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
24 * Per memcg lru locking
25 * Copyright (C) 2020 Alibaba, Inc, Alex Shi
28 #include <linux/page_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
31 #include <linux/pagewalk.h>
32 #include <linux/sched/mm.h>
33 #include <linux/shmem_fs.h>
34 #include <linux/hugetlb.h>
35 #include <linux/pagemap.h>
36 #include <linux/vm_event_item.h>
37 #include <linux/smp.h>
38 #include <linux/page-flags.h>
39 #include <linux/backing-dev.h>
40 #include <linux/bit_spinlock.h>
41 #include <linux/rcupdate.h>
42 #include <linux/limits.h>
43 #include <linux/export.h>
44 #include <linux/mutex.h>
45 #include <linux/rbtree.h>
46 #include <linux/slab.h>
47 #include <linux/swap.h>
48 #include <linux/swapops.h>
49 #include <linux/spinlock.h>
50 #include <linux/eventfd.h>
51 #include <linux/poll.h>
52 #include <linux/sort.h>
54 #include <linux/seq_file.h>
55 #include <linux/vmpressure.h>
56 #include <linux/mm_inline.h>
57 #include <linux/swap_cgroup.h>
58 #include <linux/cpu.h>
59 #include <linux/oom.h>
60 #include <linux/lockdep.h>
61 #include <linux/file.h>
62 #include <linux/tracehook.h>
63 #include <linux/psi.h>
64 #include <linux/seq_buf.h>
70 #include <linux/uaccess.h>
72 #include <trace/events/vmscan.h>
74 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
75 EXPORT_SYMBOL(memory_cgrp_subsys);
77 struct mem_cgroup *root_mem_cgroup __read_mostly;
79 /* Active memory cgroup to use from an interrupt context */
80 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
82 /* Socket memory accounting disabled? */
83 static bool cgroup_memory_nosocket;
85 /* Kernel memory accounting disabled? */
86 static bool cgroup_memory_nokmem;
88 /* Whether the swap controller is active */
89 #ifdef CONFIG_MEMCG_SWAP
90 bool cgroup_memory_noswap __read_mostly;
92 #define cgroup_memory_noswap 1
95 #ifdef CONFIG_CGROUP_WRITEBACK
96 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
99 /* Whether legacy memory+swap accounting is active */
100 static bool do_memsw_account(void)
102 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
105 #define THRESHOLDS_EVENTS_TARGET 128
106 #define SOFTLIMIT_EVENTS_TARGET 1024
109 * Cgroups above their limits are maintained in a RB-Tree, independent of
110 * their hierarchy representation
113 struct mem_cgroup_tree_per_node {
114 struct rb_root rb_root;
115 struct rb_node *rb_rightmost;
119 struct mem_cgroup_tree {
120 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
123 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
126 struct mem_cgroup_eventfd_list {
127 struct list_head list;
128 struct eventfd_ctx *eventfd;
132 * cgroup_event represents events which userspace want to receive.
134 struct mem_cgroup_event {
136 * memcg which the event belongs to.
138 struct mem_cgroup *memcg;
140 * eventfd to signal userspace about the event.
142 struct eventfd_ctx *eventfd;
144 * Each of these stored in a list by the cgroup.
146 struct list_head list;
148 * register_event() callback will be used to add new userspace
149 * waiter for changes related to this event. Use eventfd_signal()
150 * on eventfd to send notification to userspace.
152 int (*register_event)(struct mem_cgroup *memcg,
153 struct eventfd_ctx *eventfd, const char *args);
155 * unregister_event() callback will be called when userspace closes
156 * the eventfd or on cgroup removing. This callback must be set,
157 * if you want provide notification functionality.
159 void (*unregister_event)(struct mem_cgroup *memcg,
160 struct eventfd_ctx *eventfd);
162 * All fields below needed to unregister event when
163 * userspace closes eventfd.
166 wait_queue_head_t *wqh;
167 wait_queue_entry_t wait;
168 struct work_struct remove;
171 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
172 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
174 /* Stuffs for move charges at task migration. */
176 * Types of charges to be moved.
178 #define MOVE_ANON 0x1U
179 #define MOVE_FILE 0x2U
180 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
182 /* "mc" and its members are protected by cgroup_mutex */
183 static struct move_charge_struct {
184 spinlock_t lock; /* for from, to */
185 struct mm_struct *mm;
186 struct mem_cgroup *from;
187 struct mem_cgroup *to;
189 unsigned long precharge;
190 unsigned long moved_charge;
191 unsigned long moved_swap;
192 struct task_struct *moving_task; /* a task moving charges */
193 wait_queue_head_t waitq; /* a waitq for other context */
195 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
196 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
200 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
201 * limit reclaim to prevent infinite loops, if they ever occur.
203 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
204 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
206 /* for encoding cft->private value on file */
215 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
216 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
217 #define MEMFILE_ATTR(val) ((val) & 0xffff)
218 /* Used for OOM nofiier */
219 #define OOM_CONTROL (0)
222 * Iteration constructs for visiting all cgroups (under a tree). If
223 * loops are exited prematurely (break), mem_cgroup_iter_break() must
224 * be used for reference counting.
226 #define for_each_mem_cgroup_tree(iter, root) \
227 for (iter = mem_cgroup_iter(root, NULL, NULL); \
229 iter = mem_cgroup_iter(root, iter, NULL))
231 #define for_each_mem_cgroup(iter) \
232 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
234 iter = mem_cgroup_iter(NULL, iter, NULL))
236 static inline bool should_force_charge(void)
238 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
239 (current->flags & PF_EXITING);
242 /* Some nice accessors for the vmpressure. */
243 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
246 memcg = root_mem_cgroup;
247 return &memcg->vmpressure;
250 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
252 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
255 #ifdef CONFIG_MEMCG_KMEM
256 extern spinlock_t css_set_lock;
258 static int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
259 unsigned int nr_pages);
260 static void __memcg_kmem_uncharge(struct mem_cgroup *memcg,
261 unsigned int nr_pages);
263 static void obj_cgroup_release(struct percpu_ref *ref)
265 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
266 struct mem_cgroup *memcg;
267 unsigned int nr_bytes;
268 unsigned int nr_pages;
272 * At this point all allocated objects are freed, and
273 * objcg->nr_charged_bytes can't have an arbitrary byte value.
274 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
276 * The following sequence can lead to it:
277 * 1) CPU0: objcg == stock->cached_objcg
278 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
279 * PAGE_SIZE bytes are charged
280 * 3) CPU1: a process from another memcg is allocating something,
281 * the stock if flushed,
282 * objcg->nr_charged_bytes = PAGE_SIZE - 92
283 * 5) CPU0: we do release this object,
284 * 92 bytes are added to stock->nr_bytes
285 * 6) CPU0: stock is flushed,
286 * 92 bytes are added to objcg->nr_charged_bytes
288 * In the result, nr_charged_bytes == PAGE_SIZE.
289 * This page will be uncharged in obj_cgroup_release().
291 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
292 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
293 nr_pages = nr_bytes >> PAGE_SHIFT;
295 spin_lock_irqsave(&css_set_lock, flags);
296 memcg = obj_cgroup_memcg(objcg);
298 __memcg_kmem_uncharge(memcg, nr_pages);
299 list_del(&objcg->list);
300 mem_cgroup_put(memcg);
301 spin_unlock_irqrestore(&css_set_lock, flags);
303 percpu_ref_exit(ref);
304 kfree_rcu(objcg, rcu);
307 static struct obj_cgroup *obj_cgroup_alloc(void)
309 struct obj_cgroup *objcg;
312 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
316 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
322 INIT_LIST_HEAD(&objcg->list);
326 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
327 struct mem_cgroup *parent)
329 struct obj_cgroup *objcg, *iter;
331 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
333 spin_lock_irq(&css_set_lock);
335 /* Move active objcg to the parent's list */
336 xchg(&objcg->memcg, parent);
337 css_get(&parent->css);
338 list_add(&objcg->list, &parent->objcg_list);
340 /* Move already reparented objcgs to the parent's list */
341 list_for_each_entry(iter, &memcg->objcg_list, list) {
342 css_get(&parent->css);
343 xchg(&iter->memcg, parent);
344 css_put(&memcg->css);
346 list_splice(&memcg->objcg_list, &parent->objcg_list);
348 spin_unlock_irq(&css_set_lock);
350 percpu_ref_kill(&objcg->refcnt);
354 * This will be used as a shrinker list's index.
355 * The main reason for not using cgroup id for this:
356 * this works better in sparse environments, where we have a lot of memcgs,
357 * but only a few kmem-limited. Or also, if we have, for instance, 200
358 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
359 * 200 entry array for that.
361 * The current size of the caches array is stored in memcg_nr_cache_ids. It
362 * will double each time we have to increase it.
364 static DEFINE_IDA(memcg_cache_ida);
365 int memcg_nr_cache_ids;
367 /* Protects memcg_nr_cache_ids */
368 static DECLARE_RWSEM(memcg_cache_ids_sem);
370 void memcg_get_cache_ids(void)
372 down_read(&memcg_cache_ids_sem);
375 void memcg_put_cache_ids(void)
377 up_read(&memcg_cache_ids_sem);
381 * MIN_SIZE is different than 1, because we would like to avoid going through
382 * the alloc/free process all the time. In a small machine, 4 kmem-limited
383 * cgroups is a reasonable guess. In the future, it could be a parameter or
384 * tunable, but that is strictly not necessary.
386 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
387 * this constant directly from cgroup, but it is understandable that this is
388 * better kept as an internal representation in cgroup.c. In any case, the
389 * cgrp_id space is not getting any smaller, and we don't have to necessarily
390 * increase ours as well if it increases.
392 #define MEMCG_CACHES_MIN_SIZE 4
393 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
396 * A lot of the calls to the cache allocation functions are expected to be
397 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
398 * conditional to this static branch, we'll have to allow modules that does
399 * kmem_cache_alloc and the such to see this symbol as well
401 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
402 EXPORT_SYMBOL(memcg_kmem_enabled_key);
405 static int memcg_shrinker_map_size;
406 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
408 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
410 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
413 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
414 int size, int old_size)
416 struct memcg_shrinker_map *new, *old;
419 lockdep_assert_held(&memcg_shrinker_map_mutex);
422 old = rcu_dereference_protected(
423 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
424 /* Not yet online memcg */
428 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
432 /* Set all old bits, clear all new bits */
433 memset(new->map, (int)0xff, old_size);
434 memset((void *)new->map + old_size, 0, size - old_size);
436 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
437 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
443 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
445 struct mem_cgroup_per_node *pn;
446 struct memcg_shrinker_map *map;
449 if (mem_cgroup_is_root(memcg))
453 pn = mem_cgroup_nodeinfo(memcg, nid);
454 map = rcu_dereference_protected(pn->shrinker_map, true);
456 rcu_assign_pointer(pn->shrinker_map, NULL);
460 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
462 struct memcg_shrinker_map *map;
463 int nid, size, ret = 0;
465 if (mem_cgroup_is_root(memcg))
468 mutex_lock(&memcg_shrinker_map_mutex);
469 size = memcg_shrinker_map_size;
471 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
473 memcg_free_shrinker_maps(memcg);
477 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
479 mutex_unlock(&memcg_shrinker_map_mutex);
484 int memcg_expand_shrinker_maps(int new_id)
486 int size, old_size, ret = 0;
487 struct mem_cgroup *memcg;
489 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
490 old_size = memcg_shrinker_map_size;
491 if (size <= old_size)
494 mutex_lock(&memcg_shrinker_map_mutex);
495 if (!root_mem_cgroup)
498 for_each_mem_cgroup(memcg) {
499 if (mem_cgroup_is_root(memcg))
501 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
503 mem_cgroup_iter_break(NULL, memcg);
509 memcg_shrinker_map_size = size;
510 mutex_unlock(&memcg_shrinker_map_mutex);
514 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
516 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
517 struct memcg_shrinker_map *map;
520 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
521 /* Pairs with smp mb in shrink_slab() */
522 smp_mb__before_atomic();
523 set_bit(shrinker_id, map->map);
529 * mem_cgroup_css_from_page - css of the memcg associated with a page
530 * @page: page of interest
532 * If memcg is bound to the default hierarchy, css of the memcg associated
533 * with @page is returned. The returned css remains associated with @page
534 * until it is released.
536 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
539 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
541 struct mem_cgroup *memcg;
543 memcg = page_memcg(page);
545 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
546 memcg = root_mem_cgroup;
552 * page_cgroup_ino - return inode number of the memcg a page is charged to
555 * Look up the closest online ancestor of the memory cgroup @page is charged to
556 * and return its inode number or 0 if @page is not charged to any cgroup. It
557 * is safe to call this function without holding a reference to @page.
559 * Note, this function is inherently racy, because there is nothing to prevent
560 * the cgroup inode from getting torn down and potentially reallocated a moment
561 * after page_cgroup_ino() returns, so it only should be used by callers that
562 * do not care (such as procfs interfaces).
564 ino_t page_cgroup_ino(struct page *page)
566 struct mem_cgroup *memcg;
567 unsigned long ino = 0;
570 memcg = page_memcg_check(page);
572 while (memcg && !(memcg->css.flags & CSS_ONLINE))
573 memcg = parent_mem_cgroup(memcg);
575 ino = cgroup_ino(memcg->css.cgroup);
580 static struct mem_cgroup_per_node *
581 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
583 int nid = page_to_nid(page);
585 return memcg->nodeinfo[nid];
588 static struct mem_cgroup_tree_per_node *
589 soft_limit_tree_node(int nid)
591 return soft_limit_tree.rb_tree_per_node[nid];
594 static struct mem_cgroup_tree_per_node *
595 soft_limit_tree_from_page(struct page *page)
597 int nid = page_to_nid(page);
599 return soft_limit_tree.rb_tree_per_node[nid];
602 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
603 struct mem_cgroup_tree_per_node *mctz,
604 unsigned long new_usage_in_excess)
606 struct rb_node **p = &mctz->rb_root.rb_node;
607 struct rb_node *parent = NULL;
608 struct mem_cgroup_per_node *mz_node;
609 bool rightmost = true;
614 mz->usage_in_excess = new_usage_in_excess;
615 if (!mz->usage_in_excess)
619 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
621 if (mz->usage_in_excess < mz_node->usage_in_excess) {
630 mctz->rb_rightmost = &mz->tree_node;
632 rb_link_node(&mz->tree_node, parent, p);
633 rb_insert_color(&mz->tree_node, &mctz->rb_root);
637 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
638 struct mem_cgroup_tree_per_node *mctz)
643 if (&mz->tree_node == mctz->rb_rightmost)
644 mctz->rb_rightmost = rb_prev(&mz->tree_node);
646 rb_erase(&mz->tree_node, &mctz->rb_root);
650 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
651 struct mem_cgroup_tree_per_node *mctz)
655 spin_lock_irqsave(&mctz->lock, flags);
656 __mem_cgroup_remove_exceeded(mz, mctz);
657 spin_unlock_irqrestore(&mctz->lock, flags);
660 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
662 unsigned long nr_pages = page_counter_read(&memcg->memory);
663 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
664 unsigned long excess = 0;
666 if (nr_pages > soft_limit)
667 excess = nr_pages - soft_limit;
672 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
674 unsigned long excess;
675 struct mem_cgroup_per_node *mz;
676 struct mem_cgroup_tree_per_node *mctz;
678 mctz = soft_limit_tree_from_page(page);
682 * Necessary to update all ancestors when hierarchy is used.
683 * because their event counter is not touched.
685 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
686 mz = mem_cgroup_page_nodeinfo(memcg, page);
687 excess = soft_limit_excess(memcg);
689 * We have to update the tree if mz is on RB-tree or
690 * mem is over its softlimit.
692 if (excess || mz->on_tree) {
695 spin_lock_irqsave(&mctz->lock, flags);
696 /* if on-tree, remove it */
698 __mem_cgroup_remove_exceeded(mz, mctz);
700 * Insert again. mz->usage_in_excess will be updated.
701 * If excess is 0, no tree ops.
703 __mem_cgroup_insert_exceeded(mz, mctz, excess);
704 spin_unlock_irqrestore(&mctz->lock, flags);
709 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
711 struct mem_cgroup_tree_per_node *mctz;
712 struct mem_cgroup_per_node *mz;
716 mz = mem_cgroup_nodeinfo(memcg, nid);
717 mctz = soft_limit_tree_node(nid);
719 mem_cgroup_remove_exceeded(mz, mctz);
723 static struct mem_cgroup_per_node *
724 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
726 struct mem_cgroup_per_node *mz;
730 if (!mctz->rb_rightmost)
731 goto done; /* Nothing to reclaim from */
733 mz = rb_entry(mctz->rb_rightmost,
734 struct mem_cgroup_per_node, tree_node);
736 * Remove the node now but someone else can add it back,
737 * we will to add it back at the end of reclaim to its correct
738 * position in the tree.
740 __mem_cgroup_remove_exceeded(mz, mctz);
741 if (!soft_limit_excess(mz->memcg) ||
742 !css_tryget(&mz->memcg->css))
748 static struct mem_cgroup_per_node *
749 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
751 struct mem_cgroup_per_node *mz;
753 spin_lock_irq(&mctz->lock);
754 mz = __mem_cgroup_largest_soft_limit_node(mctz);
755 spin_unlock_irq(&mctz->lock);
760 * __mod_memcg_state - update cgroup memory statistics
761 * @memcg: the memory cgroup
762 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
763 * @val: delta to add to the counter, can be negative
765 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
767 long x, threshold = MEMCG_CHARGE_BATCH;
769 if (mem_cgroup_disabled())
772 if (memcg_stat_item_in_bytes(idx))
773 threshold <<= PAGE_SHIFT;
775 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
776 if (unlikely(abs(x) > threshold)) {
777 struct mem_cgroup *mi;
780 * Batch local counters to keep them in sync with
781 * the hierarchical ones.
783 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
784 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
785 atomic_long_add(x, &mi->vmstats[idx]);
788 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
791 static struct mem_cgroup_per_node *
792 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
794 struct mem_cgroup *parent;
796 parent = parent_mem_cgroup(pn->memcg);
799 return mem_cgroup_nodeinfo(parent, nid);
802 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
805 struct mem_cgroup_per_node *pn;
806 struct mem_cgroup *memcg;
807 long x, threshold = MEMCG_CHARGE_BATCH;
809 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
813 __mod_memcg_state(memcg, idx, val);
816 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
818 if (vmstat_item_in_bytes(idx))
819 threshold <<= PAGE_SHIFT;
821 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
822 if (unlikely(abs(x) > threshold)) {
823 pg_data_t *pgdat = lruvec_pgdat(lruvec);
824 struct mem_cgroup_per_node *pi;
826 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
827 atomic_long_add(x, &pi->lruvec_stat[idx]);
830 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
834 * __mod_lruvec_state - update lruvec memory statistics
835 * @lruvec: the lruvec
836 * @idx: the stat item
837 * @val: delta to add to the counter, can be negative
839 * The lruvec is the intersection of the NUMA node and a cgroup. This
840 * function updates the all three counters that are affected by a
841 * change of state at this level: per-node, per-cgroup, per-lruvec.
843 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
847 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
849 /* Update memcg and lruvec */
850 if (!mem_cgroup_disabled())
851 __mod_memcg_lruvec_state(lruvec, idx, val);
854 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
857 struct page *head = compound_head(page); /* rmap on tail pages */
858 struct mem_cgroup *memcg = page_memcg(head);
859 pg_data_t *pgdat = page_pgdat(page);
860 struct lruvec *lruvec;
862 /* Untracked pages have no memcg, no lruvec. Update only the node */
864 __mod_node_page_state(pgdat, idx, val);
868 lruvec = mem_cgroup_lruvec(memcg, pgdat);
869 __mod_lruvec_state(lruvec, idx, val);
871 EXPORT_SYMBOL(__mod_lruvec_page_state);
873 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
875 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
876 struct mem_cgroup *memcg;
877 struct lruvec *lruvec;
880 memcg = mem_cgroup_from_obj(p);
883 * Untracked pages have no memcg, no lruvec. Update only the
884 * node. If we reparent the slab objects to the root memcg,
885 * when we free the slab object, we need to update the per-memcg
886 * vmstats to keep it correct for the root memcg.
889 __mod_node_page_state(pgdat, idx, val);
891 lruvec = mem_cgroup_lruvec(memcg, pgdat);
892 __mod_lruvec_state(lruvec, idx, val);
898 * __count_memcg_events - account VM events in a cgroup
899 * @memcg: the memory cgroup
900 * @idx: the event item
901 * @count: the number of events that occured
903 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
908 if (mem_cgroup_disabled())
911 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
912 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
913 struct mem_cgroup *mi;
916 * Batch local counters to keep them in sync with
917 * the hierarchical ones.
919 __this_cpu_add(memcg->vmstats_local->events[idx], x);
920 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
921 atomic_long_add(x, &mi->vmevents[idx]);
924 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
927 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
929 return atomic_long_read(&memcg->vmevents[event]);
932 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
937 for_each_possible_cpu(cpu)
938 x += per_cpu(memcg->vmstats_local->events[event], cpu);
942 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
946 /* pagein of a big page is an event. So, ignore page size */
948 __count_memcg_events(memcg, PGPGIN, 1);
950 __count_memcg_events(memcg, PGPGOUT, 1);
951 nr_pages = -nr_pages; /* for event */
954 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
957 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
958 enum mem_cgroup_events_target target)
960 unsigned long val, next;
962 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
963 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
964 /* from time_after() in jiffies.h */
965 if ((long)(next - val) < 0) {
967 case MEM_CGROUP_TARGET_THRESH:
968 next = val + THRESHOLDS_EVENTS_TARGET;
970 case MEM_CGROUP_TARGET_SOFTLIMIT:
971 next = val + SOFTLIMIT_EVENTS_TARGET;
976 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
983 * Check events in order.
986 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
988 /* threshold event is triggered in finer grain than soft limit */
989 if (unlikely(mem_cgroup_event_ratelimit(memcg,
990 MEM_CGROUP_TARGET_THRESH))) {
993 do_softlimit = mem_cgroup_event_ratelimit(memcg,
994 MEM_CGROUP_TARGET_SOFTLIMIT);
995 mem_cgroup_threshold(memcg);
996 if (unlikely(do_softlimit))
997 mem_cgroup_update_tree(memcg, page);
1001 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1004 * mm_update_next_owner() may clear mm->owner to NULL
1005 * if it races with swapoff, page migration, etc.
1006 * So this can be called with p == NULL.
1011 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1013 EXPORT_SYMBOL(mem_cgroup_from_task);
1016 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1017 * @mm: mm from which memcg should be extracted. It can be NULL.
1019 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1020 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1023 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1025 struct mem_cgroup *memcg;
1027 if (mem_cgroup_disabled())
1033 * Page cache insertions can happen withou an
1034 * actual mm context, e.g. during disk probing
1035 * on boot, loopback IO, acct() writes etc.
1038 memcg = root_mem_cgroup;
1040 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1041 if (unlikely(!memcg))
1042 memcg = root_mem_cgroup;
1044 } while (!css_tryget(&memcg->css));
1048 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1050 static __always_inline struct mem_cgroup *active_memcg(void)
1053 return this_cpu_read(int_active_memcg);
1055 return current->active_memcg;
1058 static __always_inline struct mem_cgroup *get_active_memcg(void)
1060 struct mem_cgroup *memcg;
1063 memcg = active_memcg();
1064 /* remote memcg must hold a ref. */
1065 if (memcg && WARN_ON_ONCE(!css_tryget(&memcg->css)))
1066 memcg = root_mem_cgroup;
1072 static __always_inline bool memcg_kmem_bypass(void)
1074 /* Allow remote memcg charging from any context. */
1075 if (unlikely(active_memcg()))
1078 /* Memcg to charge can't be determined. */
1079 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
1086 * If active memcg is set, do not fallback to current->mm->memcg.
1088 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1090 if (memcg_kmem_bypass())
1093 if (unlikely(active_memcg()))
1094 return get_active_memcg();
1096 return get_mem_cgroup_from_mm(current->mm);
1100 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1101 * @root: hierarchy root
1102 * @prev: previously returned memcg, NULL on first invocation
1103 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1105 * Returns references to children of the hierarchy below @root, or
1106 * @root itself, or %NULL after a full round-trip.
1108 * Caller must pass the return value in @prev on subsequent
1109 * invocations for reference counting, or use mem_cgroup_iter_break()
1110 * to cancel a hierarchy walk before the round-trip is complete.
1112 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1113 * in the hierarchy among all concurrent reclaimers operating on the
1116 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1117 struct mem_cgroup *prev,
1118 struct mem_cgroup_reclaim_cookie *reclaim)
1120 struct mem_cgroup_reclaim_iter *iter;
1121 struct cgroup_subsys_state *css = NULL;
1122 struct mem_cgroup *memcg = NULL;
1123 struct mem_cgroup *pos = NULL;
1125 if (mem_cgroup_disabled())
1129 root = root_mem_cgroup;
1131 if (prev && !reclaim)
1137 struct mem_cgroup_per_node *mz;
1139 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1142 if (prev && reclaim->generation != iter->generation)
1146 pos = READ_ONCE(iter->position);
1147 if (!pos || css_tryget(&pos->css))
1150 * css reference reached zero, so iter->position will
1151 * be cleared by ->css_released. However, we should not
1152 * rely on this happening soon, because ->css_released
1153 * is called from a work queue, and by busy-waiting we
1154 * might block it. So we clear iter->position right
1157 (void)cmpxchg(&iter->position, pos, NULL);
1165 css = css_next_descendant_pre(css, &root->css);
1168 * Reclaimers share the hierarchy walk, and a
1169 * new one might jump in right at the end of
1170 * the hierarchy - make sure they see at least
1171 * one group and restart from the beginning.
1179 * Verify the css and acquire a reference. The root
1180 * is provided by the caller, so we know it's alive
1181 * and kicking, and don't take an extra reference.
1183 memcg = mem_cgroup_from_css(css);
1185 if (css == &root->css)
1188 if (css_tryget(css))
1196 * The position could have already been updated by a competing
1197 * thread, so check that the value hasn't changed since we read
1198 * it to avoid reclaiming from the same cgroup twice.
1200 (void)cmpxchg(&iter->position, pos, memcg);
1208 reclaim->generation = iter->generation;
1213 if (prev && prev != root)
1214 css_put(&prev->css);
1220 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1221 * @root: hierarchy root
1222 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1224 void mem_cgroup_iter_break(struct mem_cgroup *root,
1225 struct mem_cgroup *prev)
1228 root = root_mem_cgroup;
1229 if (prev && prev != root)
1230 css_put(&prev->css);
1233 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1234 struct mem_cgroup *dead_memcg)
1236 struct mem_cgroup_reclaim_iter *iter;
1237 struct mem_cgroup_per_node *mz;
1240 for_each_node(nid) {
1241 mz = mem_cgroup_nodeinfo(from, nid);
1243 cmpxchg(&iter->position, dead_memcg, NULL);
1247 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1249 struct mem_cgroup *memcg = dead_memcg;
1250 struct mem_cgroup *last;
1253 __invalidate_reclaim_iterators(memcg, dead_memcg);
1255 } while ((memcg = parent_mem_cgroup(memcg)));
1258 * When cgruop1 non-hierarchy mode is used,
1259 * parent_mem_cgroup() does not walk all the way up to the
1260 * cgroup root (root_mem_cgroup). So we have to handle
1261 * dead_memcg from cgroup root separately.
1263 if (last != root_mem_cgroup)
1264 __invalidate_reclaim_iterators(root_mem_cgroup,
1269 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1270 * @memcg: hierarchy root
1271 * @fn: function to call for each task
1272 * @arg: argument passed to @fn
1274 * This function iterates over tasks attached to @memcg or to any of its
1275 * descendants and calls @fn for each task. If @fn returns a non-zero
1276 * value, the function breaks the iteration loop and returns the value.
1277 * Otherwise, it will iterate over all tasks and return 0.
1279 * This function must not be called for the root memory cgroup.
1281 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1282 int (*fn)(struct task_struct *, void *), void *arg)
1284 struct mem_cgroup *iter;
1287 BUG_ON(memcg == root_mem_cgroup);
1289 for_each_mem_cgroup_tree(iter, memcg) {
1290 struct css_task_iter it;
1291 struct task_struct *task;
1293 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1294 while (!ret && (task = css_task_iter_next(&it)))
1295 ret = fn(task, arg);
1296 css_task_iter_end(&it);
1298 mem_cgroup_iter_break(memcg, iter);
1305 #ifdef CONFIG_DEBUG_VM
1306 void lruvec_memcg_debug(struct lruvec *lruvec, struct page *page)
1308 struct mem_cgroup *memcg;
1310 if (mem_cgroup_disabled())
1313 memcg = page_memcg(page);
1316 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != root_mem_cgroup, page);
1318 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != memcg, page);
1323 * lock_page_lruvec - lock and return lruvec for a given page.
1326 * These functions are safe to use under any of the following conditions:
1329 * - lock_page_memcg()
1330 * - page->_refcount is zero
1332 struct lruvec *lock_page_lruvec(struct page *page)
1334 struct lruvec *lruvec;
1335 struct pglist_data *pgdat = page_pgdat(page);
1337 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1338 spin_lock(&lruvec->lru_lock);
1340 lruvec_memcg_debug(lruvec, page);
1345 struct lruvec *lock_page_lruvec_irq(struct page *page)
1347 struct lruvec *lruvec;
1348 struct pglist_data *pgdat = page_pgdat(page);
1350 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1351 spin_lock_irq(&lruvec->lru_lock);
1353 lruvec_memcg_debug(lruvec, page);
1358 struct lruvec *lock_page_lruvec_irqsave(struct page *page, unsigned long *flags)
1360 struct lruvec *lruvec;
1361 struct pglist_data *pgdat = page_pgdat(page);
1363 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1364 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1366 lruvec_memcg_debug(lruvec, page);
1372 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1373 * @lruvec: mem_cgroup per zone lru vector
1374 * @lru: index of lru list the page is sitting on
1375 * @zid: zone id of the accounted pages
1376 * @nr_pages: positive when adding or negative when removing
1378 * This function must be called under lru_lock, just before a page is added
1379 * to or just after a page is removed from an lru list (that ordering being
1380 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1382 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1383 int zid, int nr_pages)
1385 struct mem_cgroup_per_node *mz;
1386 unsigned long *lru_size;
1389 if (mem_cgroup_disabled())
1392 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1393 lru_size = &mz->lru_zone_size[zid][lru];
1396 *lru_size += nr_pages;
1399 if (WARN_ONCE(size < 0,
1400 "%s(%p, %d, %d): lru_size %ld\n",
1401 __func__, lruvec, lru, nr_pages, size)) {
1407 *lru_size += nr_pages;
1411 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1412 * @memcg: the memory cgroup
1414 * Returns the maximum amount of memory @mem can be charged with, in
1417 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1419 unsigned long margin = 0;
1420 unsigned long count;
1421 unsigned long limit;
1423 count = page_counter_read(&memcg->memory);
1424 limit = READ_ONCE(memcg->memory.max);
1426 margin = limit - count;
1428 if (do_memsw_account()) {
1429 count = page_counter_read(&memcg->memsw);
1430 limit = READ_ONCE(memcg->memsw.max);
1432 margin = min(margin, limit - count);
1441 * A routine for checking "mem" is under move_account() or not.
1443 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1444 * moving cgroups. This is for waiting at high-memory pressure
1447 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1449 struct mem_cgroup *from;
1450 struct mem_cgroup *to;
1453 * Unlike task_move routines, we access mc.to, mc.from not under
1454 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1456 spin_lock(&mc.lock);
1462 ret = mem_cgroup_is_descendant(from, memcg) ||
1463 mem_cgroup_is_descendant(to, memcg);
1465 spin_unlock(&mc.lock);
1469 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1471 if (mc.moving_task && current != mc.moving_task) {
1472 if (mem_cgroup_under_move(memcg)) {
1474 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1475 /* moving charge context might have finished. */
1478 finish_wait(&mc.waitq, &wait);
1485 struct memory_stat {
1490 static const struct memory_stat memory_stats[] = {
1491 { "anon", NR_ANON_MAPPED },
1492 { "file", NR_FILE_PAGES },
1493 { "kernel_stack", NR_KERNEL_STACK_KB },
1494 { "pagetables", NR_PAGETABLE },
1495 { "percpu", MEMCG_PERCPU_B },
1496 { "sock", MEMCG_SOCK },
1497 { "shmem", NR_SHMEM },
1498 { "file_mapped", NR_FILE_MAPPED },
1499 { "file_dirty", NR_FILE_DIRTY },
1500 { "file_writeback", NR_WRITEBACK },
1502 { "swapcached", NR_SWAPCACHE },
1504 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1505 { "anon_thp", NR_ANON_THPS },
1506 { "file_thp", NR_FILE_THPS },
1507 { "shmem_thp", NR_SHMEM_THPS },
1509 { "inactive_anon", NR_INACTIVE_ANON },
1510 { "active_anon", NR_ACTIVE_ANON },
1511 { "inactive_file", NR_INACTIVE_FILE },
1512 { "active_file", NR_ACTIVE_FILE },
1513 { "unevictable", NR_UNEVICTABLE },
1514 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B },
1515 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B },
1517 /* The memory events */
1518 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON },
1519 { "workingset_refault_file", WORKINGSET_REFAULT_FILE },
1520 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON },
1521 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE },
1522 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON },
1523 { "workingset_restore_file", WORKINGSET_RESTORE_FILE },
1524 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM },
1527 /* Translate stat items to the correct unit for memory.stat output */
1528 static int memcg_page_state_unit(int item)
1531 case MEMCG_PERCPU_B:
1532 case NR_SLAB_RECLAIMABLE_B:
1533 case NR_SLAB_UNRECLAIMABLE_B:
1534 case WORKINGSET_REFAULT_ANON:
1535 case WORKINGSET_REFAULT_FILE:
1536 case WORKINGSET_ACTIVATE_ANON:
1537 case WORKINGSET_ACTIVATE_FILE:
1538 case WORKINGSET_RESTORE_ANON:
1539 case WORKINGSET_RESTORE_FILE:
1540 case WORKINGSET_NODERECLAIM:
1542 case NR_KERNEL_STACK_KB:
1549 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1552 return memcg_page_state(memcg, item) * memcg_page_state_unit(item);
1555 static char *memory_stat_format(struct mem_cgroup *memcg)
1560 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1565 * Provide statistics on the state of the memory subsystem as
1566 * well as cumulative event counters that show past behavior.
1568 * This list is ordered following a combination of these gradients:
1569 * 1) generic big picture -> specifics and details
1570 * 2) reflecting userspace activity -> reflecting kernel heuristics
1572 * Current memory state:
1575 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1578 size = memcg_page_state_output(memcg, memory_stats[i].idx);
1579 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1581 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1582 size += memcg_page_state_output(memcg,
1583 NR_SLAB_RECLAIMABLE_B);
1584 seq_buf_printf(&s, "slab %llu\n", size);
1588 /* Accumulated memory events */
1590 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1591 memcg_events(memcg, PGFAULT));
1592 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1593 memcg_events(memcg, PGMAJFAULT));
1594 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1595 memcg_events(memcg, PGREFILL));
1596 seq_buf_printf(&s, "pgscan %lu\n",
1597 memcg_events(memcg, PGSCAN_KSWAPD) +
1598 memcg_events(memcg, PGSCAN_DIRECT));
1599 seq_buf_printf(&s, "pgsteal %lu\n",
1600 memcg_events(memcg, PGSTEAL_KSWAPD) +
1601 memcg_events(memcg, PGSTEAL_DIRECT));
1602 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1603 memcg_events(memcg, PGACTIVATE));
1604 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1605 memcg_events(memcg, PGDEACTIVATE));
1606 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1607 memcg_events(memcg, PGLAZYFREE));
1608 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1609 memcg_events(memcg, PGLAZYFREED));
1611 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1612 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1613 memcg_events(memcg, THP_FAULT_ALLOC));
1614 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1615 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1616 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1618 /* The above should easily fit into one page */
1619 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1624 #define K(x) ((x) << (PAGE_SHIFT-10))
1626 * mem_cgroup_print_oom_context: Print OOM information relevant to
1627 * memory controller.
1628 * @memcg: The memory cgroup that went over limit
1629 * @p: Task that is going to be killed
1631 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1634 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1639 pr_cont(",oom_memcg=");
1640 pr_cont_cgroup_path(memcg->css.cgroup);
1642 pr_cont(",global_oom");
1644 pr_cont(",task_memcg=");
1645 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1651 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1652 * memory controller.
1653 * @memcg: The memory cgroup that went over limit
1655 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1659 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1660 K((u64)page_counter_read(&memcg->memory)),
1661 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1662 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1663 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1664 K((u64)page_counter_read(&memcg->swap)),
1665 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1667 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1668 K((u64)page_counter_read(&memcg->memsw)),
1669 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1670 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1671 K((u64)page_counter_read(&memcg->kmem)),
1672 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1675 pr_info("Memory cgroup stats for ");
1676 pr_cont_cgroup_path(memcg->css.cgroup);
1678 buf = memory_stat_format(memcg);
1686 * Return the memory (and swap, if configured) limit for a memcg.
1688 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1690 unsigned long max = READ_ONCE(memcg->memory.max);
1692 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1693 if (mem_cgroup_swappiness(memcg))
1694 max += min(READ_ONCE(memcg->swap.max),
1695 (unsigned long)total_swap_pages);
1697 if (mem_cgroup_swappiness(memcg)) {
1698 /* Calculate swap excess capacity from memsw limit */
1699 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1701 max += min(swap, (unsigned long)total_swap_pages);
1707 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1709 return page_counter_read(&memcg->memory);
1712 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1715 struct oom_control oc = {
1719 .gfp_mask = gfp_mask,
1724 if (mutex_lock_killable(&oom_lock))
1727 if (mem_cgroup_margin(memcg) >= (1 << order))
1731 * A few threads which were not waiting at mutex_lock_killable() can
1732 * fail to bail out. Therefore, check again after holding oom_lock.
1734 ret = should_force_charge() || out_of_memory(&oc);
1737 mutex_unlock(&oom_lock);
1741 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1744 unsigned long *total_scanned)
1746 struct mem_cgroup *victim = NULL;
1749 unsigned long excess;
1750 unsigned long nr_scanned;
1751 struct mem_cgroup_reclaim_cookie reclaim = {
1755 excess = soft_limit_excess(root_memcg);
1758 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1763 * If we have not been able to reclaim
1764 * anything, it might because there are
1765 * no reclaimable pages under this hierarchy
1770 * We want to do more targeted reclaim.
1771 * excess >> 2 is not to excessive so as to
1772 * reclaim too much, nor too less that we keep
1773 * coming back to reclaim from this cgroup
1775 if (total >= (excess >> 2) ||
1776 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1781 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1782 pgdat, &nr_scanned);
1783 *total_scanned += nr_scanned;
1784 if (!soft_limit_excess(root_memcg))
1787 mem_cgroup_iter_break(root_memcg, victim);
1791 #ifdef CONFIG_LOCKDEP
1792 static struct lockdep_map memcg_oom_lock_dep_map = {
1793 .name = "memcg_oom_lock",
1797 static DEFINE_SPINLOCK(memcg_oom_lock);
1800 * Check OOM-Killer is already running under our hierarchy.
1801 * If someone is running, return false.
1803 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1805 struct mem_cgroup *iter, *failed = NULL;
1807 spin_lock(&memcg_oom_lock);
1809 for_each_mem_cgroup_tree(iter, memcg) {
1810 if (iter->oom_lock) {
1812 * this subtree of our hierarchy is already locked
1813 * so we cannot give a lock.
1816 mem_cgroup_iter_break(memcg, iter);
1819 iter->oom_lock = true;
1824 * OK, we failed to lock the whole subtree so we have
1825 * to clean up what we set up to the failing subtree
1827 for_each_mem_cgroup_tree(iter, memcg) {
1828 if (iter == failed) {
1829 mem_cgroup_iter_break(memcg, iter);
1832 iter->oom_lock = false;
1835 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1837 spin_unlock(&memcg_oom_lock);
1842 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1844 struct mem_cgroup *iter;
1846 spin_lock(&memcg_oom_lock);
1847 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1848 for_each_mem_cgroup_tree(iter, memcg)
1849 iter->oom_lock = false;
1850 spin_unlock(&memcg_oom_lock);
1853 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1855 struct mem_cgroup *iter;
1857 spin_lock(&memcg_oom_lock);
1858 for_each_mem_cgroup_tree(iter, memcg)
1860 spin_unlock(&memcg_oom_lock);
1863 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1865 struct mem_cgroup *iter;
1868 * Be careful about under_oom underflows becase a child memcg
1869 * could have been added after mem_cgroup_mark_under_oom.
1871 spin_lock(&memcg_oom_lock);
1872 for_each_mem_cgroup_tree(iter, memcg)
1873 if (iter->under_oom > 0)
1875 spin_unlock(&memcg_oom_lock);
1878 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1880 struct oom_wait_info {
1881 struct mem_cgroup *memcg;
1882 wait_queue_entry_t wait;
1885 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1886 unsigned mode, int sync, void *arg)
1888 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1889 struct mem_cgroup *oom_wait_memcg;
1890 struct oom_wait_info *oom_wait_info;
1892 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1893 oom_wait_memcg = oom_wait_info->memcg;
1895 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1896 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1898 return autoremove_wake_function(wait, mode, sync, arg);
1901 static void memcg_oom_recover(struct mem_cgroup *memcg)
1904 * For the following lockless ->under_oom test, the only required
1905 * guarantee is that it must see the state asserted by an OOM when
1906 * this function is called as a result of userland actions
1907 * triggered by the notification of the OOM. This is trivially
1908 * achieved by invoking mem_cgroup_mark_under_oom() before
1909 * triggering notification.
1911 if (memcg && memcg->under_oom)
1912 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1922 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1924 enum oom_status ret;
1927 if (order > PAGE_ALLOC_COSTLY_ORDER)
1930 memcg_memory_event(memcg, MEMCG_OOM);
1933 * We are in the middle of the charge context here, so we
1934 * don't want to block when potentially sitting on a callstack
1935 * that holds all kinds of filesystem and mm locks.
1937 * cgroup1 allows disabling the OOM killer and waiting for outside
1938 * handling until the charge can succeed; remember the context and put
1939 * the task to sleep at the end of the page fault when all locks are
1942 * On the other hand, in-kernel OOM killer allows for an async victim
1943 * memory reclaim (oom_reaper) and that means that we are not solely
1944 * relying on the oom victim to make a forward progress and we can
1945 * invoke the oom killer here.
1947 * Please note that mem_cgroup_out_of_memory might fail to find a
1948 * victim and then we have to bail out from the charge path.
1950 if (memcg->oom_kill_disable) {
1951 if (!current->in_user_fault)
1953 css_get(&memcg->css);
1954 current->memcg_in_oom = memcg;
1955 current->memcg_oom_gfp_mask = mask;
1956 current->memcg_oom_order = order;
1961 mem_cgroup_mark_under_oom(memcg);
1963 locked = mem_cgroup_oom_trylock(memcg);
1966 mem_cgroup_oom_notify(memcg);
1968 mem_cgroup_unmark_under_oom(memcg);
1969 if (mem_cgroup_out_of_memory(memcg, mask, order))
1975 mem_cgroup_oom_unlock(memcg);
1981 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1982 * @handle: actually kill/wait or just clean up the OOM state
1984 * This has to be called at the end of a page fault if the memcg OOM
1985 * handler was enabled.
1987 * Memcg supports userspace OOM handling where failed allocations must
1988 * sleep on a waitqueue until the userspace task resolves the
1989 * situation. Sleeping directly in the charge context with all kinds
1990 * of locks held is not a good idea, instead we remember an OOM state
1991 * in the task and mem_cgroup_oom_synchronize() has to be called at
1992 * the end of the page fault to complete the OOM handling.
1994 * Returns %true if an ongoing memcg OOM situation was detected and
1995 * completed, %false otherwise.
1997 bool mem_cgroup_oom_synchronize(bool handle)
1999 struct mem_cgroup *memcg = current->memcg_in_oom;
2000 struct oom_wait_info owait;
2003 /* OOM is global, do not handle */
2010 owait.memcg = memcg;
2011 owait.wait.flags = 0;
2012 owait.wait.func = memcg_oom_wake_function;
2013 owait.wait.private = current;
2014 INIT_LIST_HEAD(&owait.wait.entry);
2016 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2017 mem_cgroup_mark_under_oom(memcg);
2019 locked = mem_cgroup_oom_trylock(memcg);
2022 mem_cgroup_oom_notify(memcg);
2024 if (locked && !memcg->oom_kill_disable) {
2025 mem_cgroup_unmark_under_oom(memcg);
2026 finish_wait(&memcg_oom_waitq, &owait.wait);
2027 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
2028 current->memcg_oom_order);
2031 mem_cgroup_unmark_under_oom(memcg);
2032 finish_wait(&memcg_oom_waitq, &owait.wait);
2036 mem_cgroup_oom_unlock(memcg);
2038 * There is no guarantee that an OOM-lock contender
2039 * sees the wakeups triggered by the OOM kill
2040 * uncharges. Wake any sleepers explicitely.
2042 memcg_oom_recover(memcg);
2045 current->memcg_in_oom = NULL;
2046 css_put(&memcg->css);
2051 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2052 * @victim: task to be killed by the OOM killer
2053 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2055 * Returns a pointer to a memory cgroup, which has to be cleaned up
2056 * by killing all belonging OOM-killable tasks.
2058 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2060 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2061 struct mem_cgroup *oom_domain)
2063 struct mem_cgroup *oom_group = NULL;
2064 struct mem_cgroup *memcg;
2066 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2070 oom_domain = root_mem_cgroup;
2074 memcg = mem_cgroup_from_task(victim);
2075 if (memcg == root_mem_cgroup)
2079 * If the victim task has been asynchronously moved to a different
2080 * memory cgroup, we might end up killing tasks outside oom_domain.
2081 * In this case it's better to ignore memory.group.oom.
2083 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2087 * Traverse the memory cgroup hierarchy from the victim task's
2088 * cgroup up to the OOMing cgroup (or root) to find the
2089 * highest-level memory cgroup with oom.group set.
2091 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2092 if (memcg->oom_group)
2095 if (memcg == oom_domain)
2100 css_get(&oom_group->css);
2107 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2109 pr_info("Tasks in ");
2110 pr_cont_cgroup_path(memcg->css.cgroup);
2111 pr_cont(" are going to be killed due to memory.oom.group set\n");
2115 * lock_page_memcg - lock a page and memcg binding
2118 * This function protects unlocked LRU pages from being moved to
2121 * It ensures lifetime of the locked memcg. Caller is responsible
2122 * for the lifetime of the page.
2124 void lock_page_memcg(struct page *page)
2126 struct page *head = compound_head(page); /* rmap on tail pages */
2127 struct mem_cgroup *memcg;
2128 unsigned long flags;
2131 * The RCU lock is held throughout the transaction. The fast
2132 * path can get away without acquiring the memcg->move_lock
2133 * because page moving starts with an RCU grace period.
2137 if (mem_cgroup_disabled())
2140 memcg = page_memcg(head);
2141 if (unlikely(!memcg))
2144 #ifdef CONFIG_PROVE_LOCKING
2145 local_irq_save(flags);
2146 might_lock(&memcg->move_lock);
2147 local_irq_restore(flags);
2150 if (atomic_read(&memcg->moving_account) <= 0)
2153 spin_lock_irqsave(&memcg->move_lock, flags);
2154 if (memcg != page_memcg(head)) {
2155 spin_unlock_irqrestore(&memcg->move_lock, flags);
2160 * When charge migration first begins, we can have multiple
2161 * critical sections holding the fast-path RCU lock and one
2162 * holding the slowpath move_lock. Track the task who has the
2163 * move_lock for unlock_page_memcg().
2165 memcg->move_lock_task = current;
2166 memcg->move_lock_flags = flags;
2168 EXPORT_SYMBOL(lock_page_memcg);
2170 static void __unlock_page_memcg(struct mem_cgroup *memcg)
2172 if (memcg && memcg->move_lock_task == current) {
2173 unsigned long flags = memcg->move_lock_flags;
2175 memcg->move_lock_task = NULL;
2176 memcg->move_lock_flags = 0;
2178 spin_unlock_irqrestore(&memcg->move_lock, flags);
2185 * unlock_page_memcg - unlock a page and memcg binding
2188 void unlock_page_memcg(struct page *page)
2190 struct page *head = compound_head(page);
2192 __unlock_page_memcg(page_memcg(head));
2194 EXPORT_SYMBOL(unlock_page_memcg);
2196 struct memcg_stock_pcp {
2197 struct mem_cgroup *cached; /* this never be root cgroup */
2198 unsigned int nr_pages;
2200 #ifdef CONFIG_MEMCG_KMEM
2201 struct obj_cgroup *cached_objcg;
2202 unsigned int nr_bytes;
2205 struct work_struct work;
2206 unsigned long flags;
2207 #define FLUSHING_CACHED_CHARGE 0
2209 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2210 static DEFINE_MUTEX(percpu_charge_mutex);
2212 #ifdef CONFIG_MEMCG_KMEM
2213 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2214 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2215 struct mem_cgroup *root_memcg);
2218 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2221 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2222 struct mem_cgroup *root_memcg)
2229 * consume_stock: Try to consume stocked charge on this cpu.
2230 * @memcg: memcg to consume from.
2231 * @nr_pages: how many pages to charge.
2233 * The charges will only happen if @memcg matches the current cpu's memcg
2234 * stock, and at least @nr_pages are available in that stock. Failure to
2235 * service an allocation will refill the stock.
2237 * returns true if successful, false otherwise.
2239 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2241 struct memcg_stock_pcp *stock;
2242 unsigned long flags;
2245 if (nr_pages > MEMCG_CHARGE_BATCH)
2248 local_irq_save(flags);
2250 stock = this_cpu_ptr(&memcg_stock);
2251 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2252 stock->nr_pages -= nr_pages;
2256 local_irq_restore(flags);
2262 * Returns stocks cached in percpu and reset cached information.
2264 static void drain_stock(struct memcg_stock_pcp *stock)
2266 struct mem_cgroup *old = stock->cached;
2271 if (stock->nr_pages) {
2272 page_counter_uncharge(&old->memory, stock->nr_pages);
2273 if (do_memsw_account())
2274 page_counter_uncharge(&old->memsw, stock->nr_pages);
2275 stock->nr_pages = 0;
2279 stock->cached = NULL;
2282 static void drain_local_stock(struct work_struct *dummy)
2284 struct memcg_stock_pcp *stock;
2285 unsigned long flags;
2288 * The only protection from memory hotplug vs. drain_stock races is
2289 * that we always operate on local CPU stock here with IRQ disabled
2291 local_irq_save(flags);
2293 stock = this_cpu_ptr(&memcg_stock);
2294 drain_obj_stock(stock);
2296 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2298 local_irq_restore(flags);
2302 * Cache charges(val) to local per_cpu area.
2303 * This will be consumed by consume_stock() function, later.
2305 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2307 struct memcg_stock_pcp *stock;
2308 unsigned long flags;
2310 local_irq_save(flags);
2312 stock = this_cpu_ptr(&memcg_stock);
2313 if (stock->cached != memcg) { /* reset if necessary */
2315 css_get(&memcg->css);
2316 stock->cached = memcg;
2318 stock->nr_pages += nr_pages;
2320 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2323 local_irq_restore(flags);
2327 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2328 * of the hierarchy under it.
2330 static void drain_all_stock(struct mem_cgroup *root_memcg)
2334 /* If someone's already draining, avoid adding running more workers. */
2335 if (!mutex_trylock(&percpu_charge_mutex))
2338 * Notify other cpus that system-wide "drain" is running
2339 * We do not care about races with the cpu hotplug because cpu down
2340 * as well as workers from this path always operate on the local
2341 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2344 for_each_online_cpu(cpu) {
2345 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2346 struct mem_cgroup *memcg;
2350 memcg = stock->cached;
2351 if (memcg && stock->nr_pages &&
2352 mem_cgroup_is_descendant(memcg, root_memcg))
2354 if (obj_stock_flush_required(stock, root_memcg))
2359 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2361 drain_local_stock(&stock->work);
2363 schedule_work_on(cpu, &stock->work);
2367 mutex_unlock(&percpu_charge_mutex);
2370 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2372 struct memcg_stock_pcp *stock;
2373 struct mem_cgroup *memcg, *mi;
2375 stock = &per_cpu(memcg_stock, cpu);
2378 for_each_mem_cgroup(memcg) {
2381 for (i = 0; i < MEMCG_NR_STAT; i++) {
2385 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2387 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2388 atomic_long_add(x, &memcg->vmstats[i]);
2390 if (i >= NR_VM_NODE_STAT_ITEMS)
2393 for_each_node(nid) {
2394 struct mem_cgroup_per_node *pn;
2396 pn = mem_cgroup_nodeinfo(memcg, nid);
2397 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2400 atomic_long_add(x, &pn->lruvec_stat[i]);
2401 } while ((pn = parent_nodeinfo(pn, nid)));
2405 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2408 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2410 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2411 atomic_long_add(x, &memcg->vmevents[i]);
2418 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2419 unsigned int nr_pages,
2422 unsigned long nr_reclaimed = 0;
2425 unsigned long pflags;
2427 if (page_counter_read(&memcg->memory) <=
2428 READ_ONCE(memcg->memory.high))
2431 memcg_memory_event(memcg, MEMCG_HIGH);
2433 psi_memstall_enter(&pflags);
2434 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2436 psi_memstall_leave(&pflags);
2437 } while ((memcg = parent_mem_cgroup(memcg)) &&
2438 !mem_cgroup_is_root(memcg));
2440 return nr_reclaimed;
2443 static void high_work_func(struct work_struct *work)
2445 struct mem_cgroup *memcg;
2447 memcg = container_of(work, struct mem_cgroup, high_work);
2448 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2452 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2453 * enough to still cause a significant slowdown in most cases, while still
2454 * allowing diagnostics and tracing to proceed without becoming stuck.
2456 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2459 * When calculating the delay, we use these either side of the exponentiation to
2460 * maintain precision and scale to a reasonable number of jiffies (see the table
2463 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2464 * overage ratio to a delay.
2465 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2466 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2467 * to produce a reasonable delay curve.
2469 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2470 * reasonable delay curve compared to precision-adjusted overage, not
2471 * penalising heavily at first, but still making sure that growth beyond the
2472 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2473 * example, with a high of 100 megabytes:
2475 * +-------+------------------------+
2476 * | usage | time to allocate in ms |
2477 * +-------+------------------------+
2499 * +-------+------------------------+
2501 #define MEMCG_DELAY_PRECISION_SHIFT 20
2502 #define MEMCG_DELAY_SCALING_SHIFT 14
2504 static u64 calculate_overage(unsigned long usage, unsigned long high)
2512 * Prevent division by 0 in overage calculation by acting as if
2513 * it was a threshold of 1 page
2515 high = max(high, 1UL);
2517 overage = usage - high;
2518 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2519 return div64_u64(overage, high);
2522 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2524 u64 overage, max_overage = 0;
2527 overage = calculate_overage(page_counter_read(&memcg->memory),
2528 READ_ONCE(memcg->memory.high));
2529 max_overage = max(overage, max_overage);
2530 } while ((memcg = parent_mem_cgroup(memcg)) &&
2531 !mem_cgroup_is_root(memcg));
2536 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2538 u64 overage, max_overage = 0;
2541 overage = calculate_overage(page_counter_read(&memcg->swap),
2542 READ_ONCE(memcg->swap.high));
2544 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2545 max_overage = max(overage, max_overage);
2546 } while ((memcg = parent_mem_cgroup(memcg)) &&
2547 !mem_cgroup_is_root(memcg));
2553 * Get the number of jiffies that we should penalise a mischievous cgroup which
2554 * is exceeding its memory.high by checking both it and its ancestors.
2556 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2557 unsigned int nr_pages,
2560 unsigned long penalty_jiffies;
2566 * We use overage compared to memory.high to calculate the number of
2567 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2568 * fairly lenient on small overages, and increasingly harsh when the
2569 * memcg in question makes it clear that it has no intention of stopping
2570 * its crazy behaviour, so we exponentially increase the delay based on
2573 penalty_jiffies = max_overage * max_overage * HZ;
2574 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2575 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2578 * Factor in the task's own contribution to the overage, such that four
2579 * N-sized allocations are throttled approximately the same as one
2580 * 4N-sized allocation.
2582 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2583 * larger the current charge patch is than that.
2585 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2589 * Scheduled by try_charge() to be executed from the userland return path
2590 * and reclaims memory over the high limit.
2592 void mem_cgroup_handle_over_high(void)
2594 unsigned long penalty_jiffies;
2595 unsigned long pflags;
2596 unsigned long nr_reclaimed;
2597 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2598 int nr_retries = MAX_RECLAIM_RETRIES;
2599 struct mem_cgroup *memcg;
2600 bool in_retry = false;
2602 if (likely(!nr_pages))
2605 memcg = get_mem_cgroup_from_mm(current->mm);
2606 current->memcg_nr_pages_over_high = 0;
2610 * The allocating task should reclaim at least the batch size, but for
2611 * subsequent retries we only want to do what's necessary to prevent oom
2612 * or breaching resource isolation.
2614 * This is distinct from memory.max or page allocator behaviour because
2615 * memory.high is currently batched, whereas memory.max and the page
2616 * allocator run every time an allocation is made.
2618 nr_reclaimed = reclaim_high(memcg,
2619 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2623 * memory.high is breached and reclaim is unable to keep up. Throttle
2624 * allocators proactively to slow down excessive growth.
2626 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2627 mem_find_max_overage(memcg));
2629 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2630 swap_find_max_overage(memcg));
2633 * Clamp the max delay per usermode return so as to still keep the
2634 * application moving forwards and also permit diagnostics, albeit
2637 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2640 * Don't sleep if the amount of jiffies this memcg owes us is so low
2641 * that it's not even worth doing, in an attempt to be nice to those who
2642 * go only a small amount over their memory.high value and maybe haven't
2643 * been aggressively reclaimed enough yet.
2645 if (penalty_jiffies <= HZ / 100)
2649 * If reclaim is making forward progress but we're still over
2650 * memory.high, we want to encourage that rather than doing allocator
2653 if (nr_reclaimed || nr_retries--) {
2659 * If we exit early, we're guaranteed to die (since
2660 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2661 * need to account for any ill-begotten jiffies to pay them off later.
2663 psi_memstall_enter(&pflags);
2664 schedule_timeout_killable(penalty_jiffies);
2665 psi_memstall_leave(&pflags);
2668 css_put(&memcg->css);
2671 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2672 unsigned int nr_pages)
2674 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2675 int nr_retries = MAX_RECLAIM_RETRIES;
2676 struct mem_cgroup *mem_over_limit;
2677 struct page_counter *counter;
2678 enum oom_status oom_status;
2679 unsigned long nr_reclaimed;
2680 bool may_swap = true;
2681 bool drained = false;
2682 unsigned long pflags;
2684 if (mem_cgroup_is_root(memcg))
2687 if (consume_stock(memcg, nr_pages))
2690 if (!do_memsw_account() ||
2691 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2692 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2694 if (do_memsw_account())
2695 page_counter_uncharge(&memcg->memsw, batch);
2696 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2698 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2702 if (batch > nr_pages) {
2708 * Memcg doesn't have a dedicated reserve for atomic
2709 * allocations. But like the global atomic pool, we need to
2710 * put the burden of reclaim on regular allocation requests
2711 * and let these go through as privileged allocations.
2713 if (gfp_mask & __GFP_ATOMIC)
2717 * Unlike in global OOM situations, memcg is not in a physical
2718 * memory shortage. Allow dying and OOM-killed tasks to
2719 * bypass the last charges so that they can exit quickly and
2720 * free their memory.
2722 if (unlikely(should_force_charge()))
2726 * Prevent unbounded recursion when reclaim operations need to
2727 * allocate memory. This might exceed the limits temporarily,
2728 * but we prefer facilitating memory reclaim and getting back
2729 * under the limit over triggering OOM kills in these cases.
2731 if (unlikely(current->flags & PF_MEMALLOC))
2734 if (unlikely(task_in_memcg_oom(current)))
2737 if (!gfpflags_allow_blocking(gfp_mask))
2740 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2742 psi_memstall_enter(&pflags);
2743 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2744 gfp_mask, may_swap);
2745 psi_memstall_leave(&pflags);
2747 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2751 drain_all_stock(mem_over_limit);
2756 if (gfp_mask & __GFP_NORETRY)
2759 * Even though the limit is exceeded at this point, reclaim
2760 * may have been able to free some pages. Retry the charge
2761 * before killing the task.
2763 * Only for regular pages, though: huge pages are rather
2764 * unlikely to succeed so close to the limit, and we fall back
2765 * to regular pages anyway in case of failure.
2767 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2770 * At task move, charge accounts can be doubly counted. So, it's
2771 * better to wait until the end of task_move if something is going on.
2773 if (mem_cgroup_wait_acct_move(mem_over_limit))
2779 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2782 if (gfp_mask & __GFP_NOFAIL)
2785 if (fatal_signal_pending(current))
2789 * keep retrying as long as the memcg oom killer is able to make
2790 * a forward progress or bypass the charge if the oom killer
2791 * couldn't make any progress.
2793 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2794 get_order(nr_pages * PAGE_SIZE));
2795 switch (oom_status) {
2797 nr_retries = MAX_RECLAIM_RETRIES;
2805 if (!(gfp_mask & __GFP_NOFAIL))
2809 * The allocation either can't fail or will lead to more memory
2810 * being freed very soon. Allow memory usage go over the limit
2811 * temporarily by force charging it.
2813 page_counter_charge(&memcg->memory, nr_pages);
2814 if (do_memsw_account())
2815 page_counter_charge(&memcg->memsw, nr_pages);
2820 if (batch > nr_pages)
2821 refill_stock(memcg, batch - nr_pages);
2824 * If the hierarchy is above the normal consumption range, schedule
2825 * reclaim on returning to userland. We can perform reclaim here
2826 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2827 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2828 * not recorded as it most likely matches current's and won't
2829 * change in the meantime. As high limit is checked again before
2830 * reclaim, the cost of mismatch is negligible.
2833 bool mem_high, swap_high;
2835 mem_high = page_counter_read(&memcg->memory) >
2836 READ_ONCE(memcg->memory.high);
2837 swap_high = page_counter_read(&memcg->swap) >
2838 READ_ONCE(memcg->swap.high);
2840 /* Don't bother a random interrupted task */
2841 if (in_interrupt()) {
2843 schedule_work(&memcg->high_work);
2849 if (mem_high || swap_high) {
2851 * The allocating tasks in this cgroup will need to do
2852 * reclaim or be throttled to prevent further growth
2853 * of the memory or swap footprints.
2855 * Target some best-effort fairness between the tasks,
2856 * and distribute reclaim work and delay penalties
2857 * based on how much each task is actually allocating.
2859 current->memcg_nr_pages_over_high += batch;
2860 set_notify_resume(current);
2863 } while ((memcg = parent_mem_cgroup(memcg)));
2868 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2869 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2871 if (mem_cgroup_is_root(memcg))
2874 page_counter_uncharge(&memcg->memory, nr_pages);
2875 if (do_memsw_account())
2876 page_counter_uncharge(&memcg->memsw, nr_pages);
2880 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2882 VM_BUG_ON_PAGE(page_memcg(page), page);
2884 * Any of the following ensures page's memcg stability:
2888 * - lock_page_memcg()
2889 * - exclusive reference
2891 page->memcg_data = (unsigned long)memcg;
2894 #ifdef CONFIG_MEMCG_KMEM
2895 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2896 gfp_t gfp, bool new_page)
2898 unsigned int objects = objs_per_slab_page(s, page);
2899 unsigned long memcg_data;
2902 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2907 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2910 * If the slab page is brand new and nobody can yet access
2911 * it's memcg_data, no synchronization is required and
2912 * memcg_data can be simply assigned.
2914 page->memcg_data = memcg_data;
2915 } else if (cmpxchg(&page->memcg_data, 0, memcg_data)) {
2917 * If the slab page is already in use, somebody can allocate
2918 * and assign obj_cgroups in parallel. In this case the existing
2919 * objcg vector should be reused.
2925 kmemleak_not_leak(vec);
2930 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2932 * A passed kernel object can be a slab object or a generic kernel page, so
2933 * different mechanisms for getting the memory cgroup pointer should be used.
2934 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2935 * can not know for sure how the kernel object is implemented.
2936 * mem_cgroup_from_obj() can be safely used in such cases.
2938 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2939 * cgroup_mutex, etc.
2941 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2945 if (mem_cgroup_disabled())
2948 page = virt_to_head_page(p);
2951 * Slab objects are accounted individually, not per-page.
2952 * Memcg membership data for each individual object is saved in
2953 * the page->obj_cgroups.
2955 if (page_objcgs_check(page)) {
2956 struct obj_cgroup *objcg;
2959 off = obj_to_index(page->slab_cache, page, p);
2960 objcg = page_objcgs(page)[off];
2962 return obj_cgroup_memcg(objcg);
2968 * page_memcg_check() is used here, because page_has_obj_cgroups()
2969 * check above could fail because the object cgroups vector wasn't set
2970 * at that moment, but it can be set concurrently.
2971 * page_memcg_check(page) will guarantee that a proper memory
2972 * cgroup pointer or NULL will be returned.
2974 return page_memcg_check(page);
2977 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2979 struct obj_cgroup *objcg = NULL;
2980 struct mem_cgroup *memcg;
2982 if (memcg_kmem_bypass())
2986 if (unlikely(active_memcg()))
2987 memcg = active_memcg();
2989 memcg = mem_cgroup_from_task(current);
2991 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2992 objcg = rcu_dereference(memcg->objcg);
2993 if (objcg && obj_cgroup_tryget(objcg))
3002 static int memcg_alloc_cache_id(void)
3007 id = ida_simple_get(&memcg_cache_ida,
3008 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3012 if (id < memcg_nr_cache_ids)
3016 * There's no space for the new id in memcg_caches arrays,
3017 * so we have to grow them.
3019 down_write(&memcg_cache_ids_sem);
3021 size = 2 * (id + 1);
3022 if (size < MEMCG_CACHES_MIN_SIZE)
3023 size = MEMCG_CACHES_MIN_SIZE;
3024 else if (size > MEMCG_CACHES_MAX_SIZE)
3025 size = MEMCG_CACHES_MAX_SIZE;
3027 err = memcg_update_all_list_lrus(size);
3029 memcg_nr_cache_ids = size;
3031 up_write(&memcg_cache_ids_sem);
3034 ida_simple_remove(&memcg_cache_ida, id);
3040 static void memcg_free_cache_id(int id)
3042 ida_simple_remove(&memcg_cache_ida, id);
3046 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3047 * @memcg: memory cgroup to charge
3048 * @gfp: reclaim mode
3049 * @nr_pages: number of pages to charge
3051 * Returns 0 on success, an error code on failure.
3053 static int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
3054 unsigned int nr_pages)
3056 struct page_counter *counter;
3059 ret = try_charge(memcg, gfp, nr_pages);
3063 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3064 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3067 * Enforce __GFP_NOFAIL allocation because callers are not
3068 * prepared to see failures and likely do not have any failure
3071 if (gfp & __GFP_NOFAIL) {
3072 page_counter_charge(&memcg->kmem, nr_pages);
3075 cancel_charge(memcg, nr_pages);
3082 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3083 * @memcg: memcg to uncharge
3084 * @nr_pages: number of pages to uncharge
3086 static void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3088 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3089 page_counter_uncharge(&memcg->kmem, nr_pages);
3091 refill_stock(memcg, nr_pages);
3095 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3096 * @page: page to charge
3097 * @gfp: reclaim mode
3098 * @order: allocation order
3100 * Returns 0 on success, an error code on failure.
3102 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3104 struct mem_cgroup *memcg;
3107 memcg = get_mem_cgroup_from_current();
3108 if (memcg && !mem_cgroup_is_root(memcg)) {
3109 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3111 page->memcg_data = (unsigned long)memcg |
3115 css_put(&memcg->css);
3121 * __memcg_kmem_uncharge_page: uncharge a kmem page
3122 * @page: page to uncharge
3123 * @order: allocation order
3125 void __memcg_kmem_uncharge_page(struct page *page, int order)
3127 struct mem_cgroup *memcg = page_memcg(page);
3128 unsigned int nr_pages = 1 << order;
3133 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3134 __memcg_kmem_uncharge(memcg, nr_pages);
3135 page->memcg_data = 0;
3136 css_put(&memcg->css);
3139 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3141 struct memcg_stock_pcp *stock;
3142 unsigned long flags;
3145 local_irq_save(flags);
3147 stock = this_cpu_ptr(&memcg_stock);
3148 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3149 stock->nr_bytes -= nr_bytes;
3153 local_irq_restore(flags);
3158 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3160 struct obj_cgroup *old = stock->cached_objcg;
3165 if (stock->nr_bytes) {
3166 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3167 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3171 __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
3176 * The leftover is flushed to the centralized per-memcg value.
3177 * On the next attempt to refill obj stock it will be moved
3178 * to a per-cpu stock (probably, on an other CPU), see
3179 * refill_obj_stock().
3181 * How often it's flushed is a trade-off between the memory
3182 * limit enforcement accuracy and potential CPU contention,
3183 * so it might be changed in the future.
3185 atomic_add(nr_bytes, &old->nr_charged_bytes);
3186 stock->nr_bytes = 0;
3189 obj_cgroup_put(old);
3190 stock->cached_objcg = NULL;
3193 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3194 struct mem_cgroup *root_memcg)
3196 struct mem_cgroup *memcg;
3198 if (stock->cached_objcg) {
3199 memcg = obj_cgroup_memcg(stock->cached_objcg);
3200 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3207 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3209 struct memcg_stock_pcp *stock;
3210 unsigned long flags;
3212 local_irq_save(flags);
3214 stock = this_cpu_ptr(&memcg_stock);
3215 if (stock->cached_objcg != objcg) { /* reset if necessary */
3216 drain_obj_stock(stock);
3217 obj_cgroup_get(objcg);
3218 stock->cached_objcg = objcg;
3219 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3221 stock->nr_bytes += nr_bytes;
3223 if (stock->nr_bytes > PAGE_SIZE)
3224 drain_obj_stock(stock);
3226 local_irq_restore(flags);
3229 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3231 struct mem_cgroup *memcg;
3232 unsigned int nr_pages, nr_bytes;
3235 if (consume_obj_stock(objcg, size))
3239 * In theory, memcg->nr_charged_bytes can have enough
3240 * pre-charged bytes to satisfy the allocation. However,
3241 * flushing memcg->nr_charged_bytes requires two atomic
3242 * operations, and memcg->nr_charged_bytes can't be big,
3243 * so it's better to ignore it and try grab some new pages.
3244 * memcg->nr_charged_bytes will be flushed in
3245 * refill_obj_stock(), called from this function or
3246 * independently later.
3250 memcg = obj_cgroup_memcg(objcg);
3251 if (unlikely(!css_tryget(&memcg->css)))
3255 nr_pages = size >> PAGE_SHIFT;
3256 nr_bytes = size & (PAGE_SIZE - 1);
3261 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3262 if (!ret && nr_bytes)
3263 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3265 css_put(&memcg->css);
3269 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3271 refill_obj_stock(objcg, size);
3274 #endif /* CONFIG_MEMCG_KMEM */
3277 * Because page_memcg(head) is not set on tails, set it now.
3279 void split_page_memcg(struct page *head, unsigned int nr)
3281 struct mem_cgroup *memcg = page_memcg(head);
3284 if (mem_cgroup_disabled() || !memcg)
3287 for (i = 1; i < nr; i++)
3288 head[i].memcg_data = head->memcg_data;
3289 css_get_many(&memcg->css, nr - 1);
3292 #ifdef CONFIG_MEMCG_SWAP
3294 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3295 * @entry: swap entry to be moved
3296 * @from: mem_cgroup which the entry is moved from
3297 * @to: mem_cgroup which the entry is moved to
3299 * It succeeds only when the swap_cgroup's record for this entry is the same
3300 * as the mem_cgroup's id of @from.
3302 * Returns 0 on success, -EINVAL on failure.
3304 * The caller must have charged to @to, IOW, called page_counter_charge() about
3305 * both res and memsw, and called css_get().
3307 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3308 struct mem_cgroup *from, struct mem_cgroup *to)
3310 unsigned short old_id, new_id;
3312 old_id = mem_cgroup_id(from);
3313 new_id = mem_cgroup_id(to);
3315 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3316 mod_memcg_state(from, MEMCG_SWAP, -1);
3317 mod_memcg_state(to, MEMCG_SWAP, 1);
3323 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3324 struct mem_cgroup *from, struct mem_cgroup *to)
3330 static DEFINE_MUTEX(memcg_max_mutex);
3332 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3333 unsigned long max, bool memsw)
3335 bool enlarge = false;
3336 bool drained = false;
3338 bool limits_invariant;
3339 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3342 if (signal_pending(current)) {
3347 mutex_lock(&memcg_max_mutex);
3349 * Make sure that the new limit (memsw or memory limit) doesn't
3350 * break our basic invariant rule memory.max <= memsw.max.
3352 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3353 max <= memcg->memsw.max;
3354 if (!limits_invariant) {
3355 mutex_unlock(&memcg_max_mutex);
3359 if (max > counter->max)
3361 ret = page_counter_set_max(counter, max);
3362 mutex_unlock(&memcg_max_mutex);
3368 drain_all_stock(memcg);
3373 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3374 GFP_KERNEL, !memsw)) {
3380 if (!ret && enlarge)
3381 memcg_oom_recover(memcg);
3386 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3388 unsigned long *total_scanned)
3390 unsigned long nr_reclaimed = 0;
3391 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3392 unsigned long reclaimed;
3394 struct mem_cgroup_tree_per_node *mctz;
3395 unsigned long excess;
3396 unsigned long nr_scanned;
3401 mctz = soft_limit_tree_node(pgdat->node_id);
3404 * Do not even bother to check the largest node if the root
3405 * is empty. Do it lockless to prevent lock bouncing. Races
3406 * are acceptable as soft limit is best effort anyway.
3408 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3412 * This loop can run a while, specially if mem_cgroup's continuously
3413 * keep exceeding their soft limit and putting the system under
3420 mz = mem_cgroup_largest_soft_limit_node(mctz);
3425 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3426 gfp_mask, &nr_scanned);
3427 nr_reclaimed += reclaimed;
3428 *total_scanned += nr_scanned;
3429 spin_lock_irq(&mctz->lock);
3430 __mem_cgroup_remove_exceeded(mz, mctz);
3433 * If we failed to reclaim anything from this memory cgroup
3434 * it is time to move on to the next cgroup
3438 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3440 excess = soft_limit_excess(mz->memcg);
3442 * One school of thought says that we should not add
3443 * back the node to the tree if reclaim returns 0.
3444 * But our reclaim could return 0, simply because due
3445 * to priority we are exposing a smaller subset of
3446 * memory to reclaim from. Consider this as a longer
3449 /* If excess == 0, no tree ops */
3450 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3451 spin_unlock_irq(&mctz->lock);
3452 css_put(&mz->memcg->css);
3455 * Could not reclaim anything and there are no more
3456 * mem cgroups to try or we seem to be looping without
3457 * reclaiming anything.
3459 if (!nr_reclaimed &&
3461 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3463 } while (!nr_reclaimed);
3465 css_put(&next_mz->memcg->css);
3466 return nr_reclaimed;
3470 * Reclaims as many pages from the given memcg as possible.
3472 * Caller is responsible for holding css reference for memcg.
3474 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3476 int nr_retries = MAX_RECLAIM_RETRIES;
3478 /* we call try-to-free pages for make this cgroup empty */
3479 lru_add_drain_all();
3481 drain_all_stock(memcg);
3483 /* try to free all pages in this cgroup */
3484 while (nr_retries && page_counter_read(&memcg->memory)) {
3487 if (signal_pending(current))
3490 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3494 /* maybe some writeback is necessary */
3495 congestion_wait(BLK_RW_ASYNC, HZ/10);
3503 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3504 char *buf, size_t nbytes,
3507 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3509 if (mem_cgroup_is_root(memcg))
3511 return mem_cgroup_force_empty(memcg) ?: nbytes;
3514 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3520 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3521 struct cftype *cft, u64 val)
3526 pr_warn_once("Non-hierarchical mode is deprecated. "
3527 "Please report your usecase to linux-mm@kvack.org if you "
3528 "depend on this functionality.\n");
3533 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3537 if (mem_cgroup_is_root(memcg)) {
3538 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3539 memcg_page_state(memcg, NR_ANON_MAPPED);
3541 val += memcg_page_state(memcg, MEMCG_SWAP);
3544 val = page_counter_read(&memcg->memory);
3546 val = page_counter_read(&memcg->memsw);
3559 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3562 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3563 struct page_counter *counter;
3565 switch (MEMFILE_TYPE(cft->private)) {
3567 counter = &memcg->memory;
3570 counter = &memcg->memsw;
3573 counter = &memcg->kmem;
3576 counter = &memcg->tcpmem;
3582 switch (MEMFILE_ATTR(cft->private)) {
3584 if (counter == &memcg->memory)
3585 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3586 if (counter == &memcg->memsw)
3587 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3588 return (u64)page_counter_read(counter) * PAGE_SIZE;
3590 return (u64)counter->max * PAGE_SIZE;
3592 return (u64)counter->watermark * PAGE_SIZE;
3594 return counter->failcnt;
3595 case RES_SOFT_LIMIT:
3596 return (u64)memcg->soft_limit * PAGE_SIZE;
3602 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3604 unsigned long stat[MEMCG_NR_STAT] = {0};
3605 struct mem_cgroup *mi;
3608 for_each_online_cpu(cpu)
3609 for (i = 0; i < MEMCG_NR_STAT; i++)
3610 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3612 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3613 for (i = 0; i < MEMCG_NR_STAT; i++)
3614 atomic_long_add(stat[i], &mi->vmstats[i]);
3616 for_each_node(node) {
3617 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3618 struct mem_cgroup_per_node *pi;
3620 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3623 for_each_online_cpu(cpu)
3624 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3626 pn->lruvec_stat_cpu->count[i], cpu);
3628 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3629 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3630 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3634 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3636 unsigned long events[NR_VM_EVENT_ITEMS];
3637 struct mem_cgroup *mi;
3640 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3643 for_each_online_cpu(cpu)
3644 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3645 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3648 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3649 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3650 atomic_long_add(events[i], &mi->vmevents[i]);
3653 #ifdef CONFIG_MEMCG_KMEM
3654 static int memcg_online_kmem(struct mem_cgroup *memcg)
3656 struct obj_cgroup *objcg;
3659 if (cgroup_memory_nokmem)
3662 BUG_ON(memcg->kmemcg_id >= 0);
3663 BUG_ON(memcg->kmem_state);
3665 memcg_id = memcg_alloc_cache_id();
3669 objcg = obj_cgroup_alloc();
3671 memcg_free_cache_id(memcg_id);
3674 objcg->memcg = memcg;
3675 rcu_assign_pointer(memcg->objcg, objcg);
3677 static_branch_enable(&memcg_kmem_enabled_key);
3679 memcg->kmemcg_id = memcg_id;
3680 memcg->kmem_state = KMEM_ONLINE;
3685 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3687 struct cgroup_subsys_state *css;
3688 struct mem_cgroup *parent, *child;
3691 if (memcg->kmem_state != KMEM_ONLINE)
3694 memcg->kmem_state = KMEM_ALLOCATED;
3696 parent = parent_mem_cgroup(memcg);
3698 parent = root_mem_cgroup;
3700 memcg_reparent_objcgs(memcg, parent);
3702 kmemcg_id = memcg->kmemcg_id;
3703 BUG_ON(kmemcg_id < 0);
3706 * Change kmemcg_id of this cgroup and all its descendants to the
3707 * parent's id, and then move all entries from this cgroup's list_lrus
3708 * to ones of the parent. After we have finished, all list_lrus
3709 * corresponding to this cgroup are guaranteed to remain empty. The
3710 * ordering is imposed by list_lru_node->lock taken by
3711 * memcg_drain_all_list_lrus().
3713 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3714 css_for_each_descendant_pre(css, &memcg->css) {
3715 child = mem_cgroup_from_css(css);
3716 BUG_ON(child->kmemcg_id != kmemcg_id);
3717 child->kmemcg_id = parent->kmemcg_id;
3721 memcg_drain_all_list_lrus(kmemcg_id, parent);
3723 memcg_free_cache_id(kmemcg_id);
3726 static void memcg_free_kmem(struct mem_cgroup *memcg)
3728 /* css_alloc() failed, offlining didn't happen */
3729 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3730 memcg_offline_kmem(memcg);
3733 static int memcg_online_kmem(struct mem_cgroup *memcg)
3737 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3740 static void memcg_free_kmem(struct mem_cgroup *memcg)
3743 #endif /* CONFIG_MEMCG_KMEM */
3745 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3750 mutex_lock(&memcg_max_mutex);
3751 ret = page_counter_set_max(&memcg->kmem, max);
3752 mutex_unlock(&memcg_max_mutex);
3756 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3760 mutex_lock(&memcg_max_mutex);
3762 ret = page_counter_set_max(&memcg->tcpmem, max);
3766 if (!memcg->tcpmem_active) {
3768 * The active flag needs to be written after the static_key
3769 * update. This is what guarantees that the socket activation
3770 * function is the last one to run. See mem_cgroup_sk_alloc()
3771 * for details, and note that we don't mark any socket as
3772 * belonging to this memcg until that flag is up.
3774 * We need to do this, because static_keys will span multiple
3775 * sites, but we can't control their order. If we mark a socket
3776 * as accounted, but the accounting functions are not patched in
3777 * yet, we'll lose accounting.
3779 * We never race with the readers in mem_cgroup_sk_alloc(),
3780 * because when this value change, the code to process it is not
3783 static_branch_inc(&memcg_sockets_enabled_key);
3784 memcg->tcpmem_active = true;
3787 mutex_unlock(&memcg_max_mutex);
3792 * The user of this function is...
3795 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3796 char *buf, size_t nbytes, loff_t off)
3798 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3799 unsigned long nr_pages;
3802 buf = strstrip(buf);
3803 ret = page_counter_memparse(buf, "-1", &nr_pages);
3807 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3809 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3813 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3815 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3818 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3821 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3822 "Please report your usecase to linux-mm@kvack.org if you "
3823 "depend on this functionality.\n");
3824 ret = memcg_update_kmem_max(memcg, nr_pages);
3827 ret = memcg_update_tcp_max(memcg, nr_pages);
3831 case RES_SOFT_LIMIT:
3832 memcg->soft_limit = nr_pages;
3836 return ret ?: nbytes;
3839 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3840 size_t nbytes, loff_t off)
3842 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3843 struct page_counter *counter;
3845 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3847 counter = &memcg->memory;
3850 counter = &memcg->memsw;
3853 counter = &memcg->kmem;
3856 counter = &memcg->tcpmem;
3862 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3864 page_counter_reset_watermark(counter);
3867 counter->failcnt = 0;
3876 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3879 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3883 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3884 struct cftype *cft, u64 val)
3886 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3888 if (val & ~MOVE_MASK)
3892 * No kind of locking is needed in here, because ->can_attach() will
3893 * check this value once in the beginning of the process, and then carry
3894 * on with stale data. This means that changes to this value will only
3895 * affect task migrations starting after the change.
3897 memcg->move_charge_at_immigrate = val;
3901 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3902 struct cftype *cft, u64 val)
3910 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3911 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3912 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3914 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3915 int nid, unsigned int lru_mask, bool tree)
3917 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3918 unsigned long nr = 0;
3921 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3924 if (!(BIT(lru) & lru_mask))
3927 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3929 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3934 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3935 unsigned int lru_mask,
3938 unsigned long nr = 0;
3942 if (!(BIT(lru) & lru_mask))
3945 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3947 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3952 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3956 unsigned int lru_mask;
3959 static const struct numa_stat stats[] = {
3960 { "total", LRU_ALL },
3961 { "file", LRU_ALL_FILE },
3962 { "anon", LRU_ALL_ANON },
3963 { "unevictable", BIT(LRU_UNEVICTABLE) },
3965 const struct numa_stat *stat;
3967 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3969 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3970 seq_printf(m, "%s=%lu", stat->name,
3971 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3973 for_each_node_state(nid, N_MEMORY)
3974 seq_printf(m, " N%d=%lu", nid,
3975 mem_cgroup_node_nr_lru_pages(memcg, nid,
3976 stat->lru_mask, false));
3980 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3982 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3983 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3985 for_each_node_state(nid, N_MEMORY)
3986 seq_printf(m, " N%d=%lu", nid,
3987 mem_cgroup_node_nr_lru_pages(memcg, nid,
3988 stat->lru_mask, true));
3994 #endif /* CONFIG_NUMA */
3996 static const unsigned int memcg1_stats[] = {
3999 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4009 static const char *const memcg1_stat_names[] = {
4012 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4022 /* Universal VM events cgroup1 shows, original sort order */
4023 static const unsigned int memcg1_events[] = {
4030 static int memcg_stat_show(struct seq_file *m, void *v)
4032 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4033 unsigned long memory, memsw;
4034 struct mem_cgroup *mi;
4037 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4039 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4042 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4044 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4045 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4048 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4049 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4050 memcg_events_local(memcg, memcg1_events[i]));
4052 for (i = 0; i < NR_LRU_LISTS; i++)
4053 seq_printf(m, "%s %lu\n", lru_list_name(i),
4054 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4057 /* Hierarchical information */
4058 memory = memsw = PAGE_COUNTER_MAX;
4059 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4060 memory = min(memory, READ_ONCE(mi->memory.max));
4061 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4063 seq_printf(m, "hierarchical_memory_limit %llu\n",
4064 (u64)memory * PAGE_SIZE);
4065 if (do_memsw_account())
4066 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4067 (u64)memsw * PAGE_SIZE);
4069 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4072 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4074 nr = memcg_page_state(memcg, memcg1_stats[i]);
4075 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4076 (u64)nr * PAGE_SIZE);
4079 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4080 seq_printf(m, "total_%s %llu\n",
4081 vm_event_name(memcg1_events[i]),
4082 (u64)memcg_events(memcg, memcg1_events[i]));
4084 for (i = 0; i < NR_LRU_LISTS; i++)
4085 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4086 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4089 #ifdef CONFIG_DEBUG_VM
4092 struct mem_cgroup_per_node *mz;
4093 unsigned long anon_cost = 0;
4094 unsigned long file_cost = 0;
4096 for_each_online_pgdat(pgdat) {
4097 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4099 anon_cost += mz->lruvec.anon_cost;
4100 file_cost += mz->lruvec.file_cost;
4102 seq_printf(m, "anon_cost %lu\n", anon_cost);
4103 seq_printf(m, "file_cost %lu\n", file_cost);
4110 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4113 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4115 return mem_cgroup_swappiness(memcg);
4118 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4119 struct cftype *cft, u64 val)
4121 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4127 memcg->swappiness = val;
4129 vm_swappiness = val;
4134 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4136 struct mem_cgroup_threshold_ary *t;
4137 unsigned long usage;
4142 t = rcu_dereference(memcg->thresholds.primary);
4144 t = rcu_dereference(memcg->memsw_thresholds.primary);
4149 usage = mem_cgroup_usage(memcg, swap);
4152 * current_threshold points to threshold just below or equal to usage.
4153 * If it's not true, a threshold was crossed after last
4154 * call of __mem_cgroup_threshold().
4156 i = t->current_threshold;
4159 * Iterate backward over array of thresholds starting from
4160 * current_threshold and check if a threshold is crossed.
4161 * If none of thresholds below usage is crossed, we read
4162 * only one element of the array here.
4164 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4165 eventfd_signal(t->entries[i].eventfd, 1);
4167 /* i = current_threshold + 1 */
4171 * Iterate forward over array of thresholds starting from
4172 * current_threshold+1 and check if a threshold is crossed.
4173 * If none of thresholds above usage is crossed, we read
4174 * only one element of the array here.
4176 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4177 eventfd_signal(t->entries[i].eventfd, 1);
4179 /* Update current_threshold */
4180 t->current_threshold = i - 1;
4185 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4188 __mem_cgroup_threshold(memcg, false);
4189 if (do_memsw_account())
4190 __mem_cgroup_threshold(memcg, true);
4192 memcg = parent_mem_cgroup(memcg);
4196 static int compare_thresholds(const void *a, const void *b)
4198 const struct mem_cgroup_threshold *_a = a;
4199 const struct mem_cgroup_threshold *_b = b;
4201 if (_a->threshold > _b->threshold)
4204 if (_a->threshold < _b->threshold)
4210 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4212 struct mem_cgroup_eventfd_list *ev;
4214 spin_lock(&memcg_oom_lock);
4216 list_for_each_entry(ev, &memcg->oom_notify, list)
4217 eventfd_signal(ev->eventfd, 1);
4219 spin_unlock(&memcg_oom_lock);
4223 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4225 struct mem_cgroup *iter;
4227 for_each_mem_cgroup_tree(iter, memcg)
4228 mem_cgroup_oom_notify_cb(iter);
4231 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4232 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4234 struct mem_cgroup_thresholds *thresholds;
4235 struct mem_cgroup_threshold_ary *new;
4236 unsigned long threshold;
4237 unsigned long usage;
4240 ret = page_counter_memparse(args, "-1", &threshold);
4244 mutex_lock(&memcg->thresholds_lock);
4247 thresholds = &memcg->thresholds;
4248 usage = mem_cgroup_usage(memcg, false);
4249 } else if (type == _MEMSWAP) {
4250 thresholds = &memcg->memsw_thresholds;
4251 usage = mem_cgroup_usage(memcg, true);
4255 /* Check if a threshold crossed before adding a new one */
4256 if (thresholds->primary)
4257 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4259 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4261 /* Allocate memory for new array of thresholds */
4262 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4269 /* Copy thresholds (if any) to new array */
4270 if (thresholds->primary)
4271 memcpy(new->entries, thresholds->primary->entries,
4272 flex_array_size(new, entries, size - 1));
4274 /* Add new threshold */
4275 new->entries[size - 1].eventfd = eventfd;
4276 new->entries[size - 1].threshold = threshold;
4278 /* Sort thresholds. Registering of new threshold isn't time-critical */
4279 sort(new->entries, size, sizeof(*new->entries),
4280 compare_thresholds, NULL);
4282 /* Find current threshold */
4283 new->current_threshold = -1;
4284 for (i = 0; i < size; i++) {
4285 if (new->entries[i].threshold <= usage) {
4287 * new->current_threshold will not be used until
4288 * rcu_assign_pointer(), so it's safe to increment
4291 ++new->current_threshold;
4296 /* Free old spare buffer and save old primary buffer as spare */
4297 kfree(thresholds->spare);
4298 thresholds->spare = thresholds->primary;
4300 rcu_assign_pointer(thresholds->primary, new);
4302 /* To be sure that nobody uses thresholds */
4306 mutex_unlock(&memcg->thresholds_lock);
4311 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4312 struct eventfd_ctx *eventfd, const char *args)
4314 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4317 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4318 struct eventfd_ctx *eventfd, const char *args)
4320 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4323 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4324 struct eventfd_ctx *eventfd, enum res_type type)
4326 struct mem_cgroup_thresholds *thresholds;
4327 struct mem_cgroup_threshold_ary *new;
4328 unsigned long usage;
4329 int i, j, size, entries;
4331 mutex_lock(&memcg->thresholds_lock);
4334 thresholds = &memcg->thresholds;
4335 usage = mem_cgroup_usage(memcg, false);
4336 } else if (type == _MEMSWAP) {
4337 thresholds = &memcg->memsw_thresholds;
4338 usage = mem_cgroup_usage(memcg, true);
4342 if (!thresholds->primary)
4345 /* Check if a threshold crossed before removing */
4346 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4348 /* Calculate new number of threshold */
4350 for (i = 0; i < thresholds->primary->size; i++) {
4351 if (thresholds->primary->entries[i].eventfd != eventfd)
4357 new = thresholds->spare;
4359 /* If no items related to eventfd have been cleared, nothing to do */
4363 /* Set thresholds array to NULL if we don't have thresholds */
4372 /* Copy thresholds and find current threshold */
4373 new->current_threshold = -1;
4374 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4375 if (thresholds->primary->entries[i].eventfd == eventfd)
4378 new->entries[j] = thresholds->primary->entries[i];
4379 if (new->entries[j].threshold <= usage) {
4381 * new->current_threshold will not be used
4382 * until rcu_assign_pointer(), so it's safe to increment
4385 ++new->current_threshold;
4391 /* Swap primary and spare array */
4392 thresholds->spare = thresholds->primary;
4394 rcu_assign_pointer(thresholds->primary, new);
4396 /* To be sure that nobody uses thresholds */
4399 /* If all events are unregistered, free the spare array */
4401 kfree(thresholds->spare);
4402 thresholds->spare = NULL;
4405 mutex_unlock(&memcg->thresholds_lock);
4408 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4409 struct eventfd_ctx *eventfd)
4411 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4414 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4415 struct eventfd_ctx *eventfd)
4417 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4420 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4421 struct eventfd_ctx *eventfd, const char *args)
4423 struct mem_cgroup_eventfd_list *event;
4425 event = kmalloc(sizeof(*event), GFP_KERNEL);
4429 spin_lock(&memcg_oom_lock);
4431 event->eventfd = eventfd;
4432 list_add(&event->list, &memcg->oom_notify);
4434 /* already in OOM ? */
4435 if (memcg->under_oom)
4436 eventfd_signal(eventfd, 1);
4437 spin_unlock(&memcg_oom_lock);
4442 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4443 struct eventfd_ctx *eventfd)
4445 struct mem_cgroup_eventfd_list *ev, *tmp;
4447 spin_lock(&memcg_oom_lock);
4449 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4450 if (ev->eventfd == eventfd) {
4451 list_del(&ev->list);
4456 spin_unlock(&memcg_oom_lock);
4459 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4461 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4463 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4464 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4465 seq_printf(sf, "oom_kill %lu\n",
4466 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4470 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4471 struct cftype *cft, u64 val)
4473 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4475 /* cannot set to root cgroup and only 0 and 1 are allowed */
4476 if (!css->parent || !((val == 0) || (val == 1)))
4479 memcg->oom_kill_disable = val;
4481 memcg_oom_recover(memcg);
4486 #ifdef CONFIG_CGROUP_WRITEBACK
4488 #include <trace/events/writeback.h>
4490 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4492 return wb_domain_init(&memcg->cgwb_domain, gfp);
4495 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4497 wb_domain_exit(&memcg->cgwb_domain);
4500 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4502 wb_domain_size_changed(&memcg->cgwb_domain);
4505 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4507 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4509 if (!memcg->css.parent)
4512 return &memcg->cgwb_domain;
4516 * idx can be of type enum memcg_stat_item or node_stat_item.
4517 * Keep in sync with memcg_exact_page().
4519 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4521 long x = atomic_long_read(&memcg->vmstats[idx]);
4524 for_each_online_cpu(cpu)
4525 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4532 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4533 * @wb: bdi_writeback in question
4534 * @pfilepages: out parameter for number of file pages
4535 * @pheadroom: out parameter for number of allocatable pages according to memcg
4536 * @pdirty: out parameter for number of dirty pages
4537 * @pwriteback: out parameter for number of pages under writeback
4539 * Determine the numbers of file, headroom, dirty, and writeback pages in
4540 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4541 * is a bit more involved.
4543 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4544 * headroom is calculated as the lowest headroom of itself and the
4545 * ancestors. Note that this doesn't consider the actual amount of
4546 * available memory in the system. The caller should further cap
4547 * *@pheadroom accordingly.
4549 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4550 unsigned long *pheadroom, unsigned long *pdirty,
4551 unsigned long *pwriteback)
4553 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4554 struct mem_cgroup *parent;
4556 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4558 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4559 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4560 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4561 *pheadroom = PAGE_COUNTER_MAX;
4563 while ((parent = parent_mem_cgroup(memcg))) {
4564 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4565 READ_ONCE(memcg->memory.high));
4566 unsigned long used = page_counter_read(&memcg->memory);
4568 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4574 * Foreign dirty flushing
4576 * There's an inherent mismatch between memcg and writeback. The former
4577 * trackes ownership per-page while the latter per-inode. This was a
4578 * deliberate design decision because honoring per-page ownership in the
4579 * writeback path is complicated, may lead to higher CPU and IO overheads
4580 * and deemed unnecessary given that write-sharing an inode across
4581 * different cgroups isn't a common use-case.
4583 * Combined with inode majority-writer ownership switching, this works well
4584 * enough in most cases but there are some pathological cases. For
4585 * example, let's say there are two cgroups A and B which keep writing to
4586 * different but confined parts of the same inode. B owns the inode and
4587 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4588 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4589 * triggering background writeback. A will be slowed down without a way to
4590 * make writeback of the dirty pages happen.
4592 * Conditions like the above can lead to a cgroup getting repatedly and
4593 * severely throttled after making some progress after each
4594 * dirty_expire_interval while the underyling IO device is almost
4597 * Solving this problem completely requires matching the ownership tracking
4598 * granularities between memcg and writeback in either direction. However,
4599 * the more egregious behaviors can be avoided by simply remembering the
4600 * most recent foreign dirtying events and initiating remote flushes on
4601 * them when local writeback isn't enough to keep the memory clean enough.
4603 * The following two functions implement such mechanism. When a foreign
4604 * page - a page whose memcg and writeback ownerships don't match - is
4605 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4606 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4607 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4608 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4609 * foreign bdi_writebacks which haven't expired. Both the numbers of
4610 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4611 * limited to MEMCG_CGWB_FRN_CNT.
4613 * The mechanism only remembers IDs and doesn't hold any object references.
4614 * As being wrong occasionally doesn't matter, updates and accesses to the
4615 * records are lockless and racy.
4617 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4618 struct bdi_writeback *wb)
4620 struct mem_cgroup *memcg = page_memcg(page);
4621 struct memcg_cgwb_frn *frn;
4622 u64 now = get_jiffies_64();
4623 u64 oldest_at = now;
4627 trace_track_foreign_dirty(page, wb);
4630 * Pick the slot to use. If there is already a slot for @wb, keep
4631 * using it. If not replace the oldest one which isn't being
4634 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4635 frn = &memcg->cgwb_frn[i];
4636 if (frn->bdi_id == wb->bdi->id &&
4637 frn->memcg_id == wb->memcg_css->id)
4639 if (time_before64(frn->at, oldest_at) &&
4640 atomic_read(&frn->done.cnt) == 1) {
4642 oldest_at = frn->at;
4646 if (i < MEMCG_CGWB_FRN_CNT) {
4648 * Re-using an existing one. Update timestamp lazily to
4649 * avoid making the cacheline hot. We want them to be
4650 * reasonably up-to-date and significantly shorter than
4651 * dirty_expire_interval as that's what expires the record.
4652 * Use the shorter of 1s and dirty_expire_interval / 8.
4654 unsigned long update_intv =
4655 min_t(unsigned long, HZ,
4656 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4658 if (time_before64(frn->at, now - update_intv))
4660 } else if (oldest >= 0) {
4661 /* replace the oldest free one */
4662 frn = &memcg->cgwb_frn[oldest];
4663 frn->bdi_id = wb->bdi->id;
4664 frn->memcg_id = wb->memcg_css->id;
4669 /* issue foreign writeback flushes for recorded foreign dirtying events */
4670 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4672 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4673 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4674 u64 now = jiffies_64;
4677 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4678 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4681 * If the record is older than dirty_expire_interval,
4682 * writeback on it has already started. No need to kick it
4683 * off again. Also, don't start a new one if there's
4684 * already one in flight.
4686 if (time_after64(frn->at, now - intv) &&
4687 atomic_read(&frn->done.cnt) == 1) {
4689 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4690 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4691 WB_REASON_FOREIGN_FLUSH,
4697 #else /* CONFIG_CGROUP_WRITEBACK */
4699 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4704 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4708 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4712 #endif /* CONFIG_CGROUP_WRITEBACK */
4715 * DO NOT USE IN NEW FILES.
4717 * "cgroup.event_control" implementation.
4719 * This is way over-engineered. It tries to support fully configurable
4720 * events for each user. Such level of flexibility is completely
4721 * unnecessary especially in the light of the planned unified hierarchy.
4723 * Please deprecate this and replace with something simpler if at all
4728 * Unregister event and free resources.
4730 * Gets called from workqueue.
4732 static void memcg_event_remove(struct work_struct *work)
4734 struct mem_cgroup_event *event =
4735 container_of(work, struct mem_cgroup_event, remove);
4736 struct mem_cgroup *memcg = event->memcg;
4738 remove_wait_queue(event->wqh, &event->wait);
4740 event->unregister_event(memcg, event->eventfd);
4742 /* Notify userspace the event is going away. */
4743 eventfd_signal(event->eventfd, 1);
4745 eventfd_ctx_put(event->eventfd);
4747 css_put(&memcg->css);
4751 * Gets called on EPOLLHUP on eventfd when user closes it.
4753 * Called with wqh->lock held and interrupts disabled.
4755 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4756 int sync, void *key)
4758 struct mem_cgroup_event *event =
4759 container_of(wait, struct mem_cgroup_event, wait);
4760 struct mem_cgroup *memcg = event->memcg;
4761 __poll_t flags = key_to_poll(key);
4763 if (flags & EPOLLHUP) {
4765 * If the event has been detached at cgroup removal, we
4766 * can simply return knowing the other side will cleanup
4769 * We can't race against event freeing since the other
4770 * side will require wqh->lock via remove_wait_queue(),
4773 spin_lock(&memcg->event_list_lock);
4774 if (!list_empty(&event->list)) {
4775 list_del_init(&event->list);
4777 * We are in atomic context, but cgroup_event_remove()
4778 * may sleep, so we have to call it in workqueue.
4780 schedule_work(&event->remove);
4782 spin_unlock(&memcg->event_list_lock);
4788 static void memcg_event_ptable_queue_proc(struct file *file,
4789 wait_queue_head_t *wqh, poll_table *pt)
4791 struct mem_cgroup_event *event =
4792 container_of(pt, struct mem_cgroup_event, pt);
4795 add_wait_queue(wqh, &event->wait);
4799 * DO NOT USE IN NEW FILES.
4801 * Parse input and register new cgroup event handler.
4803 * Input must be in format '<event_fd> <control_fd> <args>'.
4804 * Interpretation of args is defined by control file implementation.
4806 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4807 char *buf, size_t nbytes, loff_t off)
4809 struct cgroup_subsys_state *css = of_css(of);
4810 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4811 struct mem_cgroup_event *event;
4812 struct cgroup_subsys_state *cfile_css;
4813 unsigned int efd, cfd;
4820 buf = strstrip(buf);
4822 efd = simple_strtoul(buf, &endp, 10);
4827 cfd = simple_strtoul(buf, &endp, 10);
4828 if ((*endp != ' ') && (*endp != '\0'))
4832 event = kzalloc(sizeof(*event), GFP_KERNEL);
4836 event->memcg = memcg;
4837 INIT_LIST_HEAD(&event->list);
4838 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4839 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4840 INIT_WORK(&event->remove, memcg_event_remove);
4848 event->eventfd = eventfd_ctx_fileget(efile.file);
4849 if (IS_ERR(event->eventfd)) {
4850 ret = PTR_ERR(event->eventfd);
4857 goto out_put_eventfd;
4860 /* the process need read permission on control file */
4861 /* AV: shouldn't we check that it's been opened for read instead? */
4862 ret = file_permission(cfile.file, MAY_READ);
4867 * Determine the event callbacks and set them in @event. This used
4868 * to be done via struct cftype but cgroup core no longer knows
4869 * about these events. The following is crude but the whole thing
4870 * is for compatibility anyway.
4872 * DO NOT ADD NEW FILES.
4874 name = cfile.file->f_path.dentry->d_name.name;
4876 if (!strcmp(name, "memory.usage_in_bytes")) {
4877 event->register_event = mem_cgroup_usage_register_event;
4878 event->unregister_event = mem_cgroup_usage_unregister_event;
4879 } else if (!strcmp(name, "memory.oom_control")) {
4880 event->register_event = mem_cgroup_oom_register_event;
4881 event->unregister_event = mem_cgroup_oom_unregister_event;
4882 } else if (!strcmp(name, "memory.pressure_level")) {
4883 event->register_event = vmpressure_register_event;
4884 event->unregister_event = vmpressure_unregister_event;
4885 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4886 event->register_event = memsw_cgroup_usage_register_event;
4887 event->unregister_event = memsw_cgroup_usage_unregister_event;
4894 * Verify @cfile should belong to @css. Also, remaining events are
4895 * automatically removed on cgroup destruction but the removal is
4896 * asynchronous, so take an extra ref on @css.
4898 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4899 &memory_cgrp_subsys);
4901 if (IS_ERR(cfile_css))
4903 if (cfile_css != css) {
4908 ret = event->register_event(memcg, event->eventfd, buf);
4912 vfs_poll(efile.file, &event->pt);
4914 spin_lock(&memcg->event_list_lock);
4915 list_add(&event->list, &memcg->event_list);
4916 spin_unlock(&memcg->event_list_lock);
4928 eventfd_ctx_put(event->eventfd);
4937 static struct cftype mem_cgroup_legacy_files[] = {
4939 .name = "usage_in_bytes",
4940 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4941 .read_u64 = mem_cgroup_read_u64,
4944 .name = "max_usage_in_bytes",
4945 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4946 .write = mem_cgroup_reset,
4947 .read_u64 = mem_cgroup_read_u64,
4950 .name = "limit_in_bytes",
4951 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4952 .write = mem_cgroup_write,
4953 .read_u64 = mem_cgroup_read_u64,
4956 .name = "soft_limit_in_bytes",
4957 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4958 .write = mem_cgroup_write,
4959 .read_u64 = mem_cgroup_read_u64,
4963 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4964 .write = mem_cgroup_reset,
4965 .read_u64 = mem_cgroup_read_u64,
4969 .seq_show = memcg_stat_show,
4972 .name = "force_empty",
4973 .write = mem_cgroup_force_empty_write,
4976 .name = "use_hierarchy",
4977 .write_u64 = mem_cgroup_hierarchy_write,
4978 .read_u64 = mem_cgroup_hierarchy_read,
4981 .name = "cgroup.event_control", /* XXX: for compat */
4982 .write = memcg_write_event_control,
4983 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4986 .name = "swappiness",
4987 .read_u64 = mem_cgroup_swappiness_read,
4988 .write_u64 = mem_cgroup_swappiness_write,
4991 .name = "move_charge_at_immigrate",
4992 .read_u64 = mem_cgroup_move_charge_read,
4993 .write_u64 = mem_cgroup_move_charge_write,
4996 .name = "oom_control",
4997 .seq_show = mem_cgroup_oom_control_read,
4998 .write_u64 = mem_cgroup_oom_control_write,
4999 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5002 .name = "pressure_level",
5006 .name = "numa_stat",
5007 .seq_show = memcg_numa_stat_show,
5011 .name = "kmem.limit_in_bytes",
5012 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5013 .write = mem_cgroup_write,
5014 .read_u64 = mem_cgroup_read_u64,
5017 .name = "kmem.usage_in_bytes",
5018 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5019 .read_u64 = mem_cgroup_read_u64,
5022 .name = "kmem.failcnt",
5023 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5024 .write = mem_cgroup_reset,
5025 .read_u64 = mem_cgroup_read_u64,
5028 .name = "kmem.max_usage_in_bytes",
5029 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5030 .write = mem_cgroup_reset,
5031 .read_u64 = mem_cgroup_read_u64,
5033 #if defined(CONFIG_MEMCG_KMEM) && \
5034 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5036 .name = "kmem.slabinfo",
5037 .seq_show = memcg_slab_show,
5041 .name = "kmem.tcp.limit_in_bytes",
5042 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5043 .write = mem_cgroup_write,
5044 .read_u64 = mem_cgroup_read_u64,
5047 .name = "kmem.tcp.usage_in_bytes",
5048 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5049 .read_u64 = mem_cgroup_read_u64,
5052 .name = "kmem.tcp.failcnt",
5053 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5054 .write = mem_cgroup_reset,
5055 .read_u64 = mem_cgroup_read_u64,
5058 .name = "kmem.tcp.max_usage_in_bytes",
5059 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5060 .write = mem_cgroup_reset,
5061 .read_u64 = mem_cgroup_read_u64,
5063 { }, /* terminate */
5067 * Private memory cgroup IDR
5069 * Swap-out records and page cache shadow entries need to store memcg
5070 * references in constrained space, so we maintain an ID space that is
5071 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5072 * memory-controlled cgroups to 64k.
5074 * However, there usually are many references to the offline CSS after
5075 * the cgroup has been destroyed, such as page cache or reclaimable
5076 * slab objects, that don't need to hang on to the ID. We want to keep
5077 * those dead CSS from occupying IDs, or we might quickly exhaust the
5078 * relatively small ID space and prevent the creation of new cgroups
5079 * even when there are much fewer than 64k cgroups - possibly none.
5081 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5082 * be freed and recycled when it's no longer needed, which is usually
5083 * when the CSS is offlined.
5085 * The only exception to that are records of swapped out tmpfs/shmem
5086 * pages that need to be attributed to live ancestors on swapin. But
5087 * those references are manageable from userspace.
5090 static DEFINE_IDR(mem_cgroup_idr);
5092 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5094 if (memcg->id.id > 0) {
5095 idr_remove(&mem_cgroup_idr, memcg->id.id);
5100 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5103 refcount_add(n, &memcg->id.ref);
5106 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5108 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5109 mem_cgroup_id_remove(memcg);
5111 /* Memcg ID pins CSS */
5112 css_put(&memcg->css);
5116 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5118 mem_cgroup_id_put_many(memcg, 1);
5122 * mem_cgroup_from_id - look up a memcg from a memcg id
5123 * @id: the memcg id to look up
5125 * Caller must hold rcu_read_lock().
5127 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5129 WARN_ON_ONCE(!rcu_read_lock_held());
5130 return idr_find(&mem_cgroup_idr, id);
5133 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5135 struct mem_cgroup_per_node *pn;
5138 * This routine is called against possible nodes.
5139 * But it's BUG to call kmalloc() against offline node.
5141 * TODO: this routine can waste much memory for nodes which will
5142 * never be onlined. It's better to use memory hotplug callback
5145 if (!node_state(node, N_NORMAL_MEMORY))
5147 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5151 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5152 GFP_KERNEL_ACCOUNT);
5153 if (!pn->lruvec_stat_local) {
5158 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct batched_lruvec_stat,
5159 GFP_KERNEL_ACCOUNT);
5160 if (!pn->lruvec_stat_cpu) {
5161 free_percpu(pn->lruvec_stat_local);
5166 lruvec_init(&pn->lruvec);
5167 pn->usage_in_excess = 0;
5168 pn->on_tree = false;
5171 memcg->nodeinfo[node] = pn;
5175 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5177 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5182 free_percpu(pn->lruvec_stat_cpu);
5183 free_percpu(pn->lruvec_stat_local);
5187 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5192 free_mem_cgroup_per_node_info(memcg, node);
5193 free_percpu(memcg->vmstats_percpu);
5194 free_percpu(memcg->vmstats_local);
5198 static void mem_cgroup_free(struct mem_cgroup *memcg)
5200 memcg_wb_domain_exit(memcg);
5202 * Flush percpu vmstats and vmevents to guarantee the value correctness
5203 * on parent's and all ancestor levels.
5205 memcg_flush_percpu_vmstats(memcg);
5206 memcg_flush_percpu_vmevents(memcg);
5207 __mem_cgroup_free(memcg);
5210 static struct mem_cgroup *mem_cgroup_alloc(void)
5212 struct mem_cgroup *memcg;
5215 int __maybe_unused i;
5216 long error = -ENOMEM;
5218 size = sizeof(struct mem_cgroup);
5219 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5221 memcg = kzalloc(size, GFP_KERNEL);
5223 return ERR_PTR(error);
5225 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5226 1, MEM_CGROUP_ID_MAX,
5228 if (memcg->id.id < 0) {
5229 error = memcg->id.id;
5233 memcg->vmstats_local = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5234 GFP_KERNEL_ACCOUNT);
5235 if (!memcg->vmstats_local)
5238 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5239 GFP_KERNEL_ACCOUNT);
5240 if (!memcg->vmstats_percpu)
5244 if (alloc_mem_cgroup_per_node_info(memcg, node))
5247 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5250 INIT_WORK(&memcg->high_work, high_work_func);
5251 INIT_LIST_HEAD(&memcg->oom_notify);
5252 mutex_init(&memcg->thresholds_lock);
5253 spin_lock_init(&memcg->move_lock);
5254 vmpressure_init(&memcg->vmpressure);
5255 INIT_LIST_HEAD(&memcg->event_list);
5256 spin_lock_init(&memcg->event_list_lock);
5257 memcg->socket_pressure = jiffies;
5258 #ifdef CONFIG_MEMCG_KMEM
5259 memcg->kmemcg_id = -1;
5260 INIT_LIST_HEAD(&memcg->objcg_list);
5262 #ifdef CONFIG_CGROUP_WRITEBACK
5263 INIT_LIST_HEAD(&memcg->cgwb_list);
5264 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5265 memcg->cgwb_frn[i].done =
5266 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5268 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5269 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5270 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5271 memcg->deferred_split_queue.split_queue_len = 0;
5273 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5276 mem_cgroup_id_remove(memcg);
5277 __mem_cgroup_free(memcg);
5278 return ERR_PTR(error);
5281 static struct cgroup_subsys_state * __ref
5282 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5284 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5285 struct mem_cgroup *memcg, *old_memcg;
5286 long error = -ENOMEM;
5288 old_memcg = set_active_memcg(parent);
5289 memcg = mem_cgroup_alloc();
5290 set_active_memcg(old_memcg);
5292 return ERR_CAST(memcg);
5294 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5295 memcg->soft_limit = PAGE_COUNTER_MAX;
5296 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5298 memcg->swappiness = mem_cgroup_swappiness(parent);
5299 memcg->oom_kill_disable = parent->oom_kill_disable;
5301 page_counter_init(&memcg->memory, &parent->memory);
5302 page_counter_init(&memcg->swap, &parent->swap);
5303 page_counter_init(&memcg->kmem, &parent->kmem);
5304 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5306 page_counter_init(&memcg->memory, NULL);
5307 page_counter_init(&memcg->swap, NULL);
5308 page_counter_init(&memcg->kmem, NULL);
5309 page_counter_init(&memcg->tcpmem, NULL);
5311 root_mem_cgroup = memcg;
5315 /* The following stuff does not apply to the root */
5316 error = memcg_online_kmem(memcg);
5320 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5321 static_branch_inc(&memcg_sockets_enabled_key);
5325 mem_cgroup_id_remove(memcg);
5326 mem_cgroup_free(memcg);
5327 return ERR_PTR(error);
5330 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5332 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5335 * A memcg must be visible for memcg_expand_shrinker_maps()
5336 * by the time the maps are allocated. So, we allocate maps
5337 * here, when for_each_mem_cgroup() can't skip it.
5339 if (memcg_alloc_shrinker_maps(memcg)) {
5340 mem_cgroup_id_remove(memcg);
5344 /* Online state pins memcg ID, memcg ID pins CSS */
5345 refcount_set(&memcg->id.ref, 1);
5350 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5352 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5353 struct mem_cgroup_event *event, *tmp;
5356 * Unregister events and notify userspace.
5357 * Notify userspace about cgroup removing only after rmdir of cgroup
5358 * directory to avoid race between userspace and kernelspace.
5360 spin_lock(&memcg->event_list_lock);
5361 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5362 list_del_init(&event->list);
5363 schedule_work(&event->remove);
5365 spin_unlock(&memcg->event_list_lock);
5367 page_counter_set_min(&memcg->memory, 0);
5368 page_counter_set_low(&memcg->memory, 0);
5370 memcg_offline_kmem(memcg);
5371 wb_memcg_offline(memcg);
5373 drain_all_stock(memcg);
5375 mem_cgroup_id_put(memcg);
5378 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5380 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5382 invalidate_reclaim_iterators(memcg);
5385 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5387 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5388 int __maybe_unused i;
5390 #ifdef CONFIG_CGROUP_WRITEBACK
5391 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5392 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5394 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5395 static_branch_dec(&memcg_sockets_enabled_key);
5397 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5398 static_branch_dec(&memcg_sockets_enabled_key);
5400 vmpressure_cleanup(&memcg->vmpressure);
5401 cancel_work_sync(&memcg->high_work);
5402 mem_cgroup_remove_from_trees(memcg);
5403 memcg_free_shrinker_maps(memcg);
5404 memcg_free_kmem(memcg);
5405 mem_cgroup_free(memcg);
5409 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5410 * @css: the target css
5412 * Reset the states of the mem_cgroup associated with @css. This is
5413 * invoked when the userland requests disabling on the default hierarchy
5414 * but the memcg is pinned through dependency. The memcg should stop
5415 * applying policies and should revert to the vanilla state as it may be
5416 * made visible again.
5418 * The current implementation only resets the essential configurations.
5419 * This needs to be expanded to cover all the visible parts.
5421 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5423 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5425 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5426 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5427 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5428 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5429 page_counter_set_min(&memcg->memory, 0);
5430 page_counter_set_low(&memcg->memory, 0);
5431 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5432 memcg->soft_limit = PAGE_COUNTER_MAX;
5433 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5434 memcg_wb_domain_size_changed(memcg);
5438 /* Handlers for move charge at task migration. */
5439 static int mem_cgroup_do_precharge(unsigned long count)
5443 /* Try a single bulk charge without reclaim first, kswapd may wake */
5444 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5446 mc.precharge += count;
5450 /* Try charges one by one with reclaim, but do not retry */
5452 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5466 enum mc_target_type {
5473 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5474 unsigned long addr, pte_t ptent)
5476 struct page *page = vm_normal_page(vma, addr, ptent);
5478 if (!page || !page_mapped(page))
5480 if (PageAnon(page)) {
5481 if (!(mc.flags & MOVE_ANON))
5484 if (!(mc.flags & MOVE_FILE))
5487 if (!get_page_unless_zero(page))
5493 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5494 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5495 pte_t ptent, swp_entry_t *entry)
5497 struct page *page = NULL;
5498 swp_entry_t ent = pte_to_swp_entry(ptent);
5500 if (!(mc.flags & MOVE_ANON))
5504 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5505 * a device and because they are not accessible by CPU they are store
5506 * as special swap entry in the CPU page table.
5508 if (is_device_private_entry(ent)) {
5509 page = device_private_entry_to_page(ent);
5511 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5512 * a refcount of 1 when free (unlike normal page)
5514 if (!page_ref_add_unless(page, 1, 1))
5519 if (non_swap_entry(ent))
5523 * Because lookup_swap_cache() updates some statistics counter,
5524 * we call find_get_page() with swapper_space directly.
5526 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5527 entry->val = ent.val;
5532 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5533 pte_t ptent, swp_entry_t *entry)
5539 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5540 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5542 if (!vma->vm_file) /* anonymous vma */
5544 if (!(mc.flags & MOVE_FILE))
5547 /* page is moved even if it's not RSS of this task(page-faulted). */
5548 /* shmem/tmpfs may report page out on swap: account for that too. */
5549 return find_get_incore_page(vma->vm_file->f_mapping,
5550 linear_page_index(vma, addr));
5554 * mem_cgroup_move_account - move account of the page
5556 * @compound: charge the page as compound or small page
5557 * @from: mem_cgroup which the page is moved from.
5558 * @to: mem_cgroup which the page is moved to. @from != @to.
5560 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5562 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5565 static int mem_cgroup_move_account(struct page *page,
5567 struct mem_cgroup *from,
5568 struct mem_cgroup *to)
5570 struct lruvec *from_vec, *to_vec;
5571 struct pglist_data *pgdat;
5572 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5575 VM_BUG_ON(from == to);
5576 VM_BUG_ON_PAGE(PageLRU(page), page);
5577 VM_BUG_ON(compound && !PageTransHuge(page));
5580 * Prevent mem_cgroup_migrate() from looking at
5581 * page's memory cgroup of its source page while we change it.
5584 if (!trylock_page(page))
5588 if (page_memcg(page) != from)
5591 pgdat = page_pgdat(page);
5592 from_vec = mem_cgroup_lruvec(from, pgdat);
5593 to_vec = mem_cgroup_lruvec(to, pgdat);
5595 lock_page_memcg(page);
5597 if (PageAnon(page)) {
5598 if (page_mapped(page)) {
5599 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5600 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5601 if (PageTransHuge(page)) {
5602 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5604 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5609 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5610 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5612 if (PageSwapBacked(page)) {
5613 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5614 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5617 if (page_mapped(page)) {
5618 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5619 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5622 if (PageDirty(page)) {
5623 struct address_space *mapping = page_mapping(page);
5625 if (mapping_can_writeback(mapping)) {
5626 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5628 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5634 if (PageWriteback(page)) {
5635 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5636 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5640 * All state has been migrated, let's switch to the new memcg.
5642 * It is safe to change page's memcg here because the page
5643 * is referenced, charged, isolated, and locked: we can't race
5644 * with (un)charging, migration, LRU putback, or anything else
5645 * that would rely on a stable page's memory cgroup.
5647 * Note that lock_page_memcg is a memcg lock, not a page lock,
5648 * to save space. As soon as we switch page's memory cgroup to a
5649 * new memcg that isn't locked, the above state can change
5650 * concurrently again. Make sure we're truly done with it.
5655 css_put(&from->css);
5657 page->memcg_data = (unsigned long)to;
5659 __unlock_page_memcg(from);
5663 local_irq_disable();
5664 mem_cgroup_charge_statistics(to, page, nr_pages);
5665 memcg_check_events(to, page);
5666 mem_cgroup_charge_statistics(from, page, -nr_pages);
5667 memcg_check_events(from, page);
5676 * get_mctgt_type - get target type of moving charge
5677 * @vma: the vma the pte to be checked belongs
5678 * @addr: the address corresponding to the pte to be checked
5679 * @ptent: the pte to be checked
5680 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5683 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5684 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5685 * move charge. if @target is not NULL, the page is stored in target->page
5686 * with extra refcnt got(Callers should handle it).
5687 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5688 * target for charge migration. if @target is not NULL, the entry is stored
5690 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5691 * (so ZONE_DEVICE page and thus not on the lru).
5692 * For now we such page is charge like a regular page would be as for all
5693 * intent and purposes it is just special memory taking the place of a
5696 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5698 * Called with pte lock held.
5701 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5702 unsigned long addr, pte_t ptent, union mc_target *target)
5704 struct page *page = NULL;
5705 enum mc_target_type ret = MC_TARGET_NONE;
5706 swp_entry_t ent = { .val = 0 };
5708 if (pte_present(ptent))
5709 page = mc_handle_present_pte(vma, addr, ptent);
5710 else if (is_swap_pte(ptent))
5711 page = mc_handle_swap_pte(vma, ptent, &ent);
5712 else if (pte_none(ptent))
5713 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5715 if (!page && !ent.val)
5719 * Do only loose check w/o serialization.
5720 * mem_cgroup_move_account() checks the page is valid or
5721 * not under LRU exclusion.
5723 if (page_memcg(page) == mc.from) {
5724 ret = MC_TARGET_PAGE;
5725 if (is_device_private_page(page))
5726 ret = MC_TARGET_DEVICE;
5728 target->page = page;
5730 if (!ret || !target)
5734 * There is a swap entry and a page doesn't exist or isn't charged.
5735 * But we cannot move a tail-page in a THP.
5737 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5738 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5739 ret = MC_TARGET_SWAP;
5746 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5748 * We don't consider PMD mapped swapping or file mapped pages because THP does
5749 * not support them for now.
5750 * Caller should make sure that pmd_trans_huge(pmd) is true.
5752 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5753 unsigned long addr, pmd_t pmd, union mc_target *target)
5755 struct page *page = NULL;
5756 enum mc_target_type ret = MC_TARGET_NONE;
5758 if (unlikely(is_swap_pmd(pmd))) {
5759 VM_BUG_ON(thp_migration_supported() &&
5760 !is_pmd_migration_entry(pmd));
5763 page = pmd_page(pmd);
5764 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5765 if (!(mc.flags & MOVE_ANON))
5767 if (page_memcg(page) == mc.from) {
5768 ret = MC_TARGET_PAGE;
5771 target->page = page;
5777 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5778 unsigned long addr, pmd_t pmd, union mc_target *target)
5780 return MC_TARGET_NONE;
5784 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5785 unsigned long addr, unsigned long end,
5786 struct mm_walk *walk)
5788 struct vm_area_struct *vma = walk->vma;
5792 ptl = pmd_trans_huge_lock(pmd, vma);
5795 * Note their can not be MC_TARGET_DEVICE for now as we do not
5796 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5797 * this might change.
5799 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5800 mc.precharge += HPAGE_PMD_NR;
5805 if (pmd_trans_unstable(pmd))
5807 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5808 for (; addr != end; pte++, addr += PAGE_SIZE)
5809 if (get_mctgt_type(vma, addr, *pte, NULL))
5810 mc.precharge++; /* increment precharge temporarily */
5811 pte_unmap_unlock(pte - 1, ptl);
5817 static const struct mm_walk_ops precharge_walk_ops = {
5818 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5821 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5823 unsigned long precharge;
5826 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5827 mmap_read_unlock(mm);
5829 precharge = mc.precharge;
5835 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5837 unsigned long precharge = mem_cgroup_count_precharge(mm);
5839 VM_BUG_ON(mc.moving_task);
5840 mc.moving_task = current;
5841 return mem_cgroup_do_precharge(precharge);
5844 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5845 static void __mem_cgroup_clear_mc(void)
5847 struct mem_cgroup *from = mc.from;
5848 struct mem_cgroup *to = mc.to;
5850 /* we must uncharge all the leftover precharges from mc.to */
5852 cancel_charge(mc.to, mc.precharge);
5856 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5857 * we must uncharge here.
5859 if (mc.moved_charge) {
5860 cancel_charge(mc.from, mc.moved_charge);
5861 mc.moved_charge = 0;
5863 /* we must fixup refcnts and charges */
5864 if (mc.moved_swap) {
5865 /* uncharge swap account from the old cgroup */
5866 if (!mem_cgroup_is_root(mc.from))
5867 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5869 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5872 * we charged both to->memory and to->memsw, so we
5873 * should uncharge to->memory.
5875 if (!mem_cgroup_is_root(mc.to))
5876 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5880 memcg_oom_recover(from);
5881 memcg_oom_recover(to);
5882 wake_up_all(&mc.waitq);
5885 static void mem_cgroup_clear_mc(void)
5887 struct mm_struct *mm = mc.mm;
5890 * we must clear moving_task before waking up waiters at the end of
5893 mc.moving_task = NULL;
5894 __mem_cgroup_clear_mc();
5895 spin_lock(&mc.lock);
5899 spin_unlock(&mc.lock);
5904 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5906 struct cgroup_subsys_state *css;
5907 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5908 struct mem_cgroup *from;
5909 struct task_struct *leader, *p;
5910 struct mm_struct *mm;
5911 unsigned long move_flags;
5914 /* charge immigration isn't supported on the default hierarchy */
5915 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5919 * Multi-process migrations only happen on the default hierarchy
5920 * where charge immigration is not used. Perform charge
5921 * immigration if @tset contains a leader and whine if there are
5925 cgroup_taskset_for_each_leader(leader, css, tset) {
5928 memcg = mem_cgroup_from_css(css);
5934 * We are now commited to this value whatever it is. Changes in this
5935 * tunable will only affect upcoming migrations, not the current one.
5936 * So we need to save it, and keep it going.
5938 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5942 from = mem_cgroup_from_task(p);
5944 VM_BUG_ON(from == memcg);
5946 mm = get_task_mm(p);
5949 /* We move charges only when we move a owner of the mm */
5950 if (mm->owner == p) {
5953 VM_BUG_ON(mc.precharge);
5954 VM_BUG_ON(mc.moved_charge);
5955 VM_BUG_ON(mc.moved_swap);
5957 spin_lock(&mc.lock);
5961 mc.flags = move_flags;
5962 spin_unlock(&mc.lock);
5963 /* We set mc.moving_task later */
5965 ret = mem_cgroup_precharge_mc(mm);
5967 mem_cgroup_clear_mc();
5974 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5977 mem_cgroup_clear_mc();
5980 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5981 unsigned long addr, unsigned long end,
5982 struct mm_walk *walk)
5985 struct vm_area_struct *vma = walk->vma;
5988 enum mc_target_type target_type;
5989 union mc_target target;
5992 ptl = pmd_trans_huge_lock(pmd, vma);
5994 if (mc.precharge < HPAGE_PMD_NR) {
5998 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5999 if (target_type == MC_TARGET_PAGE) {
6001 if (!isolate_lru_page(page)) {
6002 if (!mem_cgroup_move_account(page, true,
6004 mc.precharge -= HPAGE_PMD_NR;
6005 mc.moved_charge += HPAGE_PMD_NR;
6007 putback_lru_page(page);
6010 } else if (target_type == MC_TARGET_DEVICE) {
6012 if (!mem_cgroup_move_account(page, true,
6014 mc.precharge -= HPAGE_PMD_NR;
6015 mc.moved_charge += HPAGE_PMD_NR;
6023 if (pmd_trans_unstable(pmd))
6026 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6027 for (; addr != end; addr += PAGE_SIZE) {
6028 pte_t ptent = *(pte++);
6029 bool device = false;
6035 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6036 case MC_TARGET_DEVICE:
6039 case MC_TARGET_PAGE:
6042 * We can have a part of the split pmd here. Moving it
6043 * can be done but it would be too convoluted so simply
6044 * ignore such a partial THP and keep it in original
6045 * memcg. There should be somebody mapping the head.
6047 if (PageTransCompound(page))
6049 if (!device && isolate_lru_page(page))
6051 if (!mem_cgroup_move_account(page, false,
6054 /* we uncharge from mc.from later. */
6058 putback_lru_page(page);
6059 put: /* get_mctgt_type() gets the page */
6062 case MC_TARGET_SWAP:
6064 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6066 mem_cgroup_id_get_many(mc.to, 1);
6067 /* we fixup other refcnts and charges later. */
6075 pte_unmap_unlock(pte - 1, ptl);
6080 * We have consumed all precharges we got in can_attach().
6081 * We try charge one by one, but don't do any additional
6082 * charges to mc.to if we have failed in charge once in attach()
6085 ret = mem_cgroup_do_precharge(1);
6093 static const struct mm_walk_ops charge_walk_ops = {
6094 .pmd_entry = mem_cgroup_move_charge_pte_range,
6097 static void mem_cgroup_move_charge(void)
6099 lru_add_drain_all();
6101 * Signal lock_page_memcg() to take the memcg's move_lock
6102 * while we're moving its pages to another memcg. Then wait
6103 * for already started RCU-only updates to finish.
6105 atomic_inc(&mc.from->moving_account);
6108 if (unlikely(!mmap_read_trylock(mc.mm))) {
6110 * Someone who are holding the mmap_lock might be waiting in
6111 * waitq. So we cancel all extra charges, wake up all waiters,
6112 * and retry. Because we cancel precharges, we might not be able
6113 * to move enough charges, but moving charge is a best-effort
6114 * feature anyway, so it wouldn't be a big problem.
6116 __mem_cgroup_clear_mc();
6121 * When we have consumed all precharges and failed in doing
6122 * additional charge, the page walk just aborts.
6124 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6127 mmap_read_unlock(mc.mm);
6128 atomic_dec(&mc.from->moving_account);
6131 static void mem_cgroup_move_task(void)
6134 mem_cgroup_move_charge();
6135 mem_cgroup_clear_mc();
6138 #else /* !CONFIG_MMU */
6139 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6143 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6146 static void mem_cgroup_move_task(void)
6151 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6153 if (value == PAGE_COUNTER_MAX)
6154 seq_puts(m, "max\n");
6156 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6161 static u64 memory_current_read(struct cgroup_subsys_state *css,
6164 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6166 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6169 static int memory_min_show(struct seq_file *m, void *v)
6171 return seq_puts_memcg_tunable(m,
6172 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6175 static ssize_t memory_min_write(struct kernfs_open_file *of,
6176 char *buf, size_t nbytes, loff_t off)
6178 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6182 buf = strstrip(buf);
6183 err = page_counter_memparse(buf, "max", &min);
6187 page_counter_set_min(&memcg->memory, min);
6192 static int memory_low_show(struct seq_file *m, void *v)
6194 return seq_puts_memcg_tunable(m,
6195 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6198 static ssize_t memory_low_write(struct kernfs_open_file *of,
6199 char *buf, size_t nbytes, loff_t off)
6201 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6205 buf = strstrip(buf);
6206 err = page_counter_memparse(buf, "max", &low);
6210 page_counter_set_low(&memcg->memory, low);
6215 static int memory_high_show(struct seq_file *m, void *v)
6217 return seq_puts_memcg_tunable(m,
6218 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6221 static ssize_t memory_high_write(struct kernfs_open_file *of,
6222 char *buf, size_t nbytes, loff_t off)
6224 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6225 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6226 bool drained = false;
6230 buf = strstrip(buf);
6231 err = page_counter_memparse(buf, "max", &high);
6235 page_counter_set_high(&memcg->memory, high);
6238 unsigned long nr_pages = page_counter_read(&memcg->memory);
6239 unsigned long reclaimed;
6241 if (nr_pages <= high)
6244 if (signal_pending(current))
6248 drain_all_stock(memcg);
6253 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6256 if (!reclaimed && !nr_retries--)
6260 memcg_wb_domain_size_changed(memcg);
6264 static int memory_max_show(struct seq_file *m, void *v)
6266 return seq_puts_memcg_tunable(m,
6267 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6270 static ssize_t memory_max_write(struct kernfs_open_file *of,
6271 char *buf, size_t nbytes, loff_t off)
6273 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6274 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6275 bool drained = false;
6279 buf = strstrip(buf);
6280 err = page_counter_memparse(buf, "max", &max);
6284 xchg(&memcg->memory.max, max);
6287 unsigned long nr_pages = page_counter_read(&memcg->memory);
6289 if (nr_pages <= max)
6292 if (signal_pending(current))
6296 drain_all_stock(memcg);
6302 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6308 memcg_memory_event(memcg, MEMCG_OOM);
6309 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6313 memcg_wb_domain_size_changed(memcg);
6317 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6319 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6320 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6321 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6322 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6323 seq_printf(m, "oom_kill %lu\n",
6324 atomic_long_read(&events[MEMCG_OOM_KILL]));
6327 static int memory_events_show(struct seq_file *m, void *v)
6329 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6331 __memory_events_show(m, memcg->memory_events);
6335 static int memory_events_local_show(struct seq_file *m, void *v)
6337 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6339 __memory_events_show(m, memcg->memory_events_local);
6343 static int memory_stat_show(struct seq_file *m, void *v)
6345 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6348 buf = memory_stat_format(memcg);
6357 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6360 return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6363 static int memory_numa_stat_show(struct seq_file *m, void *v)
6366 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6368 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6371 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6374 seq_printf(m, "%s", memory_stats[i].name);
6375 for_each_node_state(nid, N_MEMORY) {
6377 struct lruvec *lruvec;
6379 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6380 size = lruvec_page_state_output(lruvec,
6381 memory_stats[i].idx);
6382 seq_printf(m, " N%d=%llu", nid, size);
6391 static int memory_oom_group_show(struct seq_file *m, void *v)
6393 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6395 seq_printf(m, "%d\n", memcg->oom_group);
6400 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6401 char *buf, size_t nbytes, loff_t off)
6403 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6406 buf = strstrip(buf);
6410 ret = kstrtoint(buf, 0, &oom_group);
6414 if (oom_group != 0 && oom_group != 1)
6417 memcg->oom_group = oom_group;
6422 static struct cftype memory_files[] = {
6425 .flags = CFTYPE_NOT_ON_ROOT,
6426 .read_u64 = memory_current_read,
6430 .flags = CFTYPE_NOT_ON_ROOT,
6431 .seq_show = memory_min_show,
6432 .write = memory_min_write,
6436 .flags = CFTYPE_NOT_ON_ROOT,
6437 .seq_show = memory_low_show,
6438 .write = memory_low_write,
6442 .flags = CFTYPE_NOT_ON_ROOT,
6443 .seq_show = memory_high_show,
6444 .write = memory_high_write,
6448 .flags = CFTYPE_NOT_ON_ROOT,
6449 .seq_show = memory_max_show,
6450 .write = memory_max_write,
6454 .flags = CFTYPE_NOT_ON_ROOT,
6455 .file_offset = offsetof(struct mem_cgroup, events_file),
6456 .seq_show = memory_events_show,
6459 .name = "events.local",
6460 .flags = CFTYPE_NOT_ON_ROOT,
6461 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6462 .seq_show = memory_events_local_show,
6466 .seq_show = memory_stat_show,
6470 .name = "numa_stat",
6471 .seq_show = memory_numa_stat_show,
6475 .name = "oom.group",
6476 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6477 .seq_show = memory_oom_group_show,
6478 .write = memory_oom_group_write,
6483 struct cgroup_subsys memory_cgrp_subsys = {
6484 .css_alloc = mem_cgroup_css_alloc,
6485 .css_online = mem_cgroup_css_online,
6486 .css_offline = mem_cgroup_css_offline,
6487 .css_released = mem_cgroup_css_released,
6488 .css_free = mem_cgroup_css_free,
6489 .css_reset = mem_cgroup_css_reset,
6490 .can_attach = mem_cgroup_can_attach,
6491 .cancel_attach = mem_cgroup_cancel_attach,
6492 .post_attach = mem_cgroup_move_task,
6493 .dfl_cftypes = memory_files,
6494 .legacy_cftypes = mem_cgroup_legacy_files,
6499 * This function calculates an individual cgroup's effective
6500 * protection which is derived from its own memory.min/low, its
6501 * parent's and siblings' settings, as well as the actual memory
6502 * distribution in the tree.
6504 * The following rules apply to the effective protection values:
6506 * 1. At the first level of reclaim, effective protection is equal to
6507 * the declared protection in memory.min and memory.low.
6509 * 2. To enable safe delegation of the protection configuration, at
6510 * subsequent levels the effective protection is capped to the
6511 * parent's effective protection.
6513 * 3. To make complex and dynamic subtrees easier to configure, the
6514 * user is allowed to overcommit the declared protection at a given
6515 * level. If that is the case, the parent's effective protection is
6516 * distributed to the children in proportion to how much protection
6517 * they have declared and how much of it they are utilizing.
6519 * This makes distribution proportional, but also work-conserving:
6520 * if one cgroup claims much more protection than it uses memory,
6521 * the unused remainder is available to its siblings.
6523 * 4. Conversely, when the declared protection is undercommitted at a
6524 * given level, the distribution of the larger parental protection
6525 * budget is NOT proportional. A cgroup's protection from a sibling
6526 * is capped to its own memory.min/low setting.
6528 * 5. However, to allow protecting recursive subtrees from each other
6529 * without having to declare each individual cgroup's fixed share
6530 * of the ancestor's claim to protection, any unutilized -
6531 * "floating" - protection from up the tree is distributed in
6532 * proportion to each cgroup's *usage*. This makes the protection
6533 * neutral wrt sibling cgroups and lets them compete freely over
6534 * the shared parental protection budget, but it protects the
6535 * subtree as a whole from neighboring subtrees.
6537 * Note that 4. and 5. are not in conflict: 4. is about protecting
6538 * against immediate siblings whereas 5. is about protecting against
6539 * neighboring subtrees.
6541 static unsigned long effective_protection(unsigned long usage,
6542 unsigned long parent_usage,
6543 unsigned long setting,
6544 unsigned long parent_effective,
6545 unsigned long siblings_protected)
6547 unsigned long protected;
6550 protected = min(usage, setting);
6552 * If all cgroups at this level combined claim and use more
6553 * protection then what the parent affords them, distribute
6554 * shares in proportion to utilization.
6556 * We are using actual utilization rather than the statically
6557 * claimed protection in order to be work-conserving: claimed
6558 * but unused protection is available to siblings that would
6559 * otherwise get a smaller chunk than what they claimed.
6561 if (siblings_protected > parent_effective)
6562 return protected * parent_effective / siblings_protected;
6565 * Ok, utilized protection of all children is within what the
6566 * parent affords them, so we know whatever this child claims
6567 * and utilizes is effectively protected.
6569 * If there is unprotected usage beyond this value, reclaim
6570 * will apply pressure in proportion to that amount.
6572 * If there is unutilized protection, the cgroup will be fully
6573 * shielded from reclaim, but we do return a smaller value for
6574 * protection than what the group could enjoy in theory. This
6575 * is okay. With the overcommit distribution above, effective
6576 * protection is always dependent on how memory is actually
6577 * consumed among the siblings anyway.
6582 * If the children aren't claiming (all of) the protection
6583 * afforded to them by the parent, distribute the remainder in
6584 * proportion to the (unprotected) memory of each cgroup. That
6585 * way, cgroups that aren't explicitly prioritized wrt each
6586 * other compete freely over the allowance, but they are
6587 * collectively protected from neighboring trees.
6589 * We're using unprotected memory for the weight so that if
6590 * some cgroups DO claim explicit protection, we don't protect
6591 * the same bytes twice.
6593 * Check both usage and parent_usage against the respective
6594 * protected values. One should imply the other, but they
6595 * aren't read atomically - make sure the division is sane.
6597 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6599 if (parent_effective > siblings_protected &&
6600 parent_usage > siblings_protected &&
6601 usage > protected) {
6602 unsigned long unclaimed;
6604 unclaimed = parent_effective - siblings_protected;
6605 unclaimed *= usage - protected;
6606 unclaimed /= parent_usage - siblings_protected;
6615 * mem_cgroup_protected - check if memory consumption is in the normal range
6616 * @root: the top ancestor of the sub-tree being checked
6617 * @memcg: the memory cgroup to check
6619 * WARNING: This function is not stateless! It can only be used as part
6620 * of a top-down tree iteration, not for isolated queries.
6622 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6623 struct mem_cgroup *memcg)
6625 unsigned long usage, parent_usage;
6626 struct mem_cgroup *parent;
6628 if (mem_cgroup_disabled())
6632 root = root_mem_cgroup;
6635 * Effective values of the reclaim targets are ignored so they
6636 * can be stale. Have a look at mem_cgroup_protection for more
6638 * TODO: calculation should be more robust so that we do not need
6639 * that special casing.
6644 usage = page_counter_read(&memcg->memory);
6648 parent = parent_mem_cgroup(memcg);
6649 /* No parent means a non-hierarchical mode on v1 memcg */
6653 if (parent == root) {
6654 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6655 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6659 parent_usage = page_counter_read(&parent->memory);
6661 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6662 READ_ONCE(memcg->memory.min),
6663 READ_ONCE(parent->memory.emin),
6664 atomic_long_read(&parent->memory.children_min_usage)));
6666 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6667 READ_ONCE(memcg->memory.low),
6668 READ_ONCE(parent->memory.elow),
6669 atomic_long_read(&parent->memory.children_low_usage)));
6673 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6674 * @page: page to charge
6675 * @mm: mm context of the victim
6676 * @gfp_mask: reclaim mode
6678 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6679 * pages according to @gfp_mask if necessary.
6681 * Returns 0 on success. Otherwise, an error code is returned.
6683 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6685 unsigned int nr_pages = thp_nr_pages(page);
6686 struct mem_cgroup *memcg = NULL;
6689 if (mem_cgroup_disabled())
6692 if (PageSwapCache(page)) {
6693 swp_entry_t ent = { .val = page_private(page), };
6697 * Every swap fault against a single page tries to charge the
6698 * page, bail as early as possible. shmem_unuse() encounters
6699 * already charged pages, too. page and memcg binding is
6700 * protected by the page lock, which serializes swap cache
6701 * removal, which in turn serializes uncharging.
6703 VM_BUG_ON_PAGE(!PageLocked(page), page);
6704 if (page_memcg(compound_head(page)))
6707 id = lookup_swap_cgroup_id(ent);
6709 memcg = mem_cgroup_from_id(id);
6710 if (memcg && !css_tryget_online(&memcg->css))
6716 memcg = get_mem_cgroup_from_mm(mm);
6718 ret = try_charge(memcg, gfp_mask, nr_pages);
6722 css_get(&memcg->css);
6723 commit_charge(page, memcg);
6725 local_irq_disable();
6726 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6727 memcg_check_events(memcg, page);
6731 * Cgroup1's unified memory+swap counter has been charged with the
6732 * new swapcache page, finish the transfer by uncharging the swap
6733 * slot. The swap slot would also get uncharged when it dies, but
6734 * it can stick around indefinitely and we'd count the page twice
6737 * Cgroup2 has separate resource counters for memory and swap,
6738 * so this is a non-issue here. Memory and swap charge lifetimes
6739 * correspond 1:1 to page and swap slot lifetimes: we charge the
6740 * page to memory here, and uncharge swap when the slot is freed.
6742 if (do_memsw_account() && PageSwapCache(page)) {
6743 swp_entry_t entry = { .val = page_private(page) };
6745 * The swap entry might not get freed for a long time,
6746 * let's not wait for it. The page already received a
6747 * memory+swap charge, drop the swap entry duplicate.
6749 mem_cgroup_uncharge_swap(entry, nr_pages);
6753 css_put(&memcg->css);
6758 struct uncharge_gather {
6759 struct mem_cgroup *memcg;
6760 unsigned long nr_pages;
6761 unsigned long pgpgout;
6762 unsigned long nr_kmem;
6763 struct page *dummy_page;
6766 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6768 memset(ug, 0, sizeof(*ug));
6771 static void uncharge_batch(const struct uncharge_gather *ug)
6773 unsigned long flags;
6775 if (!mem_cgroup_is_root(ug->memcg)) {
6776 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6777 if (do_memsw_account())
6778 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6779 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6780 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6781 memcg_oom_recover(ug->memcg);
6784 local_irq_save(flags);
6785 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6786 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6787 memcg_check_events(ug->memcg, ug->dummy_page);
6788 local_irq_restore(flags);
6790 /* drop reference from uncharge_page */
6791 css_put(&ug->memcg->css);
6794 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6796 unsigned long nr_pages;
6798 VM_BUG_ON_PAGE(PageLRU(page), page);
6800 if (!page_memcg(page))
6804 * Nobody should be changing or seriously looking at
6805 * page_memcg(page) at this point, we have fully
6806 * exclusive access to the page.
6809 if (ug->memcg != page_memcg(page)) {
6812 uncharge_gather_clear(ug);
6814 ug->memcg = page_memcg(page);
6816 /* pairs with css_put in uncharge_batch */
6817 css_get(&ug->memcg->css);
6820 nr_pages = compound_nr(page);
6821 ug->nr_pages += nr_pages;
6823 if (PageMemcgKmem(page))
6824 ug->nr_kmem += nr_pages;
6828 ug->dummy_page = page;
6829 page->memcg_data = 0;
6830 css_put(&ug->memcg->css);
6834 * mem_cgroup_uncharge - uncharge a page
6835 * @page: page to uncharge
6837 * Uncharge a page previously charged with mem_cgroup_charge().
6839 void mem_cgroup_uncharge(struct page *page)
6841 struct uncharge_gather ug;
6843 if (mem_cgroup_disabled())
6846 /* Don't touch page->lru of any random page, pre-check: */
6847 if (!page_memcg(page))
6850 uncharge_gather_clear(&ug);
6851 uncharge_page(page, &ug);
6852 uncharge_batch(&ug);
6856 * mem_cgroup_uncharge_list - uncharge a list of page
6857 * @page_list: list of pages to uncharge
6859 * Uncharge a list of pages previously charged with
6860 * mem_cgroup_charge().
6862 void mem_cgroup_uncharge_list(struct list_head *page_list)
6864 struct uncharge_gather ug;
6867 if (mem_cgroup_disabled())
6870 uncharge_gather_clear(&ug);
6871 list_for_each_entry(page, page_list, lru)
6872 uncharge_page(page, &ug);
6874 uncharge_batch(&ug);
6878 * mem_cgroup_migrate - charge a page's replacement
6879 * @oldpage: currently circulating page
6880 * @newpage: replacement page
6882 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6883 * be uncharged upon free.
6885 * Both pages must be locked, @newpage->mapping must be set up.
6887 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6889 struct mem_cgroup *memcg;
6890 unsigned int nr_pages;
6891 unsigned long flags;
6893 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6894 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6895 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6896 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6899 if (mem_cgroup_disabled())
6902 /* Page cache replacement: new page already charged? */
6903 if (page_memcg(newpage))
6906 memcg = page_memcg(oldpage);
6907 VM_WARN_ON_ONCE_PAGE(!memcg, oldpage);
6911 /* Force-charge the new page. The old one will be freed soon */
6912 nr_pages = thp_nr_pages(newpage);
6914 page_counter_charge(&memcg->memory, nr_pages);
6915 if (do_memsw_account())
6916 page_counter_charge(&memcg->memsw, nr_pages);
6918 css_get(&memcg->css);
6919 commit_charge(newpage, memcg);
6921 local_irq_save(flags);
6922 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6923 memcg_check_events(memcg, newpage);
6924 local_irq_restore(flags);
6927 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6928 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6930 void mem_cgroup_sk_alloc(struct sock *sk)
6932 struct mem_cgroup *memcg;
6934 if (!mem_cgroup_sockets_enabled)
6937 /* Do not associate the sock with unrelated interrupted task's memcg. */
6942 memcg = mem_cgroup_from_task(current);
6943 if (memcg == root_mem_cgroup)
6945 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6947 if (css_tryget(&memcg->css))
6948 sk->sk_memcg = memcg;
6953 void mem_cgroup_sk_free(struct sock *sk)
6956 css_put(&sk->sk_memcg->css);
6960 * mem_cgroup_charge_skmem - charge socket memory
6961 * @memcg: memcg to charge
6962 * @nr_pages: number of pages to charge
6964 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6965 * @memcg's configured limit, %false if the charge had to be forced.
6967 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6969 gfp_t gfp_mask = GFP_KERNEL;
6971 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6972 struct page_counter *fail;
6974 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6975 memcg->tcpmem_pressure = 0;
6978 page_counter_charge(&memcg->tcpmem, nr_pages);
6979 memcg->tcpmem_pressure = 1;
6983 /* Don't block in the packet receive path */
6985 gfp_mask = GFP_NOWAIT;
6987 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6989 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6992 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6997 * mem_cgroup_uncharge_skmem - uncharge socket memory
6998 * @memcg: memcg to uncharge
6999 * @nr_pages: number of pages to uncharge
7001 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7003 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7004 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7008 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7010 refill_stock(memcg, nr_pages);
7013 static int __init cgroup_memory(char *s)
7017 while ((token = strsep(&s, ",")) != NULL) {
7020 if (!strcmp(token, "nosocket"))
7021 cgroup_memory_nosocket = true;
7022 if (!strcmp(token, "nokmem"))
7023 cgroup_memory_nokmem = true;
7027 __setup("cgroup.memory=", cgroup_memory);
7030 * subsys_initcall() for memory controller.
7032 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7033 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7034 * basically everything that doesn't depend on a specific mem_cgroup structure
7035 * should be initialized from here.
7037 static int __init mem_cgroup_init(void)
7042 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7043 * used for per-memcg-per-cpu caching of per-node statistics. In order
7044 * to work fine, we should make sure that the overfill threshold can't
7045 * exceed S32_MAX / PAGE_SIZE.
7047 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7049 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7050 memcg_hotplug_cpu_dead);
7052 for_each_possible_cpu(cpu)
7053 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7056 for_each_node(node) {
7057 struct mem_cgroup_tree_per_node *rtpn;
7059 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7060 node_online(node) ? node : NUMA_NO_NODE);
7062 rtpn->rb_root = RB_ROOT;
7063 rtpn->rb_rightmost = NULL;
7064 spin_lock_init(&rtpn->lock);
7065 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7070 subsys_initcall(mem_cgroup_init);
7072 #ifdef CONFIG_MEMCG_SWAP
7073 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7075 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7077 * The root cgroup cannot be destroyed, so it's refcount must
7080 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7084 memcg = parent_mem_cgroup(memcg);
7086 memcg = root_mem_cgroup;
7092 * mem_cgroup_swapout - transfer a memsw charge to swap
7093 * @page: page whose memsw charge to transfer
7094 * @entry: swap entry to move the charge to
7096 * Transfer the memsw charge of @page to @entry.
7098 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7100 struct mem_cgroup *memcg, *swap_memcg;
7101 unsigned int nr_entries;
7102 unsigned short oldid;
7104 VM_BUG_ON_PAGE(PageLRU(page), page);
7105 VM_BUG_ON_PAGE(page_count(page), page);
7107 if (mem_cgroup_disabled())
7110 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7113 memcg = page_memcg(page);
7115 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7120 * In case the memcg owning these pages has been offlined and doesn't
7121 * have an ID allocated to it anymore, charge the closest online
7122 * ancestor for the swap instead and transfer the memory+swap charge.
7124 swap_memcg = mem_cgroup_id_get_online(memcg);
7125 nr_entries = thp_nr_pages(page);
7126 /* Get references for the tail pages, too */
7128 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7129 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7131 VM_BUG_ON_PAGE(oldid, page);
7132 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7134 page->memcg_data = 0;
7136 if (!mem_cgroup_is_root(memcg))
7137 page_counter_uncharge(&memcg->memory, nr_entries);
7139 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7140 if (!mem_cgroup_is_root(swap_memcg))
7141 page_counter_charge(&swap_memcg->memsw, nr_entries);
7142 page_counter_uncharge(&memcg->memsw, nr_entries);
7146 * Interrupts should be disabled here because the caller holds the
7147 * i_pages lock which is taken with interrupts-off. It is
7148 * important here to have the interrupts disabled because it is the
7149 * only synchronisation we have for updating the per-CPU variables.
7151 VM_BUG_ON(!irqs_disabled());
7152 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7153 memcg_check_events(memcg, page);
7155 css_put(&memcg->css);
7159 * mem_cgroup_try_charge_swap - try charging swap space for a page
7160 * @page: page being added to swap
7161 * @entry: swap entry to charge
7163 * Try to charge @page's memcg for the swap space at @entry.
7165 * Returns 0 on success, -ENOMEM on failure.
7167 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7169 unsigned int nr_pages = thp_nr_pages(page);
7170 struct page_counter *counter;
7171 struct mem_cgroup *memcg;
7172 unsigned short oldid;
7174 if (mem_cgroup_disabled())
7177 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7180 memcg = page_memcg(page);
7182 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7187 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7191 memcg = mem_cgroup_id_get_online(memcg);
7193 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7194 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7195 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7196 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7197 mem_cgroup_id_put(memcg);
7201 /* Get references for the tail pages, too */
7203 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7204 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7205 VM_BUG_ON_PAGE(oldid, page);
7206 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7212 * mem_cgroup_uncharge_swap - uncharge swap space
7213 * @entry: swap entry to uncharge
7214 * @nr_pages: the amount of swap space to uncharge
7216 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7218 struct mem_cgroup *memcg;
7221 id = swap_cgroup_record(entry, 0, nr_pages);
7223 memcg = mem_cgroup_from_id(id);
7225 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7226 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7227 page_counter_uncharge(&memcg->swap, nr_pages);
7229 page_counter_uncharge(&memcg->memsw, nr_pages);
7231 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7232 mem_cgroup_id_put_many(memcg, nr_pages);
7237 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7239 long nr_swap_pages = get_nr_swap_pages();
7241 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7242 return nr_swap_pages;
7243 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7244 nr_swap_pages = min_t(long, nr_swap_pages,
7245 READ_ONCE(memcg->swap.max) -
7246 page_counter_read(&memcg->swap));
7247 return nr_swap_pages;
7250 bool mem_cgroup_swap_full(struct page *page)
7252 struct mem_cgroup *memcg;
7254 VM_BUG_ON_PAGE(!PageLocked(page), page);
7258 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7261 memcg = page_memcg(page);
7265 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7266 unsigned long usage = page_counter_read(&memcg->swap);
7268 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7269 usage * 2 >= READ_ONCE(memcg->swap.max))
7276 static int __init setup_swap_account(char *s)
7278 if (!strcmp(s, "1"))
7279 cgroup_memory_noswap = false;
7280 else if (!strcmp(s, "0"))
7281 cgroup_memory_noswap = true;
7284 __setup("swapaccount=", setup_swap_account);
7286 static u64 swap_current_read(struct cgroup_subsys_state *css,
7289 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7291 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7294 static int swap_high_show(struct seq_file *m, void *v)
7296 return seq_puts_memcg_tunable(m,
7297 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7300 static ssize_t swap_high_write(struct kernfs_open_file *of,
7301 char *buf, size_t nbytes, loff_t off)
7303 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7307 buf = strstrip(buf);
7308 err = page_counter_memparse(buf, "max", &high);
7312 page_counter_set_high(&memcg->swap, high);
7317 static int swap_max_show(struct seq_file *m, void *v)
7319 return seq_puts_memcg_tunable(m,
7320 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7323 static ssize_t swap_max_write(struct kernfs_open_file *of,
7324 char *buf, size_t nbytes, loff_t off)
7326 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7330 buf = strstrip(buf);
7331 err = page_counter_memparse(buf, "max", &max);
7335 xchg(&memcg->swap.max, max);
7340 static int swap_events_show(struct seq_file *m, void *v)
7342 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7344 seq_printf(m, "high %lu\n",
7345 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7346 seq_printf(m, "max %lu\n",
7347 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7348 seq_printf(m, "fail %lu\n",
7349 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7354 static struct cftype swap_files[] = {
7356 .name = "swap.current",
7357 .flags = CFTYPE_NOT_ON_ROOT,
7358 .read_u64 = swap_current_read,
7361 .name = "swap.high",
7362 .flags = CFTYPE_NOT_ON_ROOT,
7363 .seq_show = swap_high_show,
7364 .write = swap_high_write,
7368 .flags = CFTYPE_NOT_ON_ROOT,
7369 .seq_show = swap_max_show,
7370 .write = swap_max_write,
7373 .name = "swap.events",
7374 .flags = CFTYPE_NOT_ON_ROOT,
7375 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7376 .seq_show = swap_events_show,
7381 static struct cftype memsw_files[] = {
7383 .name = "memsw.usage_in_bytes",
7384 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7385 .read_u64 = mem_cgroup_read_u64,
7388 .name = "memsw.max_usage_in_bytes",
7389 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7390 .write = mem_cgroup_reset,
7391 .read_u64 = mem_cgroup_read_u64,
7394 .name = "memsw.limit_in_bytes",
7395 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7396 .write = mem_cgroup_write,
7397 .read_u64 = mem_cgroup_read_u64,
7400 .name = "memsw.failcnt",
7401 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7402 .write = mem_cgroup_reset,
7403 .read_u64 = mem_cgroup_read_u64,
7405 { }, /* terminate */
7409 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7410 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7411 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7412 * boot parameter. This may result in premature OOPS inside
7413 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7415 static int __init mem_cgroup_swap_init(void)
7417 /* No memory control -> no swap control */
7418 if (mem_cgroup_disabled())
7419 cgroup_memory_noswap = true;
7421 if (cgroup_memory_noswap)
7424 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7425 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7429 core_initcall(mem_cgroup_swap_init);
7431 #endif /* CONFIG_MEMCG_SWAP */