1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
25 #include <linux/page_counter.h>
26 #include <linux/memcontrol.h>
27 #include <linux/cgroup.h>
28 #include <linux/pagewalk.h>
29 #include <linux/sched/mm.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/hugetlb.h>
32 #include <linux/pagemap.h>
33 #include <linux/vm_event_item.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
59 #include <linux/tracehook.h>
60 #include <linux/psi.h>
61 #include <linux/seq_buf.h>
67 #include <linux/uaccess.h>
69 #include <trace/events/vmscan.h>
71 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
72 EXPORT_SYMBOL(memory_cgrp_subsys);
74 struct mem_cgroup *root_mem_cgroup __read_mostly;
76 /* Socket memory accounting disabled? */
77 static bool cgroup_memory_nosocket;
79 /* Kernel memory accounting disabled? */
80 static bool cgroup_memory_nokmem;
82 /* Whether the swap controller is active */
83 #ifdef CONFIG_MEMCG_SWAP
84 bool cgroup_memory_noswap __read_mostly;
86 #define cgroup_memory_noswap 1
89 #ifdef CONFIG_CGROUP_WRITEBACK
90 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
93 /* Whether legacy memory+swap accounting is active */
94 static bool do_memsw_account(void)
96 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
99 #define THRESHOLDS_EVENTS_TARGET 128
100 #define SOFTLIMIT_EVENTS_TARGET 1024
103 * Cgroups above their limits are maintained in a RB-Tree, independent of
104 * their hierarchy representation
107 struct mem_cgroup_tree_per_node {
108 struct rb_root rb_root;
109 struct rb_node *rb_rightmost;
113 struct mem_cgroup_tree {
114 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
117 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
120 struct mem_cgroup_eventfd_list {
121 struct list_head list;
122 struct eventfd_ctx *eventfd;
126 * cgroup_event represents events which userspace want to receive.
128 struct mem_cgroup_event {
130 * memcg which the event belongs to.
132 struct mem_cgroup *memcg;
134 * eventfd to signal userspace about the event.
136 struct eventfd_ctx *eventfd;
138 * Each of these stored in a list by the cgroup.
140 struct list_head list;
142 * register_event() callback will be used to add new userspace
143 * waiter for changes related to this event. Use eventfd_signal()
144 * on eventfd to send notification to userspace.
146 int (*register_event)(struct mem_cgroup *memcg,
147 struct eventfd_ctx *eventfd, const char *args);
149 * unregister_event() callback will be called when userspace closes
150 * the eventfd or on cgroup removing. This callback must be set,
151 * if you want provide notification functionality.
153 void (*unregister_event)(struct mem_cgroup *memcg,
154 struct eventfd_ctx *eventfd);
156 * All fields below needed to unregister event when
157 * userspace closes eventfd.
160 wait_queue_head_t *wqh;
161 wait_queue_entry_t wait;
162 struct work_struct remove;
165 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
166 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
168 /* Stuffs for move charges at task migration. */
170 * Types of charges to be moved.
172 #define MOVE_ANON 0x1U
173 #define MOVE_FILE 0x2U
174 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
176 /* "mc" and its members are protected by cgroup_mutex */
177 static struct move_charge_struct {
178 spinlock_t lock; /* for from, to */
179 struct mm_struct *mm;
180 struct mem_cgroup *from;
181 struct mem_cgroup *to;
183 unsigned long precharge;
184 unsigned long moved_charge;
185 unsigned long moved_swap;
186 struct task_struct *moving_task; /* a task moving charges */
187 wait_queue_head_t waitq; /* a waitq for other context */
189 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
190 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
194 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
195 * limit reclaim to prevent infinite loops, if they ever occur.
197 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
198 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
200 /* for encoding cft->private value on file */
209 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
210 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
211 #define MEMFILE_ATTR(val) ((val) & 0xffff)
212 /* Used for OOM nofiier */
213 #define OOM_CONTROL (0)
216 * Iteration constructs for visiting all cgroups (under a tree). If
217 * loops are exited prematurely (break), mem_cgroup_iter_break() must
218 * be used for reference counting.
220 #define for_each_mem_cgroup_tree(iter, root) \
221 for (iter = mem_cgroup_iter(root, NULL, NULL); \
223 iter = mem_cgroup_iter(root, iter, NULL))
225 #define for_each_mem_cgroup(iter) \
226 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
228 iter = mem_cgroup_iter(NULL, iter, NULL))
230 static inline bool should_force_charge(void)
232 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
233 (current->flags & PF_EXITING);
236 /* Some nice accessors for the vmpressure. */
237 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
240 memcg = root_mem_cgroup;
241 return &memcg->vmpressure;
244 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
246 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
249 #ifdef CONFIG_MEMCG_KMEM
250 extern spinlock_t css_set_lock;
252 static void obj_cgroup_release(struct percpu_ref *ref)
254 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
255 struct mem_cgroup *memcg;
256 unsigned int nr_bytes;
257 unsigned int nr_pages;
261 * At this point all allocated objects are freed, and
262 * objcg->nr_charged_bytes can't have an arbitrary byte value.
263 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
265 * The following sequence can lead to it:
266 * 1) CPU0: objcg == stock->cached_objcg
267 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
268 * PAGE_SIZE bytes are charged
269 * 3) CPU1: a process from another memcg is allocating something,
270 * the stock if flushed,
271 * objcg->nr_charged_bytes = PAGE_SIZE - 92
272 * 5) CPU0: we do release this object,
273 * 92 bytes are added to stock->nr_bytes
274 * 6) CPU0: stock is flushed,
275 * 92 bytes are added to objcg->nr_charged_bytes
277 * In the result, nr_charged_bytes == PAGE_SIZE.
278 * This page will be uncharged in obj_cgroup_release().
280 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
281 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
282 nr_pages = nr_bytes >> PAGE_SHIFT;
284 spin_lock_irqsave(&css_set_lock, flags);
285 memcg = obj_cgroup_memcg(objcg);
287 __memcg_kmem_uncharge(memcg, nr_pages);
288 list_del(&objcg->list);
289 mem_cgroup_put(memcg);
290 spin_unlock_irqrestore(&css_set_lock, flags);
292 percpu_ref_exit(ref);
293 kfree_rcu(objcg, rcu);
296 static struct obj_cgroup *obj_cgroup_alloc(void)
298 struct obj_cgroup *objcg;
301 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
305 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
311 INIT_LIST_HEAD(&objcg->list);
315 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
316 struct mem_cgroup *parent)
318 struct obj_cgroup *objcg, *iter;
320 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
322 spin_lock_irq(&css_set_lock);
324 /* Move active objcg to the parent's list */
325 xchg(&objcg->memcg, parent);
326 css_get(&parent->css);
327 list_add(&objcg->list, &parent->objcg_list);
329 /* Move already reparented objcgs to the parent's list */
330 list_for_each_entry(iter, &memcg->objcg_list, list) {
331 css_get(&parent->css);
332 xchg(&iter->memcg, parent);
333 css_put(&memcg->css);
335 list_splice(&memcg->objcg_list, &parent->objcg_list);
337 spin_unlock_irq(&css_set_lock);
339 percpu_ref_kill(&objcg->refcnt);
343 * This will be used as a shrinker list's index.
344 * The main reason for not using cgroup id for this:
345 * this works better in sparse environments, where we have a lot of memcgs,
346 * but only a few kmem-limited. Or also, if we have, for instance, 200
347 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
348 * 200 entry array for that.
350 * The current size of the caches array is stored in memcg_nr_cache_ids. It
351 * will double each time we have to increase it.
353 static DEFINE_IDA(memcg_cache_ida);
354 int memcg_nr_cache_ids;
356 /* Protects memcg_nr_cache_ids */
357 static DECLARE_RWSEM(memcg_cache_ids_sem);
359 void memcg_get_cache_ids(void)
361 down_read(&memcg_cache_ids_sem);
364 void memcg_put_cache_ids(void)
366 up_read(&memcg_cache_ids_sem);
370 * MIN_SIZE is different than 1, because we would like to avoid going through
371 * the alloc/free process all the time. In a small machine, 4 kmem-limited
372 * cgroups is a reasonable guess. In the future, it could be a parameter or
373 * tunable, but that is strictly not necessary.
375 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
376 * this constant directly from cgroup, but it is understandable that this is
377 * better kept as an internal representation in cgroup.c. In any case, the
378 * cgrp_id space is not getting any smaller, and we don't have to necessarily
379 * increase ours as well if it increases.
381 #define MEMCG_CACHES_MIN_SIZE 4
382 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
385 * A lot of the calls to the cache allocation functions are expected to be
386 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
387 * conditional to this static branch, we'll have to allow modules that does
388 * kmem_cache_alloc and the such to see this symbol as well
390 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
391 EXPORT_SYMBOL(memcg_kmem_enabled_key);
394 static int memcg_shrinker_map_size;
395 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
397 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
399 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
402 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
403 int size, int old_size)
405 struct memcg_shrinker_map *new, *old;
408 lockdep_assert_held(&memcg_shrinker_map_mutex);
411 old = rcu_dereference_protected(
412 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
413 /* Not yet online memcg */
417 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
421 /* Set all old bits, clear all new bits */
422 memset(new->map, (int)0xff, old_size);
423 memset((void *)new->map + old_size, 0, size - old_size);
425 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
426 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
432 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
434 struct mem_cgroup_per_node *pn;
435 struct memcg_shrinker_map *map;
438 if (mem_cgroup_is_root(memcg))
442 pn = mem_cgroup_nodeinfo(memcg, nid);
443 map = rcu_dereference_protected(pn->shrinker_map, true);
446 rcu_assign_pointer(pn->shrinker_map, NULL);
450 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
452 struct memcg_shrinker_map *map;
453 int nid, size, ret = 0;
455 if (mem_cgroup_is_root(memcg))
458 mutex_lock(&memcg_shrinker_map_mutex);
459 size = memcg_shrinker_map_size;
461 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
463 memcg_free_shrinker_maps(memcg);
467 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
469 mutex_unlock(&memcg_shrinker_map_mutex);
474 int memcg_expand_shrinker_maps(int new_id)
476 int size, old_size, ret = 0;
477 struct mem_cgroup *memcg;
479 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
480 old_size = memcg_shrinker_map_size;
481 if (size <= old_size)
484 mutex_lock(&memcg_shrinker_map_mutex);
485 if (!root_mem_cgroup)
488 for_each_mem_cgroup(memcg) {
489 if (mem_cgroup_is_root(memcg))
491 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
493 mem_cgroup_iter_break(NULL, memcg);
499 memcg_shrinker_map_size = size;
500 mutex_unlock(&memcg_shrinker_map_mutex);
504 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
506 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
507 struct memcg_shrinker_map *map;
510 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
511 /* Pairs with smp mb in shrink_slab() */
512 smp_mb__before_atomic();
513 set_bit(shrinker_id, map->map);
519 * mem_cgroup_css_from_page - css of the memcg associated with a page
520 * @page: page of interest
522 * If memcg is bound to the default hierarchy, css of the memcg associated
523 * with @page is returned. The returned css remains associated with @page
524 * until it is released.
526 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
529 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
531 struct mem_cgroup *memcg;
533 memcg = page->mem_cgroup;
535 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
536 memcg = root_mem_cgroup;
542 * page_cgroup_ino - return inode number of the memcg a page is charged to
545 * Look up the closest online ancestor of the memory cgroup @page is charged to
546 * and return its inode number or 0 if @page is not charged to any cgroup. It
547 * is safe to call this function without holding a reference to @page.
549 * Note, this function is inherently racy, because there is nothing to prevent
550 * the cgroup inode from getting torn down and potentially reallocated a moment
551 * after page_cgroup_ino() returns, so it only should be used by callers that
552 * do not care (such as procfs interfaces).
554 ino_t page_cgroup_ino(struct page *page)
556 struct mem_cgroup *memcg;
557 unsigned long ino = 0;
560 memcg = page->mem_cgroup;
563 * The lowest bit set means that memcg isn't a valid
564 * memcg pointer, but a obj_cgroups pointer.
565 * In this case the page is shared and doesn't belong
566 * to any specific memory cgroup.
568 if ((unsigned long) memcg & 0x1UL)
571 while (memcg && !(memcg->css.flags & CSS_ONLINE))
572 memcg = parent_mem_cgroup(memcg);
574 ino = cgroup_ino(memcg->css.cgroup);
579 static struct mem_cgroup_per_node *
580 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
582 int nid = page_to_nid(page);
584 return memcg->nodeinfo[nid];
587 static struct mem_cgroup_tree_per_node *
588 soft_limit_tree_node(int nid)
590 return soft_limit_tree.rb_tree_per_node[nid];
593 static struct mem_cgroup_tree_per_node *
594 soft_limit_tree_from_page(struct page *page)
596 int nid = page_to_nid(page);
598 return soft_limit_tree.rb_tree_per_node[nid];
601 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
602 struct mem_cgroup_tree_per_node *mctz,
603 unsigned long new_usage_in_excess)
605 struct rb_node **p = &mctz->rb_root.rb_node;
606 struct rb_node *parent = NULL;
607 struct mem_cgroup_per_node *mz_node;
608 bool rightmost = true;
613 mz->usage_in_excess = new_usage_in_excess;
614 if (!mz->usage_in_excess)
618 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
620 if (mz->usage_in_excess < mz_node->usage_in_excess) {
626 * We can't avoid mem cgroups that are over their soft
627 * limit by the same amount
629 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
634 mctz->rb_rightmost = &mz->tree_node;
636 rb_link_node(&mz->tree_node, parent, p);
637 rb_insert_color(&mz->tree_node, &mctz->rb_root);
641 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
642 struct mem_cgroup_tree_per_node *mctz)
647 if (&mz->tree_node == mctz->rb_rightmost)
648 mctz->rb_rightmost = rb_prev(&mz->tree_node);
650 rb_erase(&mz->tree_node, &mctz->rb_root);
654 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
655 struct mem_cgroup_tree_per_node *mctz)
659 spin_lock_irqsave(&mctz->lock, flags);
660 __mem_cgroup_remove_exceeded(mz, mctz);
661 spin_unlock_irqrestore(&mctz->lock, flags);
664 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
666 unsigned long nr_pages = page_counter_read(&memcg->memory);
667 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
668 unsigned long excess = 0;
670 if (nr_pages > soft_limit)
671 excess = nr_pages - soft_limit;
676 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
678 unsigned long excess;
679 struct mem_cgroup_per_node *mz;
680 struct mem_cgroup_tree_per_node *mctz;
682 mctz = soft_limit_tree_from_page(page);
686 * Necessary to update all ancestors when hierarchy is used.
687 * because their event counter is not touched.
689 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
690 mz = mem_cgroup_page_nodeinfo(memcg, page);
691 excess = soft_limit_excess(memcg);
693 * We have to update the tree if mz is on RB-tree or
694 * mem is over its softlimit.
696 if (excess || mz->on_tree) {
699 spin_lock_irqsave(&mctz->lock, flags);
700 /* if on-tree, remove it */
702 __mem_cgroup_remove_exceeded(mz, mctz);
704 * Insert again. mz->usage_in_excess will be updated.
705 * If excess is 0, no tree ops.
707 __mem_cgroup_insert_exceeded(mz, mctz, excess);
708 spin_unlock_irqrestore(&mctz->lock, flags);
713 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
715 struct mem_cgroup_tree_per_node *mctz;
716 struct mem_cgroup_per_node *mz;
720 mz = mem_cgroup_nodeinfo(memcg, nid);
721 mctz = soft_limit_tree_node(nid);
723 mem_cgroup_remove_exceeded(mz, mctz);
727 static struct mem_cgroup_per_node *
728 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
730 struct mem_cgroup_per_node *mz;
734 if (!mctz->rb_rightmost)
735 goto done; /* Nothing to reclaim from */
737 mz = rb_entry(mctz->rb_rightmost,
738 struct mem_cgroup_per_node, tree_node);
740 * Remove the node now but someone else can add it back,
741 * we will to add it back at the end of reclaim to its correct
742 * position in the tree.
744 __mem_cgroup_remove_exceeded(mz, mctz);
745 if (!soft_limit_excess(mz->memcg) ||
746 !css_tryget(&mz->memcg->css))
752 static struct mem_cgroup_per_node *
753 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
755 struct mem_cgroup_per_node *mz;
757 spin_lock_irq(&mctz->lock);
758 mz = __mem_cgroup_largest_soft_limit_node(mctz);
759 spin_unlock_irq(&mctz->lock);
764 * __mod_memcg_state - update cgroup memory statistics
765 * @memcg: the memory cgroup
766 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
767 * @val: delta to add to the counter, can be negative
769 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
771 long x, threshold = MEMCG_CHARGE_BATCH;
773 if (mem_cgroup_disabled())
776 if (memcg_stat_item_in_bytes(idx))
777 threshold <<= PAGE_SHIFT;
779 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
780 if (unlikely(abs(x) > threshold)) {
781 struct mem_cgroup *mi;
784 * Batch local counters to keep them in sync with
785 * the hierarchical ones.
787 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
788 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
789 atomic_long_add(x, &mi->vmstats[idx]);
792 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
795 static struct mem_cgroup_per_node *
796 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
798 struct mem_cgroup *parent;
800 parent = parent_mem_cgroup(pn->memcg);
803 return mem_cgroup_nodeinfo(parent, nid);
806 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
809 struct mem_cgroup_per_node *pn;
810 struct mem_cgroup *memcg;
811 long x, threshold = MEMCG_CHARGE_BATCH;
813 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
817 __mod_memcg_state(memcg, idx, val);
820 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
822 if (vmstat_item_in_bytes(idx))
823 threshold <<= PAGE_SHIFT;
825 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
826 if (unlikely(abs(x) > threshold)) {
827 pg_data_t *pgdat = lruvec_pgdat(lruvec);
828 struct mem_cgroup_per_node *pi;
830 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
831 atomic_long_add(x, &pi->lruvec_stat[idx]);
834 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
838 * __mod_lruvec_state - update lruvec memory statistics
839 * @lruvec: the lruvec
840 * @idx: the stat item
841 * @val: delta to add to the counter, can be negative
843 * The lruvec is the intersection of the NUMA node and a cgroup. This
844 * function updates the all three counters that are affected by a
845 * change of state at this level: per-node, per-cgroup, per-lruvec.
847 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
851 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
853 /* Update memcg and lruvec */
854 if (!mem_cgroup_disabled())
855 __mod_memcg_lruvec_state(lruvec, idx, val);
858 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
860 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
861 struct mem_cgroup *memcg;
862 struct lruvec *lruvec;
865 memcg = mem_cgroup_from_obj(p);
867 /* Untracked pages have no memcg, no lruvec. Update only the node */
868 if (!memcg || memcg == root_mem_cgroup) {
869 __mod_node_page_state(pgdat, idx, val);
871 lruvec = mem_cgroup_lruvec(memcg, pgdat);
872 __mod_lruvec_state(lruvec, idx, val);
877 void mod_memcg_obj_state(void *p, int idx, int val)
879 struct mem_cgroup *memcg;
882 memcg = mem_cgroup_from_obj(p);
884 mod_memcg_state(memcg, idx, val);
889 * __count_memcg_events - account VM events in a cgroup
890 * @memcg: the memory cgroup
891 * @idx: the event item
892 * @count: the number of events that occured
894 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
899 if (mem_cgroup_disabled())
902 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
903 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
904 struct mem_cgroup *mi;
907 * Batch local counters to keep them in sync with
908 * the hierarchical ones.
910 __this_cpu_add(memcg->vmstats_local->events[idx], x);
911 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
912 atomic_long_add(x, &mi->vmevents[idx]);
915 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
918 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
920 return atomic_long_read(&memcg->vmevents[event]);
923 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
928 for_each_possible_cpu(cpu)
929 x += per_cpu(memcg->vmstats_local->events[event], cpu);
933 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
937 /* pagein of a big page is an event. So, ignore page size */
939 __count_memcg_events(memcg, PGPGIN, 1);
941 __count_memcg_events(memcg, PGPGOUT, 1);
942 nr_pages = -nr_pages; /* for event */
945 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
948 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
949 enum mem_cgroup_events_target target)
951 unsigned long val, next;
953 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
954 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
955 /* from time_after() in jiffies.h */
956 if ((long)(next - val) < 0) {
958 case MEM_CGROUP_TARGET_THRESH:
959 next = val + THRESHOLDS_EVENTS_TARGET;
961 case MEM_CGROUP_TARGET_SOFTLIMIT:
962 next = val + SOFTLIMIT_EVENTS_TARGET;
967 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
974 * Check events in order.
977 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
979 /* threshold event is triggered in finer grain than soft limit */
980 if (unlikely(mem_cgroup_event_ratelimit(memcg,
981 MEM_CGROUP_TARGET_THRESH))) {
984 do_softlimit = mem_cgroup_event_ratelimit(memcg,
985 MEM_CGROUP_TARGET_SOFTLIMIT);
986 mem_cgroup_threshold(memcg);
987 if (unlikely(do_softlimit))
988 mem_cgroup_update_tree(memcg, page);
992 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
995 * mm_update_next_owner() may clear mm->owner to NULL
996 * if it races with swapoff, page migration, etc.
997 * So this can be called with p == NULL.
1002 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1004 EXPORT_SYMBOL(mem_cgroup_from_task);
1007 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1008 * @mm: mm from which memcg should be extracted. It can be NULL.
1010 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1011 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1014 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1016 struct mem_cgroup *memcg;
1018 if (mem_cgroup_disabled())
1024 * Page cache insertions can happen withou an
1025 * actual mm context, e.g. during disk probing
1026 * on boot, loopback IO, acct() writes etc.
1029 memcg = root_mem_cgroup;
1031 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1032 if (unlikely(!memcg))
1033 memcg = root_mem_cgroup;
1035 } while (!css_tryget(&memcg->css));
1039 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1042 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1043 * @page: page from which memcg should be extracted.
1045 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1046 * root_mem_cgroup is returned.
1048 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
1050 struct mem_cgroup *memcg = page->mem_cgroup;
1052 if (mem_cgroup_disabled())
1056 /* Page should not get uncharged and freed memcg under us. */
1057 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
1058 memcg = root_mem_cgroup;
1062 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1065 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
1067 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1069 if (unlikely(current->active_memcg)) {
1070 struct mem_cgroup *memcg;
1073 /* current->active_memcg must hold a ref. */
1074 if (WARN_ON_ONCE(!css_tryget(¤t->active_memcg->css)))
1075 memcg = root_mem_cgroup;
1077 memcg = current->active_memcg;
1081 return get_mem_cgroup_from_mm(current->mm);
1085 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1086 * @root: hierarchy root
1087 * @prev: previously returned memcg, NULL on first invocation
1088 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1090 * Returns references to children of the hierarchy below @root, or
1091 * @root itself, or %NULL after a full round-trip.
1093 * Caller must pass the return value in @prev on subsequent
1094 * invocations for reference counting, or use mem_cgroup_iter_break()
1095 * to cancel a hierarchy walk before the round-trip is complete.
1097 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1098 * in the hierarchy among all concurrent reclaimers operating on the
1101 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1102 struct mem_cgroup *prev,
1103 struct mem_cgroup_reclaim_cookie *reclaim)
1105 struct mem_cgroup_reclaim_iter *iter;
1106 struct cgroup_subsys_state *css = NULL;
1107 struct mem_cgroup *memcg = NULL;
1108 struct mem_cgroup *pos = NULL;
1110 if (mem_cgroup_disabled())
1114 root = root_mem_cgroup;
1116 if (prev && !reclaim)
1119 if (!root->use_hierarchy && root != root_mem_cgroup) {
1128 struct mem_cgroup_per_node *mz;
1130 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1133 if (prev && reclaim->generation != iter->generation)
1137 pos = READ_ONCE(iter->position);
1138 if (!pos || css_tryget(&pos->css))
1141 * css reference reached zero, so iter->position will
1142 * be cleared by ->css_released. However, we should not
1143 * rely on this happening soon, because ->css_released
1144 * is called from a work queue, and by busy-waiting we
1145 * might block it. So we clear iter->position right
1148 (void)cmpxchg(&iter->position, pos, NULL);
1156 css = css_next_descendant_pre(css, &root->css);
1159 * Reclaimers share the hierarchy walk, and a
1160 * new one might jump in right at the end of
1161 * the hierarchy - make sure they see at least
1162 * one group and restart from the beginning.
1170 * Verify the css and acquire a reference. The root
1171 * is provided by the caller, so we know it's alive
1172 * and kicking, and don't take an extra reference.
1174 memcg = mem_cgroup_from_css(css);
1176 if (css == &root->css)
1179 if (css_tryget(css))
1187 * The position could have already been updated by a competing
1188 * thread, so check that the value hasn't changed since we read
1189 * it to avoid reclaiming from the same cgroup twice.
1191 (void)cmpxchg(&iter->position, pos, memcg);
1199 reclaim->generation = iter->generation;
1205 if (prev && prev != root)
1206 css_put(&prev->css);
1212 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1213 * @root: hierarchy root
1214 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1216 void mem_cgroup_iter_break(struct mem_cgroup *root,
1217 struct mem_cgroup *prev)
1220 root = root_mem_cgroup;
1221 if (prev && prev != root)
1222 css_put(&prev->css);
1225 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1226 struct mem_cgroup *dead_memcg)
1228 struct mem_cgroup_reclaim_iter *iter;
1229 struct mem_cgroup_per_node *mz;
1232 for_each_node(nid) {
1233 mz = mem_cgroup_nodeinfo(from, nid);
1235 cmpxchg(&iter->position, dead_memcg, NULL);
1239 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1241 struct mem_cgroup *memcg = dead_memcg;
1242 struct mem_cgroup *last;
1245 __invalidate_reclaim_iterators(memcg, dead_memcg);
1247 } while ((memcg = parent_mem_cgroup(memcg)));
1250 * When cgruop1 non-hierarchy mode is used,
1251 * parent_mem_cgroup() does not walk all the way up to the
1252 * cgroup root (root_mem_cgroup). So we have to handle
1253 * dead_memcg from cgroup root separately.
1255 if (last != root_mem_cgroup)
1256 __invalidate_reclaim_iterators(root_mem_cgroup,
1261 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1262 * @memcg: hierarchy root
1263 * @fn: function to call for each task
1264 * @arg: argument passed to @fn
1266 * This function iterates over tasks attached to @memcg or to any of its
1267 * descendants and calls @fn for each task. If @fn returns a non-zero
1268 * value, the function breaks the iteration loop and returns the value.
1269 * Otherwise, it will iterate over all tasks and return 0.
1271 * This function must not be called for the root memory cgroup.
1273 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1274 int (*fn)(struct task_struct *, void *), void *arg)
1276 struct mem_cgroup *iter;
1279 BUG_ON(memcg == root_mem_cgroup);
1281 for_each_mem_cgroup_tree(iter, memcg) {
1282 struct css_task_iter it;
1283 struct task_struct *task;
1285 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1286 while (!ret && (task = css_task_iter_next(&it)))
1287 ret = fn(task, arg);
1288 css_task_iter_end(&it);
1290 mem_cgroup_iter_break(memcg, iter);
1298 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1300 * @pgdat: pgdat of the page
1302 * This function relies on page->mem_cgroup being stable - see the
1303 * access rules in commit_charge().
1305 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1307 struct mem_cgroup_per_node *mz;
1308 struct mem_cgroup *memcg;
1309 struct lruvec *lruvec;
1311 if (mem_cgroup_disabled()) {
1312 lruvec = &pgdat->__lruvec;
1316 memcg = page->mem_cgroup;
1318 * Swapcache readahead pages are added to the LRU - and
1319 * possibly migrated - before they are charged.
1322 memcg = root_mem_cgroup;
1324 mz = mem_cgroup_page_nodeinfo(memcg, page);
1325 lruvec = &mz->lruvec;
1328 * Since a node can be onlined after the mem_cgroup was created,
1329 * we have to be prepared to initialize lruvec->zone here;
1330 * and if offlined then reonlined, we need to reinitialize it.
1332 if (unlikely(lruvec->pgdat != pgdat))
1333 lruvec->pgdat = pgdat;
1338 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1339 * @lruvec: mem_cgroup per zone lru vector
1340 * @lru: index of lru list the page is sitting on
1341 * @zid: zone id of the accounted pages
1342 * @nr_pages: positive when adding or negative when removing
1344 * This function must be called under lru_lock, just before a page is added
1345 * to or just after a page is removed from an lru list (that ordering being
1346 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1348 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1349 int zid, int nr_pages)
1351 struct mem_cgroup_per_node *mz;
1352 unsigned long *lru_size;
1355 if (mem_cgroup_disabled())
1358 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1359 lru_size = &mz->lru_zone_size[zid][lru];
1362 *lru_size += nr_pages;
1365 if (WARN_ONCE(size < 0,
1366 "%s(%p, %d, %d): lru_size %ld\n",
1367 __func__, lruvec, lru, nr_pages, size)) {
1373 *lru_size += nr_pages;
1377 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1378 * @memcg: the memory cgroup
1380 * Returns the maximum amount of memory @mem can be charged with, in
1383 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1385 unsigned long margin = 0;
1386 unsigned long count;
1387 unsigned long limit;
1389 count = page_counter_read(&memcg->memory);
1390 limit = READ_ONCE(memcg->memory.max);
1392 margin = limit - count;
1394 if (do_memsw_account()) {
1395 count = page_counter_read(&memcg->memsw);
1396 limit = READ_ONCE(memcg->memsw.max);
1398 margin = min(margin, limit - count);
1407 * A routine for checking "mem" is under move_account() or not.
1409 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1410 * moving cgroups. This is for waiting at high-memory pressure
1413 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1415 struct mem_cgroup *from;
1416 struct mem_cgroup *to;
1419 * Unlike task_move routines, we access mc.to, mc.from not under
1420 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1422 spin_lock(&mc.lock);
1428 ret = mem_cgroup_is_descendant(from, memcg) ||
1429 mem_cgroup_is_descendant(to, memcg);
1431 spin_unlock(&mc.lock);
1435 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1437 if (mc.moving_task && current != mc.moving_task) {
1438 if (mem_cgroup_under_move(memcg)) {
1440 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1441 /* moving charge context might have finished. */
1444 finish_wait(&mc.waitq, &wait);
1451 struct memory_stat {
1457 static struct memory_stat memory_stats[] = {
1458 { "anon", PAGE_SIZE, NR_ANON_MAPPED },
1459 { "file", PAGE_SIZE, NR_FILE_PAGES },
1460 { "kernel_stack", 1024, NR_KERNEL_STACK_KB },
1461 { "percpu", 1, MEMCG_PERCPU_B },
1462 { "sock", PAGE_SIZE, MEMCG_SOCK },
1463 { "shmem", PAGE_SIZE, NR_SHMEM },
1464 { "file_mapped", PAGE_SIZE, NR_FILE_MAPPED },
1465 { "file_dirty", PAGE_SIZE, NR_FILE_DIRTY },
1466 { "file_writeback", PAGE_SIZE, NR_WRITEBACK },
1467 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1469 * The ratio will be initialized in memory_stats_init(). Because
1470 * on some architectures, the macro of HPAGE_PMD_SIZE is not
1471 * constant(e.g. powerpc).
1473 { "anon_thp", 0, NR_ANON_THPS },
1475 { "inactive_anon", PAGE_SIZE, NR_INACTIVE_ANON },
1476 { "active_anon", PAGE_SIZE, NR_ACTIVE_ANON },
1477 { "inactive_file", PAGE_SIZE, NR_INACTIVE_FILE },
1478 { "active_file", PAGE_SIZE, NR_ACTIVE_FILE },
1479 { "unevictable", PAGE_SIZE, NR_UNEVICTABLE },
1482 * Note: The slab_reclaimable and slab_unreclaimable must be
1483 * together and slab_reclaimable must be in front.
1485 { "slab_reclaimable", 1, NR_SLAB_RECLAIMABLE_B },
1486 { "slab_unreclaimable", 1, NR_SLAB_UNRECLAIMABLE_B },
1488 /* The memory events */
1489 { "workingset_refault_anon", 1, WORKINGSET_REFAULT_ANON },
1490 { "workingset_refault_file", 1, WORKINGSET_REFAULT_FILE },
1491 { "workingset_activate_anon", 1, WORKINGSET_ACTIVATE_ANON },
1492 { "workingset_activate_file", 1, WORKINGSET_ACTIVATE_FILE },
1493 { "workingset_restore_anon", 1, WORKINGSET_RESTORE_ANON },
1494 { "workingset_restore_file", 1, WORKINGSET_RESTORE_FILE },
1495 { "workingset_nodereclaim", 1, WORKINGSET_NODERECLAIM },
1498 static int __init memory_stats_init(void)
1502 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1503 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1504 if (memory_stats[i].idx == NR_ANON_THPS)
1505 memory_stats[i].ratio = HPAGE_PMD_SIZE;
1507 VM_BUG_ON(!memory_stats[i].ratio);
1508 VM_BUG_ON(memory_stats[i].idx >= MEMCG_NR_STAT);
1513 pure_initcall(memory_stats_init);
1515 static char *memory_stat_format(struct mem_cgroup *memcg)
1520 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1525 * Provide statistics on the state of the memory subsystem as
1526 * well as cumulative event counters that show past behavior.
1528 * This list is ordered following a combination of these gradients:
1529 * 1) generic big picture -> specifics and details
1530 * 2) reflecting userspace activity -> reflecting kernel heuristics
1532 * Current memory state:
1535 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1538 size = memcg_page_state(memcg, memory_stats[i].idx);
1539 size *= memory_stats[i].ratio;
1540 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1542 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1543 size = memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B) +
1544 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B);
1545 seq_buf_printf(&s, "slab %llu\n", size);
1549 /* Accumulated memory events */
1551 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1552 memcg_events(memcg, PGFAULT));
1553 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1554 memcg_events(memcg, PGMAJFAULT));
1555 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1556 memcg_events(memcg, PGREFILL));
1557 seq_buf_printf(&s, "pgscan %lu\n",
1558 memcg_events(memcg, PGSCAN_KSWAPD) +
1559 memcg_events(memcg, PGSCAN_DIRECT));
1560 seq_buf_printf(&s, "pgsteal %lu\n",
1561 memcg_events(memcg, PGSTEAL_KSWAPD) +
1562 memcg_events(memcg, PGSTEAL_DIRECT));
1563 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1564 memcg_events(memcg, PGACTIVATE));
1565 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1566 memcg_events(memcg, PGDEACTIVATE));
1567 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1568 memcg_events(memcg, PGLAZYFREE));
1569 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1570 memcg_events(memcg, PGLAZYFREED));
1572 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1573 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1574 memcg_events(memcg, THP_FAULT_ALLOC));
1575 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1576 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1577 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1579 /* The above should easily fit into one page */
1580 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1585 #define K(x) ((x) << (PAGE_SHIFT-10))
1587 * mem_cgroup_print_oom_context: Print OOM information relevant to
1588 * memory controller.
1589 * @memcg: The memory cgroup that went over limit
1590 * @p: Task that is going to be killed
1592 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1595 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1600 pr_cont(",oom_memcg=");
1601 pr_cont_cgroup_path(memcg->css.cgroup);
1603 pr_cont(",global_oom");
1605 pr_cont(",task_memcg=");
1606 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1612 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1613 * memory controller.
1614 * @memcg: The memory cgroup that went over limit
1616 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1620 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1621 K((u64)page_counter_read(&memcg->memory)),
1622 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1623 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1624 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1625 K((u64)page_counter_read(&memcg->swap)),
1626 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1628 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1629 K((u64)page_counter_read(&memcg->memsw)),
1630 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1631 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1632 K((u64)page_counter_read(&memcg->kmem)),
1633 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1636 pr_info("Memory cgroup stats for ");
1637 pr_cont_cgroup_path(memcg->css.cgroup);
1639 buf = memory_stat_format(memcg);
1647 * Return the memory (and swap, if configured) limit for a memcg.
1649 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1651 unsigned long max = READ_ONCE(memcg->memory.max);
1653 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1654 if (mem_cgroup_swappiness(memcg))
1655 max += min(READ_ONCE(memcg->swap.max),
1656 (unsigned long)total_swap_pages);
1658 if (mem_cgroup_swappiness(memcg)) {
1659 /* Calculate swap excess capacity from memsw limit */
1660 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1662 max += min(swap, (unsigned long)total_swap_pages);
1668 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1670 return page_counter_read(&memcg->memory);
1673 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1676 struct oom_control oc = {
1680 .gfp_mask = gfp_mask,
1685 if (mutex_lock_killable(&oom_lock))
1688 if (mem_cgroup_margin(memcg) >= (1 << order))
1692 * A few threads which were not waiting at mutex_lock_killable() can
1693 * fail to bail out. Therefore, check again after holding oom_lock.
1695 ret = should_force_charge() || out_of_memory(&oc);
1698 mutex_unlock(&oom_lock);
1702 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1705 unsigned long *total_scanned)
1707 struct mem_cgroup *victim = NULL;
1710 unsigned long excess;
1711 unsigned long nr_scanned;
1712 struct mem_cgroup_reclaim_cookie reclaim = {
1716 excess = soft_limit_excess(root_memcg);
1719 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1724 * If we have not been able to reclaim
1725 * anything, it might because there are
1726 * no reclaimable pages under this hierarchy
1731 * We want to do more targeted reclaim.
1732 * excess >> 2 is not to excessive so as to
1733 * reclaim too much, nor too less that we keep
1734 * coming back to reclaim from this cgroup
1736 if (total >= (excess >> 2) ||
1737 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1742 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1743 pgdat, &nr_scanned);
1744 *total_scanned += nr_scanned;
1745 if (!soft_limit_excess(root_memcg))
1748 mem_cgroup_iter_break(root_memcg, victim);
1752 #ifdef CONFIG_LOCKDEP
1753 static struct lockdep_map memcg_oom_lock_dep_map = {
1754 .name = "memcg_oom_lock",
1758 static DEFINE_SPINLOCK(memcg_oom_lock);
1761 * Check OOM-Killer is already running under our hierarchy.
1762 * If someone is running, return false.
1764 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1766 struct mem_cgroup *iter, *failed = NULL;
1768 spin_lock(&memcg_oom_lock);
1770 for_each_mem_cgroup_tree(iter, memcg) {
1771 if (iter->oom_lock) {
1773 * this subtree of our hierarchy is already locked
1774 * so we cannot give a lock.
1777 mem_cgroup_iter_break(memcg, iter);
1780 iter->oom_lock = true;
1785 * OK, we failed to lock the whole subtree so we have
1786 * to clean up what we set up to the failing subtree
1788 for_each_mem_cgroup_tree(iter, memcg) {
1789 if (iter == failed) {
1790 mem_cgroup_iter_break(memcg, iter);
1793 iter->oom_lock = false;
1796 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1798 spin_unlock(&memcg_oom_lock);
1803 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1805 struct mem_cgroup *iter;
1807 spin_lock(&memcg_oom_lock);
1808 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1809 for_each_mem_cgroup_tree(iter, memcg)
1810 iter->oom_lock = false;
1811 spin_unlock(&memcg_oom_lock);
1814 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1816 struct mem_cgroup *iter;
1818 spin_lock(&memcg_oom_lock);
1819 for_each_mem_cgroup_tree(iter, memcg)
1821 spin_unlock(&memcg_oom_lock);
1824 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1826 struct mem_cgroup *iter;
1829 * Be careful about under_oom underflows becase a child memcg
1830 * could have been added after mem_cgroup_mark_under_oom.
1832 spin_lock(&memcg_oom_lock);
1833 for_each_mem_cgroup_tree(iter, memcg)
1834 if (iter->under_oom > 0)
1836 spin_unlock(&memcg_oom_lock);
1839 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1841 struct oom_wait_info {
1842 struct mem_cgroup *memcg;
1843 wait_queue_entry_t wait;
1846 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1847 unsigned mode, int sync, void *arg)
1849 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1850 struct mem_cgroup *oom_wait_memcg;
1851 struct oom_wait_info *oom_wait_info;
1853 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1854 oom_wait_memcg = oom_wait_info->memcg;
1856 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1857 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1859 return autoremove_wake_function(wait, mode, sync, arg);
1862 static void memcg_oom_recover(struct mem_cgroup *memcg)
1865 * For the following lockless ->under_oom test, the only required
1866 * guarantee is that it must see the state asserted by an OOM when
1867 * this function is called as a result of userland actions
1868 * triggered by the notification of the OOM. This is trivially
1869 * achieved by invoking mem_cgroup_mark_under_oom() before
1870 * triggering notification.
1872 if (memcg && memcg->under_oom)
1873 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1883 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1885 enum oom_status ret;
1888 if (order > PAGE_ALLOC_COSTLY_ORDER)
1891 memcg_memory_event(memcg, MEMCG_OOM);
1894 * We are in the middle of the charge context here, so we
1895 * don't want to block when potentially sitting on a callstack
1896 * that holds all kinds of filesystem and mm locks.
1898 * cgroup1 allows disabling the OOM killer and waiting for outside
1899 * handling until the charge can succeed; remember the context and put
1900 * the task to sleep at the end of the page fault when all locks are
1903 * On the other hand, in-kernel OOM killer allows for an async victim
1904 * memory reclaim (oom_reaper) and that means that we are not solely
1905 * relying on the oom victim to make a forward progress and we can
1906 * invoke the oom killer here.
1908 * Please note that mem_cgroup_out_of_memory might fail to find a
1909 * victim and then we have to bail out from the charge path.
1911 if (memcg->oom_kill_disable) {
1912 if (!current->in_user_fault)
1914 css_get(&memcg->css);
1915 current->memcg_in_oom = memcg;
1916 current->memcg_oom_gfp_mask = mask;
1917 current->memcg_oom_order = order;
1922 mem_cgroup_mark_under_oom(memcg);
1924 locked = mem_cgroup_oom_trylock(memcg);
1927 mem_cgroup_oom_notify(memcg);
1929 mem_cgroup_unmark_under_oom(memcg);
1930 if (mem_cgroup_out_of_memory(memcg, mask, order))
1936 mem_cgroup_oom_unlock(memcg);
1942 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1943 * @handle: actually kill/wait or just clean up the OOM state
1945 * This has to be called at the end of a page fault if the memcg OOM
1946 * handler was enabled.
1948 * Memcg supports userspace OOM handling where failed allocations must
1949 * sleep on a waitqueue until the userspace task resolves the
1950 * situation. Sleeping directly in the charge context with all kinds
1951 * of locks held is not a good idea, instead we remember an OOM state
1952 * in the task and mem_cgroup_oom_synchronize() has to be called at
1953 * the end of the page fault to complete the OOM handling.
1955 * Returns %true if an ongoing memcg OOM situation was detected and
1956 * completed, %false otherwise.
1958 bool mem_cgroup_oom_synchronize(bool handle)
1960 struct mem_cgroup *memcg = current->memcg_in_oom;
1961 struct oom_wait_info owait;
1964 /* OOM is global, do not handle */
1971 owait.memcg = memcg;
1972 owait.wait.flags = 0;
1973 owait.wait.func = memcg_oom_wake_function;
1974 owait.wait.private = current;
1975 INIT_LIST_HEAD(&owait.wait.entry);
1977 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1978 mem_cgroup_mark_under_oom(memcg);
1980 locked = mem_cgroup_oom_trylock(memcg);
1983 mem_cgroup_oom_notify(memcg);
1985 if (locked && !memcg->oom_kill_disable) {
1986 mem_cgroup_unmark_under_oom(memcg);
1987 finish_wait(&memcg_oom_waitq, &owait.wait);
1988 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1989 current->memcg_oom_order);
1992 mem_cgroup_unmark_under_oom(memcg);
1993 finish_wait(&memcg_oom_waitq, &owait.wait);
1997 mem_cgroup_oom_unlock(memcg);
1999 * There is no guarantee that an OOM-lock contender
2000 * sees the wakeups triggered by the OOM kill
2001 * uncharges. Wake any sleepers explicitely.
2003 memcg_oom_recover(memcg);
2006 current->memcg_in_oom = NULL;
2007 css_put(&memcg->css);
2012 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2013 * @victim: task to be killed by the OOM killer
2014 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2016 * Returns a pointer to a memory cgroup, which has to be cleaned up
2017 * by killing all belonging OOM-killable tasks.
2019 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2021 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2022 struct mem_cgroup *oom_domain)
2024 struct mem_cgroup *oom_group = NULL;
2025 struct mem_cgroup *memcg;
2027 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2031 oom_domain = root_mem_cgroup;
2035 memcg = mem_cgroup_from_task(victim);
2036 if (memcg == root_mem_cgroup)
2040 * If the victim task has been asynchronously moved to a different
2041 * memory cgroup, we might end up killing tasks outside oom_domain.
2042 * In this case it's better to ignore memory.group.oom.
2044 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2048 * Traverse the memory cgroup hierarchy from the victim task's
2049 * cgroup up to the OOMing cgroup (or root) to find the
2050 * highest-level memory cgroup with oom.group set.
2052 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2053 if (memcg->oom_group)
2056 if (memcg == oom_domain)
2061 css_get(&oom_group->css);
2068 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2070 pr_info("Tasks in ");
2071 pr_cont_cgroup_path(memcg->css.cgroup);
2072 pr_cont(" are going to be killed due to memory.oom.group set\n");
2076 * lock_page_memcg - lock a page->mem_cgroup binding
2079 * This function protects unlocked LRU pages from being moved to
2082 * It ensures lifetime of the returned memcg. Caller is responsible
2083 * for the lifetime of the page; __unlock_page_memcg() is available
2084 * when @page might get freed inside the locked section.
2086 struct mem_cgroup *lock_page_memcg(struct page *page)
2088 struct page *head = compound_head(page); /* rmap on tail pages */
2089 struct mem_cgroup *memcg;
2090 unsigned long flags;
2093 * The RCU lock is held throughout the transaction. The fast
2094 * path can get away without acquiring the memcg->move_lock
2095 * because page moving starts with an RCU grace period.
2097 * The RCU lock also protects the memcg from being freed when
2098 * the page state that is going to change is the only thing
2099 * preventing the page itself from being freed. E.g. writeback
2100 * doesn't hold a page reference and relies on PG_writeback to
2101 * keep off truncation, migration and so forth.
2105 if (mem_cgroup_disabled())
2108 memcg = head->mem_cgroup;
2109 if (unlikely(!memcg))
2112 if (atomic_read(&memcg->moving_account) <= 0)
2115 spin_lock_irqsave(&memcg->move_lock, flags);
2116 if (memcg != head->mem_cgroup) {
2117 spin_unlock_irqrestore(&memcg->move_lock, flags);
2122 * When charge migration first begins, we can have locked and
2123 * unlocked page stat updates happening concurrently. Track
2124 * the task who has the lock for unlock_page_memcg().
2126 memcg->move_lock_task = current;
2127 memcg->move_lock_flags = flags;
2131 EXPORT_SYMBOL(lock_page_memcg);
2134 * __unlock_page_memcg - unlock and unpin a memcg
2137 * Unlock and unpin a memcg returned by lock_page_memcg().
2139 void __unlock_page_memcg(struct mem_cgroup *memcg)
2141 if (memcg && memcg->move_lock_task == current) {
2142 unsigned long flags = memcg->move_lock_flags;
2144 memcg->move_lock_task = NULL;
2145 memcg->move_lock_flags = 0;
2147 spin_unlock_irqrestore(&memcg->move_lock, flags);
2154 * unlock_page_memcg - unlock a page->mem_cgroup binding
2157 void unlock_page_memcg(struct page *page)
2159 struct page *head = compound_head(page);
2161 __unlock_page_memcg(head->mem_cgroup);
2163 EXPORT_SYMBOL(unlock_page_memcg);
2165 struct memcg_stock_pcp {
2166 struct mem_cgroup *cached; /* this never be root cgroup */
2167 unsigned int nr_pages;
2169 #ifdef CONFIG_MEMCG_KMEM
2170 struct obj_cgroup *cached_objcg;
2171 unsigned int nr_bytes;
2174 struct work_struct work;
2175 unsigned long flags;
2176 #define FLUSHING_CACHED_CHARGE 0
2178 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2179 static DEFINE_MUTEX(percpu_charge_mutex);
2181 #ifdef CONFIG_MEMCG_KMEM
2182 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2183 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2184 struct mem_cgroup *root_memcg);
2187 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2190 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2191 struct mem_cgroup *root_memcg)
2198 * consume_stock: Try to consume stocked charge on this cpu.
2199 * @memcg: memcg to consume from.
2200 * @nr_pages: how many pages to charge.
2202 * The charges will only happen if @memcg matches the current cpu's memcg
2203 * stock, and at least @nr_pages are available in that stock. Failure to
2204 * service an allocation will refill the stock.
2206 * returns true if successful, false otherwise.
2208 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2210 struct memcg_stock_pcp *stock;
2211 unsigned long flags;
2214 if (nr_pages > MEMCG_CHARGE_BATCH)
2217 local_irq_save(flags);
2219 stock = this_cpu_ptr(&memcg_stock);
2220 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2221 stock->nr_pages -= nr_pages;
2225 local_irq_restore(flags);
2231 * Returns stocks cached in percpu and reset cached information.
2233 static void drain_stock(struct memcg_stock_pcp *stock)
2235 struct mem_cgroup *old = stock->cached;
2240 if (stock->nr_pages) {
2241 page_counter_uncharge(&old->memory, stock->nr_pages);
2242 if (do_memsw_account())
2243 page_counter_uncharge(&old->memsw, stock->nr_pages);
2244 stock->nr_pages = 0;
2248 stock->cached = NULL;
2251 static void drain_local_stock(struct work_struct *dummy)
2253 struct memcg_stock_pcp *stock;
2254 unsigned long flags;
2257 * The only protection from memory hotplug vs. drain_stock races is
2258 * that we always operate on local CPU stock here with IRQ disabled
2260 local_irq_save(flags);
2262 stock = this_cpu_ptr(&memcg_stock);
2263 drain_obj_stock(stock);
2265 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2267 local_irq_restore(flags);
2271 * Cache charges(val) to local per_cpu area.
2272 * This will be consumed by consume_stock() function, later.
2274 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2276 struct memcg_stock_pcp *stock;
2277 unsigned long flags;
2279 local_irq_save(flags);
2281 stock = this_cpu_ptr(&memcg_stock);
2282 if (stock->cached != memcg) { /* reset if necessary */
2284 css_get(&memcg->css);
2285 stock->cached = memcg;
2287 stock->nr_pages += nr_pages;
2289 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2292 local_irq_restore(flags);
2296 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2297 * of the hierarchy under it.
2299 static void drain_all_stock(struct mem_cgroup *root_memcg)
2303 /* If someone's already draining, avoid adding running more workers. */
2304 if (!mutex_trylock(&percpu_charge_mutex))
2307 * Notify other cpus that system-wide "drain" is running
2308 * We do not care about races with the cpu hotplug because cpu down
2309 * as well as workers from this path always operate on the local
2310 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2313 for_each_online_cpu(cpu) {
2314 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2315 struct mem_cgroup *memcg;
2319 memcg = stock->cached;
2320 if (memcg && stock->nr_pages &&
2321 mem_cgroup_is_descendant(memcg, root_memcg))
2323 if (obj_stock_flush_required(stock, root_memcg))
2328 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2330 drain_local_stock(&stock->work);
2332 schedule_work_on(cpu, &stock->work);
2336 mutex_unlock(&percpu_charge_mutex);
2339 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2341 struct memcg_stock_pcp *stock;
2342 struct mem_cgroup *memcg, *mi;
2344 stock = &per_cpu(memcg_stock, cpu);
2347 for_each_mem_cgroup(memcg) {
2350 for (i = 0; i < MEMCG_NR_STAT; i++) {
2354 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2356 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2357 atomic_long_add(x, &memcg->vmstats[i]);
2359 if (i >= NR_VM_NODE_STAT_ITEMS)
2362 for_each_node(nid) {
2363 struct mem_cgroup_per_node *pn;
2365 pn = mem_cgroup_nodeinfo(memcg, nid);
2366 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2369 atomic_long_add(x, &pn->lruvec_stat[i]);
2370 } while ((pn = parent_nodeinfo(pn, nid)));
2374 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2377 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2379 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2380 atomic_long_add(x, &memcg->vmevents[i]);
2387 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2388 unsigned int nr_pages,
2391 unsigned long nr_reclaimed = 0;
2394 unsigned long pflags;
2396 if (page_counter_read(&memcg->memory) <=
2397 READ_ONCE(memcg->memory.high))
2400 memcg_memory_event(memcg, MEMCG_HIGH);
2402 psi_memstall_enter(&pflags);
2403 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2405 psi_memstall_leave(&pflags);
2406 } while ((memcg = parent_mem_cgroup(memcg)) &&
2407 !mem_cgroup_is_root(memcg));
2409 return nr_reclaimed;
2412 static void high_work_func(struct work_struct *work)
2414 struct mem_cgroup *memcg;
2416 memcg = container_of(work, struct mem_cgroup, high_work);
2417 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2421 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2422 * enough to still cause a significant slowdown in most cases, while still
2423 * allowing diagnostics and tracing to proceed without becoming stuck.
2425 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2428 * When calculating the delay, we use these either side of the exponentiation to
2429 * maintain precision and scale to a reasonable number of jiffies (see the table
2432 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2433 * overage ratio to a delay.
2434 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2435 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2436 * to produce a reasonable delay curve.
2438 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2439 * reasonable delay curve compared to precision-adjusted overage, not
2440 * penalising heavily at first, but still making sure that growth beyond the
2441 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2442 * example, with a high of 100 megabytes:
2444 * +-------+------------------------+
2445 * | usage | time to allocate in ms |
2446 * +-------+------------------------+
2468 * +-------+------------------------+
2470 #define MEMCG_DELAY_PRECISION_SHIFT 20
2471 #define MEMCG_DELAY_SCALING_SHIFT 14
2473 static u64 calculate_overage(unsigned long usage, unsigned long high)
2481 * Prevent division by 0 in overage calculation by acting as if
2482 * it was a threshold of 1 page
2484 high = max(high, 1UL);
2486 overage = usage - high;
2487 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2488 return div64_u64(overage, high);
2491 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2493 u64 overage, max_overage = 0;
2496 overage = calculate_overage(page_counter_read(&memcg->memory),
2497 READ_ONCE(memcg->memory.high));
2498 max_overage = max(overage, max_overage);
2499 } while ((memcg = parent_mem_cgroup(memcg)) &&
2500 !mem_cgroup_is_root(memcg));
2505 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2507 u64 overage, max_overage = 0;
2510 overage = calculate_overage(page_counter_read(&memcg->swap),
2511 READ_ONCE(memcg->swap.high));
2513 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2514 max_overage = max(overage, max_overage);
2515 } while ((memcg = parent_mem_cgroup(memcg)) &&
2516 !mem_cgroup_is_root(memcg));
2522 * Get the number of jiffies that we should penalise a mischievous cgroup which
2523 * is exceeding its memory.high by checking both it and its ancestors.
2525 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2526 unsigned int nr_pages,
2529 unsigned long penalty_jiffies;
2535 * We use overage compared to memory.high to calculate the number of
2536 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2537 * fairly lenient on small overages, and increasingly harsh when the
2538 * memcg in question makes it clear that it has no intention of stopping
2539 * its crazy behaviour, so we exponentially increase the delay based on
2542 penalty_jiffies = max_overage * max_overage * HZ;
2543 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2544 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2547 * Factor in the task's own contribution to the overage, such that four
2548 * N-sized allocations are throttled approximately the same as one
2549 * 4N-sized allocation.
2551 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2552 * larger the current charge patch is than that.
2554 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2558 * Scheduled by try_charge() to be executed from the userland return path
2559 * and reclaims memory over the high limit.
2561 void mem_cgroup_handle_over_high(void)
2563 unsigned long penalty_jiffies;
2564 unsigned long pflags;
2565 unsigned long nr_reclaimed;
2566 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2567 int nr_retries = MAX_RECLAIM_RETRIES;
2568 struct mem_cgroup *memcg;
2569 bool in_retry = false;
2571 if (likely(!nr_pages))
2574 memcg = get_mem_cgroup_from_mm(current->mm);
2575 current->memcg_nr_pages_over_high = 0;
2579 * The allocating task should reclaim at least the batch size, but for
2580 * subsequent retries we only want to do what's necessary to prevent oom
2581 * or breaching resource isolation.
2583 * This is distinct from memory.max or page allocator behaviour because
2584 * memory.high is currently batched, whereas memory.max and the page
2585 * allocator run every time an allocation is made.
2587 nr_reclaimed = reclaim_high(memcg,
2588 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2592 * memory.high is breached and reclaim is unable to keep up. Throttle
2593 * allocators proactively to slow down excessive growth.
2595 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2596 mem_find_max_overage(memcg));
2598 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2599 swap_find_max_overage(memcg));
2602 * Clamp the max delay per usermode return so as to still keep the
2603 * application moving forwards and also permit diagnostics, albeit
2606 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2609 * Don't sleep if the amount of jiffies this memcg owes us is so low
2610 * that it's not even worth doing, in an attempt to be nice to those who
2611 * go only a small amount over their memory.high value and maybe haven't
2612 * been aggressively reclaimed enough yet.
2614 if (penalty_jiffies <= HZ / 100)
2618 * If reclaim is making forward progress but we're still over
2619 * memory.high, we want to encourage that rather than doing allocator
2622 if (nr_reclaimed || nr_retries--) {
2628 * If we exit early, we're guaranteed to die (since
2629 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2630 * need to account for any ill-begotten jiffies to pay them off later.
2632 psi_memstall_enter(&pflags);
2633 schedule_timeout_killable(penalty_jiffies);
2634 psi_memstall_leave(&pflags);
2637 css_put(&memcg->css);
2640 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2641 unsigned int nr_pages)
2643 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2644 int nr_retries = MAX_RECLAIM_RETRIES;
2645 struct mem_cgroup *mem_over_limit;
2646 struct page_counter *counter;
2647 enum oom_status oom_status;
2648 unsigned long nr_reclaimed;
2649 bool may_swap = true;
2650 bool drained = false;
2651 unsigned long pflags;
2653 if (mem_cgroup_is_root(memcg))
2656 if (consume_stock(memcg, nr_pages))
2659 if (!do_memsw_account() ||
2660 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2661 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2663 if (do_memsw_account())
2664 page_counter_uncharge(&memcg->memsw, batch);
2665 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2667 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2671 if (batch > nr_pages) {
2677 * Memcg doesn't have a dedicated reserve for atomic
2678 * allocations. But like the global atomic pool, we need to
2679 * put the burden of reclaim on regular allocation requests
2680 * and let these go through as privileged allocations.
2682 if (gfp_mask & __GFP_ATOMIC)
2686 * Unlike in global OOM situations, memcg is not in a physical
2687 * memory shortage. Allow dying and OOM-killed tasks to
2688 * bypass the last charges so that they can exit quickly and
2689 * free their memory.
2691 if (unlikely(should_force_charge()))
2695 * Prevent unbounded recursion when reclaim operations need to
2696 * allocate memory. This might exceed the limits temporarily,
2697 * but we prefer facilitating memory reclaim and getting back
2698 * under the limit over triggering OOM kills in these cases.
2700 if (unlikely(current->flags & PF_MEMALLOC))
2703 if (unlikely(task_in_memcg_oom(current)))
2706 if (!gfpflags_allow_blocking(gfp_mask))
2709 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2711 psi_memstall_enter(&pflags);
2712 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2713 gfp_mask, may_swap);
2714 psi_memstall_leave(&pflags);
2716 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2720 drain_all_stock(mem_over_limit);
2725 if (gfp_mask & __GFP_NORETRY)
2728 * Even though the limit is exceeded at this point, reclaim
2729 * may have been able to free some pages. Retry the charge
2730 * before killing the task.
2732 * Only for regular pages, though: huge pages are rather
2733 * unlikely to succeed so close to the limit, and we fall back
2734 * to regular pages anyway in case of failure.
2736 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2739 * At task move, charge accounts can be doubly counted. So, it's
2740 * better to wait until the end of task_move if something is going on.
2742 if (mem_cgroup_wait_acct_move(mem_over_limit))
2748 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2751 if (gfp_mask & __GFP_NOFAIL)
2754 if (fatal_signal_pending(current))
2758 * keep retrying as long as the memcg oom killer is able to make
2759 * a forward progress or bypass the charge if the oom killer
2760 * couldn't make any progress.
2762 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2763 get_order(nr_pages * PAGE_SIZE));
2764 switch (oom_status) {
2766 nr_retries = MAX_RECLAIM_RETRIES;
2774 if (!(gfp_mask & __GFP_NOFAIL))
2778 * The allocation either can't fail or will lead to more memory
2779 * being freed very soon. Allow memory usage go over the limit
2780 * temporarily by force charging it.
2782 page_counter_charge(&memcg->memory, nr_pages);
2783 if (do_memsw_account())
2784 page_counter_charge(&memcg->memsw, nr_pages);
2789 if (batch > nr_pages)
2790 refill_stock(memcg, batch - nr_pages);
2793 * If the hierarchy is above the normal consumption range, schedule
2794 * reclaim on returning to userland. We can perform reclaim here
2795 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2796 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2797 * not recorded as it most likely matches current's and won't
2798 * change in the meantime. As high limit is checked again before
2799 * reclaim, the cost of mismatch is negligible.
2802 bool mem_high, swap_high;
2804 mem_high = page_counter_read(&memcg->memory) >
2805 READ_ONCE(memcg->memory.high);
2806 swap_high = page_counter_read(&memcg->swap) >
2807 READ_ONCE(memcg->swap.high);
2809 /* Don't bother a random interrupted task */
2810 if (in_interrupt()) {
2812 schedule_work(&memcg->high_work);
2818 if (mem_high || swap_high) {
2820 * The allocating tasks in this cgroup will need to do
2821 * reclaim or be throttled to prevent further growth
2822 * of the memory or swap footprints.
2824 * Target some best-effort fairness between the tasks,
2825 * and distribute reclaim work and delay penalties
2826 * based on how much each task is actually allocating.
2828 current->memcg_nr_pages_over_high += batch;
2829 set_notify_resume(current);
2832 } while ((memcg = parent_mem_cgroup(memcg)));
2837 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2838 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2840 if (mem_cgroup_is_root(memcg))
2843 page_counter_uncharge(&memcg->memory, nr_pages);
2844 if (do_memsw_account())
2845 page_counter_uncharge(&memcg->memsw, nr_pages);
2849 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2851 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2853 * Any of the following ensures page->mem_cgroup stability:
2857 * - lock_page_memcg()
2858 * - exclusive reference
2860 page->mem_cgroup = memcg;
2863 #ifdef CONFIG_MEMCG_KMEM
2864 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2867 unsigned int objects = objs_per_slab_page(s, page);
2870 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2875 if (cmpxchg(&page->obj_cgroups, NULL,
2876 (struct obj_cgroup **) ((unsigned long)vec | 0x1UL)))
2879 kmemleak_not_leak(vec);
2885 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2887 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2888 * cgroup_mutex, etc.
2890 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2894 if (mem_cgroup_disabled())
2897 page = virt_to_head_page(p);
2900 * If page->mem_cgroup is set, it's either a simple mem_cgroup pointer
2901 * or a pointer to obj_cgroup vector. In the latter case the lowest
2902 * bit of the pointer is set.
2903 * The page->mem_cgroup pointer can be asynchronously changed
2904 * from NULL to (obj_cgroup_vec | 0x1UL), but can't be changed
2905 * from a valid memcg pointer to objcg vector or back.
2907 if (!page->mem_cgroup)
2911 * Slab objects are accounted individually, not per-page.
2912 * Memcg membership data for each individual object is saved in
2913 * the page->obj_cgroups.
2915 if (page_has_obj_cgroups(page)) {
2916 struct obj_cgroup *objcg;
2919 off = obj_to_index(page->slab_cache, page, p);
2920 objcg = page_obj_cgroups(page)[off];
2922 return obj_cgroup_memcg(objcg);
2927 /* All other pages use page->mem_cgroup */
2928 return page->mem_cgroup;
2931 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2933 struct obj_cgroup *objcg = NULL;
2934 struct mem_cgroup *memcg;
2936 if (unlikely(!current->mm && !current->active_memcg))
2940 if (unlikely(current->active_memcg))
2941 memcg = rcu_dereference(current->active_memcg);
2943 memcg = mem_cgroup_from_task(current);
2945 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2946 objcg = rcu_dereference(memcg->objcg);
2947 if (objcg && obj_cgroup_tryget(objcg))
2955 static int memcg_alloc_cache_id(void)
2960 id = ida_simple_get(&memcg_cache_ida,
2961 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2965 if (id < memcg_nr_cache_ids)
2969 * There's no space for the new id in memcg_caches arrays,
2970 * so we have to grow them.
2972 down_write(&memcg_cache_ids_sem);
2974 size = 2 * (id + 1);
2975 if (size < MEMCG_CACHES_MIN_SIZE)
2976 size = MEMCG_CACHES_MIN_SIZE;
2977 else if (size > MEMCG_CACHES_MAX_SIZE)
2978 size = MEMCG_CACHES_MAX_SIZE;
2980 err = memcg_update_all_list_lrus(size);
2982 memcg_nr_cache_ids = size;
2984 up_write(&memcg_cache_ids_sem);
2987 ida_simple_remove(&memcg_cache_ida, id);
2993 static void memcg_free_cache_id(int id)
2995 ida_simple_remove(&memcg_cache_ida, id);
2999 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3000 * @memcg: memory cgroup to charge
3001 * @gfp: reclaim mode
3002 * @nr_pages: number of pages to charge
3004 * Returns 0 on success, an error code on failure.
3006 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
3007 unsigned int nr_pages)
3009 struct page_counter *counter;
3012 ret = try_charge(memcg, gfp, nr_pages);
3016 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3017 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3020 * Enforce __GFP_NOFAIL allocation because callers are not
3021 * prepared to see failures and likely do not have any failure
3024 if (gfp & __GFP_NOFAIL) {
3025 page_counter_charge(&memcg->kmem, nr_pages);
3028 cancel_charge(memcg, nr_pages);
3035 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3036 * @memcg: memcg to uncharge
3037 * @nr_pages: number of pages to uncharge
3039 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3041 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3042 page_counter_uncharge(&memcg->kmem, nr_pages);
3044 page_counter_uncharge(&memcg->memory, nr_pages);
3045 if (do_memsw_account())
3046 page_counter_uncharge(&memcg->memsw, nr_pages);
3050 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3051 * @page: page to charge
3052 * @gfp: reclaim mode
3053 * @order: allocation order
3055 * Returns 0 on success, an error code on failure.
3057 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3059 struct mem_cgroup *memcg;
3062 if (memcg_kmem_bypass())
3065 memcg = get_mem_cgroup_from_current();
3066 if (!mem_cgroup_is_root(memcg)) {
3067 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3069 page->mem_cgroup = memcg;
3070 __SetPageKmemcg(page);
3074 css_put(&memcg->css);
3079 * __memcg_kmem_uncharge_page: uncharge a kmem page
3080 * @page: page to uncharge
3081 * @order: allocation order
3083 void __memcg_kmem_uncharge_page(struct page *page, int order)
3085 struct mem_cgroup *memcg = page->mem_cgroup;
3086 unsigned int nr_pages = 1 << order;
3091 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3092 __memcg_kmem_uncharge(memcg, nr_pages);
3093 page->mem_cgroup = NULL;
3094 css_put(&memcg->css);
3096 /* slab pages do not have PageKmemcg flag set */
3097 if (PageKmemcg(page))
3098 __ClearPageKmemcg(page);
3101 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3103 struct memcg_stock_pcp *stock;
3104 unsigned long flags;
3107 local_irq_save(flags);
3109 stock = this_cpu_ptr(&memcg_stock);
3110 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3111 stock->nr_bytes -= nr_bytes;
3115 local_irq_restore(flags);
3120 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3122 struct obj_cgroup *old = stock->cached_objcg;
3127 if (stock->nr_bytes) {
3128 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3129 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3133 __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
3138 * The leftover is flushed to the centralized per-memcg value.
3139 * On the next attempt to refill obj stock it will be moved
3140 * to a per-cpu stock (probably, on an other CPU), see
3141 * refill_obj_stock().
3143 * How often it's flushed is a trade-off between the memory
3144 * limit enforcement accuracy and potential CPU contention,
3145 * so it might be changed in the future.
3147 atomic_add(nr_bytes, &old->nr_charged_bytes);
3148 stock->nr_bytes = 0;
3151 obj_cgroup_put(old);
3152 stock->cached_objcg = NULL;
3155 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3156 struct mem_cgroup *root_memcg)
3158 struct mem_cgroup *memcg;
3160 if (stock->cached_objcg) {
3161 memcg = obj_cgroup_memcg(stock->cached_objcg);
3162 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3169 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3171 struct memcg_stock_pcp *stock;
3172 unsigned long flags;
3174 local_irq_save(flags);
3176 stock = this_cpu_ptr(&memcg_stock);
3177 if (stock->cached_objcg != objcg) { /* reset if necessary */
3178 drain_obj_stock(stock);
3179 obj_cgroup_get(objcg);
3180 stock->cached_objcg = objcg;
3181 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3183 stock->nr_bytes += nr_bytes;
3185 if (stock->nr_bytes > PAGE_SIZE)
3186 drain_obj_stock(stock);
3188 local_irq_restore(flags);
3191 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3193 struct mem_cgroup *memcg;
3194 unsigned int nr_pages, nr_bytes;
3197 if (consume_obj_stock(objcg, size))
3201 * In theory, memcg->nr_charged_bytes can have enough
3202 * pre-charged bytes to satisfy the allocation. However,
3203 * flushing memcg->nr_charged_bytes requires two atomic
3204 * operations, and memcg->nr_charged_bytes can't be big,
3205 * so it's better to ignore it and try grab some new pages.
3206 * memcg->nr_charged_bytes will be flushed in
3207 * refill_obj_stock(), called from this function or
3208 * independently later.
3211 memcg = obj_cgroup_memcg(objcg);
3212 css_get(&memcg->css);
3215 nr_pages = size >> PAGE_SHIFT;
3216 nr_bytes = size & (PAGE_SIZE - 1);
3221 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3222 if (!ret && nr_bytes)
3223 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3225 css_put(&memcg->css);
3229 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3231 refill_obj_stock(objcg, size);
3234 #endif /* CONFIG_MEMCG_KMEM */
3236 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3239 * Because tail pages are not marked as "used", set it. We're under
3240 * pgdat->lru_lock and migration entries setup in all page mappings.
3242 void mem_cgroup_split_huge_fixup(struct page *head)
3244 struct mem_cgroup *memcg = head->mem_cgroup;
3247 if (mem_cgroup_disabled())
3250 for (i = 1; i < HPAGE_PMD_NR; i++) {
3251 css_get(&memcg->css);
3252 head[i].mem_cgroup = memcg;
3255 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3257 #ifdef CONFIG_MEMCG_SWAP
3259 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3260 * @entry: swap entry to be moved
3261 * @from: mem_cgroup which the entry is moved from
3262 * @to: mem_cgroup which the entry is moved to
3264 * It succeeds only when the swap_cgroup's record for this entry is the same
3265 * as the mem_cgroup's id of @from.
3267 * Returns 0 on success, -EINVAL on failure.
3269 * The caller must have charged to @to, IOW, called page_counter_charge() about
3270 * both res and memsw, and called css_get().
3272 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3273 struct mem_cgroup *from, struct mem_cgroup *to)
3275 unsigned short old_id, new_id;
3277 old_id = mem_cgroup_id(from);
3278 new_id = mem_cgroup_id(to);
3280 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3281 mod_memcg_state(from, MEMCG_SWAP, -1);
3282 mod_memcg_state(to, MEMCG_SWAP, 1);
3288 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3289 struct mem_cgroup *from, struct mem_cgroup *to)
3295 static DEFINE_MUTEX(memcg_max_mutex);
3297 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3298 unsigned long max, bool memsw)
3300 bool enlarge = false;
3301 bool drained = false;
3303 bool limits_invariant;
3304 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3307 if (signal_pending(current)) {
3312 mutex_lock(&memcg_max_mutex);
3314 * Make sure that the new limit (memsw or memory limit) doesn't
3315 * break our basic invariant rule memory.max <= memsw.max.
3317 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3318 max <= memcg->memsw.max;
3319 if (!limits_invariant) {
3320 mutex_unlock(&memcg_max_mutex);
3324 if (max > counter->max)
3326 ret = page_counter_set_max(counter, max);
3327 mutex_unlock(&memcg_max_mutex);
3333 drain_all_stock(memcg);
3338 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3339 GFP_KERNEL, !memsw)) {
3345 if (!ret && enlarge)
3346 memcg_oom_recover(memcg);
3351 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3353 unsigned long *total_scanned)
3355 unsigned long nr_reclaimed = 0;
3356 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3357 unsigned long reclaimed;
3359 struct mem_cgroup_tree_per_node *mctz;
3360 unsigned long excess;
3361 unsigned long nr_scanned;
3366 mctz = soft_limit_tree_node(pgdat->node_id);
3369 * Do not even bother to check the largest node if the root
3370 * is empty. Do it lockless to prevent lock bouncing. Races
3371 * are acceptable as soft limit is best effort anyway.
3373 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3377 * This loop can run a while, specially if mem_cgroup's continuously
3378 * keep exceeding their soft limit and putting the system under
3385 mz = mem_cgroup_largest_soft_limit_node(mctz);
3390 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3391 gfp_mask, &nr_scanned);
3392 nr_reclaimed += reclaimed;
3393 *total_scanned += nr_scanned;
3394 spin_lock_irq(&mctz->lock);
3395 __mem_cgroup_remove_exceeded(mz, mctz);
3398 * If we failed to reclaim anything from this memory cgroup
3399 * it is time to move on to the next cgroup
3403 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3405 excess = soft_limit_excess(mz->memcg);
3407 * One school of thought says that we should not add
3408 * back the node to the tree if reclaim returns 0.
3409 * But our reclaim could return 0, simply because due
3410 * to priority we are exposing a smaller subset of
3411 * memory to reclaim from. Consider this as a longer
3414 /* If excess == 0, no tree ops */
3415 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3416 spin_unlock_irq(&mctz->lock);
3417 css_put(&mz->memcg->css);
3420 * Could not reclaim anything and there are no more
3421 * mem cgroups to try or we seem to be looping without
3422 * reclaiming anything.
3424 if (!nr_reclaimed &&
3426 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3428 } while (!nr_reclaimed);
3430 css_put(&next_mz->memcg->css);
3431 return nr_reclaimed;
3435 * Test whether @memcg has children, dead or alive. Note that this
3436 * function doesn't care whether @memcg has use_hierarchy enabled and
3437 * returns %true if there are child csses according to the cgroup
3438 * hierarchy. Testing use_hierarchy is the caller's responsibility.
3440 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3445 ret = css_next_child(NULL, &memcg->css);
3451 * Reclaims as many pages from the given memcg as possible.
3453 * Caller is responsible for holding css reference for memcg.
3455 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3457 int nr_retries = MAX_RECLAIM_RETRIES;
3459 /* we call try-to-free pages for make this cgroup empty */
3460 lru_add_drain_all();
3462 drain_all_stock(memcg);
3464 /* try to free all pages in this cgroup */
3465 while (nr_retries && page_counter_read(&memcg->memory)) {
3468 if (signal_pending(current))
3471 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3475 /* maybe some writeback is necessary */
3476 congestion_wait(BLK_RW_ASYNC, HZ/10);
3484 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3485 char *buf, size_t nbytes,
3488 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3490 if (mem_cgroup_is_root(memcg))
3492 return mem_cgroup_force_empty(memcg) ?: nbytes;
3495 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3498 return mem_cgroup_from_css(css)->use_hierarchy;
3501 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3502 struct cftype *cft, u64 val)
3505 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3506 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3508 if (memcg->use_hierarchy == val)
3512 * If parent's use_hierarchy is set, we can't make any modifications
3513 * in the child subtrees. If it is unset, then the change can
3514 * occur, provided the current cgroup has no children.
3516 * For the root cgroup, parent_mem is NULL, we allow value to be
3517 * set if there are no children.
3519 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3520 (val == 1 || val == 0)) {
3521 if (!memcg_has_children(memcg))
3522 memcg->use_hierarchy = val;
3531 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3535 if (mem_cgroup_is_root(memcg)) {
3536 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3537 memcg_page_state(memcg, NR_ANON_MAPPED);
3539 val += memcg_page_state(memcg, MEMCG_SWAP);
3542 val = page_counter_read(&memcg->memory);
3544 val = page_counter_read(&memcg->memsw);
3557 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3560 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3561 struct page_counter *counter;
3563 switch (MEMFILE_TYPE(cft->private)) {
3565 counter = &memcg->memory;
3568 counter = &memcg->memsw;
3571 counter = &memcg->kmem;
3574 counter = &memcg->tcpmem;
3580 switch (MEMFILE_ATTR(cft->private)) {
3582 if (counter == &memcg->memory)
3583 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3584 if (counter == &memcg->memsw)
3585 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3586 return (u64)page_counter_read(counter) * PAGE_SIZE;
3588 return (u64)counter->max * PAGE_SIZE;
3590 return (u64)counter->watermark * PAGE_SIZE;
3592 return counter->failcnt;
3593 case RES_SOFT_LIMIT:
3594 return (u64)memcg->soft_limit * PAGE_SIZE;
3600 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3602 unsigned long stat[MEMCG_NR_STAT] = {0};
3603 struct mem_cgroup *mi;
3606 for_each_online_cpu(cpu)
3607 for (i = 0; i < MEMCG_NR_STAT; i++)
3608 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3610 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3611 for (i = 0; i < MEMCG_NR_STAT; i++)
3612 atomic_long_add(stat[i], &mi->vmstats[i]);
3614 for_each_node(node) {
3615 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3616 struct mem_cgroup_per_node *pi;
3618 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3621 for_each_online_cpu(cpu)
3622 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3624 pn->lruvec_stat_cpu->count[i], cpu);
3626 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3627 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3628 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3632 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3634 unsigned long events[NR_VM_EVENT_ITEMS];
3635 struct mem_cgroup *mi;
3638 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3641 for_each_online_cpu(cpu)
3642 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3643 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3646 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3647 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3648 atomic_long_add(events[i], &mi->vmevents[i]);
3651 #ifdef CONFIG_MEMCG_KMEM
3652 static int memcg_online_kmem(struct mem_cgroup *memcg)
3654 struct obj_cgroup *objcg;
3657 if (cgroup_memory_nokmem)
3660 BUG_ON(memcg->kmemcg_id >= 0);
3661 BUG_ON(memcg->kmem_state);
3663 memcg_id = memcg_alloc_cache_id();
3667 objcg = obj_cgroup_alloc();
3669 memcg_free_cache_id(memcg_id);
3672 objcg->memcg = memcg;
3673 rcu_assign_pointer(memcg->objcg, objcg);
3675 static_branch_enable(&memcg_kmem_enabled_key);
3678 * A memory cgroup is considered kmem-online as soon as it gets
3679 * kmemcg_id. Setting the id after enabling static branching will
3680 * guarantee no one starts accounting before all call sites are
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;
3722 if (!memcg->use_hierarchy)
3727 memcg_drain_all_list_lrus(kmemcg_id, parent);
3729 memcg_free_cache_id(kmemcg_id);
3732 static void memcg_free_kmem(struct mem_cgroup *memcg)
3734 /* css_alloc() failed, offlining didn't happen */
3735 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3736 memcg_offline_kmem(memcg);
3739 static int memcg_online_kmem(struct mem_cgroup *memcg)
3743 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3746 static void memcg_free_kmem(struct mem_cgroup *memcg)
3749 #endif /* CONFIG_MEMCG_KMEM */
3751 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3756 mutex_lock(&memcg_max_mutex);
3757 ret = page_counter_set_max(&memcg->kmem, max);
3758 mutex_unlock(&memcg_max_mutex);
3762 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3766 mutex_lock(&memcg_max_mutex);
3768 ret = page_counter_set_max(&memcg->tcpmem, max);
3772 if (!memcg->tcpmem_active) {
3774 * The active flag needs to be written after the static_key
3775 * update. This is what guarantees that the socket activation
3776 * function is the last one to run. See mem_cgroup_sk_alloc()
3777 * for details, and note that we don't mark any socket as
3778 * belonging to this memcg until that flag is up.
3780 * We need to do this, because static_keys will span multiple
3781 * sites, but we can't control their order. If we mark a socket
3782 * as accounted, but the accounting functions are not patched in
3783 * yet, we'll lose accounting.
3785 * We never race with the readers in mem_cgroup_sk_alloc(),
3786 * because when this value change, the code to process it is not
3789 static_branch_inc(&memcg_sockets_enabled_key);
3790 memcg->tcpmem_active = true;
3793 mutex_unlock(&memcg_max_mutex);
3798 * The user of this function is...
3801 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3802 char *buf, size_t nbytes, loff_t off)
3804 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3805 unsigned long nr_pages;
3808 buf = strstrip(buf);
3809 ret = page_counter_memparse(buf, "-1", &nr_pages);
3813 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3815 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3819 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3821 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3824 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3827 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3828 "Please report your usecase to linux-mm@kvack.org if you "
3829 "depend on this functionality.\n");
3830 ret = memcg_update_kmem_max(memcg, nr_pages);
3833 ret = memcg_update_tcp_max(memcg, nr_pages);
3837 case RES_SOFT_LIMIT:
3838 memcg->soft_limit = nr_pages;
3842 return ret ?: nbytes;
3845 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3846 size_t nbytes, loff_t off)
3848 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3849 struct page_counter *counter;
3851 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3853 counter = &memcg->memory;
3856 counter = &memcg->memsw;
3859 counter = &memcg->kmem;
3862 counter = &memcg->tcpmem;
3868 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3870 page_counter_reset_watermark(counter);
3873 counter->failcnt = 0;
3882 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3885 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3889 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3890 struct cftype *cft, u64 val)
3892 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3894 if (val & ~MOVE_MASK)
3898 * No kind of locking is needed in here, because ->can_attach() will
3899 * check this value once in the beginning of the process, and then carry
3900 * on with stale data. This means that changes to this value will only
3901 * affect task migrations starting after the change.
3903 memcg->move_charge_at_immigrate = val;
3907 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3908 struct cftype *cft, u64 val)
3916 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3917 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3918 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3920 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3921 int nid, unsigned int lru_mask, bool tree)
3923 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3924 unsigned long nr = 0;
3927 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3930 if (!(BIT(lru) & lru_mask))
3933 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3935 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3940 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3941 unsigned int lru_mask,
3944 unsigned long nr = 0;
3948 if (!(BIT(lru) & lru_mask))
3951 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3953 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3958 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3962 unsigned int lru_mask;
3965 static const struct numa_stat stats[] = {
3966 { "total", LRU_ALL },
3967 { "file", LRU_ALL_FILE },
3968 { "anon", LRU_ALL_ANON },
3969 { "unevictable", BIT(LRU_UNEVICTABLE) },
3971 const struct numa_stat *stat;
3973 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3975 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3976 seq_printf(m, "%s=%lu", stat->name,
3977 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3979 for_each_node_state(nid, N_MEMORY)
3980 seq_printf(m, " N%d=%lu", nid,
3981 mem_cgroup_node_nr_lru_pages(memcg, nid,
3982 stat->lru_mask, false));
3986 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3988 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3989 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3991 for_each_node_state(nid, N_MEMORY)
3992 seq_printf(m, " N%d=%lu", nid,
3993 mem_cgroup_node_nr_lru_pages(memcg, nid,
3994 stat->lru_mask, true));
4000 #endif /* CONFIG_NUMA */
4002 static const unsigned int memcg1_stats[] = {
4005 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4015 static const char *const memcg1_stat_names[] = {
4018 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4028 /* Universal VM events cgroup1 shows, original sort order */
4029 static const unsigned int memcg1_events[] = {
4036 static int memcg_stat_show(struct seq_file *m, void *v)
4038 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4039 unsigned long memory, memsw;
4040 struct mem_cgroup *mi;
4043 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4045 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4048 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4050 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4051 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4052 if (memcg1_stats[i] == NR_ANON_THPS)
4055 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4058 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4059 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4060 memcg_events_local(memcg, memcg1_events[i]));
4062 for (i = 0; i < NR_LRU_LISTS; i++)
4063 seq_printf(m, "%s %lu\n", lru_list_name(i),
4064 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4067 /* Hierarchical information */
4068 memory = memsw = PAGE_COUNTER_MAX;
4069 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4070 memory = min(memory, READ_ONCE(mi->memory.max));
4071 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4073 seq_printf(m, "hierarchical_memory_limit %llu\n",
4074 (u64)memory * PAGE_SIZE);
4075 if (do_memsw_account())
4076 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4077 (u64)memsw * PAGE_SIZE);
4079 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4080 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4082 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4083 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
4087 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4088 seq_printf(m, "total_%s %llu\n",
4089 vm_event_name(memcg1_events[i]),
4090 (u64)memcg_events(memcg, memcg1_events[i]));
4092 for (i = 0; i < NR_LRU_LISTS; i++)
4093 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4094 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4097 #ifdef CONFIG_DEBUG_VM
4100 struct mem_cgroup_per_node *mz;
4101 unsigned long anon_cost = 0;
4102 unsigned long file_cost = 0;
4104 for_each_online_pgdat(pgdat) {
4105 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4107 anon_cost += mz->lruvec.anon_cost;
4108 file_cost += mz->lruvec.file_cost;
4110 seq_printf(m, "anon_cost %lu\n", anon_cost);
4111 seq_printf(m, "file_cost %lu\n", file_cost);
4118 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4121 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4123 return mem_cgroup_swappiness(memcg);
4126 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4127 struct cftype *cft, u64 val)
4129 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4135 memcg->swappiness = val;
4137 vm_swappiness = val;
4142 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4144 struct mem_cgroup_threshold_ary *t;
4145 unsigned long usage;
4150 t = rcu_dereference(memcg->thresholds.primary);
4152 t = rcu_dereference(memcg->memsw_thresholds.primary);
4157 usage = mem_cgroup_usage(memcg, swap);
4160 * current_threshold points to threshold just below or equal to usage.
4161 * If it's not true, a threshold was crossed after last
4162 * call of __mem_cgroup_threshold().
4164 i = t->current_threshold;
4167 * Iterate backward over array of thresholds starting from
4168 * current_threshold and check if a threshold is crossed.
4169 * If none of thresholds below usage is crossed, we read
4170 * only one element of the array here.
4172 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4173 eventfd_signal(t->entries[i].eventfd, 1);
4175 /* i = current_threshold + 1 */
4179 * Iterate forward over array of thresholds starting from
4180 * current_threshold+1 and check if a threshold is crossed.
4181 * If none of thresholds above usage is crossed, we read
4182 * only one element of the array here.
4184 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4185 eventfd_signal(t->entries[i].eventfd, 1);
4187 /* Update current_threshold */
4188 t->current_threshold = i - 1;
4193 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4196 __mem_cgroup_threshold(memcg, false);
4197 if (do_memsw_account())
4198 __mem_cgroup_threshold(memcg, true);
4200 memcg = parent_mem_cgroup(memcg);
4204 static int compare_thresholds(const void *a, const void *b)
4206 const struct mem_cgroup_threshold *_a = a;
4207 const struct mem_cgroup_threshold *_b = b;
4209 if (_a->threshold > _b->threshold)
4212 if (_a->threshold < _b->threshold)
4218 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4220 struct mem_cgroup_eventfd_list *ev;
4222 spin_lock(&memcg_oom_lock);
4224 list_for_each_entry(ev, &memcg->oom_notify, list)
4225 eventfd_signal(ev->eventfd, 1);
4227 spin_unlock(&memcg_oom_lock);
4231 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4233 struct mem_cgroup *iter;
4235 for_each_mem_cgroup_tree(iter, memcg)
4236 mem_cgroup_oom_notify_cb(iter);
4239 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4240 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4242 struct mem_cgroup_thresholds *thresholds;
4243 struct mem_cgroup_threshold_ary *new;
4244 unsigned long threshold;
4245 unsigned long usage;
4248 ret = page_counter_memparse(args, "-1", &threshold);
4252 mutex_lock(&memcg->thresholds_lock);
4255 thresholds = &memcg->thresholds;
4256 usage = mem_cgroup_usage(memcg, false);
4257 } else if (type == _MEMSWAP) {
4258 thresholds = &memcg->memsw_thresholds;
4259 usage = mem_cgroup_usage(memcg, true);
4263 /* Check if a threshold crossed before adding a new one */
4264 if (thresholds->primary)
4265 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4267 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4269 /* Allocate memory for new array of thresholds */
4270 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4277 /* Copy thresholds (if any) to new array */
4278 if (thresholds->primary)
4279 memcpy(new->entries, thresholds->primary->entries,
4280 flex_array_size(new, entries, size - 1));
4282 /* Add new threshold */
4283 new->entries[size - 1].eventfd = eventfd;
4284 new->entries[size - 1].threshold = threshold;
4286 /* Sort thresholds. Registering of new threshold isn't time-critical */
4287 sort(new->entries, size, sizeof(*new->entries),
4288 compare_thresholds, NULL);
4290 /* Find current threshold */
4291 new->current_threshold = -1;
4292 for (i = 0; i < size; i++) {
4293 if (new->entries[i].threshold <= usage) {
4295 * new->current_threshold will not be used until
4296 * rcu_assign_pointer(), so it's safe to increment
4299 ++new->current_threshold;
4304 /* Free old spare buffer and save old primary buffer as spare */
4305 kfree(thresholds->spare);
4306 thresholds->spare = thresholds->primary;
4308 rcu_assign_pointer(thresholds->primary, new);
4310 /* To be sure that nobody uses thresholds */
4314 mutex_unlock(&memcg->thresholds_lock);
4319 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4320 struct eventfd_ctx *eventfd, const char *args)
4322 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4325 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4326 struct eventfd_ctx *eventfd, const char *args)
4328 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4331 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4332 struct eventfd_ctx *eventfd, enum res_type type)
4334 struct mem_cgroup_thresholds *thresholds;
4335 struct mem_cgroup_threshold_ary *new;
4336 unsigned long usage;
4337 int i, j, size, entries;
4339 mutex_lock(&memcg->thresholds_lock);
4342 thresholds = &memcg->thresholds;
4343 usage = mem_cgroup_usage(memcg, false);
4344 } else if (type == _MEMSWAP) {
4345 thresholds = &memcg->memsw_thresholds;
4346 usage = mem_cgroup_usage(memcg, true);
4350 if (!thresholds->primary)
4353 /* Check if a threshold crossed before removing */
4354 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4356 /* Calculate new number of threshold */
4358 for (i = 0; i < thresholds->primary->size; i++) {
4359 if (thresholds->primary->entries[i].eventfd != eventfd)
4365 new = thresholds->spare;
4367 /* If no items related to eventfd have been cleared, nothing to do */
4371 /* Set thresholds array to NULL if we don't have thresholds */
4380 /* Copy thresholds and find current threshold */
4381 new->current_threshold = -1;
4382 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4383 if (thresholds->primary->entries[i].eventfd == eventfd)
4386 new->entries[j] = thresholds->primary->entries[i];
4387 if (new->entries[j].threshold <= usage) {
4389 * new->current_threshold will not be used
4390 * until rcu_assign_pointer(), so it's safe to increment
4393 ++new->current_threshold;
4399 /* Swap primary and spare array */
4400 thresholds->spare = thresholds->primary;
4402 rcu_assign_pointer(thresholds->primary, new);
4404 /* To be sure that nobody uses thresholds */
4407 /* If all events are unregistered, free the spare array */
4409 kfree(thresholds->spare);
4410 thresholds->spare = NULL;
4413 mutex_unlock(&memcg->thresholds_lock);
4416 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4417 struct eventfd_ctx *eventfd)
4419 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4422 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4423 struct eventfd_ctx *eventfd)
4425 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4428 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4429 struct eventfd_ctx *eventfd, const char *args)
4431 struct mem_cgroup_eventfd_list *event;
4433 event = kmalloc(sizeof(*event), GFP_KERNEL);
4437 spin_lock(&memcg_oom_lock);
4439 event->eventfd = eventfd;
4440 list_add(&event->list, &memcg->oom_notify);
4442 /* already in OOM ? */
4443 if (memcg->under_oom)
4444 eventfd_signal(eventfd, 1);
4445 spin_unlock(&memcg_oom_lock);
4450 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4451 struct eventfd_ctx *eventfd)
4453 struct mem_cgroup_eventfd_list *ev, *tmp;
4455 spin_lock(&memcg_oom_lock);
4457 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4458 if (ev->eventfd == eventfd) {
4459 list_del(&ev->list);
4464 spin_unlock(&memcg_oom_lock);
4467 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4469 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4471 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4472 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4473 seq_printf(sf, "oom_kill %lu\n",
4474 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4478 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4479 struct cftype *cft, u64 val)
4481 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4483 /* cannot set to root cgroup and only 0 and 1 are allowed */
4484 if (!css->parent || !((val == 0) || (val == 1)))
4487 memcg->oom_kill_disable = val;
4489 memcg_oom_recover(memcg);
4494 #ifdef CONFIG_CGROUP_WRITEBACK
4496 #include <trace/events/writeback.h>
4498 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4500 return wb_domain_init(&memcg->cgwb_domain, gfp);
4503 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4505 wb_domain_exit(&memcg->cgwb_domain);
4508 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4510 wb_domain_size_changed(&memcg->cgwb_domain);
4513 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4515 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4517 if (!memcg->css.parent)
4520 return &memcg->cgwb_domain;
4524 * idx can be of type enum memcg_stat_item or node_stat_item.
4525 * Keep in sync with memcg_exact_page().
4527 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4529 long x = atomic_long_read(&memcg->vmstats[idx]);
4532 for_each_online_cpu(cpu)
4533 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4540 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4541 * @wb: bdi_writeback in question
4542 * @pfilepages: out parameter for number of file pages
4543 * @pheadroom: out parameter for number of allocatable pages according to memcg
4544 * @pdirty: out parameter for number of dirty pages
4545 * @pwriteback: out parameter for number of pages under writeback
4547 * Determine the numbers of file, headroom, dirty, and writeback pages in
4548 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4549 * is a bit more involved.
4551 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4552 * headroom is calculated as the lowest headroom of itself and the
4553 * ancestors. Note that this doesn't consider the actual amount of
4554 * available memory in the system. The caller should further cap
4555 * *@pheadroom accordingly.
4557 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4558 unsigned long *pheadroom, unsigned long *pdirty,
4559 unsigned long *pwriteback)
4561 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4562 struct mem_cgroup *parent;
4564 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4566 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4567 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4568 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4569 *pheadroom = PAGE_COUNTER_MAX;
4571 while ((parent = parent_mem_cgroup(memcg))) {
4572 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4573 READ_ONCE(memcg->memory.high));
4574 unsigned long used = page_counter_read(&memcg->memory);
4576 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4582 * Foreign dirty flushing
4584 * There's an inherent mismatch between memcg and writeback. The former
4585 * trackes ownership per-page while the latter per-inode. This was a
4586 * deliberate design decision because honoring per-page ownership in the
4587 * writeback path is complicated, may lead to higher CPU and IO overheads
4588 * and deemed unnecessary given that write-sharing an inode across
4589 * different cgroups isn't a common use-case.
4591 * Combined with inode majority-writer ownership switching, this works well
4592 * enough in most cases but there are some pathological cases. For
4593 * example, let's say there are two cgroups A and B which keep writing to
4594 * different but confined parts of the same inode. B owns the inode and
4595 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4596 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4597 * triggering background writeback. A will be slowed down without a way to
4598 * make writeback of the dirty pages happen.
4600 * Conditions like the above can lead to a cgroup getting repatedly and
4601 * severely throttled after making some progress after each
4602 * dirty_expire_interval while the underyling IO device is almost
4605 * Solving this problem completely requires matching the ownership tracking
4606 * granularities between memcg and writeback in either direction. However,
4607 * the more egregious behaviors can be avoided by simply remembering the
4608 * most recent foreign dirtying events and initiating remote flushes on
4609 * them when local writeback isn't enough to keep the memory clean enough.
4611 * The following two functions implement such mechanism. When a foreign
4612 * page - a page whose memcg and writeback ownerships don't match - is
4613 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4614 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4615 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4616 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4617 * foreign bdi_writebacks which haven't expired. Both the numbers of
4618 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4619 * limited to MEMCG_CGWB_FRN_CNT.
4621 * The mechanism only remembers IDs and doesn't hold any object references.
4622 * As being wrong occasionally doesn't matter, updates and accesses to the
4623 * records are lockless and racy.
4625 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4626 struct bdi_writeback *wb)
4628 struct mem_cgroup *memcg = page->mem_cgroup;
4629 struct memcg_cgwb_frn *frn;
4630 u64 now = get_jiffies_64();
4631 u64 oldest_at = now;
4635 trace_track_foreign_dirty(page, wb);
4638 * Pick the slot to use. If there is already a slot for @wb, keep
4639 * using it. If not replace the oldest one which isn't being
4642 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4643 frn = &memcg->cgwb_frn[i];
4644 if (frn->bdi_id == wb->bdi->id &&
4645 frn->memcg_id == wb->memcg_css->id)
4647 if (time_before64(frn->at, oldest_at) &&
4648 atomic_read(&frn->done.cnt) == 1) {
4650 oldest_at = frn->at;
4654 if (i < MEMCG_CGWB_FRN_CNT) {
4656 * Re-using an existing one. Update timestamp lazily to
4657 * avoid making the cacheline hot. We want them to be
4658 * reasonably up-to-date and significantly shorter than
4659 * dirty_expire_interval as that's what expires the record.
4660 * Use the shorter of 1s and dirty_expire_interval / 8.
4662 unsigned long update_intv =
4663 min_t(unsigned long, HZ,
4664 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4666 if (time_before64(frn->at, now - update_intv))
4668 } else if (oldest >= 0) {
4669 /* replace the oldest free one */
4670 frn = &memcg->cgwb_frn[oldest];
4671 frn->bdi_id = wb->bdi->id;
4672 frn->memcg_id = wb->memcg_css->id;
4677 /* issue foreign writeback flushes for recorded foreign dirtying events */
4678 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4680 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4681 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4682 u64 now = jiffies_64;
4685 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4686 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4689 * If the record is older than dirty_expire_interval,
4690 * writeback on it has already started. No need to kick it
4691 * off again. Also, don't start a new one if there's
4692 * already one in flight.
4694 if (time_after64(frn->at, now - intv) &&
4695 atomic_read(&frn->done.cnt) == 1) {
4697 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4698 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4699 WB_REASON_FOREIGN_FLUSH,
4705 #else /* CONFIG_CGROUP_WRITEBACK */
4707 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4712 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4716 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4720 #endif /* CONFIG_CGROUP_WRITEBACK */
4723 * DO NOT USE IN NEW FILES.
4725 * "cgroup.event_control" implementation.
4727 * This is way over-engineered. It tries to support fully configurable
4728 * events for each user. Such level of flexibility is completely
4729 * unnecessary especially in the light of the planned unified hierarchy.
4731 * Please deprecate this and replace with something simpler if at all
4736 * Unregister event and free resources.
4738 * Gets called from workqueue.
4740 static void memcg_event_remove(struct work_struct *work)
4742 struct mem_cgroup_event *event =
4743 container_of(work, struct mem_cgroup_event, remove);
4744 struct mem_cgroup *memcg = event->memcg;
4746 remove_wait_queue(event->wqh, &event->wait);
4748 event->unregister_event(memcg, event->eventfd);
4750 /* Notify userspace the event is going away. */
4751 eventfd_signal(event->eventfd, 1);
4753 eventfd_ctx_put(event->eventfd);
4755 css_put(&memcg->css);
4759 * Gets called on EPOLLHUP on eventfd when user closes it.
4761 * Called with wqh->lock held and interrupts disabled.
4763 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4764 int sync, void *key)
4766 struct mem_cgroup_event *event =
4767 container_of(wait, struct mem_cgroup_event, wait);
4768 struct mem_cgroup *memcg = event->memcg;
4769 __poll_t flags = key_to_poll(key);
4771 if (flags & EPOLLHUP) {
4773 * If the event has been detached at cgroup removal, we
4774 * can simply return knowing the other side will cleanup
4777 * We can't race against event freeing since the other
4778 * side will require wqh->lock via remove_wait_queue(),
4781 spin_lock(&memcg->event_list_lock);
4782 if (!list_empty(&event->list)) {
4783 list_del_init(&event->list);
4785 * We are in atomic context, but cgroup_event_remove()
4786 * may sleep, so we have to call it in workqueue.
4788 schedule_work(&event->remove);
4790 spin_unlock(&memcg->event_list_lock);
4796 static void memcg_event_ptable_queue_proc(struct file *file,
4797 wait_queue_head_t *wqh, poll_table *pt)
4799 struct mem_cgroup_event *event =
4800 container_of(pt, struct mem_cgroup_event, pt);
4803 add_wait_queue(wqh, &event->wait);
4807 * DO NOT USE IN NEW FILES.
4809 * Parse input and register new cgroup event handler.
4811 * Input must be in format '<event_fd> <control_fd> <args>'.
4812 * Interpretation of args is defined by control file implementation.
4814 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4815 char *buf, size_t nbytes, loff_t off)
4817 struct cgroup_subsys_state *css = of_css(of);
4818 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4819 struct mem_cgroup_event *event;
4820 struct cgroup_subsys_state *cfile_css;
4821 unsigned int efd, cfd;
4828 buf = strstrip(buf);
4830 efd = simple_strtoul(buf, &endp, 10);
4835 cfd = simple_strtoul(buf, &endp, 10);
4836 if ((*endp != ' ') && (*endp != '\0'))
4840 event = kzalloc(sizeof(*event), GFP_KERNEL);
4844 event->memcg = memcg;
4845 INIT_LIST_HEAD(&event->list);
4846 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4847 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4848 INIT_WORK(&event->remove, memcg_event_remove);
4856 event->eventfd = eventfd_ctx_fileget(efile.file);
4857 if (IS_ERR(event->eventfd)) {
4858 ret = PTR_ERR(event->eventfd);
4865 goto out_put_eventfd;
4868 /* the process need read permission on control file */
4869 /* AV: shouldn't we check that it's been opened for read instead? */
4870 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4875 * Determine the event callbacks and set them in @event. This used
4876 * to be done via struct cftype but cgroup core no longer knows
4877 * about these events. The following is crude but the whole thing
4878 * is for compatibility anyway.
4880 * DO NOT ADD NEW FILES.
4882 name = cfile.file->f_path.dentry->d_name.name;
4884 if (!strcmp(name, "memory.usage_in_bytes")) {
4885 event->register_event = mem_cgroup_usage_register_event;
4886 event->unregister_event = mem_cgroup_usage_unregister_event;
4887 } else if (!strcmp(name, "memory.oom_control")) {
4888 event->register_event = mem_cgroup_oom_register_event;
4889 event->unregister_event = mem_cgroup_oom_unregister_event;
4890 } else if (!strcmp(name, "memory.pressure_level")) {
4891 event->register_event = vmpressure_register_event;
4892 event->unregister_event = vmpressure_unregister_event;
4893 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4894 event->register_event = memsw_cgroup_usage_register_event;
4895 event->unregister_event = memsw_cgroup_usage_unregister_event;
4902 * Verify @cfile should belong to @css. Also, remaining events are
4903 * automatically removed on cgroup destruction but the removal is
4904 * asynchronous, so take an extra ref on @css.
4906 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4907 &memory_cgrp_subsys);
4909 if (IS_ERR(cfile_css))
4911 if (cfile_css != css) {
4916 ret = event->register_event(memcg, event->eventfd, buf);
4920 vfs_poll(efile.file, &event->pt);
4922 spin_lock(&memcg->event_list_lock);
4923 list_add(&event->list, &memcg->event_list);
4924 spin_unlock(&memcg->event_list_lock);
4936 eventfd_ctx_put(event->eventfd);
4945 static struct cftype mem_cgroup_legacy_files[] = {
4947 .name = "usage_in_bytes",
4948 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4949 .read_u64 = mem_cgroup_read_u64,
4952 .name = "max_usage_in_bytes",
4953 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4954 .write = mem_cgroup_reset,
4955 .read_u64 = mem_cgroup_read_u64,
4958 .name = "limit_in_bytes",
4959 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4960 .write = mem_cgroup_write,
4961 .read_u64 = mem_cgroup_read_u64,
4964 .name = "soft_limit_in_bytes",
4965 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4966 .write = mem_cgroup_write,
4967 .read_u64 = mem_cgroup_read_u64,
4971 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4972 .write = mem_cgroup_reset,
4973 .read_u64 = mem_cgroup_read_u64,
4977 .seq_show = memcg_stat_show,
4980 .name = "force_empty",
4981 .write = mem_cgroup_force_empty_write,
4984 .name = "use_hierarchy",
4985 .write_u64 = mem_cgroup_hierarchy_write,
4986 .read_u64 = mem_cgroup_hierarchy_read,
4989 .name = "cgroup.event_control", /* XXX: for compat */
4990 .write = memcg_write_event_control,
4991 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4994 .name = "swappiness",
4995 .read_u64 = mem_cgroup_swappiness_read,
4996 .write_u64 = mem_cgroup_swappiness_write,
4999 .name = "move_charge_at_immigrate",
5000 .read_u64 = mem_cgroup_move_charge_read,
5001 .write_u64 = mem_cgroup_move_charge_write,
5004 .name = "oom_control",
5005 .seq_show = mem_cgroup_oom_control_read,
5006 .write_u64 = mem_cgroup_oom_control_write,
5007 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5010 .name = "pressure_level",
5014 .name = "numa_stat",
5015 .seq_show = memcg_numa_stat_show,
5019 .name = "kmem.limit_in_bytes",
5020 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5021 .write = mem_cgroup_write,
5022 .read_u64 = mem_cgroup_read_u64,
5025 .name = "kmem.usage_in_bytes",
5026 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5027 .read_u64 = mem_cgroup_read_u64,
5030 .name = "kmem.failcnt",
5031 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5032 .write = mem_cgroup_reset,
5033 .read_u64 = mem_cgroup_read_u64,
5036 .name = "kmem.max_usage_in_bytes",
5037 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5038 .write = mem_cgroup_reset,
5039 .read_u64 = mem_cgroup_read_u64,
5041 #if defined(CONFIG_MEMCG_KMEM) && \
5042 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5044 .name = "kmem.slabinfo",
5045 .seq_show = memcg_slab_show,
5049 .name = "kmem.tcp.limit_in_bytes",
5050 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5051 .write = mem_cgroup_write,
5052 .read_u64 = mem_cgroup_read_u64,
5055 .name = "kmem.tcp.usage_in_bytes",
5056 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5057 .read_u64 = mem_cgroup_read_u64,
5060 .name = "kmem.tcp.failcnt",
5061 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5062 .write = mem_cgroup_reset,
5063 .read_u64 = mem_cgroup_read_u64,
5066 .name = "kmem.tcp.max_usage_in_bytes",
5067 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5068 .write = mem_cgroup_reset,
5069 .read_u64 = mem_cgroup_read_u64,
5071 { }, /* terminate */
5075 * Private memory cgroup IDR
5077 * Swap-out records and page cache shadow entries need to store memcg
5078 * references in constrained space, so we maintain an ID space that is
5079 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5080 * memory-controlled cgroups to 64k.
5082 * However, there usually are many references to the offline CSS after
5083 * the cgroup has been destroyed, such as page cache or reclaimable
5084 * slab objects, that don't need to hang on to the ID. We want to keep
5085 * those dead CSS from occupying IDs, or we might quickly exhaust the
5086 * relatively small ID space and prevent the creation of new cgroups
5087 * even when there are much fewer than 64k cgroups - possibly none.
5089 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5090 * be freed and recycled when it's no longer needed, which is usually
5091 * when the CSS is offlined.
5093 * The only exception to that are records of swapped out tmpfs/shmem
5094 * pages that need to be attributed to live ancestors on swapin. But
5095 * those references are manageable from userspace.
5098 static DEFINE_IDR(mem_cgroup_idr);
5100 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5102 if (memcg->id.id > 0) {
5103 idr_remove(&mem_cgroup_idr, memcg->id.id);
5108 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5111 refcount_add(n, &memcg->id.ref);
5114 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5116 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5117 mem_cgroup_id_remove(memcg);
5119 /* Memcg ID pins CSS */
5120 css_put(&memcg->css);
5124 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5126 mem_cgroup_id_put_many(memcg, 1);
5130 * mem_cgroup_from_id - look up a memcg from a memcg id
5131 * @id: the memcg id to look up
5133 * Caller must hold rcu_read_lock().
5135 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5137 WARN_ON_ONCE(!rcu_read_lock_held());
5138 return idr_find(&mem_cgroup_idr, id);
5141 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5143 struct mem_cgroup_per_node *pn;
5146 * This routine is called against possible nodes.
5147 * But it's BUG to call kmalloc() against offline node.
5149 * TODO: this routine can waste much memory for nodes which will
5150 * never be onlined. It's better to use memory hotplug callback
5153 if (!node_state(node, N_NORMAL_MEMORY))
5155 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5159 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5160 GFP_KERNEL_ACCOUNT);
5161 if (!pn->lruvec_stat_local) {
5166 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct lruvec_stat,
5167 GFP_KERNEL_ACCOUNT);
5168 if (!pn->lruvec_stat_cpu) {
5169 free_percpu(pn->lruvec_stat_local);
5174 lruvec_init(&pn->lruvec);
5175 pn->usage_in_excess = 0;
5176 pn->on_tree = false;
5179 memcg->nodeinfo[node] = pn;
5183 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5185 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5190 free_percpu(pn->lruvec_stat_cpu);
5191 free_percpu(pn->lruvec_stat_local);
5195 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5200 free_mem_cgroup_per_node_info(memcg, node);
5201 free_percpu(memcg->vmstats_percpu);
5202 free_percpu(memcg->vmstats_local);
5206 static void mem_cgroup_free(struct mem_cgroup *memcg)
5208 memcg_wb_domain_exit(memcg);
5210 * Flush percpu vmstats and vmevents to guarantee the value correctness
5211 * on parent's and all ancestor levels.
5213 memcg_flush_percpu_vmstats(memcg);
5214 memcg_flush_percpu_vmevents(memcg);
5215 __mem_cgroup_free(memcg);
5218 static struct mem_cgroup *mem_cgroup_alloc(void)
5220 struct mem_cgroup *memcg;
5223 int __maybe_unused i;
5224 long error = -ENOMEM;
5226 size = sizeof(struct mem_cgroup);
5227 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5229 memcg = kzalloc(size, GFP_KERNEL);
5231 return ERR_PTR(error);
5233 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5234 1, MEM_CGROUP_ID_MAX,
5236 if (memcg->id.id < 0) {
5237 error = memcg->id.id;
5241 memcg->vmstats_local = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5242 GFP_KERNEL_ACCOUNT);
5243 if (!memcg->vmstats_local)
5246 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5247 GFP_KERNEL_ACCOUNT);
5248 if (!memcg->vmstats_percpu)
5252 if (alloc_mem_cgroup_per_node_info(memcg, node))
5255 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5258 INIT_WORK(&memcg->high_work, high_work_func);
5259 INIT_LIST_HEAD(&memcg->oom_notify);
5260 mutex_init(&memcg->thresholds_lock);
5261 spin_lock_init(&memcg->move_lock);
5262 vmpressure_init(&memcg->vmpressure);
5263 INIT_LIST_HEAD(&memcg->event_list);
5264 spin_lock_init(&memcg->event_list_lock);
5265 memcg->socket_pressure = jiffies;
5266 #ifdef CONFIG_MEMCG_KMEM
5267 memcg->kmemcg_id = -1;
5268 INIT_LIST_HEAD(&memcg->objcg_list);
5270 #ifdef CONFIG_CGROUP_WRITEBACK
5271 INIT_LIST_HEAD(&memcg->cgwb_list);
5272 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5273 memcg->cgwb_frn[i].done =
5274 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5276 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5277 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5278 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5279 memcg->deferred_split_queue.split_queue_len = 0;
5281 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5284 mem_cgroup_id_remove(memcg);
5285 __mem_cgroup_free(memcg);
5286 return ERR_PTR(error);
5289 static struct cgroup_subsys_state * __ref
5290 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5292 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5293 struct mem_cgroup *memcg;
5294 long error = -ENOMEM;
5296 memalloc_use_memcg(parent);
5297 memcg = mem_cgroup_alloc();
5298 memalloc_unuse_memcg();
5300 return ERR_CAST(memcg);
5302 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5303 memcg->soft_limit = PAGE_COUNTER_MAX;
5304 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5306 memcg->swappiness = mem_cgroup_swappiness(parent);
5307 memcg->oom_kill_disable = parent->oom_kill_disable;
5309 if (parent && parent->use_hierarchy) {
5310 memcg->use_hierarchy = true;
5311 page_counter_init(&memcg->memory, &parent->memory);
5312 page_counter_init(&memcg->swap, &parent->swap);
5313 page_counter_init(&memcg->kmem, &parent->kmem);
5314 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5316 page_counter_init(&memcg->memory, NULL);
5317 page_counter_init(&memcg->swap, NULL);
5318 page_counter_init(&memcg->kmem, NULL);
5319 page_counter_init(&memcg->tcpmem, NULL);
5321 * Deeper hierachy with use_hierarchy == false doesn't make
5322 * much sense so let cgroup subsystem know about this
5323 * unfortunate state in our controller.
5325 if (parent != root_mem_cgroup)
5326 memory_cgrp_subsys.broken_hierarchy = true;
5329 /* The following stuff does not apply to the root */
5331 root_mem_cgroup = memcg;
5335 error = memcg_online_kmem(memcg);
5339 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5340 static_branch_inc(&memcg_sockets_enabled_key);
5344 mem_cgroup_id_remove(memcg);
5345 mem_cgroup_free(memcg);
5346 return ERR_PTR(error);
5349 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5351 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5354 * A memcg must be visible for memcg_expand_shrinker_maps()
5355 * by the time the maps are allocated. So, we allocate maps
5356 * here, when for_each_mem_cgroup() can't skip it.
5358 if (memcg_alloc_shrinker_maps(memcg)) {
5359 mem_cgroup_id_remove(memcg);
5363 /* Online state pins memcg ID, memcg ID pins CSS */
5364 refcount_set(&memcg->id.ref, 1);
5369 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5371 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5372 struct mem_cgroup_event *event, *tmp;
5375 * Unregister events and notify userspace.
5376 * Notify userspace about cgroup removing only after rmdir of cgroup
5377 * directory to avoid race between userspace and kernelspace.
5379 spin_lock(&memcg->event_list_lock);
5380 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5381 list_del_init(&event->list);
5382 schedule_work(&event->remove);
5384 spin_unlock(&memcg->event_list_lock);
5386 page_counter_set_min(&memcg->memory, 0);
5387 page_counter_set_low(&memcg->memory, 0);
5389 memcg_offline_kmem(memcg);
5390 wb_memcg_offline(memcg);
5392 drain_all_stock(memcg);
5394 mem_cgroup_id_put(memcg);
5397 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5399 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5401 invalidate_reclaim_iterators(memcg);
5404 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5406 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5407 int __maybe_unused i;
5409 #ifdef CONFIG_CGROUP_WRITEBACK
5410 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5411 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5413 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5414 static_branch_dec(&memcg_sockets_enabled_key);
5416 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5417 static_branch_dec(&memcg_sockets_enabled_key);
5419 vmpressure_cleanup(&memcg->vmpressure);
5420 cancel_work_sync(&memcg->high_work);
5421 mem_cgroup_remove_from_trees(memcg);
5422 memcg_free_shrinker_maps(memcg);
5423 memcg_free_kmem(memcg);
5424 mem_cgroup_free(memcg);
5428 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5429 * @css: the target css
5431 * Reset the states of the mem_cgroup associated with @css. This is
5432 * invoked when the userland requests disabling on the default hierarchy
5433 * but the memcg is pinned through dependency. The memcg should stop
5434 * applying policies and should revert to the vanilla state as it may be
5435 * made visible again.
5437 * The current implementation only resets the essential configurations.
5438 * This needs to be expanded to cover all the visible parts.
5440 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5442 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5444 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5445 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5446 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5447 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5448 page_counter_set_min(&memcg->memory, 0);
5449 page_counter_set_low(&memcg->memory, 0);
5450 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5451 memcg->soft_limit = PAGE_COUNTER_MAX;
5452 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5453 memcg_wb_domain_size_changed(memcg);
5457 /* Handlers for move charge at task migration. */
5458 static int mem_cgroup_do_precharge(unsigned long count)
5462 /* Try a single bulk charge without reclaim first, kswapd may wake */
5463 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5465 mc.precharge += count;
5469 /* Try charges one by one with reclaim, but do not retry */
5471 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5485 enum mc_target_type {
5492 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5493 unsigned long addr, pte_t ptent)
5495 struct page *page = vm_normal_page(vma, addr, ptent);
5497 if (!page || !page_mapped(page))
5499 if (PageAnon(page)) {
5500 if (!(mc.flags & MOVE_ANON))
5503 if (!(mc.flags & MOVE_FILE))
5506 if (!get_page_unless_zero(page))
5512 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5513 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5514 pte_t ptent, swp_entry_t *entry)
5516 struct page *page = NULL;
5517 swp_entry_t ent = pte_to_swp_entry(ptent);
5519 if (!(mc.flags & MOVE_ANON))
5523 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5524 * a device and because they are not accessible by CPU they are store
5525 * as special swap entry in the CPU page table.
5527 if (is_device_private_entry(ent)) {
5528 page = device_private_entry_to_page(ent);
5530 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5531 * a refcount of 1 when free (unlike normal page)
5533 if (!page_ref_add_unless(page, 1, 1))
5538 if (non_swap_entry(ent))
5542 * Because lookup_swap_cache() updates some statistics counter,
5543 * we call find_get_page() with swapper_space directly.
5545 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5546 entry->val = ent.val;
5551 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5552 pte_t ptent, swp_entry_t *entry)
5558 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5559 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5561 if (!vma->vm_file) /* anonymous vma */
5563 if (!(mc.flags & MOVE_FILE))
5566 /* page is moved even if it's not RSS of this task(page-faulted). */
5567 /* shmem/tmpfs may report page out on swap: account for that too. */
5568 return find_get_incore_page(vma->vm_file->f_mapping,
5569 linear_page_index(vma, addr));
5573 * mem_cgroup_move_account - move account of the page
5575 * @compound: charge the page as compound or small page
5576 * @from: mem_cgroup which the page is moved from.
5577 * @to: mem_cgroup which the page is moved to. @from != @to.
5579 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5581 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5584 static int mem_cgroup_move_account(struct page *page,
5586 struct mem_cgroup *from,
5587 struct mem_cgroup *to)
5589 struct lruvec *from_vec, *to_vec;
5590 struct pglist_data *pgdat;
5591 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5594 VM_BUG_ON(from == to);
5595 VM_BUG_ON_PAGE(PageLRU(page), page);
5596 VM_BUG_ON(compound && !PageTransHuge(page));
5599 * Prevent mem_cgroup_migrate() from looking at
5600 * page->mem_cgroup of its source page while we change it.
5603 if (!trylock_page(page))
5607 if (page->mem_cgroup != from)
5610 pgdat = page_pgdat(page);
5611 from_vec = mem_cgroup_lruvec(from, pgdat);
5612 to_vec = mem_cgroup_lruvec(to, pgdat);
5614 lock_page_memcg(page);
5616 if (PageAnon(page)) {
5617 if (page_mapped(page)) {
5618 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5619 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5620 if (PageTransHuge(page)) {
5621 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5623 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5629 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5630 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5632 if (PageSwapBacked(page)) {
5633 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5634 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5637 if (page_mapped(page)) {
5638 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5639 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5642 if (PageDirty(page)) {
5643 struct address_space *mapping = page_mapping(page);
5645 if (mapping_can_writeback(mapping)) {
5646 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5648 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5654 if (PageWriteback(page)) {
5655 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5656 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5660 * All state has been migrated, let's switch to the new memcg.
5662 * It is safe to change page->mem_cgroup here because the page
5663 * is referenced, charged, isolated, and locked: we can't race
5664 * with (un)charging, migration, LRU putback, or anything else
5665 * that would rely on a stable page->mem_cgroup.
5667 * Note that lock_page_memcg is a memcg lock, not a page lock,
5668 * to save space. As soon as we switch page->mem_cgroup to a
5669 * new memcg that isn't locked, the above state can change
5670 * concurrently again. Make sure we're truly done with it.
5675 css_put(&from->css);
5677 page->mem_cgroup = to;
5679 __unlock_page_memcg(from);
5683 local_irq_disable();
5684 mem_cgroup_charge_statistics(to, page, nr_pages);
5685 memcg_check_events(to, page);
5686 mem_cgroup_charge_statistics(from, page, -nr_pages);
5687 memcg_check_events(from, page);
5696 * get_mctgt_type - get target type of moving charge
5697 * @vma: the vma the pte to be checked belongs
5698 * @addr: the address corresponding to the pte to be checked
5699 * @ptent: the pte to be checked
5700 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5703 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5704 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5705 * move charge. if @target is not NULL, the page is stored in target->page
5706 * with extra refcnt got(Callers should handle it).
5707 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5708 * target for charge migration. if @target is not NULL, the entry is stored
5710 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5711 * (so ZONE_DEVICE page and thus not on the lru).
5712 * For now we such page is charge like a regular page would be as for all
5713 * intent and purposes it is just special memory taking the place of a
5716 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5718 * Called with pte lock held.
5721 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5722 unsigned long addr, pte_t ptent, union mc_target *target)
5724 struct page *page = NULL;
5725 enum mc_target_type ret = MC_TARGET_NONE;
5726 swp_entry_t ent = { .val = 0 };
5728 if (pte_present(ptent))
5729 page = mc_handle_present_pte(vma, addr, ptent);
5730 else if (is_swap_pte(ptent))
5731 page = mc_handle_swap_pte(vma, ptent, &ent);
5732 else if (pte_none(ptent))
5733 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5735 if (!page && !ent.val)
5739 * Do only loose check w/o serialization.
5740 * mem_cgroup_move_account() checks the page is valid or
5741 * not under LRU exclusion.
5743 if (page->mem_cgroup == mc.from) {
5744 ret = MC_TARGET_PAGE;
5745 if (is_device_private_page(page))
5746 ret = MC_TARGET_DEVICE;
5748 target->page = page;
5750 if (!ret || !target)
5754 * There is a swap entry and a page doesn't exist or isn't charged.
5755 * But we cannot move a tail-page in a THP.
5757 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5758 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5759 ret = MC_TARGET_SWAP;
5766 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5768 * We don't consider PMD mapped swapping or file mapped pages because THP does
5769 * not support them for now.
5770 * Caller should make sure that pmd_trans_huge(pmd) is true.
5772 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5773 unsigned long addr, pmd_t pmd, union mc_target *target)
5775 struct page *page = NULL;
5776 enum mc_target_type ret = MC_TARGET_NONE;
5778 if (unlikely(is_swap_pmd(pmd))) {
5779 VM_BUG_ON(thp_migration_supported() &&
5780 !is_pmd_migration_entry(pmd));
5783 page = pmd_page(pmd);
5784 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5785 if (!(mc.flags & MOVE_ANON))
5787 if (page->mem_cgroup == mc.from) {
5788 ret = MC_TARGET_PAGE;
5791 target->page = page;
5797 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5798 unsigned long addr, pmd_t pmd, union mc_target *target)
5800 return MC_TARGET_NONE;
5804 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5805 unsigned long addr, unsigned long end,
5806 struct mm_walk *walk)
5808 struct vm_area_struct *vma = walk->vma;
5812 ptl = pmd_trans_huge_lock(pmd, vma);
5815 * Note their can not be MC_TARGET_DEVICE for now as we do not
5816 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5817 * this might change.
5819 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5820 mc.precharge += HPAGE_PMD_NR;
5825 if (pmd_trans_unstable(pmd))
5827 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5828 for (; addr != end; pte++, addr += PAGE_SIZE)
5829 if (get_mctgt_type(vma, addr, *pte, NULL))
5830 mc.precharge++; /* increment precharge temporarily */
5831 pte_unmap_unlock(pte - 1, ptl);
5837 static const struct mm_walk_ops precharge_walk_ops = {
5838 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5841 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5843 unsigned long precharge;
5846 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5847 mmap_read_unlock(mm);
5849 precharge = mc.precharge;
5855 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5857 unsigned long precharge = mem_cgroup_count_precharge(mm);
5859 VM_BUG_ON(mc.moving_task);
5860 mc.moving_task = current;
5861 return mem_cgroup_do_precharge(precharge);
5864 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5865 static void __mem_cgroup_clear_mc(void)
5867 struct mem_cgroup *from = mc.from;
5868 struct mem_cgroup *to = mc.to;
5870 /* we must uncharge all the leftover precharges from mc.to */
5872 cancel_charge(mc.to, mc.precharge);
5876 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5877 * we must uncharge here.
5879 if (mc.moved_charge) {
5880 cancel_charge(mc.from, mc.moved_charge);
5881 mc.moved_charge = 0;
5883 /* we must fixup refcnts and charges */
5884 if (mc.moved_swap) {
5885 /* uncharge swap account from the old cgroup */
5886 if (!mem_cgroup_is_root(mc.from))
5887 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5889 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5892 * we charged both to->memory and to->memsw, so we
5893 * should uncharge to->memory.
5895 if (!mem_cgroup_is_root(mc.to))
5896 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5900 memcg_oom_recover(from);
5901 memcg_oom_recover(to);
5902 wake_up_all(&mc.waitq);
5905 static void mem_cgroup_clear_mc(void)
5907 struct mm_struct *mm = mc.mm;
5910 * we must clear moving_task before waking up waiters at the end of
5913 mc.moving_task = NULL;
5914 __mem_cgroup_clear_mc();
5915 spin_lock(&mc.lock);
5919 spin_unlock(&mc.lock);
5924 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5926 struct cgroup_subsys_state *css;
5927 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5928 struct mem_cgroup *from;
5929 struct task_struct *leader, *p;
5930 struct mm_struct *mm;
5931 unsigned long move_flags;
5934 /* charge immigration isn't supported on the default hierarchy */
5935 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5939 * Multi-process migrations only happen on the default hierarchy
5940 * where charge immigration is not used. Perform charge
5941 * immigration if @tset contains a leader and whine if there are
5945 cgroup_taskset_for_each_leader(leader, css, tset) {
5948 memcg = mem_cgroup_from_css(css);
5954 * We are now commited to this value whatever it is. Changes in this
5955 * tunable will only affect upcoming migrations, not the current one.
5956 * So we need to save it, and keep it going.
5958 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5962 from = mem_cgroup_from_task(p);
5964 VM_BUG_ON(from == memcg);
5966 mm = get_task_mm(p);
5969 /* We move charges only when we move a owner of the mm */
5970 if (mm->owner == p) {
5973 VM_BUG_ON(mc.precharge);
5974 VM_BUG_ON(mc.moved_charge);
5975 VM_BUG_ON(mc.moved_swap);
5977 spin_lock(&mc.lock);
5981 mc.flags = move_flags;
5982 spin_unlock(&mc.lock);
5983 /* We set mc.moving_task later */
5985 ret = mem_cgroup_precharge_mc(mm);
5987 mem_cgroup_clear_mc();
5994 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5997 mem_cgroup_clear_mc();
6000 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6001 unsigned long addr, unsigned long end,
6002 struct mm_walk *walk)
6005 struct vm_area_struct *vma = walk->vma;
6008 enum mc_target_type target_type;
6009 union mc_target target;
6012 ptl = pmd_trans_huge_lock(pmd, vma);
6014 if (mc.precharge < HPAGE_PMD_NR) {
6018 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6019 if (target_type == MC_TARGET_PAGE) {
6021 if (!isolate_lru_page(page)) {
6022 if (!mem_cgroup_move_account(page, true,
6024 mc.precharge -= HPAGE_PMD_NR;
6025 mc.moved_charge += HPAGE_PMD_NR;
6027 putback_lru_page(page);
6030 } else if (target_type == MC_TARGET_DEVICE) {
6032 if (!mem_cgroup_move_account(page, true,
6034 mc.precharge -= HPAGE_PMD_NR;
6035 mc.moved_charge += HPAGE_PMD_NR;
6043 if (pmd_trans_unstable(pmd))
6046 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6047 for (; addr != end; addr += PAGE_SIZE) {
6048 pte_t ptent = *(pte++);
6049 bool device = false;
6055 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6056 case MC_TARGET_DEVICE:
6059 case MC_TARGET_PAGE:
6062 * We can have a part of the split pmd here. Moving it
6063 * can be done but it would be too convoluted so simply
6064 * ignore such a partial THP and keep it in original
6065 * memcg. There should be somebody mapping the head.
6067 if (PageTransCompound(page))
6069 if (!device && isolate_lru_page(page))
6071 if (!mem_cgroup_move_account(page, false,
6074 /* we uncharge from mc.from later. */
6078 putback_lru_page(page);
6079 put: /* get_mctgt_type() gets the page */
6082 case MC_TARGET_SWAP:
6084 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6086 mem_cgroup_id_get_many(mc.to, 1);
6087 /* we fixup other refcnts and charges later. */
6095 pte_unmap_unlock(pte - 1, ptl);
6100 * We have consumed all precharges we got in can_attach().
6101 * We try charge one by one, but don't do any additional
6102 * charges to mc.to if we have failed in charge once in attach()
6105 ret = mem_cgroup_do_precharge(1);
6113 static const struct mm_walk_ops charge_walk_ops = {
6114 .pmd_entry = mem_cgroup_move_charge_pte_range,
6117 static void mem_cgroup_move_charge(void)
6119 lru_add_drain_all();
6121 * Signal lock_page_memcg() to take the memcg's move_lock
6122 * while we're moving its pages to another memcg. Then wait
6123 * for already started RCU-only updates to finish.
6125 atomic_inc(&mc.from->moving_account);
6128 if (unlikely(!mmap_read_trylock(mc.mm))) {
6130 * Someone who are holding the mmap_lock might be waiting in
6131 * waitq. So we cancel all extra charges, wake up all waiters,
6132 * and retry. Because we cancel precharges, we might not be able
6133 * to move enough charges, but moving charge is a best-effort
6134 * feature anyway, so it wouldn't be a big problem.
6136 __mem_cgroup_clear_mc();
6141 * When we have consumed all precharges and failed in doing
6142 * additional charge, the page walk just aborts.
6144 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6147 mmap_read_unlock(mc.mm);
6148 atomic_dec(&mc.from->moving_account);
6151 static void mem_cgroup_move_task(void)
6154 mem_cgroup_move_charge();
6155 mem_cgroup_clear_mc();
6158 #else /* !CONFIG_MMU */
6159 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6163 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6166 static void mem_cgroup_move_task(void)
6172 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6173 * to verify whether we're attached to the default hierarchy on each mount
6176 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6179 * use_hierarchy is forced on the default hierarchy. cgroup core
6180 * guarantees that @root doesn't have any children, so turning it
6181 * on for the root memcg is enough.
6183 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6184 root_mem_cgroup->use_hierarchy = true;
6186 root_mem_cgroup->use_hierarchy = false;
6189 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6191 if (value == PAGE_COUNTER_MAX)
6192 seq_puts(m, "max\n");
6194 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6199 static u64 memory_current_read(struct cgroup_subsys_state *css,
6202 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6204 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6207 static int memory_min_show(struct seq_file *m, void *v)
6209 return seq_puts_memcg_tunable(m,
6210 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6213 static ssize_t memory_min_write(struct kernfs_open_file *of,
6214 char *buf, size_t nbytes, loff_t off)
6216 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6220 buf = strstrip(buf);
6221 err = page_counter_memparse(buf, "max", &min);
6225 page_counter_set_min(&memcg->memory, min);
6230 static int memory_low_show(struct seq_file *m, void *v)
6232 return seq_puts_memcg_tunable(m,
6233 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6236 static ssize_t memory_low_write(struct kernfs_open_file *of,
6237 char *buf, size_t nbytes, loff_t off)
6239 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6243 buf = strstrip(buf);
6244 err = page_counter_memparse(buf, "max", &low);
6248 page_counter_set_low(&memcg->memory, low);
6253 static int memory_high_show(struct seq_file *m, void *v)
6255 return seq_puts_memcg_tunable(m,
6256 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6259 static ssize_t memory_high_write(struct kernfs_open_file *of,
6260 char *buf, size_t nbytes, loff_t off)
6262 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6263 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6264 bool drained = false;
6268 buf = strstrip(buf);
6269 err = page_counter_memparse(buf, "max", &high);
6274 unsigned long nr_pages = page_counter_read(&memcg->memory);
6275 unsigned long reclaimed;
6277 if (nr_pages <= high)
6280 if (signal_pending(current))
6284 drain_all_stock(memcg);
6289 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6292 if (!reclaimed && !nr_retries--)
6296 page_counter_set_high(&memcg->memory, high);
6298 memcg_wb_domain_size_changed(memcg);
6303 static int memory_max_show(struct seq_file *m, void *v)
6305 return seq_puts_memcg_tunable(m,
6306 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6309 static ssize_t memory_max_write(struct kernfs_open_file *of,
6310 char *buf, size_t nbytes, loff_t off)
6312 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6313 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6314 bool drained = false;
6318 buf = strstrip(buf);
6319 err = page_counter_memparse(buf, "max", &max);
6323 xchg(&memcg->memory.max, max);
6326 unsigned long nr_pages = page_counter_read(&memcg->memory);
6328 if (nr_pages <= max)
6331 if (signal_pending(current))
6335 drain_all_stock(memcg);
6341 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6347 memcg_memory_event(memcg, MEMCG_OOM);
6348 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6352 memcg_wb_domain_size_changed(memcg);
6356 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6358 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6359 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6360 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6361 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6362 seq_printf(m, "oom_kill %lu\n",
6363 atomic_long_read(&events[MEMCG_OOM_KILL]));
6366 static int memory_events_show(struct seq_file *m, void *v)
6368 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6370 __memory_events_show(m, memcg->memory_events);
6374 static int memory_events_local_show(struct seq_file *m, void *v)
6376 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6378 __memory_events_show(m, memcg->memory_events_local);
6382 static int memory_stat_show(struct seq_file *m, void *v)
6384 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6387 buf = memory_stat_format(memcg);
6396 static int memory_numa_stat_show(struct seq_file *m, void *v)
6399 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6401 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6404 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6407 seq_printf(m, "%s", memory_stats[i].name);
6408 for_each_node_state(nid, N_MEMORY) {
6410 struct lruvec *lruvec;
6412 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6413 size = lruvec_page_state(lruvec, memory_stats[i].idx);
6414 size *= memory_stats[i].ratio;
6415 seq_printf(m, " N%d=%llu", nid, size);
6424 static int memory_oom_group_show(struct seq_file *m, void *v)
6426 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6428 seq_printf(m, "%d\n", memcg->oom_group);
6433 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6434 char *buf, size_t nbytes, loff_t off)
6436 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6439 buf = strstrip(buf);
6443 ret = kstrtoint(buf, 0, &oom_group);
6447 if (oom_group != 0 && oom_group != 1)
6450 memcg->oom_group = oom_group;
6455 static struct cftype memory_files[] = {
6458 .flags = CFTYPE_NOT_ON_ROOT,
6459 .read_u64 = memory_current_read,
6463 .flags = CFTYPE_NOT_ON_ROOT,
6464 .seq_show = memory_min_show,
6465 .write = memory_min_write,
6469 .flags = CFTYPE_NOT_ON_ROOT,
6470 .seq_show = memory_low_show,
6471 .write = memory_low_write,
6475 .flags = CFTYPE_NOT_ON_ROOT,
6476 .seq_show = memory_high_show,
6477 .write = memory_high_write,
6481 .flags = CFTYPE_NOT_ON_ROOT,
6482 .seq_show = memory_max_show,
6483 .write = memory_max_write,
6487 .flags = CFTYPE_NOT_ON_ROOT,
6488 .file_offset = offsetof(struct mem_cgroup, events_file),
6489 .seq_show = memory_events_show,
6492 .name = "events.local",
6493 .flags = CFTYPE_NOT_ON_ROOT,
6494 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6495 .seq_show = memory_events_local_show,
6499 .seq_show = memory_stat_show,
6503 .name = "numa_stat",
6504 .seq_show = memory_numa_stat_show,
6508 .name = "oom.group",
6509 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6510 .seq_show = memory_oom_group_show,
6511 .write = memory_oom_group_write,
6516 struct cgroup_subsys memory_cgrp_subsys = {
6517 .css_alloc = mem_cgroup_css_alloc,
6518 .css_online = mem_cgroup_css_online,
6519 .css_offline = mem_cgroup_css_offline,
6520 .css_released = mem_cgroup_css_released,
6521 .css_free = mem_cgroup_css_free,
6522 .css_reset = mem_cgroup_css_reset,
6523 .can_attach = mem_cgroup_can_attach,
6524 .cancel_attach = mem_cgroup_cancel_attach,
6525 .post_attach = mem_cgroup_move_task,
6526 .bind = mem_cgroup_bind,
6527 .dfl_cftypes = memory_files,
6528 .legacy_cftypes = mem_cgroup_legacy_files,
6533 * This function calculates an individual cgroup's effective
6534 * protection which is derived from its own memory.min/low, its
6535 * parent's and siblings' settings, as well as the actual memory
6536 * distribution in the tree.
6538 * The following rules apply to the effective protection values:
6540 * 1. At the first level of reclaim, effective protection is equal to
6541 * the declared protection in memory.min and memory.low.
6543 * 2. To enable safe delegation of the protection configuration, at
6544 * subsequent levels the effective protection is capped to the
6545 * parent's effective protection.
6547 * 3. To make complex and dynamic subtrees easier to configure, the
6548 * user is allowed to overcommit the declared protection at a given
6549 * level. If that is the case, the parent's effective protection is
6550 * distributed to the children in proportion to how much protection
6551 * they have declared and how much of it they are utilizing.
6553 * This makes distribution proportional, but also work-conserving:
6554 * if one cgroup claims much more protection than it uses memory,
6555 * the unused remainder is available to its siblings.
6557 * 4. Conversely, when the declared protection is undercommitted at a
6558 * given level, the distribution of the larger parental protection
6559 * budget is NOT proportional. A cgroup's protection from a sibling
6560 * is capped to its own memory.min/low setting.
6562 * 5. However, to allow protecting recursive subtrees from each other
6563 * without having to declare each individual cgroup's fixed share
6564 * of the ancestor's claim to protection, any unutilized -
6565 * "floating" - protection from up the tree is distributed in
6566 * proportion to each cgroup's *usage*. This makes the protection
6567 * neutral wrt sibling cgroups and lets them compete freely over
6568 * the shared parental protection budget, but it protects the
6569 * subtree as a whole from neighboring subtrees.
6571 * Note that 4. and 5. are not in conflict: 4. is about protecting
6572 * against immediate siblings whereas 5. is about protecting against
6573 * neighboring subtrees.
6575 static unsigned long effective_protection(unsigned long usage,
6576 unsigned long parent_usage,
6577 unsigned long setting,
6578 unsigned long parent_effective,
6579 unsigned long siblings_protected)
6581 unsigned long protected;
6584 protected = min(usage, setting);
6586 * If all cgroups at this level combined claim and use more
6587 * protection then what the parent affords them, distribute
6588 * shares in proportion to utilization.
6590 * We are using actual utilization rather than the statically
6591 * claimed protection in order to be work-conserving: claimed
6592 * but unused protection is available to siblings that would
6593 * otherwise get a smaller chunk than what they claimed.
6595 if (siblings_protected > parent_effective)
6596 return protected * parent_effective / siblings_protected;
6599 * Ok, utilized protection of all children is within what the
6600 * parent affords them, so we know whatever this child claims
6601 * and utilizes is effectively protected.
6603 * If there is unprotected usage beyond this value, reclaim
6604 * will apply pressure in proportion to that amount.
6606 * If there is unutilized protection, the cgroup will be fully
6607 * shielded from reclaim, but we do return a smaller value for
6608 * protection than what the group could enjoy in theory. This
6609 * is okay. With the overcommit distribution above, effective
6610 * protection is always dependent on how memory is actually
6611 * consumed among the siblings anyway.
6616 * If the children aren't claiming (all of) the protection
6617 * afforded to them by the parent, distribute the remainder in
6618 * proportion to the (unprotected) memory of each cgroup. That
6619 * way, cgroups that aren't explicitly prioritized wrt each
6620 * other compete freely over the allowance, but they are
6621 * collectively protected from neighboring trees.
6623 * We're using unprotected memory for the weight so that if
6624 * some cgroups DO claim explicit protection, we don't protect
6625 * the same bytes twice.
6627 * Check both usage and parent_usage against the respective
6628 * protected values. One should imply the other, but they
6629 * aren't read atomically - make sure the division is sane.
6631 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6633 if (parent_effective > siblings_protected &&
6634 parent_usage > siblings_protected &&
6635 usage > protected) {
6636 unsigned long unclaimed;
6638 unclaimed = parent_effective - siblings_protected;
6639 unclaimed *= usage - protected;
6640 unclaimed /= parent_usage - siblings_protected;
6649 * mem_cgroup_protected - check if memory consumption is in the normal range
6650 * @root: the top ancestor of the sub-tree being checked
6651 * @memcg: the memory cgroup to check
6653 * WARNING: This function is not stateless! It can only be used as part
6654 * of a top-down tree iteration, not for isolated queries.
6656 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6657 struct mem_cgroup *memcg)
6659 unsigned long usage, parent_usage;
6660 struct mem_cgroup *parent;
6662 if (mem_cgroup_disabled())
6666 root = root_mem_cgroup;
6669 * Effective values of the reclaim targets are ignored so they
6670 * can be stale. Have a look at mem_cgroup_protection for more
6672 * TODO: calculation should be more robust so that we do not need
6673 * that special casing.
6678 usage = page_counter_read(&memcg->memory);
6682 parent = parent_mem_cgroup(memcg);
6683 /* No parent means a non-hierarchical mode on v1 memcg */
6687 if (parent == root) {
6688 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6689 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6693 parent_usage = page_counter_read(&parent->memory);
6695 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6696 READ_ONCE(memcg->memory.min),
6697 READ_ONCE(parent->memory.emin),
6698 atomic_long_read(&parent->memory.children_min_usage)));
6700 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6701 READ_ONCE(memcg->memory.low),
6702 READ_ONCE(parent->memory.elow),
6703 atomic_long_read(&parent->memory.children_low_usage)));
6707 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6708 * @page: page to charge
6709 * @mm: mm context of the victim
6710 * @gfp_mask: reclaim mode
6712 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6713 * pages according to @gfp_mask if necessary.
6715 * Returns 0 on success. Otherwise, an error code is returned.
6717 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6719 unsigned int nr_pages = thp_nr_pages(page);
6720 struct mem_cgroup *memcg = NULL;
6723 if (mem_cgroup_disabled())
6726 if (PageSwapCache(page)) {
6727 swp_entry_t ent = { .val = page_private(page), };
6731 * Every swap fault against a single page tries to charge the
6732 * page, bail as early as possible. shmem_unuse() encounters
6733 * already charged pages, too. page->mem_cgroup is protected
6734 * by the page lock, which serializes swap cache removal, which
6735 * in turn serializes uncharging.
6737 VM_BUG_ON_PAGE(!PageLocked(page), page);
6738 if (compound_head(page)->mem_cgroup)
6741 id = lookup_swap_cgroup_id(ent);
6743 memcg = mem_cgroup_from_id(id);
6744 if (memcg && !css_tryget_online(&memcg->css))
6750 memcg = get_mem_cgroup_from_mm(mm);
6752 ret = try_charge(memcg, gfp_mask, nr_pages);
6756 css_get(&memcg->css);
6757 commit_charge(page, memcg);
6759 local_irq_disable();
6760 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6761 memcg_check_events(memcg, page);
6764 if (PageSwapCache(page)) {
6765 swp_entry_t entry = { .val = page_private(page) };
6767 * The swap entry might not get freed for a long time,
6768 * let's not wait for it. The page already received a
6769 * memory+swap charge, drop the swap entry duplicate.
6771 mem_cgroup_uncharge_swap(entry, nr_pages);
6775 css_put(&memcg->css);
6780 struct uncharge_gather {
6781 struct mem_cgroup *memcg;
6782 unsigned long nr_pages;
6783 unsigned long pgpgout;
6784 unsigned long nr_kmem;
6785 struct page *dummy_page;
6788 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6790 memset(ug, 0, sizeof(*ug));
6793 static void uncharge_batch(const struct uncharge_gather *ug)
6795 unsigned long flags;
6797 if (!mem_cgroup_is_root(ug->memcg)) {
6798 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6799 if (do_memsw_account())
6800 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6801 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6802 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6803 memcg_oom_recover(ug->memcg);
6806 local_irq_save(flags);
6807 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6808 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6809 memcg_check_events(ug->memcg, ug->dummy_page);
6810 local_irq_restore(flags);
6812 /* drop reference from uncharge_page */
6813 css_put(&ug->memcg->css);
6816 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6818 unsigned long nr_pages;
6820 VM_BUG_ON_PAGE(PageLRU(page), page);
6822 if (!page->mem_cgroup)
6826 * Nobody should be changing or seriously looking at
6827 * page->mem_cgroup at this point, we have fully
6828 * exclusive access to the page.
6831 if (ug->memcg != page->mem_cgroup) {
6834 uncharge_gather_clear(ug);
6836 ug->memcg = page->mem_cgroup;
6838 /* pairs with css_put in uncharge_batch */
6839 css_get(&ug->memcg->css);
6842 nr_pages = compound_nr(page);
6843 ug->nr_pages += nr_pages;
6845 if (!PageKmemcg(page)) {
6848 ug->nr_kmem += nr_pages;
6849 __ClearPageKmemcg(page);
6852 ug->dummy_page = page;
6853 page->mem_cgroup = NULL;
6854 css_put(&ug->memcg->css);
6857 static void uncharge_list(struct list_head *page_list)
6859 struct uncharge_gather ug;
6860 struct list_head *next;
6862 uncharge_gather_clear(&ug);
6865 * Note that the list can be a single page->lru; hence the
6866 * do-while loop instead of a simple list_for_each_entry().
6868 next = page_list->next;
6872 page = list_entry(next, struct page, lru);
6873 next = page->lru.next;
6875 uncharge_page(page, &ug);
6876 } while (next != page_list);
6879 uncharge_batch(&ug);
6883 * mem_cgroup_uncharge - uncharge a page
6884 * @page: page to uncharge
6886 * Uncharge a page previously charged with mem_cgroup_charge().
6888 void mem_cgroup_uncharge(struct page *page)
6890 struct uncharge_gather ug;
6892 if (mem_cgroup_disabled())
6895 /* Don't touch page->lru of any random page, pre-check: */
6896 if (!page->mem_cgroup)
6899 uncharge_gather_clear(&ug);
6900 uncharge_page(page, &ug);
6901 uncharge_batch(&ug);
6905 * mem_cgroup_uncharge_list - uncharge a list of page
6906 * @page_list: list of pages to uncharge
6908 * Uncharge a list of pages previously charged with
6909 * mem_cgroup_charge().
6911 void mem_cgroup_uncharge_list(struct list_head *page_list)
6913 if (mem_cgroup_disabled())
6916 if (!list_empty(page_list))
6917 uncharge_list(page_list);
6921 * mem_cgroup_migrate - charge a page's replacement
6922 * @oldpage: currently circulating page
6923 * @newpage: replacement page
6925 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6926 * be uncharged upon free.
6928 * Both pages must be locked, @newpage->mapping must be set up.
6930 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6932 struct mem_cgroup *memcg;
6933 unsigned int nr_pages;
6934 unsigned long flags;
6936 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6937 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6938 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6939 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6942 if (mem_cgroup_disabled())
6945 /* Page cache replacement: new page already charged? */
6946 if (newpage->mem_cgroup)
6949 /* Swapcache readahead pages can get replaced before being charged */
6950 memcg = oldpage->mem_cgroup;
6954 /* Force-charge the new page. The old one will be freed soon */
6955 nr_pages = thp_nr_pages(newpage);
6957 page_counter_charge(&memcg->memory, nr_pages);
6958 if (do_memsw_account())
6959 page_counter_charge(&memcg->memsw, nr_pages);
6961 css_get(&memcg->css);
6962 commit_charge(newpage, memcg);
6964 local_irq_save(flags);
6965 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6966 memcg_check_events(memcg, newpage);
6967 local_irq_restore(flags);
6970 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6971 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6973 void mem_cgroup_sk_alloc(struct sock *sk)
6975 struct mem_cgroup *memcg;
6977 if (!mem_cgroup_sockets_enabled)
6980 /* Do not associate the sock with unrelated interrupted task's memcg. */
6985 memcg = mem_cgroup_from_task(current);
6986 if (memcg == root_mem_cgroup)
6988 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6990 if (css_tryget(&memcg->css))
6991 sk->sk_memcg = memcg;
6996 void mem_cgroup_sk_free(struct sock *sk)
6999 css_put(&sk->sk_memcg->css);
7003 * mem_cgroup_charge_skmem - charge socket memory
7004 * @memcg: memcg to charge
7005 * @nr_pages: number of pages to charge
7007 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7008 * @memcg's configured limit, %false if the charge had to be forced.
7010 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7012 gfp_t gfp_mask = GFP_KERNEL;
7014 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7015 struct page_counter *fail;
7017 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7018 memcg->tcpmem_pressure = 0;
7021 page_counter_charge(&memcg->tcpmem, nr_pages);
7022 memcg->tcpmem_pressure = 1;
7026 /* Don't block in the packet receive path */
7028 gfp_mask = GFP_NOWAIT;
7030 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7032 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7035 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7040 * mem_cgroup_uncharge_skmem - uncharge socket memory
7041 * @memcg: memcg to uncharge
7042 * @nr_pages: number of pages to uncharge
7044 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7046 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7047 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7051 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7053 refill_stock(memcg, nr_pages);
7056 static int __init cgroup_memory(char *s)
7060 while ((token = strsep(&s, ",")) != NULL) {
7063 if (!strcmp(token, "nosocket"))
7064 cgroup_memory_nosocket = true;
7065 if (!strcmp(token, "nokmem"))
7066 cgroup_memory_nokmem = true;
7070 __setup("cgroup.memory=", cgroup_memory);
7073 * subsys_initcall() for memory controller.
7075 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7076 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7077 * basically everything that doesn't depend on a specific mem_cgroup structure
7078 * should be initialized from here.
7080 static int __init mem_cgroup_init(void)
7084 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7085 memcg_hotplug_cpu_dead);
7087 for_each_possible_cpu(cpu)
7088 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7091 for_each_node(node) {
7092 struct mem_cgroup_tree_per_node *rtpn;
7094 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7095 node_online(node) ? node : NUMA_NO_NODE);
7097 rtpn->rb_root = RB_ROOT;
7098 rtpn->rb_rightmost = NULL;
7099 spin_lock_init(&rtpn->lock);
7100 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7105 subsys_initcall(mem_cgroup_init);
7107 #ifdef CONFIG_MEMCG_SWAP
7108 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7110 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7112 * The root cgroup cannot be destroyed, so it's refcount must
7115 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7119 memcg = parent_mem_cgroup(memcg);
7121 memcg = root_mem_cgroup;
7127 * mem_cgroup_swapout - transfer a memsw charge to swap
7128 * @page: page whose memsw charge to transfer
7129 * @entry: swap entry to move the charge to
7131 * Transfer the memsw charge of @page to @entry.
7133 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7135 struct mem_cgroup *memcg, *swap_memcg;
7136 unsigned int nr_entries;
7137 unsigned short oldid;
7139 VM_BUG_ON_PAGE(PageLRU(page), page);
7140 VM_BUG_ON_PAGE(page_count(page), page);
7142 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7145 memcg = page->mem_cgroup;
7147 /* Readahead page, never charged */
7152 * In case the memcg owning these pages has been offlined and doesn't
7153 * have an ID allocated to it anymore, charge the closest online
7154 * ancestor for the swap instead and transfer the memory+swap charge.
7156 swap_memcg = mem_cgroup_id_get_online(memcg);
7157 nr_entries = thp_nr_pages(page);
7158 /* Get references for the tail pages, too */
7160 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7161 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7163 VM_BUG_ON_PAGE(oldid, page);
7164 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7166 page->mem_cgroup = NULL;
7168 if (!mem_cgroup_is_root(memcg))
7169 page_counter_uncharge(&memcg->memory, nr_entries);
7171 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7172 if (!mem_cgroup_is_root(swap_memcg))
7173 page_counter_charge(&swap_memcg->memsw, nr_entries);
7174 page_counter_uncharge(&memcg->memsw, nr_entries);
7178 * Interrupts should be disabled here because the caller holds the
7179 * i_pages lock which is taken with interrupts-off. It is
7180 * important here to have the interrupts disabled because it is the
7181 * only synchronisation we have for updating the per-CPU variables.
7183 VM_BUG_ON(!irqs_disabled());
7184 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7185 memcg_check_events(memcg, page);
7187 css_put(&memcg->css);
7191 * mem_cgroup_try_charge_swap - try charging swap space for a page
7192 * @page: page being added to swap
7193 * @entry: swap entry to charge
7195 * Try to charge @page's memcg for the swap space at @entry.
7197 * Returns 0 on success, -ENOMEM on failure.
7199 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7201 unsigned int nr_pages = thp_nr_pages(page);
7202 struct page_counter *counter;
7203 struct mem_cgroup *memcg;
7204 unsigned short oldid;
7206 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7209 memcg = page->mem_cgroup;
7211 /* Readahead page, never charged */
7216 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7220 memcg = mem_cgroup_id_get_online(memcg);
7222 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7223 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7224 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7225 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7226 mem_cgroup_id_put(memcg);
7230 /* Get references for the tail pages, too */
7232 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7233 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7234 VM_BUG_ON_PAGE(oldid, page);
7235 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7241 * mem_cgroup_uncharge_swap - uncharge swap space
7242 * @entry: swap entry to uncharge
7243 * @nr_pages: the amount of swap space to uncharge
7245 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7247 struct mem_cgroup *memcg;
7250 id = swap_cgroup_record(entry, 0, nr_pages);
7252 memcg = mem_cgroup_from_id(id);
7254 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7255 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7256 page_counter_uncharge(&memcg->swap, nr_pages);
7258 page_counter_uncharge(&memcg->memsw, nr_pages);
7260 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7261 mem_cgroup_id_put_many(memcg, nr_pages);
7266 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7268 long nr_swap_pages = get_nr_swap_pages();
7270 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7271 return nr_swap_pages;
7272 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7273 nr_swap_pages = min_t(long, nr_swap_pages,
7274 READ_ONCE(memcg->swap.max) -
7275 page_counter_read(&memcg->swap));
7276 return nr_swap_pages;
7279 bool mem_cgroup_swap_full(struct page *page)
7281 struct mem_cgroup *memcg;
7283 VM_BUG_ON_PAGE(!PageLocked(page), page);
7287 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7290 memcg = page->mem_cgroup;
7294 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7295 unsigned long usage = page_counter_read(&memcg->swap);
7297 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7298 usage * 2 >= READ_ONCE(memcg->swap.max))
7305 static int __init setup_swap_account(char *s)
7307 if (!strcmp(s, "1"))
7308 cgroup_memory_noswap = 0;
7309 else if (!strcmp(s, "0"))
7310 cgroup_memory_noswap = 1;
7313 __setup("swapaccount=", setup_swap_account);
7315 static u64 swap_current_read(struct cgroup_subsys_state *css,
7318 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7320 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7323 static int swap_high_show(struct seq_file *m, void *v)
7325 return seq_puts_memcg_tunable(m,
7326 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7329 static ssize_t swap_high_write(struct kernfs_open_file *of,
7330 char *buf, size_t nbytes, loff_t off)
7332 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7336 buf = strstrip(buf);
7337 err = page_counter_memparse(buf, "max", &high);
7341 page_counter_set_high(&memcg->swap, high);
7346 static int swap_max_show(struct seq_file *m, void *v)
7348 return seq_puts_memcg_tunable(m,
7349 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7352 static ssize_t swap_max_write(struct kernfs_open_file *of,
7353 char *buf, size_t nbytes, loff_t off)
7355 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7359 buf = strstrip(buf);
7360 err = page_counter_memparse(buf, "max", &max);
7364 xchg(&memcg->swap.max, max);
7369 static int swap_events_show(struct seq_file *m, void *v)
7371 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7373 seq_printf(m, "high %lu\n",
7374 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7375 seq_printf(m, "max %lu\n",
7376 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7377 seq_printf(m, "fail %lu\n",
7378 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7383 static struct cftype swap_files[] = {
7385 .name = "swap.current",
7386 .flags = CFTYPE_NOT_ON_ROOT,
7387 .read_u64 = swap_current_read,
7390 .name = "swap.high",
7391 .flags = CFTYPE_NOT_ON_ROOT,
7392 .seq_show = swap_high_show,
7393 .write = swap_high_write,
7397 .flags = CFTYPE_NOT_ON_ROOT,
7398 .seq_show = swap_max_show,
7399 .write = swap_max_write,
7402 .name = "swap.events",
7403 .flags = CFTYPE_NOT_ON_ROOT,
7404 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7405 .seq_show = swap_events_show,
7410 static struct cftype memsw_files[] = {
7412 .name = "memsw.usage_in_bytes",
7413 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7414 .read_u64 = mem_cgroup_read_u64,
7417 .name = "memsw.max_usage_in_bytes",
7418 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7419 .write = mem_cgroup_reset,
7420 .read_u64 = mem_cgroup_read_u64,
7423 .name = "memsw.limit_in_bytes",
7424 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7425 .write = mem_cgroup_write,
7426 .read_u64 = mem_cgroup_read_u64,
7429 .name = "memsw.failcnt",
7430 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7431 .write = mem_cgroup_reset,
7432 .read_u64 = mem_cgroup_read_u64,
7434 { }, /* terminate */
7438 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7439 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7440 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7441 * boot parameter. This may result in premature OOPS inside
7442 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7444 static int __init mem_cgroup_swap_init(void)
7446 /* No memory control -> no swap control */
7447 if (mem_cgroup_disabled())
7448 cgroup_memory_noswap = true;
7450 if (cgroup_memory_noswap)
7453 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7454 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7458 core_initcall(mem_cgroup_swap_init);
7460 #endif /* CONFIG_MEMCG_SWAP */