHuge(page));
5592 * Prevent mem_cgroup_migrate() from looking at
5593 * page->mem_cgroup of its source page while we change it.
5596 if (!trylock_page(page))
5600 if (page->mem_cgroup != from)
5603 pgdat = page_pgdat(page);
5604 from_vec = mem_cgroup_lruvec(from, pgdat);
5605 to_vec = mem_cgroup_lruvec(to, pgdat);
5607 lock_page_memcg(page);
5609 if (PageAnon(page)) {
5610 if (page_mapped(page)) {
5611 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5612 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5613 if (PageTransHuge(page)) {
5614 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5616 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5622 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5623 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5625 if (PageSwapBacked(page)) {
5626 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5627 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5630 if (page_mapped(page)) {
5631 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5632 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5635 if (PageDirty(page)) {
5636 struct address_space *mapping = page_mapping(page);
5638 if (mapping_can_writeback(mapping)) {
5639 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5641 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5647 if (PageWriteback(page)) {
5648 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5649 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5653 * All state has been migrated, let's switch to the new memcg.
5655 * It is safe to change page->mem_cgroup here because the page
5656 * is referenced, charged, isolated, and locked: we can't race
5657 * with (un)charging, migration, LRU putback, or anything else
5658 * that would rely on a stable page->mem_cgroup.
5660 * Note that lock_page_memcg is a memcg lock, not a page lock,
5661 * to save space. As soon as we switch page->mem_cgroup to a
5662 * new memcg that isn't locked, the above state can change
5663 * concurrently again. Make sure we're truly done with it.
5668 css_put(&from->css);
5670 page->mem_cgroup = to;
5672 __unlock_page_memcg(from);
5676 local_irq_disable();
5677 mem_cgroup_charge_statistics(to, page, nr_pages);
5678 memcg_check_events(to, page);
5679 mem_cgroup_charge_statistics(from, page, -nr_pages);
5680 memcg_check_events(from, page);
5689 * get_mctgt_type - get target type of moving charge
5690 * @vma: the vma the pte to be checked belongs
5691 * @addr: the address corresponding to the pte to be checked
5692 * @ptent: the pte to be checked
5693 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5696 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5697 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5698 * move charge. if @target is not NULL, the page is stored in target->page
5699 * with extra refcnt got(Callers should handle it).
5700 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5701 * target for charge migration. if @target is not NULL, the entry is stored
5703 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5704 * (so ZONE_DEVICE page and thus not on the lru).
5705 * For now we such page is charge like a regular page would be as for all
5706 * intent and purposes it is just special memory taking the place of a
5709 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5711 * Called with pte lock held.
5714 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5715 unsigned long addr, pte_t ptent, union mc_target *target)
5717 struct page *page = NULL;
5718 enum mc_target_type ret = MC_TARGET_NONE;
5719 swp_entry_t ent = { .val = 0 };
5721 if (pte_present(ptent))
5722 page = mc_handle_present_pte(vma, addr, ptent);
5723 else if (is_swap_pte(ptent))
5724 page = mc_handle_swap_pte(vma, ptent, &ent);
5725 else if (pte_none(ptent))
5726 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5728 if (!page && !ent.val)
5732 * Do only loose check w/o serialization.
5733 * mem_cgroup_move_account() checks the page is valid or
5734 * not under LRU exclusion.
5736 if (page->mem_cgroup == mc.from) {
5737 ret = MC_TARGET_PAGE;
5738 if (is_device_private_page(page))
5739 ret = MC_TARGET_DEVICE;
5741 target->page = page;
5743 if (!ret || !target)
5747 * There is a swap entry and a page doesn't exist or isn't charged.
5748 * But we cannot move a tail-page in a THP.
5750 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5751 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5752 ret = MC_TARGET_SWAP;
5759 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5761 * We don't consider PMD mapped swapping or file mapped pages because THP does
5762 * not support them for now.
5763 * Caller should make sure that pmd_trans_huge(pmd) is true.
5765 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5766 unsigned long addr, pmd_t pmd, union mc_target *target)
5768 struct page *page = NULL;
5769 enum mc_target_type ret = MC_TARGET_NONE;
5771 if (unlikely(is_swap_pmd(pmd))) {
5772 VM_BUG_ON(thp_migration_supported() &&
5773 !is_pmd_migration_entry(pmd));
5776 page = pmd_page(pmd);
5777 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5778 if (!(mc.flags & MOVE_ANON))
5780 if (page->mem_cgroup == mc.from) {
5781 ret = MC_TARGET_PAGE;
5784 target->page = page;
5790 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5791 unsigned long addr, pmd_t pmd, union mc_target *target)
5793 return MC_TARGET_NONE;
5797 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5798 unsigned long addr, unsigned long end,
5799 struct mm_walk *walk)
5801 struct vm_area_struct *vma = walk->vma;
5805 ptl = pmd_trans_huge_lock(pmd, vma);
5808 * Note their can not be MC_TARGET_DEVICE for now as we do not
5809 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5810 * this might change.
5812 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5813 mc.precharge += HPAGE_PMD_NR;
5818 if (pmd_trans_unstable(pmd))
5820 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5821 for (; addr != end; pte++, addr += PAGE_SIZE)
5822 if (get_mctgt_type(vma, addr, *pte, NULL))
5823 mc.precharge++; /* increment precharge temporarily */
5824 pte_unmap_unlock(pte - 1, ptl);
5830 static const struct mm_walk_ops precharge_walk_ops = {
5831 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5834 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5836 unsigned long precharge;
5839 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5840 mmap_read_unlock(mm);
5842 precharge = mc.precharge;
5848 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5850 unsigned long precharge = mem_cgroup_count_precharge(mm);
5852 VM_BUG_ON(mc.moving_task);
5853 mc.moving_task = current;
5854 return mem_cgroup_do_precharge(precharge);
5857 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5858 static void __mem_cgroup_clear_mc(void)
5860 struct mem_cgroup *from = mc.from;
5861 struct mem_cgroup *to = mc.to;
5863 /* we must uncharge all the leftover precharges from mc.to */
5865 cancel_charge(mc.to, mc.precharge);
5869 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5870 * we must uncharge here.
5872 if (mc.moved_charge) {
5873 cancel_charge(mc.from, mc.moved_charge);
5874 mc.moved_charge = 0;
5876 /* we must fixup refcnts and charges */
5877 if (mc.moved_swap) {
5878 /* uncharge swap account from the old cgroup */
5879 if (!mem_cgroup_is_root(mc.from))
5880 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5882 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5885 * we charged both to->memory and to->memsw, so we
5886 * should uncharge to->memory.
5888 if (!mem_cgroup_is_root(mc.to))
5889 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5893 memcg_oom_recover(from);
5894 memcg_oom_recover(to);
5895 wake_up_all(&mc.waitq);
5898 static void mem_cgroup_clear_mc(void)
5900 struct mm_struct *mm = mc.mm;
5903 * we must clear moving_task before waking up waiters at the end of
5906 mc.moving_task = NULL;
5907 __mem_cgroup_clear_mc();
5908 spin_lock(&mc.lock);
5912 spin_unlock(&mc.lock);
5917 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5919 struct cgroup_subsys_state *css;
5920 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5921 struct mem_cgroup *from;
5922 struct task_struct *leader, *p;
5923 struct mm_struct *mm;
5924 unsigned long move_flags;
5927 /* charge immigration isn't supported on the default hierarchy */
5928 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5932 * Multi-process migrations only happen on the default hierarchy
5933 * where charge immigration is not used. Perform charge
5934 * immigration if @tset contains a leader and whine if there are
5938 cgroup_taskset_for_each_leader(leader, css, tset) {
5941 memcg = mem_cgroup_from_css(css);
5947 * We are now commited to this value whatever it is. Changes in this
5948 * tunable will only affect upcoming migrations, not the current one.
5949 * So we need to save it, and keep it going.
5951 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5955 from = mem_cgroup_from_task(p);
5957 VM_BUG_ON(from == memcg);
5959 mm = get_task_mm(p);
5962 /* We move charges only when we move a owner of the mm */
5963 if (mm->owner == p) {
5966 VM_BUG_ON(mc.precharge);
5967 VM_BUG_ON(mc.moved_charge);
5968 VM_BUG_ON(mc.moved_swap);
5970 spin_lock(&mc.lock);
5974 mc.flags = move_flags;
5975 spin_unlock(&mc.lock);
5976 /* We set mc.moving_task later */
5978 ret = mem_cgroup_precharge_mc(mm);
5980 mem_cgroup_clear_mc();
5987 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5990 mem_cgroup_clear_mc();
5993 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5994 unsigned long addr, unsigned long end,
5995 struct mm_walk *walk)
5998 struct vm_area_struct *vma = walk->vma;
6001 enum mc_target_type target_type;
6002 union mc_target target;
6005 ptl = pmd_trans_huge_lock(pmd, vma);
6007 if (mc.precharge < HPAGE_PMD_NR) {
6011 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6012 if (target_type == MC_TARGET_PAGE) {
6014 if (!isolate_lru_page(page)) {
6015 if (!mem_cgroup_move_account(page, true,
6017 mc.precharge -= HPAGE_PMD_NR;
6018 mc.moved_charge += HPAGE_PMD_NR;
6020 putback_lru_page(page);
6023 } else if (target_type == MC_TARGET_DEVICE) {
6025 if (!mem_cgroup_move_account(page, true,
6027 mc.precharge -= HPAGE_PMD_NR;
6028 mc.moved_charge += HPAGE_PMD_NR;
6036 if (pmd_trans_unstable(pmd))
6039 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6040 for (; addr != end; addr += PAGE_SIZE) {
6041 pte_t ptent = *(pte++);
6042 bool device = false;
6048 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6049 case MC_TARGET_DEVICE:
6052 case MC_TARGET_PAGE:
6055 * We can have a part of the split pmd here. Moving it
6056 * can be done but it would be too convoluted so simply
6057 * ignore such a partial THP and keep it in original
6058 * memcg. There should be somebody mapping the head.
6060 if (PageTransCompound(page))
6062 if (!device && isolate_lru_page(page))
6064 if (!mem_cgroup_move_account(page, false,
6067 /* we uncharge from mc.from later. */
6071 putback_lru_page(page);
6072 put: /* get_mctgt_type() gets the page */
6075 case MC_TARGET_SWAP:
6077 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6079 mem_cgroup_id_get_many(mc.to, 1);
6080 /* we fixup other refcnts and charges later. */
6088 pte_unmap_unlock(pte - 1, ptl);
6093 * We have consumed all precharges we got in can_attach().
6094 * We try charge one by one, but don't do any additional
6095 * charges to mc.to if we have failed in charge once in attach()
6098 ret = mem_cgroup_do_precharge(1);
6106 static const struct mm_walk_ops charge_walk_ops = {
6107 .pmd_entry = mem_cgroup_move_charge_pte_range,
6110 static void mem_cgroup_move_charge(void)
6112 lru_add_drain_all();
6114 * Signal lock_page_memcg() to take the memcg's move_lock
6115 * while we're moving its pages to another memcg. Then wait
6116 * for already started RCU-only updates to finish.
6118 atomic_inc(&mc.from->moving_account);
6121 if (unlikely(!mmap_read_trylock(mc.mm))) {
6123 * Someone who are holding the mmap_lock might be waiting in
6124 * waitq. So we cancel all extra charges, wake up all waiters,
6125 * and retry. Because we cancel precharges, we might not be able
6126 * to move enough charges, but moving charge is a best-effort
6127 * feature anyway, so it wouldn't be a big problem.
6129 __mem_cgroup_clear_mc();
6134 * When we have consumed all precharges and failed in doing
6135 * additional charge, the page walk just aborts.
6137 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6140 mmap_read_unlock(mc.mm);
6141 atomic_dec(&mc.from->moving_account);
6144 static void mem_cgroup_move_task(void)
6147 mem_cgroup_move_charge();
6148 mem_cgroup_clear_mc();
6151 #else /* !CONFIG_MMU */
6152 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6156 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6159 static void mem_cgroup_move_task(void)
6164 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6166 if (value == PAGE_COUNTER_MAX)
6167 seq_puts(m, "max\n");
6169 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6174 static u64 memory_current_read(struct cgroup_subsys_state *css,
6177 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6179 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6182 static int memory_min_show(struct seq_file *m, void *v)
6184 return seq_puts_memcg_tunable(m,
6185 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6188 static ssize_t memory_min_write(struct kernfs_open_file *of,
6189 char *buf, size_t nbytes, loff_t off)
6191 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6195 buf = strstrip(buf);
6196 err = page_counter_memparse(buf, "max", &min);
6200 page_counter_set_min(&memcg->memory, min);
6205 static int memory_low_show(struct seq_file *m, void *v)
6207 return seq_puts_memcg_tunable(m,
6208 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6211 static ssize_t memory_low_write(struct kernfs_open_file *of,
6212 char *buf, size_t nbytes, loff_t off)
6214 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6218 buf = strstrip(buf);
6219 err = page_counter_memparse(buf, "max", &low);
6223 page_counter_set_low(&memcg->memory, low);
6228 static int memory_high_show(struct seq_file *m, void *v)
6230 return seq_puts_memcg_tunable(m,
6231 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6234 static ssize_t memory_high_write(struct kernfs_open_file *of,
6235 char *buf, size_t nbytes, loff_t off)
6237 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6238 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6239 bool drained = false;
6243 buf = strstrip(buf);
6244 err = page_counter_memparse(buf, "max", &high);
6249 unsigned long nr_pages = page_counter_read(&memcg->memory);
6250 unsigned long reclaimed;
6252 if (nr_pages <= high)
6255 if (signal_pending(current))
6259 drain_all_stock(memcg);
6264 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6267 if (!reclaimed && !nr_retries--)
6271 page_counter_set_high(&memcg->memory, high);
6273 memcg_wb_domain_size_changed(memcg);
6278 static int memory_max_show(struct seq_file *m, void *v)
6280 return seq_puts_memcg_tunable(m,
6281 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6284 static ssize_t memory_max_write(struct kernfs_open_file *of,
6285 char *buf, size_t nbytes, loff_t off)
6287 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6288 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6289 bool drained = false;
6293 buf = strstrip(buf);
6294 err = page_counter_memparse(buf, "max", &max);
6298 xchg(&memcg->memory.max, max);
6301 unsigned long nr_pages = page_counter_read(&memcg->memory);
6303 if (nr_pages <= max)
6306 if (signal_pending(current))
6310 drain_all_stock(memcg);
6316 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6322 memcg_memory_event(memcg, MEMCG_OOM);
6323 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6327 memcg_wb_domain_size_changed(memcg);
6331 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6333 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6334 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6335 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6336 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6337 seq_printf(m, "oom_kill %lu\n",
6338 atomic_long_read(&events[MEMCG_OOM_KILL]));
6341 static int memory_events_show(struct seq_file *m, void *v)
6343 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6345 __memory_events_show(m, memcg->memory_events);
6349 static int memory_events_local_show(struct seq_file *m, void *v)
6351 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6353 __memory_events_show(m, memcg->memory_events_local);
6357 static int memory_stat_show(struct seq_file *m, void *v)
6359 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6362 buf = memory_stat_format(memcg);
6371 static int memory_numa_stat_show(struct seq_file *m, void *v)
6374 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6376 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6379 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6382 seq_printf(m, "%s", memory_stats[i].name);
6383 for_each_node_state(nid, N_MEMORY) {
6385 struct lruvec *lruvec;
6387 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6388 size = lruvec_page_state(lruvec, memory_stats[i].idx);
6389 size *= memory_stats[i].ratio;
6390 seq_printf(m, " N%d=%llu", nid, size);
6399 static int memory_oom_group_show(struct seq_file *m, void *v)
6401 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6403 seq_printf(m, "%d\n", memcg->oom_group);
6408 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6409 char *buf, size_t nbytes, loff_t off)
6411 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6414 buf = strstrip(buf);
6418 ret = kstrtoint(buf, 0, &oom_group);
6422 if (oom_group != 0 && oom_group != 1)
6425 memcg->oom_group = oom_group;
6430 static struct cftype memory_files[] = {
6433 .flags = CFTYPE_NOT_ON_ROOT,
6434 .read_u64 = memory_current_read,
6438 .flags = CFTYPE_NOT_ON_ROOT,
6439 .seq_show = memory_min_show,
6440 .write = memory_min_write,
6444 .flags = CFTYPE_NOT_ON_ROOT,
6445 .seq_show = memory_low_show,
6446 .write = memory_low_write,
6450 .flags = CFTYPE_NOT_ON_ROOT,
6451 .seq_show = memory_high_show,
6452 .write = memory_high_write,
6456 .flags = CFTYPE_NOT_ON_ROOT,
6457 .seq_show = memory_max_show,
6458 .write = memory_max_write,
6462 .flags = CFTYPE_NOT_ON_ROOT,
6463 .file_offset = offsetof(struct mem_cgroup, events_file),
6464 .seq_show = memory_events_show,
6467 .name = "events.local",
6468 .flags = CFTYPE_NOT_ON_ROOT,
6469 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6470 .seq_show = memory_events_local_show,
6474 .seq_show = memory_stat_show,
6478 .name = "numa_stat",
6479 .seq_show = memory_numa_stat_show,
6483 .name = "oom.group",
6484 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6485 .seq_show = memory_oom_group_show,
6486 .write = memory_oom_group_write,
6491 struct cgroup_subsys memory_cgrp_subsys = {
6492 .css_alloc = mem_cgroup_css_alloc,
6493 .css_online = mem_cgroup_css_online,
6494 .css_offline = mem_cgroup_css_offline,
6495 .css_released = mem_cgroup_css_released,
6496 .css_free = mem_cgroup_css_free,
6497 .css_reset = mem_cgroup_css_reset,
6498 .can_attach = mem_cgroup_can_attach,
6499 .cancel_attach = mem_cgroup_cancel_attach,
6500 .post_attach = mem_cgroup_move_task,
6501 .dfl_cftypes = memory_files,
6502 .legacy_cftypes = mem_cgroup_legacy_files,
6507 * This function calculates an individual cgroup's effective
6508 * protection which is derived from its own memory.min/low, its
6509 * parent's and siblings' settings, as well as the actual memory
6510 * distribution in the tree.
6512 * The following rules apply to the effective protection values:
6514 * 1. At the first level of reclaim, effective protection is equal to
6515 * the declared protection in memory.min and memory.low.
6517 * 2. To enable safe delegation of the protection configuration, at
6518 * subsequent levels the effective protection is capped to the
6519 * parent's effective protection.
6521 * 3. To make complex and dynamic subtrees easier to configure, the
6522 * user is allowed to overcommit the declared protection at a given
6523 * level. If that is the case, the parent's effective protection is
6524 * distributed to the children in proportion to how much protection
6525 * they have declared and how much of it they are utilizing.
6527 * This makes distribution proportional, but also work-conserving:
6528 * if one cgroup claims much more protection than it uses memory,
6529 * the unused remainder is available to its siblings.
6531 * 4. Conversely, when the declared protection is undercommitted at a
6532 * given level, the distribution of the larger parental protection
6533 * budget is NOT proportional. A cgroup's protection from a sibling
6534 * is capped to its own memory.min/low setting.
6536 * 5. However, to allow protecting recursive subtrees from each other
6537 * without having to declare each individual cgroup's fixed share
6538 * of the ancestor's claim to protection, any unutilized -
6539 * "floating" - protection from up the tree is distributed in
6540 * proportion to each cgroup's *usage*. This makes the protection
6541 * neutral wrt sibling cgroups and lets them compete freely over
6542 * the shared parental protection budget, but it protects the
6543 * subtree as a whole from neighboring subtrees.
6545 * Note that 4. and 5. are not in conflict: 4. is about protecting
6546 * against immediate siblings whereas 5. is about protecting against
6547 * neighboring subtrees.
6549 static unsigned long effective_protection(unsigned long usage,
6550 unsigned long parent_usage,
6551 unsigned long setting,
6552 unsigned long parent_effective,
6553 unsigned long siblings_protected)
6555 unsigned long protected;
6558 protected = min(usage, setting);
6560 * If all cgroups at this level combined claim and use more
6561 * protection then what the parent affords them, distribute
6562 * shares in proportion to utilization.
6564 * We are using actual utilization rather than the statically
6565 * claimed protection in order to be work-conserving: claimed
6566 * but unused protection is available to siblings that would
6567 * otherwise get a smaller chunk than what they claimed.
6569 if (siblings_protected > parent_effective)
6570 return protected * parent_effective / siblings_protected;
6573 * Ok, utilized protection of all children is within what the
6574 * parent affords them, so we know whatever this child claims
6575 * and utilizes is effectively protected.
6577 * If there is unprotected usage beyond this value, reclaim
6578 * will apply pressure in proportion to that amount.
6580 * If there is unutilized protection, the cgroup will be fully
6581 * shielded from reclaim, but we do return a smaller value for
6582 * protection than what the group could enjoy in theory. This
6583 * is okay. With the overcommit distribution above, effective
6584 * protection is always dependent on how memory is actually
6585 * consumed among the siblings anyway.
6590 * If the children aren't claiming (all of) the protection
6591 * afforded to them by the parent, distribute the remainder in
6592 * proportion to the (unprotected) memory of each cgroup. That
6593 * way, cgroups that aren't explicitly prioritized wrt each
6594 * other compete freely over the allowance, but they are
6595 * collectively protected from neighboring trees.
6597 * We're using unprotected memory for the weight so that if
6598 * some cgroups DO claim explicit protection, we don't protect
6599 * the same bytes twice.
6601 * Check both usage and parent_usage against the respective
6602 * protected values. One should imply the other, but they
6603 * aren't read atomically - make sure the division is sane.
6605 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6607 if (parent_effective > siblings_protected &&
6608 parent_usage > siblings_protected &&
6609 usage > protected) {
6610 unsigned long unclaimed;
6612 unclaimed = parent_effective - siblings_protected;
6613 unclaimed *= usage - protected;
6614 unclaimed /= parent_usage - siblings_protected;
6623 * mem_cgroup_protected - check if memory consumption is in the normal range
6624 * @root: the top ancestor of the sub-tree being checked
6625 * @memcg: the memory cgroup to check
6627 * WARNING: This function is not stateless! It can only be used as part
6628 * of a top-down tree iteration, not for isolated queries.
6630 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6631 struct mem_cgroup *memcg)
6633 unsigned long usage, parent_usage;
6634 struct mem_cgroup *parent;
6636 if (mem_cgroup_disabled())
6640 root = root_mem_cgroup;
6643 * Effective values of the reclaim targets are ignored so they
6644 * can be stale. Have a look at mem_cgroup_protection for more
6646 * TODO: calculation should be more robust so that we do not need
6647 * that special casing.
6652 usage = page_counter_read(&memcg->memory);
6656 parent = parent_mem_cgroup(memcg);
6657 /* No parent means a non-hierarchical mode on v1 memcg */
6661 if (parent == root) {
6662 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6663 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6667 parent_usage = page_counter_read(&parent->memory);
6669 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6670 READ_ONCE(memcg->memory.min),
6671 READ_ONCE(parent->memory.emin),
6672 atomic_long_read(&parent->memory.children_min_usage)));
6674 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6675 READ_ONCE(memcg->memory.low),
6676 READ_ONCE(parent->memory.elow),
6677 atomic_long_read(&parent->memory.children_low_usage)));
6681 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6682 * @page: page to charge
6683 * @mm: mm context of the victim
6684 * @gfp_mask: reclaim mode
6686 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6687 * pages according to @gfp_mask if necessary.
6689 * Returns 0 on success. Otherwise, an error code is returned.
6691 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6693 unsigned int nr_pages = thp_nr_pages(page);
6694 struct mem_cgroup *memcg = NULL;
6697 if (mem_cgroup_disabled())
6700 if (PageSwapCache(page)) {
6701 swp_entry_t ent = { .val = page_private(page), };
6705 * Every swap fault against a single page tries to charge the
6706 * page, bail as early as possible. shmem_unuse() encounters
6707 * already charged pages, too. page->mem_cgroup is protected
6708 * by the page lock, which serializes swap cache removal, which
6709 * in turn serializes uncharging.
6711 VM_BUG_ON_PAGE(!PageLocked(page), page);
6712 if (compound_head(page)->mem_cgroup)
6715 id = lookup_swap_cgroup_id(ent);
6717 memcg = mem_cgroup_from_id(id);
6718 if (memcg && !css_tryget_online(&memcg->css))
6724 memcg = get_mem_cgroup_from_mm(mm);
6726 ret = try_charge(memcg, gfp_mask, nr_pages);
6730 css_get(&memcg->css);
6731 commit_charge(page, memcg);
6733 local_irq_disable();
6734 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6735 memcg_check_events(memcg, page);
6738 if (PageSwapCache(page)) {
6739 swp_entry_t entry = { .val = page_private(page) };
6741 * The swap entry might not get freed for a long time,
6742 * let's not wait for it. The page already received a
6743 * memory+swap charge, drop the swap entry duplicate.
6745 mem_cgroup_uncharge_swap(entry, nr_pages);
6749 css_put(&memcg->css);
6754 struct uncharge_gather {
6755 struct mem_cgroup *memcg;
6756 unsigned long nr_pages;
6757 unsigned long pgpgout;
6758 unsigned long nr_kmem;
6759 struct page *dummy_page;
6762 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6764 memset(ug, 0, sizeof(*ug));
6767 static void uncharge_batch(const struct uncharge_gather *ug)
6769 unsigned long flags;
6771 if (!mem_cgroup_is_root(ug->memcg)) {
6772 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6773 if (do_memsw_account())
6774 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6775 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6776 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6777 memcg_oom_recover(ug->memcg);
6780 local_irq_save(flags);
6781 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6782 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6783 memcg_check_events(ug->memcg, ug->dummy_page);
6784 local_irq_restore(flags);
6786 /* drop reference from uncharge_page */
6787 css_put(&ug->memcg->css);
6790 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6792 unsigned long nr_pages;
6794 VM_BUG_ON_PAGE(PageLRU(page), page);
6796 if (!page->mem_cgroup)
6800 * Nobody should be changing or seriously looking at
6801 * page->mem_cgroup at this point, we have fully
6802 * exclusive access to the page.
6805 if (ug->memcg != page->mem_cgroup) {
6808 uncharge_gather_clear(ug);
6810 ug->memcg = page->mem_cgroup;
6812 /* pairs with css_put in uncharge_batch */
6813 css_get(&ug->memcg->css);
6816 nr_pages = compound_nr(page);
6817 ug->nr_pages += nr_pages;
6819 if (!PageKmemcg(page)) {
6822 ug->nr_kmem += nr_pages;
6823 __ClearPageKmemcg(page);
6826 ug->dummy_page = page;
6827 page->mem_cgroup = NULL;
6828 css_put(&ug->memcg->css);
6831 static void uncharge_list(struct list_head *page_list)
6833 struct uncharge_gather ug;
6834 struct list_head *next;
6836 uncharge_gather_clear(&ug);
6839 * Note that the list can be a single page->lru; hence the
6840 * do-while loop instead of a simple list_for_each_entry().
6842 next = page_list->next;
6846 page = list_entry(next, struct page, lru);
6847 next = page->lru.next;
6849 uncharge_page(page, &ug);
6850 } while (next != page_list);
6853 uncharge_batch(&ug);
6857 * mem_cgroup_uncharge - uncharge a page
6858 * @page: page to uncharge
6860 * Uncharge a page previously charged with mem_cgroup_charge().
6862 void mem_cgroup_uncharge(struct page *page)
6864 struct uncharge_gather ug;
6866 if (mem_cgroup_disabled())
6869 /* Don't touch page->lru of any random page, pre-check: */
6870 if (!page->mem_cgroup)
6873 uncharge_gather_clear(&ug);
6874 uncharge_page(page, &ug);
6875 uncharge_batch(&ug);
6879 * mem_cgroup_uncharge_list - uncharge a list of page
6880 * @page_list: list of pages to uncharge
6882 * Uncharge a list of pages previously charged with
6883 * mem_cgroup_charge().
6885 void mem_cgroup_uncharge_list(struct list_head *page_list)
6887 if (mem_cgroup_disabled())
6890 if (!list_empty(page_list))
6891 uncharge_list(page_list);
6895 * mem_cgroup_migrate - charge a page's replacement
6896 * @oldpage: currently circulating page
6897 * @newpage: replacement page
6899 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6900 * be uncharged upon free.
6902 * Both pages must be locked, @newpage->mapping must be set up.
6904 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6906 struct mem_cgroup *memcg;
6907 unsigned int nr_pages;
6908 unsigned long flags;
6910 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6911 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6912 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6913 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6916 if (mem_cgroup_disabled())
6919 /* Page cache replacement: new page already charged? */
6920 if (newpage->mem_cgroup)
6923 memcg = oldpage->mem_cgroup;
6927 /* Force-charge the new page. The old one will be freed soon */
6928 nr_pages = thp_nr_pages(newpage);
6930 page_counter_charge(&memcg->memory, nr_pages);
6931 if (do_memsw_account())
6932 page_counter_charge(&memcg->memsw, nr_pages);
6934 css_get(&memcg->css);
6935 commit_charge(newpage, memcg);
6937 local_irq_save(flags);
6938 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6939 memcg_check_events(memcg, newpage);
6940 local_irq_restore(flags);
6943 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6944 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6946 void mem_cgroup_sk_alloc(struct sock *sk)
6948 struct mem_cgroup *memcg;
6950 if (!mem_cgroup_sockets_enabled)
6953 /* Do not associate the sock with unrelated interrupted task's memcg. */
6958 memcg = mem_cgroup_from_task(current);
6959 if (memcg == root_mem_cgroup)
6961 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6963 if (css_tryget(&memcg->css))
6964 sk->sk_memcg = memcg;
6969 void mem_cgroup_sk_free(struct sock *sk)
6972 css_put(&sk->sk_memcg->css);
6976 * mem_cgroup_charge_skmem - charge socket memory
6977 * @memcg: memcg to charge
6978 * @nr_pages: number of pages to charge
6980 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6981 * @memcg's configured limit, %false if the charge had to be forced.
6983 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6985 gfp_t gfp_mask = GFP_KERNEL;
6987 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6988 struct page_counter *fail;
6990 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6991 memcg->tcpmem_pressure = 0;
6994 page_counter_charge(&memcg->tcpmem, nr_pages);
6995 memcg->tcpmem_pressure = 1;
6999 /* Don't block in the packet receive path */
7001 gfp_mask = GFP_NOWAIT;
7003 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7005 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7008 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7013 * mem_cgroup_uncharge_skmem - uncharge socket memory
7014 * @memcg: memcg to uncharge
7015 * @nr_pages: number of pages to uncharge
7017 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7019 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7020 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7024 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7026 refill_stock(memcg, nr_pages);
7029 static int __init cgroup_memory(char *s)
7033 while ((token = strsep(&s, ",")) != NULL) {
7036 if (!strcmp(token, "nosocket"))
7037 cgroup_memory_nosocket = true;
7038 if (!strcmp(token, "nokmem"))
7039 cgroup_memory_nokmem = true;
7043 __setup("cgroup.memory=", cgroup_memory);
7046 * subsys_initcall() for memory controller.
7048 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7049 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7050 * basically everything that doesn't depend on a specific mem_cgroup structure
7051 * should be initialized from here.
7053 static int __init mem_cgroup_init(void)
7057 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7058 memcg_hotplug_cpu_dead);
7060 for_each_possible_cpu(cpu)
7061 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7064 for_each_node(node) {
7065 struct mem_cgroup_tree_per_node *rtpn;
7067 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7068 node_online(node) ? node : NUMA_NO_NODE);
7070 rtpn->rb_root = RB_ROOT;
7071 rtpn->rb_rightmost = NULL;
7072 spin_lock_init(&rtpn->lock);
7073 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7078 subsys_initcall(mem_cgroup_init);
7080 #ifdef CONFIG_MEMCG_SWAP
7081 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7083 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7085 * The root cgroup cannot be destroyed, so it's refcount must
7088 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7092 memcg = parent_mem_cgroup(memcg);
7094 memcg = root_mem_cgroup;
7100 * mem_cgroup_swapout - transfer a memsw charge to swap
7101 * @page: page whose memsw charge to transfer
7102 * @entry: swap entry to move the charge to
7104 * Transfer the memsw charge of @page to @entry.
7106 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7108 struct mem_cgroup *memcg, *swap_memcg;
7109 unsigned int nr_entries;
7110 unsigned short oldid;
7112 VM_BUG_ON_PAGE(PageLRU(page), page);
7113 VM_BUG_ON_PAGE(page_count(page), page);
7115 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7118 memcg = page->mem_cgroup;
7120 /* Readahead page, never charged */
7125 * In case the memcg owning these pages has been offlined and doesn't
7126 * have an ID allocated to it anymore, charge the closest online
7127 * ancestor for the swap instead and transfer the memory+swap charge.
7129 swap_memcg = mem_cgroup_id_get_online(memcg);
7130 nr_entries = thp_nr_pages(page);
7131 /* Get references for the tail pages, too */
7133 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7134 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7136 VM_BUG_ON_PAGE(oldid, page);
7137 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7139 page->mem_cgroup = NULL;
7141 if (!mem_cgroup_is_root(memcg))
7142 page_counter_uncharge(&memcg->memory, nr_entries);
7144 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7145 if (!mem_cgroup_is_root(swap_memcg))
7146 page_counter_charge(&swap_memcg->memsw, nr_entries);
7147 page_counter_uncharge(&memcg->memsw, nr_entries);
7151 * Interrupts should be disabled here because the caller holds the
7152 * i_pages lock which is taken with interrupts-off. It is
7153 * important here to have the interrupts disabled because it is the
7154 * only synchronisation we have for updating the per-CPU variables.
7156 VM_BUG_ON(!irqs_disabled());
7157 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7158 memcg_check_events(memcg, page);
7160 css_put(&memcg->css);
7164 * mem_cgroup_try_charge_swap - try charging swap space for a page
7165 * @page: page being added to swap
7166 * @entry: swap entry to charge
7168 * Try to charge @page's memcg for the swap space at @entry.
7170 * Returns 0 on success, -ENOMEM on failure.
7172 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7174 unsigned int nr_pages = thp_nr_pages(page);
7175 struct page_counter *counter;
7176 struct mem_cgroup *memcg;
7177 unsigned short oldid;
7179 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7182 memcg = page->mem_cgroup;
7184 /* Readahead page, never charged */
7189 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7193 memcg = mem_cgroup_id_get_online(memcg);
7195 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7196 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7197 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7198 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7199 mem_cgroup_id_put(memcg);
7203 /* Get references for the tail pages, too */
7205 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7206 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7207 VM_BUG_ON_PAGE(oldid, page);
7208 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7214 * mem_cgroup_uncharge_swap - uncharge swap space
7215 * @entry: swap entry to uncharge
7216 * @nr_pages: the amount of swap space to uncharge
7218 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7220 struct mem_cgroup *memcg;
7223 id = swap_cgroup_record(entry, 0, nr_pages);
7225 memcg = mem_cgroup_from_id(id);
7227 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7228 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7229 page_counter_uncharge(&memcg->swap, nr_pages);
7231 page_counter_uncharge(&memcg->memsw, nr_pages);
7233 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7234 mem_cgroup_id_put_many(memcg, nr_pages);
7239 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7241 long nr_swap_pages = get_nr_swap_pages();
7243 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7244 return nr_swap_pages;
7245 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7246 nr_swap_pages = min_t(long, nr_swap_pages,
7247 READ_ONCE(memcg->swap.max) -
7248 page_counter_read(&memcg->swap));
7249 return nr_swap_pages;
7252 bool mem_cgroup_swap_full(struct page *page)
7254 struct mem_cgroup *memcg;
7256 VM_BUG_ON_PAGE(!PageLocked(page), page);
7260 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7263 memcg = page->mem_cgroup;
7267 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7268 unsigned long usage = page_counter_read(&memcg->swap);
7270 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7271 usage * 2 >= READ_ONCE(memcg->swap.max))
7278 static int __init setup_swap_account(char *s)
7280 if (!strcmp(s, "1"))
7281 cgroup_memory_noswap = false;
7282 else if (!strcmp(s, "0"))
7283 cgroup_memory_noswap = true;
7286 __setup("swapaccount=", setup_swap_account);
7288 static u64 swap_current_read(struct cgroup_subsys_state *css,
7291 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7293 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7296 static int swap_high_show(struct seq_file *m, void *v)
7298 return seq_puts_memcg_tunable(m,
7299 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7302 static ssize_t swap_high_write(struct kernfs_open_file *of,
7303 char *buf, size_t nbytes, loff_t off)
7305 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7309 buf = strstrip(buf);
7310 err = page_counter_memparse(buf, "max", &high);
7314 page_counter_set_high(&memcg->swap, high);
7319 static int swap_max_show(struct seq_file *m, void *v)
7321 return seq_puts_memcg_tunable(m,
7322 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7325 static ssize_t swap_max_write(struct kernfs_open_file *of,
7326 char *buf, size_t nbytes, loff_t off)
7328 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7332 buf = strstrip(buf);
7333 err = page_counter_memparse(buf, "max", &max);
7337 xchg(&memcg->swap.max, max);
7342 static int swap_events_show(struct seq_file *m, void *v)
7344 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7346 seq_printf(m, "high %lu\n",
7347 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7348 seq_printf(m, "max %lu\n",
7349 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7350 seq_printf(m, "fail %lu\n",
7351 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7356 static struct cftype swap_files[] = {
7358 .name = "swap.current",
7359 .flags = CFTYPE_NOT_ON_ROOT,
7360 .read_u64 = swap_current_read,
7363 .name = "swap.high",
7364 .flags = CFTYPE_NOT_ON_ROOT,
7365 .seq_show = swap_high_show,
7366 .write = swap_high_write,
7370 .flags = CFTYPE_NOT_ON_ROOT,
7371 .seq_show = swap_max_show,
7372 .write = swap_max_write,
7375 .name = "swap.events",
7376 .flags = CFTYPE_NOT_ON_ROOT,
7377 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7378 .seq_show = swap_events_show,
7383 static struct cftype memsw_files[] = {
7385 .name = "memsw.usage_in_bytes",
7386 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7387 .read_u64 = mem_cgroup_read_u64,
7390 .name = "memsw.max_usage_in_bytes",
7391 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7392 .write = mem_cgroup_reset,
7393 .read_u64 = mem_cgroup_read_u64,
7396 .name = "memsw.limit_in_bytes",
7397 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7398 .write = mem_cgroup_write,
7399 .read_u64 = mem_cgroup_read_u64,
7402 .name = "memsw.failcnt",
7403 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7404 .write = mem_cgroup_reset,
7405 .read_u64 = mem_cgroup_read_u64,
7407 { }, /* terminate */
7411 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7412 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7413 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7414 * boot parameter. This may result in premature OOPS inside
7415 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7417 static int __init mem_cgroup_swap_init(void)
7419 /* No memory control -> no swap control */
7420 if (mem_cgroup_disabled())
7421 cgroup_memory_noswap = true;
7423 if (cgroup_memory_noswap)
7426 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7427 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7431 core_initcall(mem_cgroup_swap_init);
7433 #endif /* CONFIG_MEMCG_SWAP */