1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
25 #include <linux/page_counter.h>
26 #include <linux/memcontrol.h>
27 #include <linux/cgroup.h>
28 #include <linux/pagewalk.h>
29 #include <linux/sched/mm.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/hugetlb.h>
32 #include <linux/pagemap.h>
33 #include <linux/vm_event_item.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
59 #include <linux/tracehook.h>
60 #include <linux/psi.h>
61 #include <linux/seq_buf.h>
67 #include <linux/uaccess.h>
69 #include <trace/events/vmscan.h>
71 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
72 EXPORT_SYMBOL(memory_cgrp_subsys);
74 struct mem_cgroup *root_mem_cgroup __read_mostly;
76 /* Active memory cgroup to use from an interrupt context */
77 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
79 /* Socket memory accounting disabled? */
80 static bool cgroup_memory_nosocket;
82 /* Kernel memory accounting disabled? */
83 static bool cgroup_memory_nokmem;
85 /* Whether the swap controller is active */
86 #ifdef CONFIG_MEMCG_SWAP
87 bool cgroup_memory_noswap __read_mostly;
89 #define cgroup_memory_noswap 1
92 #ifdef CONFIG_CGROUP_WRITEBACK
93 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
96 /* Whether legacy memory+swap accounting is active */
97 static bool do_memsw_account(void)
99 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
102 #define THRESHOLDS_EVENTS_TARGET 128
103 #define SOFTLIMIT_EVENTS_TARGET 1024
106 * Cgroups above their limits are maintained in a RB-Tree, independent of
107 * their hierarchy representation
110 struct mem_cgroup_tree_per_node {
111 struct rb_root rb_root;
112 struct rb_node *rb_rightmost;
116 struct mem_cgroup_tree {
117 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
120 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
123 struct mem_cgroup_eventfd_list {
124 struct list_head list;
125 struct eventfd_ctx *eventfd;
129 * cgroup_event represents events which userspace want to receive.
131 struct mem_cgroup_event {
133 * memcg which the event belongs to.
135 struct mem_cgroup *memcg;
137 * eventfd to signal userspace about the event.
139 struct eventfd_ctx *eventfd;
141 * Each of these stored in a list by the cgroup.
143 struct list_head list;
145 * register_event() callback will be used to add new userspace
146 * waiter for changes related to this event. Use eventfd_signal()
147 * on eventfd to send notification to userspace.
149 int (*register_event)(struct mem_cgroup *memcg,
150 struct eventfd_ctx *eventfd, const char *args);
152 * unregister_event() callback will be called when userspace closes
153 * the eventfd or on cgroup removing. This callback must be set,
154 * if you want provide notification functionality.
156 void (*unregister_event)(struct mem_cgroup *memcg,
157 struct eventfd_ctx *eventfd);
159 * All fields below needed to unregister event when
160 * userspace closes eventfd.
163 wait_queue_head_t *wqh;
164 wait_queue_entry_t wait;
165 struct work_struct remove;
168 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
169 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
171 /* Stuffs for move charges at task migration. */
173 * Types of charges to be moved.
175 #define MOVE_ANON 0x1U
176 #define MOVE_FILE 0x2U
177 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
179 /* "mc" and its members are protected by cgroup_mutex */
180 static struct move_charge_struct {
181 spinlock_t lock; /* for from, to */
182 struct mm_struct *mm;
183 struct mem_cgroup *from;
184 struct mem_cgroup *to;
186 unsigned long precharge;
187 unsigned long moved_charge;
188 unsigned long moved_swap;
189 struct task_struct *moving_task; /* a task moving charges */
190 wait_queue_head_t waitq; /* a waitq for other context */
192 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
193 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
197 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
198 * limit reclaim to prevent infinite loops, if they ever occur.
200 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
201 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
203 /* for encoding cft->private value on file */
212 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
213 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
214 #define MEMFILE_ATTR(val) ((val) & 0xffff)
215 /* Used for OOM nofiier */
216 #define OOM_CONTROL (0)
219 * Iteration constructs for visiting all cgroups (under a tree). If
220 * loops are exited prematurely (break), mem_cgroup_iter_break() must
221 * be used for reference counting.
223 #define for_each_mem_cgroup_tree(iter, root) \
224 for (iter = mem_cgroup_iter(root, NULL, NULL); \
226 iter = mem_cgroup_iter(root, iter, NULL))
228 #define for_each_mem_cgroup(iter) \
229 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
231 iter = mem_cgroup_iter(NULL, iter, NULL))
233 static inline bool should_force_charge(void)
235 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
236 (current->flags & PF_EXITING);
239 /* Some nice accessors for the vmpressure. */
240 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
243 memcg = root_mem_cgroup;
244 return &memcg->vmpressure;
247 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
249 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
252 #ifdef CONFIG_MEMCG_KMEM
253 extern spinlock_t css_set_lock;
255 static void obj_cgroup_release(struct percpu_ref *ref)
257 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
258 struct mem_cgroup *memcg;
259 unsigned int nr_bytes;
260 unsigned int nr_pages;
264 * At this point all allocated objects are freed, and
265 * objcg->nr_charged_bytes can't have an arbitrary byte value.
266 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
268 * The following sequence can lead to it:
269 * 1) CPU0: objcg == stock->cached_objcg
270 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
271 * PAGE_SIZE bytes are charged
272 * 3) CPU1: a process from another memcg is allocating something,
273 * the stock if flushed,
274 * objcg->nr_charged_bytes = PAGE_SIZE - 92
275 * 5) CPU0: we do release this object,
276 * 92 bytes are added to stock->nr_bytes
277 * 6) CPU0: stock is flushed,
278 * 92 bytes are added to objcg->nr_charged_bytes
280 * In the result, nr_charged_bytes == PAGE_SIZE.
281 * This page will be uncharged in obj_cgroup_release().
283 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
284 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
285 nr_pages = nr_bytes >> PAGE_SHIFT;
287 spin_lock_irqsave(&css_set_lock, flags);
288 memcg = obj_cgroup_memcg(objcg);
290 __memcg_kmem_uncharge(memcg, nr_pages);
291 list_del(&objcg->list);
292 mem_cgroup_put(memcg);
293 spin_unlock_irqrestore(&css_set_lock, flags);
295 percpu_ref_exit(ref);
296 kfree_rcu(objcg, rcu);
299 static struct obj_cgroup *obj_cgroup_alloc(void)
301 struct obj_cgroup *objcg;
304 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
308 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
314 INIT_LIST_HEAD(&objcg->list);
318 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
319 struct mem_cgroup *parent)
321 struct obj_cgroup *objcg, *iter;
323 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
325 spin_lock_irq(&css_set_lock);
327 /* Move active objcg to the parent's list */
328 xchg(&objcg->memcg, parent);
329 css_get(&parent->css);
330 list_add(&objcg->list, &parent->objcg_list);
332 /* Move already reparented objcgs to the parent's list */
333 list_for_each_entry(iter, &memcg->objcg_list, list) {
334 css_get(&parent->css);
335 xchg(&iter->memcg, parent);
336 css_put(&memcg->css);
338 list_splice(&memcg->objcg_list, &parent->objcg_list);
340 spin_unlock_irq(&css_set_lock);
342 percpu_ref_kill(&objcg->refcnt);
346 * This will be used as a shrinker list's index.
347 * The main reason for not using cgroup id for this:
348 * this works better in sparse environments, where we have a lot of memcgs,
349 * but only a few kmem-limited. Or also, if we have, for instance, 200
350 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
351 * 200 entry array for that.
353 * The current size of the caches array is stored in memcg_nr_cache_ids. It
354 * will double each time we have to increase it.
356 static DEFINE_IDA(memcg_cache_ida);
357 int memcg_nr_cache_ids;
359 /* Protects memcg_nr_cache_ids */
360 static DECLARE_RWSEM(memcg_cache_ids_sem);
362 void memcg_get_cache_ids(void)
364 down_read(&memcg_cache_ids_sem);
367 void memcg_put_cache_ids(void)
369 up_read(&memcg_cache_ids_sem);
373 * MIN_SIZE is different than 1, because we would like to avoid going through
374 * the alloc/free process all the time. In a small machine, 4 kmem-limited
375 * cgroups is a reasonable guess. In the future, it could be a parameter or
376 * tunable, but that is strictly not necessary.
378 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
379 * this constant directly from cgroup, but it is understandable that this is
380 * better kept as an internal representation in cgroup.c. In any case, the
381 * cgrp_id space is not getting any smaller, and we don't have to necessarily
382 * increase ours as well if it increases.
384 #define MEMCG_CACHES_MIN_SIZE 4
385 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
388 * A lot of the calls to the cache allocation functions are expected to be
389 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
390 * conditional to this static branch, we'll have to allow modules that does
391 * kmem_cache_alloc and the such to see this symbol as well
393 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
394 EXPORT_SYMBOL(memcg_kmem_enabled_key);
397 static int memcg_shrinker_map_size;
398 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
400 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
402 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
405 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
406 int size, int old_size)
408 struct memcg_shrinker_map *new, *old;
411 lockdep_assert_held(&memcg_shrinker_map_mutex);
414 old = rcu_dereference_protected(
415 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
416 /* Not yet online memcg */
420 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
424 /* Set all old bits, clear all new bits */
425 memset(new->map, (int)0xff, old_size);
426 memset((void *)new->map + old_size, 0, size - old_size);
428 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
429 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
435 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
437 struct mem_cgroup_per_node *pn;
438 struct memcg_shrinker_map *map;
441 if (mem_cgroup_is_root(memcg))
445 pn = mem_cgroup_nodeinfo(memcg, nid);
446 map = rcu_dereference_protected(pn->shrinker_map, true);
449 rcu_assign_pointer(pn->shrinker_map, NULL);
453 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
455 struct memcg_shrinker_map *map;
456 int nid, size, ret = 0;
458 if (mem_cgroup_is_root(memcg))
461 mutex_lock(&memcg_shrinker_map_mutex);
462 size = memcg_shrinker_map_size;
464 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
466 memcg_free_shrinker_maps(memcg);
470 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
472 mutex_unlock(&memcg_shrinker_map_mutex);
477 int memcg_expand_shrinker_maps(int new_id)
479 int size, old_size, ret = 0;
480 struct mem_cgroup *memcg;
482 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
483 old_size = memcg_shrinker_map_size;
484 if (size <= old_size)
487 mutex_lock(&memcg_shrinker_map_mutex);
488 if (!root_mem_cgroup)
491 for_each_mem_cgroup(memcg) {
492 if (mem_cgroup_is_root(memcg))
494 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
496 mem_cgroup_iter_break(NULL, memcg);
502 memcg_shrinker_map_size = size;
503 mutex_unlock(&memcg_shrinker_map_mutex);
507 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
509 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
510 struct memcg_shrinker_map *map;
513 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
514 /* Pairs with smp mb in shrink_slab() */
515 smp_mb__before_atomic();
516 set_bit(shrinker_id, map->map);
522 * mem_cgroup_css_from_page - css of the memcg associated with a page
523 * @page: page of interest
525 * If memcg is bound to the default hierarchy, css of the memcg associated
526 * with @page is returned. The returned css remains associated with @page
527 * until it is released.
529 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
532 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
534 struct mem_cgroup *memcg;
536 memcg = page_memcg(page);
538 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
539 memcg = root_mem_cgroup;
545 * page_cgroup_ino - return inode number of the memcg a page is charged to
548 * Look up the closest online ancestor of the memory cgroup @page is charged to
549 * and return its inode number or 0 if @page is not charged to any cgroup. It
550 * is safe to call this function without holding a reference to @page.
552 * Note, this function is inherently racy, because there is nothing to prevent
553 * the cgroup inode from getting torn down and potentially reallocated a moment
554 * after page_cgroup_ino() returns, so it only should be used by callers that
555 * do not care (such as procfs interfaces).
557 ino_t page_cgroup_ino(struct page *page)
559 struct mem_cgroup *memcg;
560 unsigned long ino = 0;
563 memcg = page_memcg_check(page);
565 while (memcg && !(memcg->css.flags & CSS_ONLINE))
566 memcg = parent_mem_cgroup(memcg);
568 ino = cgroup_ino(memcg->css.cgroup);
573 static struct mem_cgroup_per_node *
574 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
576 int nid = page_to_nid(page);
578 return memcg->nodeinfo[nid];
581 static struct mem_cgroup_tree_per_node *
582 soft_limit_tree_node(int nid)
584 return soft_limit_tree.rb_tree_per_node[nid];
587 static struct mem_cgroup_tree_per_node *
588 soft_limit_tree_from_page(struct page *page)
590 int nid = page_to_nid(page);
592 return soft_limit_tree.rb_tree_per_node[nid];
595 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
596 struct mem_cgroup_tree_per_node *mctz,
597 unsigned long new_usage_in_excess)
599 struct rb_node **p = &mctz->rb_root.rb_node;
600 struct rb_node *parent = NULL;
601 struct mem_cgroup_per_node *mz_node;
602 bool rightmost = true;
607 mz->usage_in_excess = new_usage_in_excess;
608 if (!mz->usage_in_excess)
612 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
614 if (mz->usage_in_excess < mz_node->usage_in_excess) {
623 mctz->rb_rightmost = &mz->tree_node;
625 rb_link_node(&mz->tree_node, parent, p);
626 rb_insert_color(&mz->tree_node, &mctz->rb_root);
630 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
631 struct mem_cgroup_tree_per_node *mctz)
636 if (&mz->tree_node == mctz->rb_rightmost)
637 mctz->rb_rightmost = rb_prev(&mz->tree_node);
639 rb_erase(&mz->tree_node, &mctz->rb_root);
643 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
644 struct mem_cgroup_tree_per_node *mctz)
648 spin_lock_irqsave(&mctz->lock, flags);
649 __mem_cgroup_remove_exceeded(mz, mctz);
650 spin_unlock_irqrestore(&mctz->lock, flags);
653 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
655 unsigned long nr_pages = page_counter_read(&memcg->memory);
656 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
657 unsigned long excess = 0;
659 if (nr_pages > soft_limit)
660 excess = nr_pages - soft_limit;
665 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
667 unsigned long excess;
668 struct mem_cgroup_per_node *mz;
669 struct mem_cgroup_tree_per_node *mctz;
671 mctz = soft_limit_tree_from_page(page);
675 * Necessary to update all ancestors when hierarchy is used.
676 * because their event counter is not touched.
678 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
679 mz = mem_cgroup_page_nodeinfo(memcg, page);
680 excess = soft_limit_excess(memcg);
682 * We have to update the tree if mz is on RB-tree or
683 * mem is over its softlimit.
685 if (excess || mz->on_tree) {
688 spin_lock_irqsave(&mctz->lock, flags);
689 /* if on-tree, remove it */
691 __mem_cgroup_remove_exceeded(mz, mctz);
693 * Insert again. mz->usage_in_excess will be updated.
694 * If excess is 0, no tree ops.
696 __mem_cgroup_insert_exceeded(mz, mctz, excess);
697 spin_unlock_irqrestore(&mctz->lock, flags);
702 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
704 struct mem_cgroup_tree_per_node *mctz;
705 struct mem_cgroup_per_node *mz;
709 mz = mem_cgroup_nodeinfo(memcg, nid);
710 mctz = soft_limit_tree_node(nid);
712 mem_cgroup_remove_exceeded(mz, mctz);
716 static struct mem_cgroup_per_node *
717 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
719 struct mem_cgroup_per_node *mz;
723 if (!mctz->rb_rightmost)
724 goto done; /* Nothing to reclaim from */
726 mz = rb_entry(mctz->rb_rightmost,
727 struct mem_cgroup_per_node, tree_node);
729 * Remove the node now but someone else can add it back,
730 * we will to add it back at the end of reclaim to its correct
731 * position in the tree.
733 __mem_cgroup_remove_exceeded(mz, mctz);
734 if (!soft_limit_excess(mz->memcg) ||
735 !css_tryget(&mz->memcg->css))
741 static struct mem_cgroup_per_node *
742 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
744 struct mem_cgroup_per_node *mz;
746 spin_lock_irq(&mctz->lock);
747 mz = __mem_cgroup_largest_soft_limit_node(mctz);
748 spin_unlock_irq(&mctz->lock);
753 * __mod_memcg_state - update cgroup memory statistics
754 * @memcg: the memory cgroup
755 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
756 * @val: delta to add to the counter, can be negative
758 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
760 long x, threshold = MEMCG_CHARGE_BATCH;
762 if (mem_cgroup_disabled())
765 if (memcg_stat_item_in_bytes(idx))
766 threshold <<= PAGE_SHIFT;
768 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
769 if (unlikely(abs(x) > threshold)) {
770 struct mem_cgroup *mi;
773 * Batch local counters to keep them in sync with
774 * the hierarchical ones.
776 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
777 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
778 atomic_long_add(x, &mi->vmstats[idx]);
781 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
784 static struct mem_cgroup_per_node *
785 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
787 struct mem_cgroup *parent;
789 parent = parent_mem_cgroup(pn->memcg);
792 return mem_cgroup_nodeinfo(parent, nid);
795 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
798 struct mem_cgroup_per_node *pn;
799 struct mem_cgroup *memcg;
800 long x, threshold = MEMCG_CHARGE_BATCH;
802 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
806 __mod_memcg_state(memcg, idx, val);
809 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
811 if (vmstat_item_in_bytes(idx))
812 threshold <<= PAGE_SHIFT;
814 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
815 if (unlikely(abs(x) > threshold)) {
816 pg_data_t *pgdat = lruvec_pgdat(lruvec);
817 struct mem_cgroup_per_node *pi;
819 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
820 atomic_long_add(x, &pi->lruvec_stat[idx]);
823 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
827 * __mod_lruvec_state - update lruvec memory statistics
828 * @lruvec: the lruvec
829 * @idx: the stat item
830 * @val: delta to add to the counter, can be negative
832 * The lruvec is the intersection of the NUMA node and a cgroup. This
833 * function updates the all three counters that are affected by a
834 * change of state at this level: per-node, per-cgroup, per-lruvec.
836 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
840 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
842 /* Update memcg and lruvec */
843 if (!mem_cgroup_disabled())
844 __mod_memcg_lruvec_state(lruvec, idx, val);
847 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
850 struct page *head = compound_head(page); /* rmap on tail pages */
851 struct mem_cgroup *memcg = page_memcg(head);
852 pg_data_t *pgdat = page_pgdat(page);
853 struct lruvec *lruvec;
855 /* Untracked pages have no memcg, no lruvec. Update only the node */
857 __mod_node_page_state(pgdat, idx, val);
861 lruvec = mem_cgroup_lruvec(memcg, pgdat);
862 __mod_lruvec_state(lruvec, idx, val);
864 EXPORT_SYMBOL(__mod_lruvec_page_state);
866 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
868 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
869 struct mem_cgroup *memcg;
870 struct lruvec *lruvec;
873 memcg = mem_cgroup_from_obj(p);
876 * Untracked pages have no memcg, no lruvec. Update only the
877 * node. If we reparent the slab objects to the root memcg,
878 * when we free the slab object, we need to update the per-memcg
879 * vmstats to keep it correct for the root memcg.
882 __mod_node_page_state(pgdat, idx, val);
884 lruvec = mem_cgroup_lruvec(memcg, pgdat);
885 __mod_lruvec_state(lruvec, idx, val);
891 * __count_memcg_events - account VM events in a cgroup
892 * @memcg: the memory cgroup
893 * @idx: the event item
894 * @count: the number of events that occured
896 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
901 if (mem_cgroup_disabled())
904 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
905 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
906 struct mem_cgroup *mi;
909 * Batch local counters to keep them in sync with
910 * the hierarchical ones.
912 __this_cpu_add(memcg->vmstats_local->events[idx], x);
913 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
914 atomic_long_add(x, &mi->vmevents[idx]);
917 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
920 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
922 return atomic_long_read(&memcg->vmevents[event]);
925 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
930 for_each_possible_cpu(cpu)
931 x += per_cpu(memcg->vmstats_local->events[event], cpu);
935 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
939 /* pagein of a big page is an event. So, ignore page size */
941 __count_memcg_events(memcg, PGPGIN, 1);
943 __count_memcg_events(memcg, PGPGOUT, 1);
944 nr_pages = -nr_pages; /* for event */
947 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
950 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
951 enum mem_cgroup_events_target target)
953 unsigned long val, next;
955 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
956 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
957 /* from time_after() in jiffies.h */
958 if ((long)(next - val) < 0) {
960 case MEM_CGROUP_TARGET_THRESH:
961 next = val + THRESHOLDS_EVENTS_TARGET;
963 case MEM_CGROUP_TARGET_SOFTLIMIT:
964 next = val + SOFTLIMIT_EVENTS_TARGET;
969 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
976 * Check events in order.
979 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
981 /* threshold event is triggered in finer grain than soft limit */
982 if (unlikely(mem_cgroup_event_ratelimit(memcg,
983 MEM_CGROUP_TARGET_THRESH))) {
986 do_softlimit = mem_cgroup_event_ratelimit(memcg,
987 MEM_CGROUP_TARGET_SOFTLIMIT);
988 mem_cgroup_threshold(memcg);
989 if (unlikely(do_softlimit))
990 mem_cgroup_update_tree(memcg, page);
994 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
997 * mm_update_next_owner() may clear mm->owner to NULL
998 * if it races with swapoff, page migration, etc.
999 * So this can be called with p == NULL.
1004 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1006 EXPORT_SYMBOL(mem_cgroup_from_task);
1009 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1010 * @mm: mm from which memcg should be extracted. It can be NULL.
1012 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1013 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1016 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1018 struct mem_cgroup *memcg;
1020 if (mem_cgroup_disabled())
1026 * Page cache insertions can happen withou an
1027 * actual mm context, e.g. during disk probing
1028 * on boot, loopback IO, acct() writes etc.
1031 memcg = root_mem_cgroup;
1033 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1034 if (unlikely(!memcg))
1035 memcg = root_mem_cgroup;
1037 } while (!css_tryget(&memcg->css));
1041 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1044 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1045 * @page: page from which memcg should be extracted.
1047 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1048 * root_mem_cgroup is returned.
1050 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
1052 struct mem_cgroup *memcg = page_memcg(page);
1054 if (mem_cgroup_disabled())
1058 /* Page should not get uncharged and freed memcg under us. */
1059 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
1060 memcg = root_mem_cgroup;
1064 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1066 static __always_inline struct mem_cgroup *active_memcg(void)
1069 return this_cpu_read(int_active_memcg);
1071 return current->active_memcg;
1074 static __always_inline struct mem_cgroup *get_active_memcg(void)
1076 struct mem_cgroup *memcg;
1079 memcg = active_memcg();
1081 /* current->active_memcg must hold a ref. */
1082 if (WARN_ON_ONCE(!css_tryget(&memcg->css)))
1083 memcg = root_mem_cgroup;
1085 memcg = current->active_memcg;
1092 static __always_inline bool memcg_kmem_bypass(void)
1094 /* Allow remote memcg charging from any context. */
1095 if (unlikely(active_memcg()))
1098 /* Memcg to charge can't be determined. */
1099 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
1106 * If active memcg is set, do not fallback to current->mm->memcg.
1108 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1110 if (memcg_kmem_bypass())
1113 if (unlikely(active_memcg()))
1114 return get_active_memcg();
1116 return get_mem_cgroup_from_mm(current->mm);
1120 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1121 * @root: hierarchy root
1122 * @prev: previously returned memcg, NULL on first invocation
1123 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1125 * Returns references to children of the hierarchy below @root, or
1126 * @root itself, or %NULL after a full round-trip.
1128 * Caller must pass the return value in @prev on subsequent
1129 * invocations for reference counting, or use mem_cgroup_iter_break()
1130 * to cancel a hierarchy walk before the round-trip is complete.
1132 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1133 * in the hierarchy among all concurrent reclaimers operating on the
1136 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1137 struct mem_cgroup *prev,
1138 struct mem_cgroup_reclaim_cookie *reclaim)
1140 struct mem_cgroup_reclaim_iter *iter;
1141 struct cgroup_subsys_state *css = NULL;
1142 struct mem_cgroup *memcg = NULL;
1143 struct mem_cgroup *pos = NULL;
1145 if (mem_cgroup_disabled())
1149 root = root_mem_cgroup;
1151 if (prev && !reclaim)
1157 struct mem_cgroup_per_node *mz;
1159 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1162 if (prev && reclaim->generation != iter->generation)
1166 pos = READ_ONCE(iter->position);
1167 if (!pos || css_tryget(&pos->css))
1170 * css reference reached zero, so iter->position will
1171 * be cleared by ->css_released. However, we should not
1172 * rely on this happening soon, because ->css_released
1173 * is called from a work queue, and by busy-waiting we
1174 * might block it. So we clear iter->position right
1177 (void)cmpxchg(&iter->position, pos, NULL);
1185 css = css_next_descendant_pre(css, &root->css);
1188 * Reclaimers share the hierarchy walk, and a
1189 * new one might jump in right at the end of
1190 * the hierarchy - make sure they see at least
1191 * one group and restart from the beginning.
1199 * Verify the css and acquire a reference. The root
1200 * is provided by the caller, so we know it's alive
1201 * and kicking, and don't take an extra reference.
1203 memcg = mem_cgroup_from_css(css);
1205 if (css == &root->css)
1208 if (css_tryget(css))
1216 * The position could have already been updated by a competing
1217 * thread, so check that the value hasn't changed since we read
1218 * it to avoid reclaiming from the same cgroup twice.
1220 (void)cmpxchg(&iter->position, pos, memcg);
1228 reclaim->generation = iter->generation;
1233 if (prev && prev != root)
1234 css_put(&prev->css);
1240 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1241 * @root: hierarchy root
1242 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1244 void mem_cgroup_iter_break(struct mem_cgroup *root,
1245 struct mem_cgroup *prev)
1248 root = root_mem_cgroup;
1249 if (prev && prev != root)
1250 css_put(&prev->css);
1253 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1254 struct mem_cgroup *dead_memcg)
1256 struct mem_cgroup_reclaim_iter *iter;
1257 struct mem_cgroup_per_node *mz;
1260 for_each_node(nid) {
1261 mz = mem_cgroup_nodeinfo(from, nid);
1263 cmpxchg(&iter->position, dead_memcg, NULL);
1267 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1269 struct mem_cgroup *memcg = dead_memcg;
1270 struct mem_cgroup *last;
1273 __invalidate_reclaim_iterators(memcg, dead_memcg);
1275 } while ((memcg = parent_mem_cgroup(memcg)));
1278 * When cgruop1 non-hierarchy mode is used,
1279 * parent_mem_cgroup() does not walk all the way up to the
1280 * cgroup root (root_mem_cgroup). So we have to handle
1281 * dead_memcg from cgroup root separately.
1283 if (last != root_mem_cgroup)
1284 __invalidate_reclaim_iterators(root_mem_cgroup,
1289 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1290 * @memcg: hierarchy root
1291 * @fn: function to call for each task
1292 * @arg: argument passed to @fn
1294 * This function iterates over tasks attached to @memcg or to any of its
1295 * descendants and calls @fn for each task. If @fn returns a non-zero
1296 * value, the function breaks the iteration loop and returns the value.
1297 * Otherwise, it will iterate over all tasks and return 0.
1299 * This function must not be called for the root memory cgroup.
1301 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1302 int (*fn)(struct task_struct *, void *), void *arg)
1304 struct mem_cgroup *iter;
1307 BUG_ON(memcg == root_mem_cgroup);
1309 for_each_mem_cgroup_tree(iter, memcg) {
1310 struct css_task_iter it;
1311 struct task_struct *task;
1313 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1314 while (!ret && (task = css_task_iter_next(&it)))
1315 ret = fn(task, arg);
1316 css_task_iter_end(&it);
1318 mem_cgroup_iter_break(memcg, iter);
1326 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1328 * @pgdat: pgdat of the page
1330 * This function relies on page's memcg being stable - see the
1331 * access rules in commit_charge().
1333 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1335 struct mem_cgroup_per_node *mz;
1336 struct mem_cgroup *memcg;
1337 struct lruvec *lruvec;
1339 if (mem_cgroup_disabled()) {
1340 lruvec = &pgdat->__lruvec;
1344 memcg = page_memcg(page);
1346 * Swapcache readahead pages are added to the LRU - and
1347 * possibly migrated - before they are charged.
1350 memcg = root_mem_cgroup;
1352 mz = mem_cgroup_page_nodeinfo(memcg, page);
1353 lruvec = &mz->lruvec;
1356 * Since a node can be onlined after the mem_cgroup was created,
1357 * we have to be prepared to initialize lruvec->zone here;
1358 * and if offlined then reonlined, we need to reinitialize it.
1360 if (unlikely(lruvec->pgdat != pgdat))
1361 lruvec->pgdat = pgdat;
1366 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1367 * @lruvec: mem_cgroup per zone lru vector
1368 * @lru: index of lru list the page is sitting on
1369 * @zid: zone id of the accounted pages
1370 * @nr_pages: positive when adding or negative when removing
1372 * This function must be called under lru_lock, just before a page is added
1373 * to or just after a page is removed from an lru list (that ordering being
1374 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1376 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1377 int zid, int nr_pages)
1379 struct mem_cgroup_per_node *mz;
1380 unsigned long *lru_size;
1383 if (mem_cgroup_disabled())
1386 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1387 lru_size = &mz->lru_zone_size[zid][lru];
1390 *lru_size += nr_pages;
1393 if (WARN_ONCE(size < 0,
1394 "%s(%p, %d, %d): lru_size %ld\n",
1395 __func__, lruvec, lru, nr_pages, size)) {
1401 *lru_size += nr_pages;
1405 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1406 * @memcg: the memory cgroup
1408 * Returns the maximum amount of memory @mem can be charged with, in
1411 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1413 unsigned long margin = 0;
1414 unsigned long count;
1415 unsigned long limit;
1417 count = page_counter_read(&memcg->memory);
1418 limit = READ_ONCE(memcg->memory.max);
1420 margin = limit - count;
1422 if (do_memsw_account()) {
1423 count = page_counter_read(&memcg->memsw);
1424 limit = READ_ONCE(memcg->memsw.max);
1426 margin = min(margin, limit - count);
1435 * A routine for checking "mem" is under move_account() or not.
1437 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1438 * moving cgroups. This is for waiting at high-memory pressure
1441 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1443 struct mem_cgroup *from;
1444 struct mem_cgroup *to;
1447 * Unlike task_move routines, we access mc.to, mc.from not under
1448 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1450 spin_lock(&mc.lock);
1456 ret = mem_cgroup_is_descendant(from, memcg) ||
1457 mem_cgroup_is_descendant(to, memcg);
1459 spin_unlock(&mc.lock);
1463 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1465 if (mc.moving_task && current != mc.moving_task) {
1466 if (mem_cgroup_under_move(memcg)) {
1468 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1469 /* moving charge context might have finished. */
1472 finish_wait(&mc.waitq, &wait);
1479 struct memory_stat {
1485 static struct memory_stat memory_stats[] = {
1486 { "anon", PAGE_SIZE, NR_ANON_MAPPED },
1487 { "file", PAGE_SIZE, NR_FILE_PAGES },
1488 { "kernel_stack", 1024, NR_KERNEL_STACK_KB },
1489 { "pagetables", PAGE_SIZE, NR_PAGETABLE },
1490 { "percpu", 1, MEMCG_PERCPU_B },
1491 { "sock", PAGE_SIZE, MEMCG_SOCK },
1492 { "shmem", PAGE_SIZE, NR_SHMEM },
1493 { "file_mapped", PAGE_SIZE, NR_FILE_MAPPED },
1494 { "file_dirty", PAGE_SIZE, NR_FILE_DIRTY },
1495 { "file_writeback", PAGE_SIZE, NR_WRITEBACK },
1496 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1498 * The ratio will be initialized in memory_stats_init(). Because
1499 * on some architectures, the macro of HPAGE_PMD_SIZE is not
1500 * constant(e.g. powerpc).
1502 { "anon_thp", 0, NR_ANON_THPS },
1503 { "file_thp", 0, NR_FILE_THPS },
1504 { "shmem_thp", 0, NR_SHMEM_THPS },
1506 { "inactive_anon", PAGE_SIZE, NR_INACTIVE_ANON },
1507 { "active_anon", PAGE_SIZE, NR_ACTIVE_ANON },
1508 { "inactive_file", PAGE_SIZE, NR_INACTIVE_FILE },
1509 { "active_file", PAGE_SIZE, NR_ACTIVE_FILE },
1510 { "unevictable", PAGE_SIZE, NR_UNEVICTABLE },
1513 * Note: The slab_reclaimable and slab_unreclaimable must be
1514 * together and slab_reclaimable must be in front.
1516 { "slab_reclaimable", 1, NR_SLAB_RECLAIMABLE_B },
1517 { "slab_unreclaimable", 1, NR_SLAB_UNRECLAIMABLE_B },
1519 /* The memory events */
1520 { "workingset_refault_anon", 1, WORKINGSET_REFAULT_ANON },
1521 { "workingset_refault_file", 1, WORKINGSET_REFAULT_FILE },
1522 { "workingset_activate_anon", 1, WORKINGSET_ACTIVATE_ANON },
1523 { "workingset_activate_file", 1, WORKINGSET_ACTIVATE_FILE },
1524 { "workingset_restore_anon", 1, WORKINGSET_RESTORE_ANON },
1525 { "workingset_restore_file", 1, WORKINGSET_RESTORE_FILE },
1526 { "workingset_nodereclaim", 1, WORKINGSET_NODERECLAIM },
1529 static int __init memory_stats_init(void)
1533 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1534 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1535 if (memory_stats[i].idx == NR_ANON_THPS ||
1536 memory_stats[i].idx == NR_FILE_THPS ||
1537 memory_stats[i].idx == NR_SHMEM_THPS)
1538 memory_stats[i].ratio = HPAGE_PMD_SIZE;
1540 VM_BUG_ON(!memory_stats[i].ratio);
1541 VM_BUG_ON(memory_stats[i].idx >= MEMCG_NR_STAT);
1546 pure_initcall(memory_stats_init);
1548 static char *memory_stat_format(struct mem_cgroup *memcg)
1553 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1558 * Provide statistics on the state of the memory subsystem as
1559 * well as cumulative event counters that show past behavior.
1561 * This list is ordered following a combination of these gradients:
1562 * 1) generic big picture -> specifics and details
1563 * 2) reflecting userspace activity -> reflecting kernel heuristics
1565 * Current memory state:
1568 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1571 size = memcg_page_state(memcg, memory_stats[i].idx);
1572 size *= memory_stats[i].ratio;
1573 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1575 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1576 size = memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B) +
1577 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B);
1578 seq_buf_printf(&s, "slab %llu\n", size);
1582 /* Accumulated memory events */
1584 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1585 memcg_events(memcg, PGFAULT));
1586 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1587 memcg_events(memcg, PGMAJFAULT));
1588 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1589 memcg_events(memcg, PGREFILL));
1590 seq_buf_printf(&s, "pgscan %lu\n",
1591 memcg_events(memcg, PGSCAN_KSWAPD) +
1592 memcg_events(memcg, PGSCAN_DIRECT));
1593 seq_buf_printf(&s, "pgsteal %lu\n",
1594 memcg_events(memcg, PGSTEAL_KSWAPD) +
1595 memcg_events(memcg, PGSTEAL_DIRECT));
1596 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1597 memcg_events(memcg, PGACTIVATE));
1598 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1599 memcg_events(memcg, PGDEACTIVATE));
1600 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1601 memcg_events(memcg, PGLAZYFREE));
1602 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1603 memcg_events(memcg, PGLAZYFREED));
1605 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1606 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1607 memcg_events(memcg, THP_FAULT_ALLOC));
1608 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1609 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1610 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1612 /* The above should easily fit into one page */
1613 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1618 #define K(x) ((x) << (PAGE_SHIFT-10))
1620 * mem_cgroup_print_oom_context: Print OOM information relevant to
1621 * memory controller.
1622 * @memcg: The memory cgroup that went over limit
1623 * @p: Task that is going to be killed
1625 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1628 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1633 pr_cont(",oom_memcg=");
1634 pr_cont_cgroup_path(memcg->css.cgroup);
1636 pr_cont(",global_oom");
1638 pr_cont(",task_memcg=");
1639 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1645 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1646 * memory controller.
1647 * @memcg: The memory cgroup that went over limit
1649 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1653 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1654 K((u64)page_counter_read(&memcg->memory)),
1655 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1656 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1657 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1658 K((u64)page_counter_read(&memcg->swap)),
1659 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1661 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1662 K((u64)page_counter_read(&memcg->memsw)),
1663 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1664 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1665 K((u64)page_counter_read(&memcg->kmem)),
1666 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1669 pr_info("Memory cgroup stats for ");
1670 pr_cont_cgroup_path(memcg->css.cgroup);
1672 buf = memory_stat_format(memcg);
1680 * Return the memory (and swap, if configured) limit for a memcg.
1682 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1684 unsigned long max = READ_ONCE(memcg->memory.max);
1686 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1687 if (mem_cgroup_swappiness(memcg))
1688 max += min(READ_ONCE(memcg->swap.max),
1689 (unsigned long)total_swap_pages);
1691 if (mem_cgroup_swappiness(memcg)) {
1692 /* Calculate swap excess capacity from memsw limit */
1693 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1695 max += min(swap, (unsigned long)total_swap_pages);
1701 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1703 return page_counter_read(&memcg->memory);
1706 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1709 struct oom_control oc = {
1713 .gfp_mask = gfp_mask,
1718 if (mutex_lock_killable(&oom_lock))
1721 if (mem_cgroup_margin(memcg) >= (1 << order))
1725 * A few threads which were not waiting at mutex_lock_killable() can
1726 * fail to bail out. Therefore, check again after holding oom_lock.
1728 ret = should_force_charge() || out_of_memory(&oc);
1731 mutex_unlock(&oom_lock);
1735 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1738 unsigned long *total_scanned)
1740 struct mem_cgroup *victim = NULL;
1743 unsigned long excess;
1744 unsigned long nr_scanned;
1745 struct mem_cgroup_reclaim_cookie reclaim = {
1749 excess = soft_limit_excess(root_memcg);
1752 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1757 * If we have not been able to reclaim
1758 * anything, it might because there are
1759 * no reclaimable pages under this hierarchy
1764 * We want to do more targeted reclaim.
1765 * excess >> 2 is not to excessive so as to
1766 * reclaim too much, nor too less that we keep
1767 * coming back to reclaim from this cgroup
1769 if (total >= (excess >> 2) ||
1770 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1775 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1776 pgdat, &nr_scanned);
1777 *total_scanned += nr_scanned;
1778 if (!soft_limit_excess(root_memcg))
1781 mem_cgroup_iter_break(root_memcg, victim);
1785 #ifdef CONFIG_LOCKDEP
1786 static struct lockdep_map memcg_oom_lock_dep_map = {
1787 .name = "memcg_oom_lock",
1791 static DEFINE_SPINLOCK(memcg_oom_lock);
1794 * Check OOM-Killer is already running under our hierarchy.
1795 * If someone is running, return false.
1797 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1799 struct mem_cgroup *iter, *failed = NULL;
1801 spin_lock(&memcg_oom_lock);
1803 for_each_mem_cgroup_tree(iter, memcg) {
1804 if (iter->oom_lock) {
1806 * this subtree of our hierarchy is already locked
1807 * so we cannot give a lock.
1810 mem_cgroup_iter_break(memcg, iter);
1813 iter->oom_lock = true;
1818 * OK, we failed to lock the whole subtree so we have
1819 * to clean up what we set up to the failing subtree
1821 for_each_mem_cgroup_tree(iter, memcg) {
1822 if (iter == failed) {
1823 mem_cgroup_iter_break(memcg, iter);
1826 iter->oom_lock = false;
1829 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1831 spin_unlock(&memcg_oom_lock);
1836 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1838 struct mem_cgroup *iter;
1840 spin_lock(&memcg_oom_lock);
1841 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1842 for_each_mem_cgroup_tree(iter, memcg)
1843 iter->oom_lock = false;
1844 spin_unlock(&memcg_oom_lock);
1847 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1849 struct mem_cgroup *iter;
1851 spin_lock(&memcg_oom_lock);
1852 for_each_mem_cgroup_tree(iter, memcg)
1854 spin_unlock(&memcg_oom_lock);
1857 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1859 struct mem_cgroup *iter;
1862 * Be careful about under_oom underflows becase a child memcg
1863 * could have been added after mem_cgroup_mark_under_oom.
1865 spin_lock(&memcg_oom_lock);
1866 for_each_mem_cgroup_tree(iter, memcg)
1867 if (iter->under_oom > 0)
1869 spin_unlock(&memcg_oom_lock);
1872 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1874 struct oom_wait_info {
1875 struct mem_cgroup *memcg;
1876 wait_queue_entry_t wait;
1879 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1880 unsigned mode, int sync, void *arg)
1882 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1883 struct mem_cgroup *oom_wait_memcg;
1884 struct oom_wait_info *oom_wait_info;
1886 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1887 oom_wait_memcg = oom_wait_info->memcg;
1889 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1890 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1892 return autoremove_wake_function(wait, mode, sync, arg);
1895 static void memcg_oom_recover(struct mem_cgroup *memcg)
1898 * For the following lockless ->under_oom test, the only required
1899 * guarantee is that it must see the state asserted by an OOM when
1900 * this function is called as a result of userland actions
1901 * triggered by the notification of the OOM. This is trivially
1902 * achieved by invoking mem_cgroup_mark_under_oom() before
1903 * triggering notification.
1905 if (memcg && memcg->under_oom)
1906 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1916 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1918 enum oom_status ret;
1921 if (order > PAGE_ALLOC_COSTLY_ORDER)
1924 memcg_memory_event(memcg, MEMCG_OOM);
1927 * We are in the middle of the charge context here, so we
1928 * don't want to block when potentially sitting on a callstack
1929 * that holds all kinds of filesystem and mm locks.
1931 * cgroup1 allows disabling the OOM killer and waiting for outside
1932 * handling until the charge can succeed; remember the context and put
1933 * the task to sleep at the end of the page fault when all locks are
1936 * On the other hand, in-kernel OOM killer allows for an async victim
1937 * memory reclaim (oom_reaper) and that means that we are not solely
1938 * relying on the oom victim to make a forward progress and we can
1939 * invoke the oom killer here.
1941 * Please note that mem_cgroup_out_of_memory might fail to find a
1942 * victim and then we have to bail out from the charge path.
1944 if (memcg->oom_kill_disable) {
1945 if (!current->in_user_fault)
1947 css_get(&memcg->css);
1948 current->memcg_in_oom = memcg;
1949 current->memcg_oom_gfp_mask = mask;
1950 current->memcg_oom_order = order;
1955 mem_cgroup_mark_under_oom(memcg);
1957 locked = mem_cgroup_oom_trylock(memcg);
1960 mem_cgroup_oom_notify(memcg);
1962 mem_cgroup_unmark_under_oom(memcg);
1963 if (mem_cgroup_out_of_memory(memcg, mask, order))
1969 mem_cgroup_oom_unlock(memcg);
1975 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1976 * @handle: actually kill/wait or just clean up the OOM state
1978 * This has to be called at the end of a page fault if the memcg OOM
1979 * handler was enabled.
1981 * Memcg supports userspace OOM handling where failed allocations must
1982 * sleep on a waitqueue until the userspace task resolves the
1983 * situation. Sleeping directly in the charge context with all kinds
1984 * of locks held is not a good idea, instead we remember an OOM state
1985 * in the task and mem_cgroup_oom_synchronize() has to be called at
1986 * the end of the page fault to complete the OOM handling.
1988 * Returns %true if an ongoing memcg OOM situation was detected and
1989 * completed, %false otherwise.
1991 bool mem_cgroup_oom_synchronize(bool handle)
1993 struct mem_cgroup *memcg = current->memcg_in_oom;
1994 struct oom_wait_info owait;
1997 /* OOM is global, do not handle */
2004 owait.memcg = memcg;
2005 owait.wait.flags = 0;
2006 owait.wait.func = memcg_oom_wake_function;
2007 owait.wait.private = current;
2008 INIT_LIST_HEAD(&owait.wait.entry);
2010 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2011 mem_cgroup_mark_under_oom(memcg);
2013 locked = mem_cgroup_oom_trylock(memcg);
2016 mem_cgroup_oom_notify(memcg);
2018 if (locked && !memcg->oom_kill_disable) {
2019 mem_cgroup_unmark_under_oom(memcg);
2020 finish_wait(&memcg_oom_waitq, &owait.wait);
2021 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
2022 current->memcg_oom_order);
2025 mem_cgroup_unmark_under_oom(memcg);
2026 finish_wait(&memcg_oom_waitq, &owait.wait);
2030 mem_cgroup_oom_unlock(memcg);
2032 * There is no guarantee that an OOM-lock contender
2033 * sees the wakeups triggered by the OOM kill
2034 * uncharges. Wake any sleepers explicitely.
2036 memcg_oom_recover(memcg);
2039 current->memcg_in_oom = NULL;
2040 css_put(&memcg->css);
2045 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2046 * @victim: task to be killed by the OOM killer
2047 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2049 * Returns a pointer to a memory cgroup, which has to be cleaned up
2050 * by killing all belonging OOM-killable tasks.
2052 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2054 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2055 struct mem_cgroup *oom_domain)
2057 struct mem_cgroup *oom_group = NULL;
2058 struct mem_cgroup *memcg;
2060 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2064 oom_domain = root_mem_cgroup;
2068 memcg = mem_cgroup_from_task(victim);
2069 if (memcg == root_mem_cgroup)
2073 * If the victim task has been asynchronously moved to a different
2074 * memory cgroup, we might end up killing tasks outside oom_domain.
2075 * In this case it's better to ignore memory.group.oom.
2077 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2081 * Traverse the memory cgroup hierarchy from the victim task's
2082 * cgroup up to the OOMing cgroup (or root) to find the
2083 * highest-level memory cgroup with oom.group set.
2085 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2086 if (memcg->oom_group)
2089 if (memcg == oom_domain)
2094 css_get(&oom_group->css);
2101 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2103 pr_info("Tasks in ");
2104 pr_cont_cgroup_path(memcg->css.cgroup);
2105 pr_cont(" are going to be killed due to memory.oom.group set\n");
2109 * lock_page_memcg - lock a page and memcg binding
2112 * This function protects unlocked LRU pages from being moved to
2115 * It ensures lifetime of the returned memcg. Caller is responsible
2116 * for the lifetime of the page; __unlock_page_memcg() is available
2117 * when @page might get freed inside the locked section.
2119 struct mem_cgroup *lock_page_memcg(struct page *page)
2121 struct page *head = compound_head(page); /* rmap on tail pages */
2122 struct mem_cgroup *memcg;
2123 unsigned long flags;
2126 * The RCU lock is held throughout the transaction. The fast
2127 * path can get away without acquiring the memcg->move_lock
2128 * because page moving starts with an RCU grace period.
2130 * The RCU lock also protects the memcg from being freed when
2131 * the page state that is going to change is the only thing
2132 * preventing the page itself from being freed. E.g. writeback
2133 * doesn't hold a page reference and relies on PG_writeback to
2134 * keep off truncation, migration and so forth.
2138 if (mem_cgroup_disabled())
2141 memcg = page_memcg(head);
2142 if (unlikely(!memcg))
2145 if (atomic_read(&memcg->moving_account) <= 0)
2148 spin_lock_irqsave(&memcg->move_lock, flags);
2149 if (memcg != page_memcg(head)) {
2150 spin_unlock_irqrestore(&memcg->move_lock, flags);
2155 * When charge migration first begins, we can have locked and
2156 * unlocked page stat updates happening concurrently. Track
2157 * the task who has the lock for unlock_page_memcg().
2159 memcg->move_lock_task = current;
2160 memcg->move_lock_flags = flags;
2164 EXPORT_SYMBOL(lock_page_memcg);
2167 * __unlock_page_memcg - unlock and unpin a memcg
2170 * Unlock and unpin a memcg returned by lock_page_memcg().
2172 void __unlock_page_memcg(struct mem_cgroup *memcg)
2174 if (memcg && memcg->move_lock_task == current) {
2175 unsigned long flags = memcg->move_lock_flags;
2177 memcg->move_lock_task = NULL;
2178 memcg->move_lock_flags = 0;
2180 spin_unlock_irqrestore(&memcg->move_lock, flags);
2187 * unlock_page_memcg - unlock a page and memcg binding
2190 void unlock_page_memcg(struct page *page)
2192 struct page *head = compound_head(page);
2194 __unlock_page_memcg(page_memcg(head));
2196 EXPORT_SYMBOL(unlock_page_memcg);
2198 struct memcg_stock_pcp {
2199 struct mem_cgroup *cached; /* this never be root cgroup */
2200 unsigned int nr_pages;
2202 #ifdef CONFIG_MEMCG_KMEM
2203 struct obj_cgroup *cached_objcg;
2204 unsigned int nr_bytes;
2207 struct work_struct work;
2208 unsigned long flags;
2209 #define FLUSHING_CACHED_CHARGE 0
2211 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2212 static DEFINE_MUTEX(percpu_charge_mutex);
2214 #ifdef CONFIG_MEMCG_KMEM
2215 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2216 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2217 struct mem_cgroup *root_memcg);
2220 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2223 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2224 struct mem_cgroup *root_memcg)
2231 * consume_stock: Try to consume stocked charge on this cpu.
2232 * @memcg: memcg to consume from.
2233 * @nr_pages: how many pages to charge.
2235 * The charges will only happen if @memcg matches the current cpu's memcg
2236 * stock, and at least @nr_pages are available in that stock. Failure to
2237 * service an allocation will refill the stock.
2239 * returns true if successful, false otherwise.
2241 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2243 struct memcg_stock_pcp *stock;
2244 unsigned long flags;
2247 if (nr_pages > MEMCG_CHARGE_BATCH)
2250 local_irq_save(flags);
2252 stock = this_cpu_ptr(&memcg_stock);
2253 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2254 stock->nr_pages -= nr_pages;
2258 local_irq_restore(flags);
2264 * Returns stocks cached in percpu and reset cached information.
2266 static void drain_stock(struct memcg_stock_pcp *stock)
2268 struct mem_cgroup *old = stock->cached;
2273 if (stock->nr_pages) {
2274 page_counter_uncharge(&old->memory, stock->nr_pages);
2275 if (do_memsw_account())
2276 page_counter_uncharge(&old->memsw, stock->nr_pages);
2277 stock->nr_pages = 0;
2281 stock->cached = NULL;
2284 static void drain_local_stock(struct work_struct *dummy)
2286 struct memcg_stock_pcp *stock;
2287 unsigned long flags;
2290 * The only protection from memory hotplug vs. drain_stock races is
2291 * that we always operate on local CPU stock here with IRQ disabled
2293 local_irq_save(flags);
2295 stock = this_cpu_ptr(&memcg_stock);
2296 drain_obj_stock(stock);
2298 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2300 local_irq_restore(flags);
2304 * Cache charges(val) to local per_cpu area.
2305 * This will be consumed by consume_stock() function, later.
2307 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2309 struct memcg_stock_pcp *stock;
2310 unsigned long flags;
2312 local_irq_save(flags);
2314 stock = this_cpu_ptr(&memcg_stock);
2315 if (stock->cached != memcg) { /* reset if necessary */
2317 css_get(&memcg->css);
2318 stock->cached = memcg;
2320 stock->nr_pages += nr_pages;
2322 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2325 local_irq_restore(flags);
2329 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2330 * of the hierarchy under it.
2332 static void drain_all_stock(struct mem_cgroup *root_memcg)
2336 /* If someone's already draining, avoid adding running more workers. */
2337 if (!mutex_trylock(&percpu_charge_mutex))
2340 * Notify other cpus that system-wide "drain" is running
2341 * We do not care about races with the cpu hotplug because cpu down
2342 * as well as workers from this path always operate on the local
2343 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2346 for_each_online_cpu(cpu) {
2347 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2348 struct mem_cgroup *memcg;
2352 memcg = stock->cached;
2353 if (memcg && stock->nr_pages &&
2354 mem_cgroup_is_descendant(memcg, root_memcg))
2356 if (obj_stock_flush_required(stock, root_memcg))
2361 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2363 drain_local_stock(&stock->work);
2365 schedule_work_on(cpu, &stock->work);
2369 mutex_unlock(&percpu_charge_mutex);
2372 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2374 struct memcg_stock_pcp *stock;
2375 struct mem_cgroup *memcg, *mi;
2377 stock = &per_cpu(memcg_stock, cpu);
2380 for_each_mem_cgroup(memcg) {
2383 for (i = 0; i < MEMCG_NR_STAT; i++) {
2387 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2389 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2390 atomic_long_add(x, &memcg->vmstats[i]);
2392 if (i >= NR_VM_NODE_STAT_ITEMS)
2395 for_each_node(nid) {
2396 struct mem_cgroup_per_node *pn;
2398 pn = mem_cgroup_nodeinfo(memcg, nid);
2399 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2402 atomic_long_add(x, &pn->lruvec_stat[i]);
2403 } while ((pn = parent_nodeinfo(pn, nid)));
2407 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2410 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2412 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2413 atomic_long_add(x, &memcg->vmevents[i]);
2420 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2421 unsigned int nr_pages,
2424 unsigned long nr_reclaimed = 0;
2427 unsigned long pflags;
2429 if (page_counter_read(&memcg->memory) <=
2430 READ_ONCE(memcg->memory.high))
2433 memcg_memory_event(memcg, MEMCG_HIGH);
2435 psi_memstall_enter(&pflags);
2436 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2438 psi_memstall_leave(&pflags);
2439 } while ((memcg = parent_mem_cgroup(memcg)) &&
2440 !mem_cgroup_is_root(memcg));
2442 return nr_reclaimed;
2445 static void high_work_func(struct work_struct *work)
2447 struct mem_cgroup *memcg;
2449 memcg = container_of(work, struct mem_cgroup, high_work);
2450 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2454 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2455 * enough to still cause a significant slowdown in most cases, while still
2456 * allowing diagnostics and tracing to proceed without becoming stuck.
2458 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2461 * When calculating the delay, we use these either side of the exponentiation to
2462 * maintain precision and scale to a reasonable number of jiffies (see the table
2465 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2466 * overage ratio to a delay.
2467 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2468 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2469 * to produce a reasonable delay curve.
2471 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2472 * reasonable delay curve compared to precision-adjusted overage, not
2473 * penalising heavily at first, but still making sure that growth beyond the
2474 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2475 * example, with a high of 100 megabytes:
2477 * +-------+------------------------+
2478 * | usage | time to allocate in ms |
2479 * +-------+------------------------+
2501 * +-------+------------------------+
2503 #define MEMCG_DELAY_PRECISION_SHIFT 20
2504 #define MEMCG_DELAY_SCALING_SHIFT 14
2506 static u64 calculate_overage(unsigned long usage, unsigned long high)
2514 * Prevent division by 0 in overage calculation by acting as if
2515 * it was a threshold of 1 page
2517 high = max(high, 1UL);
2519 overage = usage - high;
2520 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2521 return div64_u64(overage, high);
2524 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2526 u64 overage, max_overage = 0;
2529 overage = calculate_overage(page_counter_read(&memcg->memory),
2530 READ_ONCE(memcg->memory.high));
2531 max_overage = max(overage, max_overage);
2532 } while ((memcg = parent_mem_cgroup(memcg)) &&
2533 !mem_cgroup_is_root(memcg));
2538 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2540 u64 overage, max_overage = 0;
2543 overage = calculate_overage(page_counter_read(&memcg->swap),
2544 READ_ONCE(memcg->swap.high));
2546 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2547 max_overage = max(overage, max_overage);
2548 } while ((memcg = parent_mem_cgroup(memcg)) &&
2549 !mem_cgroup_is_root(memcg));
2555 * Get the number of jiffies that we should penalise a mischievous cgroup which
2556 * is exceeding its memory.high by checking both it and its ancestors.
2558 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2559 unsigned int nr_pages,
2562 unsigned long penalty_jiffies;
2568 * We use overage compared to memory.high to calculate the number of
2569 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2570 * fairly lenient on small overages, and increasingly harsh when the
2571 * memcg in question makes it clear that it has no intention of stopping
2572 * its crazy behaviour, so we exponentially increase the delay based on
2575 penalty_jiffies = max_overage * max_overage * HZ;
2576 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2577 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2580 * Factor in the task's own contribution to the overage, such that four
2581 * N-sized allocations are throttled approximately the same as one
2582 * 4N-sized allocation.
2584 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2585 * larger the current charge patch is than that.
2587 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2591 * Scheduled by try_charge() to be executed from the userland return path
2592 * and reclaims memory over the high limit.
2594 void mem_cgroup_handle_over_high(void)
2596 unsigned long penalty_jiffies;
2597 unsigned long pflags;
2598 unsigned long nr_reclaimed;
2599 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2600 int nr_retries = MAX_RECLAIM_RETRIES;
2601 struct mem_cgroup *memcg;
2602 bool in_retry = false;
2604 if (likely(!nr_pages))
2607 memcg = get_mem_cgroup_from_mm(current->mm);
2608 current->memcg_nr_pages_over_high = 0;
2612 * The allocating task should reclaim at least the batch size, but for
2613 * subsequent retries we only want to do what's necessary to prevent oom
2614 * or breaching resource isolation.
2616 * This is distinct from memory.max or page allocator behaviour because
2617 * memory.high is currently batched, whereas memory.max and the page
2618 * allocator run every time an allocation is made.
2620 nr_reclaimed = reclaim_high(memcg,
2621 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2625 * memory.high is breached and reclaim is unable to keep up. Throttle
2626 * allocators proactively to slow down excessive growth.
2628 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2629 mem_find_max_overage(memcg));
2631 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2632 swap_find_max_overage(memcg));
2635 * Clamp the max delay per usermode return so as to still keep the
2636 * application moving forwards and also permit diagnostics, albeit
2639 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2642 * Don't sleep if the amount of jiffies this memcg owes us is so low
2643 * that it's not even worth doing, in an attempt to be nice to those who
2644 * go only a small amount over their memory.high value and maybe haven't
2645 * been aggressively reclaimed enough yet.
2647 if (penalty_jiffies <= HZ / 100)
2651 * If reclaim is making forward progress but we're still over
2652 * memory.high, we want to encourage that rather than doing allocator
2655 if (nr_reclaimed || nr_retries--) {
2661 * If we exit early, we're guaranteed to die (since
2662 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2663 * need to account for any ill-begotten jiffies to pay them off later.
2665 psi_memstall_enter(&pflags);
2666 schedule_timeout_killable(penalty_jiffies);
2667 psi_memstall_leave(&pflags);
2670 css_put(&memcg->css);
2673 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2674 unsigned int nr_pages)
2676 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2677 int nr_retries = MAX_RECLAIM_RETRIES;
2678 struct mem_cgroup *mem_over_limit;
2679 struct page_counter *counter;
2680 enum oom_status oom_status;
2681 unsigned long nr_reclaimed;
2682 bool may_swap = true;
2683 bool drained = false;
2684 unsigned long pflags;
2686 if (mem_cgroup_is_root(memcg))
2689 if (consume_stock(memcg, nr_pages))
2692 if (!do_memsw_account() ||
2693 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2694 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2696 if (do_memsw_account())
2697 page_counter_uncharge(&memcg->memsw, batch);
2698 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2700 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2704 if (batch > nr_pages) {
2710 * Memcg doesn't have a dedicated reserve for atomic
2711 * allocations. But like the global atomic pool, we need to
2712 * put the burden of reclaim on regular allocation requests
2713 * and let these go through as privileged allocations.
2715 if (gfp_mask & __GFP_ATOMIC)
2719 * Unlike in global OOM situations, memcg is not in a physical
2720 * memory shortage. Allow dying and OOM-killed tasks to
2721 * bypass the last charges so that they can exit quickly and
2722 * free their memory.
2724 if (unlikely(should_force_charge()))
2728 * Prevent unbounded recursion when reclaim operations need to
2729 * allocate memory. This might exceed the limits temporarily,
2730 * but we prefer facilitating memory reclaim and getting back
2731 * under the limit over triggering OOM kills in these cases.
2733 if (unlikely(current->flags & PF_MEMALLOC))
2736 if (unlikely(task_in_memcg_oom(current)))
2739 if (!gfpflags_allow_blocking(gfp_mask))
2742 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2744 psi_memstall_enter(&pflags);
2745 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2746 gfp_mask, may_swap);
2747 psi_memstall_leave(&pflags);
2749 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2753 drain_all_stock(mem_over_limit);
2758 if (gfp_mask & __GFP_NORETRY)
2761 * Even though the limit is exceeded at this point, reclaim
2762 * may have been able to free some pages. Retry the charge
2763 * before killing the task.
2765 * Only for regular pages, though: huge pages are rather
2766 * unlikely to succeed so close to the limit, and we fall back
2767 * to regular pages anyway in case of failure.
2769 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2772 * At task move, charge accounts can be doubly counted. So, it's
2773 * better to wait until the end of task_move if something is going on.
2775 if (mem_cgroup_wait_acct_move(mem_over_limit))
2781 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2784 if (gfp_mask & __GFP_NOFAIL)
2787 if (fatal_signal_pending(current))
2791 * keep retrying as long as the memcg oom killer is able to make
2792 * a forward progress or bypass the charge if the oom killer
2793 * couldn't make any progress.
2795 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2796 get_order(nr_pages * PAGE_SIZE));
2797 switch (oom_status) {
2799 nr_retries = MAX_RECLAIM_RETRIES;
2807 if (!(gfp_mask & __GFP_NOFAIL))
2811 * The allocation either can't fail or will lead to more memory
2812 * being freed very soon. Allow memory usage go over the limit
2813 * temporarily by force charging it.
2815 page_counter_charge(&memcg->memory, nr_pages);
2816 if (do_memsw_account())
2817 page_counter_charge(&memcg->memsw, nr_pages);
2822 if (batch > nr_pages)
2823 refill_stock(memcg, batch - nr_pages);
2826 * If the hierarchy is above the normal consumption range, schedule
2827 * reclaim on returning to userland. We can perform reclaim here
2828 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2829 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2830 * not recorded as it most likely matches current's and won't
2831 * change in the meantime. As high limit is checked again before
2832 * reclaim, the cost of mismatch is negligible.
2835 bool mem_high, swap_high;
2837 mem_high = page_counter_read(&memcg->memory) >
2838 READ_ONCE(memcg->memory.high);
2839 swap_high = page_counter_read(&memcg->swap) >
2840 READ_ONCE(memcg->swap.high);
2842 /* Don't bother a random interrupted task */
2843 if (in_interrupt()) {
2845 schedule_work(&memcg->high_work);
2851 if (mem_high || swap_high) {
2853 * The allocating tasks in this cgroup will need to do
2854 * reclaim or be throttled to prevent further growth
2855 * of the memory or swap footprints.
2857 * Target some best-effort fairness between the tasks,
2858 * and distribute reclaim work and delay penalties
2859 * based on how much each task is actually allocating.
2861 current->memcg_nr_pages_over_high += batch;
2862 set_notify_resume(current);
2865 } while ((memcg = parent_mem_cgroup(memcg)));
2870 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2871 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2873 if (mem_cgroup_is_root(memcg))
2876 page_counter_uncharge(&memcg->memory, nr_pages);
2877 if (do_memsw_account())
2878 page_counter_uncharge(&memcg->memsw, nr_pages);
2882 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2884 VM_BUG_ON_PAGE(page_memcg(page), page);
2886 * Any of the following ensures page's memcg stability:
2890 * - lock_page_memcg()
2891 * - exclusive reference
2893 page->memcg_data = (unsigned long)memcg;
2896 #ifdef CONFIG_MEMCG_KMEM
2897 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2900 unsigned int objects = objs_per_slab_page(s, page);
2903 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2908 if (!set_page_objcgs(page, vec))
2911 kmemleak_not_leak(vec);
2917 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2919 * A passed kernel object can be a slab object or a generic kernel page, so
2920 * different mechanisms for getting the memory cgroup pointer should be used.
2921 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2922 * can not know for sure how the kernel object is implemented.
2923 * mem_cgroup_from_obj() can be safely used in such cases.
2925 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2926 * cgroup_mutex, etc.
2928 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2932 if (mem_cgroup_disabled())
2935 page = virt_to_head_page(p);
2938 * Slab objects are accounted individually, not per-page.
2939 * Memcg membership data for each individual object is saved in
2940 * the page->obj_cgroups.
2942 if (page_objcgs_check(page)) {
2943 struct obj_cgroup *objcg;
2946 off = obj_to_index(page->slab_cache, page, p);
2947 objcg = page_objcgs(page)[off];
2949 return obj_cgroup_memcg(objcg);
2955 * page_memcg_check() is used here, because page_has_obj_cgroups()
2956 * check above could fail because the object cgroups vector wasn't set
2957 * at that moment, but it can be set concurrently.
2958 * page_memcg_check(page) will guarantee that a proper memory
2959 * cgroup pointer or NULL will be returned.
2961 return page_memcg_check(page);
2964 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2966 struct obj_cgroup *objcg = NULL;
2967 struct mem_cgroup *memcg;
2969 if (memcg_kmem_bypass())
2973 if (unlikely(active_memcg()))
2974 memcg = active_memcg();
2976 memcg = mem_cgroup_from_task(current);
2978 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2979 objcg = rcu_dereference(memcg->objcg);
2980 if (objcg && obj_cgroup_tryget(objcg))
2989 static int memcg_alloc_cache_id(void)
2994 id = ida_simple_get(&memcg_cache_ida,
2995 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2999 if (id < memcg_nr_cache_ids)
3003 * There's no space for the new id in memcg_caches arrays,
3004 * so we have to grow them.
3006 down_write(&memcg_cache_ids_sem);
3008 size = 2 * (id + 1);
3009 if (size < MEMCG_CACHES_MIN_SIZE)
3010 size = MEMCG_CACHES_MIN_SIZE;
3011 else if (size > MEMCG_CACHES_MAX_SIZE)
3012 size = MEMCG_CACHES_MAX_SIZE;
3014 err = memcg_update_all_list_lrus(size);
3016 memcg_nr_cache_ids = size;
3018 up_write(&memcg_cache_ids_sem);
3021 ida_simple_remove(&memcg_cache_ida, id);
3027 static void memcg_free_cache_id(int id)
3029 ida_simple_remove(&memcg_cache_ida, id);
3033 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3034 * @memcg: memory cgroup to charge
3035 * @gfp: reclaim mode
3036 * @nr_pages: number of pages to charge
3038 * Returns 0 on success, an error code on failure.
3040 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
3041 unsigned int nr_pages)
3043 struct page_counter *counter;
3046 ret = try_charge(memcg, gfp, nr_pages);
3050 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3051 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3054 * Enforce __GFP_NOFAIL allocation because callers are not
3055 * prepared to see failures and likely do not have any failure
3058 if (gfp & __GFP_NOFAIL) {
3059 page_counter_charge(&memcg->kmem, nr_pages);
3062 cancel_charge(memcg, nr_pages);
3069 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3070 * @memcg: memcg to uncharge
3071 * @nr_pages: number of pages to uncharge
3073 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3075 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3076 page_counter_uncharge(&memcg->kmem, nr_pages);
3078 page_counter_uncharge(&memcg->memory, nr_pages);
3079 if (do_memsw_account())
3080 page_counter_uncharge(&memcg->memsw, nr_pages);
3084 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3085 * @page: page to charge
3086 * @gfp: reclaim mode
3087 * @order: allocation order
3089 * Returns 0 on success, an error code on failure.
3091 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3093 struct mem_cgroup *memcg;
3096 memcg = get_mem_cgroup_from_current();
3097 if (memcg && !mem_cgroup_is_root(memcg)) {
3098 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3100 page->memcg_data = (unsigned long)memcg |
3104 css_put(&memcg->css);
3110 * __memcg_kmem_uncharge_page: uncharge a kmem page
3111 * @page: page to uncharge
3112 * @order: allocation order
3114 void __memcg_kmem_uncharge_page(struct page *page, int order)
3116 struct mem_cgroup *memcg = page_memcg(page);
3117 unsigned int nr_pages = 1 << order;
3122 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3123 __memcg_kmem_uncharge(memcg, nr_pages);
3124 page->memcg_data = 0;
3125 css_put(&memcg->css);
3128 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3130 struct memcg_stock_pcp *stock;
3131 unsigned long flags;
3134 local_irq_save(flags);
3136 stock = this_cpu_ptr(&memcg_stock);
3137 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3138 stock->nr_bytes -= nr_bytes;
3142 local_irq_restore(flags);
3147 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3149 struct obj_cgroup *old = stock->cached_objcg;
3154 if (stock->nr_bytes) {
3155 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3156 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3160 __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
3165 * The leftover is flushed to the centralized per-memcg value.
3166 * On the next attempt to refill obj stock it will be moved
3167 * to a per-cpu stock (probably, on an other CPU), see
3168 * refill_obj_stock().
3170 * How often it's flushed is a trade-off between the memory
3171 * limit enforcement accuracy and potential CPU contention,
3172 * so it might be changed in the future.
3174 atomic_add(nr_bytes, &old->nr_charged_bytes);
3175 stock->nr_bytes = 0;
3178 obj_cgroup_put(old);
3179 stock->cached_objcg = NULL;
3182 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3183 struct mem_cgroup *root_memcg)
3185 struct mem_cgroup *memcg;
3187 if (stock->cached_objcg) {
3188 memcg = obj_cgroup_memcg(stock->cached_objcg);
3189 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3196 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3198 struct memcg_stock_pcp *stock;
3199 unsigned long flags;
3201 local_irq_save(flags);
3203 stock = this_cpu_ptr(&memcg_stock);
3204 if (stock->cached_objcg != objcg) { /* reset if necessary */
3205 drain_obj_stock(stock);
3206 obj_cgroup_get(objcg);
3207 stock->cached_objcg = objcg;
3208 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3210 stock->nr_bytes += nr_bytes;
3212 if (stock->nr_bytes > PAGE_SIZE)
3213 drain_obj_stock(stock);
3215 local_irq_restore(flags);
3218 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3220 struct mem_cgroup *memcg;
3221 unsigned int nr_pages, nr_bytes;
3224 if (consume_obj_stock(objcg, size))
3228 * In theory, memcg->nr_charged_bytes can have enough
3229 * pre-charged bytes to satisfy the allocation. However,
3230 * flushing memcg->nr_charged_bytes requires two atomic
3231 * operations, and memcg->nr_charged_bytes can't be big,
3232 * so it's better to ignore it and try grab some new pages.
3233 * memcg->nr_charged_bytes will be flushed in
3234 * refill_obj_stock(), called from this function or
3235 * independently later.
3239 memcg = obj_cgroup_memcg(objcg);
3240 if (unlikely(!css_tryget(&memcg->css)))
3244 nr_pages = size >> PAGE_SHIFT;
3245 nr_bytes = size & (PAGE_SIZE - 1);
3250 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3251 if (!ret && nr_bytes)
3252 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3254 css_put(&memcg->css);
3258 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3260 refill_obj_stock(objcg, size);
3263 #endif /* CONFIG_MEMCG_KMEM */
3265 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3268 * Because tail pages are not marked as "used", set it. We're under
3269 * pgdat->lru_lock and migration entries setup in all page mappings.
3271 void mem_cgroup_split_huge_fixup(struct page *head)
3273 struct mem_cgroup *memcg = page_memcg(head);
3276 if (mem_cgroup_disabled())
3279 for (i = 1; i < HPAGE_PMD_NR; i++) {
3280 css_get(&memcg->css);
3281 head[i].memcg_data = (unsigned long)memcg;
3284 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3286 #ifdef CONFIG_MEMCG_SWAP
3288 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3289 * @entry: swap entry to be moved
3290 * @from: mem_cgroup which the entry is moved from
3291 * @to: mem_cgroup which the entry is moved to
3293 * It succeeds only when the swap_cgroup's record for this entry is the same
3294 * as the mem_cgroup's id of @from.
3296 * Returns 0 on success, -EINVAL on failure.
3298 * The caller must have charged to @to, IOW, called page_counter_charge() about
3299 * both res and memsw, and called css_get().
3301 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3302 struct mem_cgroup *from, struct mem_cgroup *to)
3304 unsigned short old_id, new_id;
3306 old_id = mem_cgroup_id(from);
3307 new_id = mem_cgroup_id(to);
3309 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3310 mod_memcg_state(from, MEMCG_SWAP, -1);
3311 mod_memcg_state(to, MEMCG_SWAP, 1);
3317 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3318 struct mem_cgroup *from, struct mem_cgroup *to)
3324 static DEFINE_MUTEX(memcg_max_mutex);
3326 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3327 unsigned long max, bool memsw)
3329 bool enlarge = false;
3330 bool drained = false;
3332 bool limits_invariant;
3333 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3336 if (signal_pending(current)) {
3341 mutex_lock(&memcg_max_mutex);
3343 * Make sure that the new limit (memsw or memory limit) doesn't
3344 * break our basic invariant rule memory.max <= memsw.max.
3346 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3347 max <= memcg->memsw.max;
3348 if (!limits_invariant) {
3349 mutex_unlock(&memcg_max_mutex);
3353 if (max > counter->max)
3355 ret = page_counter_set_max(counter, max);
3356 mutex_unlock(&memcg_max_mutex);
3362 drain_all_stock(memcg);
3367 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3368 GFP_KERNEL, !memsw)) {
3374 if (!ret && enlarge)
3375 memcg_oom_recover(memcg);
3380 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3382 unsigned long *total_scanned)
3384 unsigned long nr_reclaimed = 0;
3385 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3386 unsigned long reclaimed;
3388 struct mem_cgroup_tree_per_node *mctz;
3389 unsigned long excess;
3390 unsigned long nr_scanned;
3395 mctz = soft_limit_tree_node(pgdat->node_id);
3398 * Do not even bother to check the largest node if the root
3399 * is empty. Do it lockless to prevent lock bouncing. Races
3400 * are acceptable as soft limit is best effort anyway.
3402 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3406 * This loop can run a while, specially if mem_cgroup's continuously
3407 * keep exceeding their soft limit and putting the system under
3414 mz = mem_cgroup_largest_soft_limit_node(mctz);
3419 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3420 gfp_mask, &nr_scanned);
3421 nr_reclaimed += reclaimed;
3422 *total_scanned += nr_scanned;
3423 spin_lock_irq(&mctz->lock);
3424 __mem_cgroup_remove_exceeded(mz, mctz);
3427 * If we failed to reclaim anything from this memory cgroup
3428 * it is time to move on to the next cgroup
3432 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3434 excess = soft_limit_excess(mz->memcg);
3436 * One school of thought says that we should not add
3437 * back the node to the tree if reclaim returns 0.
3438 * But our reclaim could return 0, simply because due
3439 * to priority we are exposing a smaller subset of
3440 * memory to reclaim from. Consider this as a longer
3443 /* If excess == 0, no tree ops */
3444 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3445 spin_unlock_irq(&mctz->lock);
3446 css_put(&mz->memcg->css);
3449 * Could not reclaim anything and there are no more
3450 * mem cgroups to try or we seem to be looping without
3451 * reclaiming anything.
3453 if (!nr_reclaimed &&
3455 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3457 } while (!nr_reclaimed);
3459 css_put(&next_mz->memcg->css);
3460 return nr_reclaimed;
3464 * Reclaims as many pages from the given memcg as possible.
3466 * Caller is responsible for holding css reference for memcg.
3468 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3470 int nr_retries = MAX_RECLAIM_RETRIES;
3472 /* we call try-to-free pages for make this cgroup empty */
3473 lru_add_drain_all();
3475 drain_all_stock(memcg);
3477 /* try to free all pages in this cgroup */
3478 while (nr_retries && page_counter_read(&memcg->memory)) {
3481 if (signal_pending(current))
3484 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3488 /* maybe some writeback is necessary */
3489 congestion_wait(BLK_RW_ASYNC, HZ/10);
3497 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3498 char *buf, size_t nbytes,
3501 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3503 if (mem_cgroup_is_root(memcg))
3505 return mem_cgroup_force_empty(memcg) ?: nbytes;
3508 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3514 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3515 struct cftype *cft, u64 val)
3520 pr_warn_once("Non-hierarchical mode is deprecated. "
3521 "Please report your usecase to linux-mm@kvack.org if you "
3522 "depend on this functionality.\n");
3527 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3531 if (mem_cgroup_is_root(memcg)) {
3532 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3533 memcg_page_state(memcg, NR_ANON_MAPPED);
3535 val += memcg_page_state(memcg, MEMCG_SWAP);
3538 val = page_counter_read(&memcg->memory);
3540 val = page_counter_read(&memcg->memsw);
3553 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3556 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3557 struct page_counter *counter;
3559 switch (MEMFILE_TYPE(cft->private)) {
3561 counter = &memcg->memory;
3564 counter = &memcg->memsw;
3567 counter = &memcg->kmem;
3570 counter = &memcg->tcpmem;
3576 switch (MEMFILE_ATTR(cft->private)) {
3578 if (counter == &memcg->memory)
3579 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3580 if (counter == &memcg->memsw)
3581 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3582 return (u64)page_counter_read(counter) * PAGE_SIZE;
3584 return (u64)counter->max * PAGE_SIZE;
3586 return (u64)counter->watermark * PAGE_SIZE;
3588 return counter->failcnt;
3589 case RES_SOFT_LIMIT:
3590 return (u64)memcg->soft_limit * PAGE_SIZE;
3596 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3598 unsigned long stat[MEMCG_NR_STAT] = {0};
3599 struct mem_cgroup *mi;
3602 for_each_online_cpu(cpu)
3603 for (i = 0; i < MEMCG_NR_STAT; i++)
3604 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3606 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3607 for (i = 0; i < MEMCG_NR_STAT; i++)
3608 atomic_long_add(stat[i], &mi->vmstats[i]);
3610 for_each_node(node) {
3611 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3612 struct mem_cgroup_per_node *pi;
3614 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3617 for_each_online_cpu(cpu)
3618 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3620 pn->lruvec_stat_cpu->count[i], cpu);
3622 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3623 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3624 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3628 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3630 unsigned long events[NR_VM_EVENT_ITEMS];
3631 struct mem_cgroup *mi;
3634 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3637 for_each_online_cpu(cpu)
3638 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3639 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3642 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3643 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3644 atomic_long_add(events[i], &mi->vmevents[i]);
3647 #ifdef CONFIG_MEMCG_KMEM
3648 static int memcg_online_kmem(struct mem_cgroup *memcg)
3650 struct obj_cgroup *objcg;
3653 if (cgroup_memory_nokmem)
3656 BUG_ON(memcg->kmemcg_id >= 0);
3657 BUG_ON(memcg->kmem_state);
3659 memcg_id = memcg_alloc_cache_id();
3663 objcg = obj_cgroup_alloc();
3665 memcg_free_cache_id(memcg_id);
3668 objcg->memcg = memcg;
3669 rcu_assign_pointer(memcg->objcg, objcg);
3671 static_branch_enable(&memcg_kmem_enabled_key);
3673 memcg->kmemcg_id = memcg_id;
3674 memcg->kmem_state = KMEM_ONLINE;
3679 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3681 struct cgroup_subsys_state *css;
3682 struct mem_cgroup *parent, *child;
3685 if (memcg->kmem_state != KMEM_ONLINE)
3688 memcg->kmem_state = KMEM_ALLOCATED;
3690 parent = parent_mem_cgroup(memcg);
3692 parent = root_mem_cgroup;
3694 memcg_reparent_objcgs(memcg, parent);
3696 kmemcg_id = memcg->kmemcg_id;
3697 BUG_ON(kmemcg_id < 0);
3700 * Change kmemcg_id of this cgroup and all its descendants to the
3701 * parent's id, and then move all entries from this cgroup's list_lrus
3702 * to ones of the parent. After we have finished, all list_lrus
3703 * corresponding to this cgroup are guaranteed to remain empty. The
3704 * ordering is imposed by list_lru_node->lock taken by
3705 * memcg_drain_all_list_lrus().
3707 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3708 css_for_each_descendant_pre(css, &memcg->css) {
3709 child = mem_cgroup_from_css(css);
3710 BUG_ON(child->kmemcg_id != kmemcg_id);
3711 child->kmemcg_id = parent->kmemcg_id;
3715 memcg_drain_all_list_lrus(kmemcg_id, parent);
3717 memcg_free_cache_id(kmemcg_id);
3720 static void memcg_free_kmem(struct mem_cgroup *memcg)
3722 /* css_alloc() failed, offlining didn't happen */
3723 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3724 memcg_offline_kmem(memcg);
3727 static int memcg_online_kmem(struct mem_cgroup *memcg)
3731 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3734 static void memcg_free_kmem(struct mem_cgroup *memcg)
3737 #endif /* CONFIG_MEMCG_KMEM */
3739 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3744 mutex_lock(&memcg_max_mutex);
3745 ret = page_counter_set_max(&memcg->kmem, max);
3746 mutex_unlock(&memcg_max_mutex);
3750 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3754 mutex_lock(&memcg_max_mutex);
3756 ret = page_counter_set_max(&memcg->tcpmem, max);
3760 if (!memcg->tcpmem_active) {
3762 * The active flag needs to be written after the static_key
3763 * update. This is what guarantees that the socket activation
3764 * function is the last one to run. See mem_cgroup_sk_alloc()
3765 * for details, and note that we don't mark any socket as
3766 * belonging to this memcg until that flag is up.
3768 * We need to do this, because static_keys will span multiple
3769 * sites, but we can't control their order. If we mark a socket
3770 * as accounted, but the accounting functions are not patched in
3771 * yet, we'll lose accounting.
3773 * We never race with the readers in mem_cgroup_sk_alloc(),
3774 * because when this value change, the code to process it is not
3777 static_branch_inc(&memcg_sockets_enabled_key);
3778 memcg->tcpmem_active = true;
3781 mutex_unlock(&memcg_max_mutex);
3786 * The user of this function is...
3789 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3790 char *buf, size_t nbytes, loff_t off)
3792 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3793 unsigned long nr_pages;
3796 buf = strstrip(buf);
3797 ret = page_counter_memparse(buf, "-1", &nr_pages);
3801 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3803 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3807 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3809 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3812 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3815 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3816 "Please report your usecase to linux-mm@kvack.org if you "
3817 "depend on this functionality.\n");
3818 ret = memcg_update_kmem_max(memcg, nr_pages);
3821 ret = memcg_update_tcp_max(memcg, nr_pages);
3825 case RES_SOFT_LIMIT:
3826 memcg->soft_limit = nr_pages;
3830 return ret ?: nbytes;
3833 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3834 size_t nbytes, loff_t off)
3836 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3837 struct page_counter *counter;
3839 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3841 counter = &memcg->memory;
3844 counter = &memcg->memsw;
3847 counter = &memcg->kmem;
3850 counter = &memcg->tcpmem;
3856 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3858 page_counter_reset_watermark(counter);
3861 counter->failcnt = 0;
3870 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3873 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3877 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3878 struct cftype *cft, u64 val)
3880 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3882 if (val & ~MOVE_MASK)
3886 * No kind of locking is needed in here, because ->can_attach() will
3887 * check this value once in the beginning of the process, and then carry
3888 * on with stale data. This means that changes to this value will only
3889 * affect task migrations starting after the change.
3891 memcg->move_charge_at_immigrate = val;
3895 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3896 struct cftype *cft, u64 val)
3904 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3905 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3906 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3908 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3909 int nid, unsigned int lru_mask, bool tree)
3911 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3912 unsigned long nr = 0;
3915 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3918 if (!(BIT(lru) & lru_mask))
3921 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3923 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3928 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3929 unsigned int lru_mask,
3932 unsigned long nr = 0;
3936 if (!(BIT(lru) & lru_mask))
3939 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3941 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3946 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3950 unsigned int lru_mask;
3953 static const struct numa_stat stats[] = {
3954 { "total", LRU_ALL },
3955 { "file", LRU_ALL_FILE },
3956 { "anon", LRU_ALL_ANON },
3957 { "unevictable", BIT(LRU_UNEVICTABLE) },
3959 const struct numa_stat *stat;
3961 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3963 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3964 seq_printf(m, "%s=%lu", stat->name,
3965 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3967 for_each_node_state(nid, N_MEMORY)
3968 seq_printf(m, " N%d=%lu", nid,
3969 mem_cgroup_node_nr_lru_pages(memcg, nid,
3970 stat->lru_mask, false));
3974 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3976 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3977 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3979 for_each_node_state(nid, N_MEMORY)
3980 seq_printf(m, " N%d=%lu", nid,
3981 mem_cgroup_node_nr_lru_pages(memcg, nid,
3982 stat->lru_mask, true));
3988 #endif /* CONFIG_NUMA */
3990 static const unsigned int memcg1_stats[] = {
3993 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4003 static const char *const memcg1_stat_names[] = {
4006 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4016 /* Universal VM events cgroup1 shows, original sort order */
4017 static const unsigned int memcg1_events[] = {
4024 static int memcg_stat_show(struct seq_file *m, void *v)
4026 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4027 unsigned long memory, memsw;
4028 struct mem_cgroup *mi;
4031 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4033 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4036 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4038 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4039 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4040 if (memcg1_stats[i] == NR_ANON_THPS)
4043 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4046 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4047 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4048 memcg_events_local(memcg, memcg1_events[i]));
4050 for (i = 0; i < NR_LRU_LISTS; i++)
4051 seq_printf(m, "%s %lu\n", lru_list_name(i),
4052 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4055 /* Hierarchical information */
4056 memory = memsw = PAGE_COUNTER_MAX;
4057 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4058 memory = min(memory, READ_ONCE(mi->memory.max));
4059 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4061 seq_printf(m, "hierarchical_memory_limit %llu\n",
4062 (u64)memory * PAGE_SIZE);
4063 if (do_memsw_account())
4064 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4065 (u64)memsw * PAGE_SIZE);
4067 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4070 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4072 nr = memcg_page_state(memcg, memcg1_stats[i]);
4073 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4074 if (memcg1_stats[i] == NR_ANON_THPS)
4077 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4078 (u64)nr * PAGE_SIZE);
4081 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4082 seq_printf(m, "total_%s %llu\n",
4083 vm_event_name(memcg1_events[i]),
4084 (u64)memcg_events(memcg, memcg1_events[i]));
4086 for (i = 0; i < NR_LRU_LISTS; i++)
4087 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4088 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4091 #ifdef CONFIG_DEBUG_VM
4094 struct mem_cgroup_per_node *mz;
4095 unsigned long anon_cost = 0;
4096 unsigned long file_cost = 0;
4098 for_each_online_pgdat(pgdat) {
4099 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4101 anon_cost += mz->lruvec.anon_cost;
4102 file_cost += mz->lruvec.file_cost;
4104 seq_printf(m, "anon_cost %lu\n", anon_cost);
4105 seq_printf(m, "file_cost %lu\n", file_cost);
4112 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4115 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4117 return mem_cgroup_swappiness(memcg);
4120 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4121 struct cftype *cft, u64 val)
4123 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4129 memcg->swappiness = val;
4131 vm_swappiness = val;
4136 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4138 struct mem_cgroup_threshold_ary *t;
4139 unsigned long usage;
4144 t = rcu_dereference(memcg->thresholds.primary);
4146 t = rcu_dereference(memcg->memsw_thresholds.primary);
4151 usage = mem_cgroup_usage(memcg, swap);
4154 * current_threshold points to threshold just below or equal to usage.
4155 * If it's not true, a threshold was crossed after last
4156 * call of __mem_cgroup_threshold().
4158 i = t->current_threshold;
4161 * Iterate backward over array of thresholds starting from
4162 * current_threshold and check if a threshold is crossed.
4163 * If none of thresholds below usage is crossed, we read
4164 * only one element of the array here.
4166 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4167 eventfd_signal(t->entries[i].eventfd, 1);
4169 /* i = current_threshold + 1 */
4173 * Iterate forward over array of thresholds starting from
4174 * current_threshold+1 and check if a threshold is crossed.
4175 * If none of thresholds above usage is crossed, we read
4176 * only one element of the array here.
4178 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4179 eventfd_signal(t->entries[i].eventfd, 1);
4181 /* Update current_threshold */
4182 t->current_threshold = i - 1;
4187 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4190 __mem_cgroup_threshold(memcg, false);
4191 if (do_memsw_account())
4192 __mem_cgroup_threshold(memcg, true);
4194 memcg = parent_mem_cgroup(memcg);
4198 static int compare_thresholds(const void *a, const void *b)
4200 const struct mem_cgroup_threshold *_a = a;
4201 const struct mem_cgroup_threshold *_b = b;
4203 if (_a->threshold > _b->threshold)
4206 if (_a->threshold < _b->threshold)
4212 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4214 struct mem_cgroup_eventfd_list *ev;
4216 spin_lock(&memcg_oom_lock);
4218 list_for_each_entry(ev, &memcg->oom_notify, list)
4219 eventfd_signal(ev->eventfd, 1);
4221 spin_unlock(&memcg_oom_lock);
4225 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4227 struct mem_cgroup *iter;
4229 for_each_mem_cgroup_tree(iter, memcg)
4230 mem_cgroup_oom_notify_cb(iter);
4233 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4234 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4236 struct mem_cgroup_thresholds *thresholds;
4237 struct mem_cgroup_threshold_ary *new;
4238 unsigned long threshold;
4239 unsigned long usage;
4242 ret = page_counter_memparse(args, "-1", &threshold);
4246 mutex_lock(&memcg->thresholds_lock);
4249 thresholds = &memcg->thresholds;
4250 usage = mem_cgroup_usage(memcg, false);
4251 } else if (type == _MEMSWAP) {
4252 thresholds = &memcg->memsw_thresholds;
4253 usage = mem_cgroup_usage(memcg, true);
4257 /* Check if a threshold crossed before adding a new one */
4258 if (thresholds->primary)
4259 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4261 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4263 /* Allocate memory for new array of thresholds */
4264 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4271 /* Copy thresholds (if any) to new array */
4272 if (thresholds->primary)
4273 memcpy(new->entries, thresholds->primary->entries,
4274 flex_array_size(new, entries, size - 1));
4276 /* Add new threshold */
4277 new->entries[size - 1].eventfd = eventfd;
4278 new->entries[size - 1].threshold = threshold;
4280 /* Sort thresholds. Registering of new threshold isn't time-critical */
4281 sort(new->entries, size, sizeof(*new->entries),
4282 compare_thresholds, NULL);
4284 /* Find current threshold */
4285 new->current_threshold = -1;
4286 for (i = 0; i < size; i++) {
4287 if (new->entries[i].threshold <= usage) {
4289 * new->current_threshold will not be used until
4290 * rcu_assign_pointer(), so it's safe to increment
4293 ++new->current_threshold;
4298 /* Free old spare buffer and save old primary buffer as spare */
4299 kfree(thresholds->spare);
4300 thresholds->spare = thresholds->primary;
4302 rcu_assign_pointer(thresholds->primary, new);
4304 /* To be sure that nobody uses thresholds */
4308 mutex_unlock(&memcg->thresholds_lock);
4313 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4314 struct eventfd_ctx *eventfd, const char *args)
4316 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4319 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4320 struct eventfd_ctx *eventfd, const char *args)
4322 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4325 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4326 struct eventfd_ctx *eventfd, enum res_type type)
4328 struct mem_cgroup_thresholds *thresholds;
4329 struct mem_cgroup_threshold_ary *new;
4330 unsigned long usage;
4331 int i, j, size, entries;
4333 mutex_lock(&memcg->thresholds_lock);
4336 thresholds = &memcg->thresholds;
4337 usage = mem_cgroup_usage(memcg, false);
4338 } else if (type == _MEMSWAP) {
4339 thresholds = &memcg->memsw_thresholds;
4340 usage = mem_cgroup_usage(memcg, true);
4344 if (!thresholds->primary)
4347 /* Check if a threshold crossed before removing */
4348 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4350 /* Calculate new number of threshold */
4352 for (i = 0; i < thresholds->primary->size; i++) {
4353 if (thresholds->primary->entries[i].eventfd != eventfd)
4359 new = thresholds->spare;
4361 /* If no items related to eventfd have been cleared, nothing to do */
4365 /* Set thresholds array to NULL if we don't have thresholds */
4374 /* Copy thresholds and find current threshold */
4375 new->current_threshold = -1;
4376 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4377 if (thresholds->primary->entries[i].eventfd == eventfd)
4380 new->entries[j] = thresholds->primary->entries[i];
4381 if (new->entries[j].threshold <= usage) {
4383 * new->current_threshold will not be used
4384 * until rcu_assign_pointer(), so it's safe to increment
4387 ++new->current_threshold;
4393 /* Swap primary and spare array */
4394 thresholds->spare = thresholds->primary;
4396 rcu_assign_pointer(thresholds->primary, new);
4398 /* To be sure that nobody uses thresholds */
4401 /* If all events are unregistered, free the spare array */
4403 kfree(thresholds->spare);
4404 thresholds->spare = NULL;
4407 mutex_unlock(&memcg->thresholds_lock);
4410 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4411 struct eventfd_ctx *eventfd)
4413 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4416 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4417 struct eventfd_ctx *eventfd)
4419 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4422 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4423 struct eventfd_ctx *eventfd, const char *args)
4425 struct mem_cgroup_eventfd_list *event;
4427 event = kmalloc(sizeof(*event), GFP_KERNEL);
4431 spin_lock(&memcg_oom_lock);
4433 event->eventfd = eventfd;
4434 list_add(&event->list, &memcg->oom_notify);
4436 /* already in OOM ? */
4437 if (memcg->under_oom)
4438 eventfd_signal(eventfd, 1);
4439 spin_unlock(&memcg_oom_lock);
4444 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4445 struct eventfd_ctx *eventfd)
4447 struct mem_cgroup_eventfd_list *ev, *tmp;
4449 spin_lock(&memcg_oom_lock);
4451 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4452 if (ev->eventfd == eventfd) {
4453 list_del(&ev->list);
4458 spin_unlock(&memcg_oom_lock);
4461 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4463 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4465 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4466 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4467 seq_printf(sf, "oom_kill %lu\n",
4468 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4472 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4473 struct cftype *cft, u64 val)
4475 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4477 /* cannot set to root cgroup and only 0 and 1 are allowed */
4478 if (!css->parent || !((val == 0) || (val == 1)))
4481 memcg->oom_kill_disable = val;
4483 memcg_oom_recover(memcg);
4488 #ifdef CONFIG_CGROUP_WRITEBACK
4490 #include <trace/events/writeback.h>
4492 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4494 return wb_domain_init(&memcg->cgwb_domain, gfp);
4497 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4499 wb_domain_exit(&memcg->cgwb_domain);
4502 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4504 wb_domain_size_changed(&memcg->cgwb_domain);
4507 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4509 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4511 if (!memcg->css.parent)
4514 return &memcg->cgwb_domain;
4518 * idx can be of type enum memcg_stat_item or node_stat_item.
4519 * Keep in sync with memcg_exact_page().
4521 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4523 long x = atomic_long_read(&memcg->vmstats[idx]);
4526 for_each_online_cpu(cpu)
4527 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4534 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4535 * @wb: bdi_writeback in question
4536 * @pfilepages: out parameter for number of file pages
4537 * @pheadroom: out parameter for number of allocatable pages according to memcg
4538 * @pdirty: out parameter for number of dirty pages
4539 * @pwriteback: out parameter for number of pages under writeback
4541 * Determine the numbers of file, headroom, dirty, and writeback pages in
4542 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4543 * is a bit more involved.
4545 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4546 * headroom is calculated as the lowest headroom of itself and the
4547 * ancestors. Note that this doesn't consider the actual amount of
4548 * available memory in the system. The caller should further cap
4549 * *@pheadroom accordingly.
4551 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4552 unsigned long *pheadroom, unsigned long *pdirty,
4553 unsigned long *pwriteback)
4555 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4556 struct mem_cgroup *parent;
4558 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4560 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4561 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4562 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4563 *pheadroom = PAGE_COUNTER_MAX;
4565 while ((parent = parent_mem_cgroup(memcg))) {
4566 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4567 READ_ONCE(memcg->memory.high));
4568 unsigned long used = page_counter_read(&memcg->memory);
4570 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4576 * Foreign dirty flushing
4578 * There's an inherent mismatch between memcg and writeback. The former
4579 * trackes ownership per-page while the latter per-inode. This was a
4580 * deliberate design decision because honoring per-page ownership in the
4581 * writeback path is complicated, may lead to higher CPU and IO overheads
4582 * and deemed unnecessary given that write-sharing an inode across
4583 * different cgroups isn't a common use-case.
4585 * Combined with inode majority-writer ownership switching, this works well
4586 * enough in most cases but there are some pathological cases. For
4587 * example, let's say there are two cgroups A and B which keep writing to
4588 * different but confined parts of the same inode. B owns the inode and
4589 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4590 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4591 * triggering background writeback. A will be slowed down without a way to
4592 * make writeback of the dirty pages happen.
4594 * Conditions like the above can lead to a cgroup getting repatedly and
4595 * severely throttled after making some progress after each
4596 * dirty_expire_interval while the underyling IO device is almost
4599 * Solving this problem completely requires matching the ownership tracking
4600 * granularities between memcg and writeback in either direction. However,
4601 * the more egregious behaviors can be avoided by simply remembering the
4602 * most recent foreign dirtying events and initiating remote flushes on
4603 * them when local writeback isn't enough to keep the memory clean enough.
4605 * The following two functions implement such mechanism. When a foreign
4606 * page - a page whose memcg and writeback ownerships don't match - is
4607 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4608 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4609 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4610 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4611 * foreign bdi_writebacks which haven't expired. Both the numbers of
4612 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4613 * limited to MEMCG_CGWB_FRN_CNT.
4615 * The mechanism only remembers IDs and doesn't hold any object references.
4616 * As being wrong occasionally doesn't matter, updates and accesses to the
4617 * records are lockless and racy.
4619 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4620 struct bdi_writeback *wb)
4622 struct mem_cgroup *memcg = page_memcg(page);
4623 struct memcg_cgwb_frn *frn;
4624 u64 now = get_jiffies_64();
4625 u64 oldest_at = now;
4629 trace_track_foreign_dirty(page, wb);
4632 * Pick the slot to use. If there is already a slot for @wb, keep
4633 * using it. If not replace the oldest one which isn't being
4636 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4637 frn = &memcg->cgwb_frn[i];
4638 if (frn->bdi_id == wb->bdi->id &&
4639 frn->memcg_id == wb->memcg_css->id)
4641 if (time_before64(frn->at, oldest_at) &&
4642 atomic_read(&frn->done.cnt) == 1) {
4644 oldest_at = frn->at;
4648 if (i < MEMCG_CGWB_FRN_CNT) {
4650 * Re-using an existing one. Update timestamp lazily to
4651 * avoid making the cacheline hot. We want them to be
4652 * reasonably up-to-date and significantly shorter than
4653 * dirty_expire_interval as that's what expires the record.
4654 * Use the shorter of 1s and dirty_expire_interval / 8.
4656 unsigned long update_intv =
4657 min_t(unsigned long, HZ,
4658 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4660 if (time_before64(frn->at, now - update_intv))
4662 } else if (oldest >= 0) {
4663 /* replace the oldest free one */
4664 frn = &memcg->cgwb_frn[oldest];
4665 frn->bdi_id = wb->bdi->id;
4666 frn->memcg_id = wb->memcg_css->id;
4671 /* issue foreign writeback flushes for recorded foreign dirtying events */
4672 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4674 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4675 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4676 u64 now = jiffies_64;
4679 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4680 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4683 * If the record is older than dirty_expire_interval,
4684 * writeback on it has already started. No need to kick it
4685 * off again. Also, don't start a new one if there's
4686 * already one in flight.
4688 if (time_after64(frn->at, now - intv) &&
4689 atomic_read(&frn->done.cnt) == 1) {
4691 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4692 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4693 WB_REASON_FOREIGN_FLUSH,
4699 #else /* CONFIG_CGROUP_WRITEBACK */
4701 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4706 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4710 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4714 #endif /* CONFIG_CGROUP_WRITEBACK */
4717 * DO NOT USE IN NEW FILES.
4719 * "cgroup.event_control" implementation.
4721 * This is way over-engineered. It tries to support fully configurable
4722 * events for each user. Such level of flexibility is completely
4723 * unnecessary especially in the light of the planned unified hierarchy.
4725 * Please deprecate this and replace with something simpler if at all
4730 * Unregister event and free resources.
4732 * Gets called from workqueue.
4734 static void memcg_event_remove(struct work_struct *work)
4736 struct mem_cgroup_event *event =
4737 container_of(work, struct mem_cgroup_event, remove);
4738 struct mem_cgroup *memcg = event->memcg;
4740 remove_wait_queue(event->wqh, &event->wait);
4742 event->unregister_event(memcg, event->eventfd);
4744 /* Notify userspace the event is going away. */
4745 eventfd_signal(event->eventfd, 1);
4747 eventfd_ctx_put(event->eventfd);
4749 css_put(&memcg->css);
4753 * Gets called on EPOLLHUP on eventfd when user closes it.
4755 * Called with wqh->lock held and interrupts disabled.
4757 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4758 int sync, void *key)
4760 struct mem_cgroup_event *event =
4761 container_of(wait, struct mem_cgroup_event, wait);
4762 struct mem_cgroup *memcg = event->memcg;
4763 __poll_t flags = key_to_poll(key);
4765 if (flags & EPOLLHUP) {
4767 * If the event has been detached at cgroup removal, we
4768 * can simply return knowing the other side will cleanup
4771 * We can't race against event freeing since the other
4772 * side will require wqh->lock via remove_wait_queue(),
4775 spin_lock(&memcg->event_list_lock);
4776 if (!list_empty(&event->list)) {
4777 list_del_init(&event->list);
4779 * We are in atomic context, but cgroup_event_remove()
4780 * may sleep, so we have to call it in workqueue.
4782 schedule_work(&event->remove);
4784 spin_unlock(&memcg->event_list_lock);
4790 static void memcg_event_ptable_queue_proc(struct file *file,
4791 wait_queue_head_t *wqh, poll_table *pt)
4793 struct mem_cgroup_event *event =
4794 container_of(pt, struct mem_cgroup_event, pt);
4797 add_wait_queue(wqh, &event->wait);
4801 * DO NOT USE IN NEW FILES.
4803 * Parse input and register new cgroup event handler.
4805 * Input must be in format '<event_fd> <control_fd> <args>'.
4806 * Interpretation of args is defined by control file implementation.
4808 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4809 char *buf, size_t nbytes, loff_t off)
4811 struct cgroup_subsys_state *css = of_css(of);
4812 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4813 struct mem_cgroup_event *event;
4814 struct cgroup_subsys_state *cfile_css;
4815 unsigned int efd, cfd;
4822 buf = strstrip(buf);
4824 efd = simple_strtoul(buf, &endp, 10);
4829 cfd = simple_strtoul(buf, &endp, 10);
4830 if ((*endp != ' ') && (*endp != '\0'))
4834 event = kzalloc(sizeof(*event), GFP_KERNEL);
4838 event->memcg = memcg;
4839 INIT_LIST_HEAD(&event->list);
4840 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4841 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4842 INIT_WORK(&event->remove, memcg_event_remove);
4850 event->eventfd = eventfd_ctx_fileget(efile.file);
4851 if (IS_ERR(event->eventfd)) {
4852 ret = PTR_ERR(event->eventfd);
4859 goto out_put_eventfd;
4862 /* the process need read permission on control file */
4863 /* AV: shouldn't we check that it's been opened for read instead? */
4864 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4869 * Determine the event callbacks and set them in @event. This used
4870 * to be done via struct cftype but cgroup core no longer knows
4871 * about these events. The following is crude but the whole thing
4872 * is for compatibility anyway.
4874 * DO NOT ADD NEW FILES.
4876 name = cfile.file->f_path.dentry->d_name.name;
4878 if (!strcmp(name, "memory.usage_in_bytes")) {
4879 event->register_event = mem_cgroup_usage_register_event;
4880 event->unregister_event = mem_cgroup_usage_unregister_event;
4881 } else if (!strcmp(name, "memory.oom_control")) {
4882 event->register_event = mem_cgroup_oom_register_event;
4883 event->unregister_event = mem_cgroup_oom_unregister_event;
4884 } else if (!strcmp(name, "memory.pressure_level")) {
4885 event->register_event = vmpressure_register_event;
4886 event->unregister_event = vmpressure_unregister_event;
4887 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4888 event->register_event = memsw_cgroup_usage_register_event;
4889 event->unregister_event = memsw_cgroup_usage_unregister_event;
4896 * Verify @cfile should belong to @css. Also, remaining events are
4897 * automatically removed on cgroup destruction but the removal is
4898 * asynchronous, so take an extra ref on @css.
4900 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4901 &memory_cgrp_subsys);
4903 if (IS_ERR(cfile_css))
4905 if (cfile_css != css) {
4910 ret = event->register_event(memcg, event->eventfd, buf);
4914 vfs_poll(efile.file, &event->pt);
4916 spin_lock(&memcg->event_list_lock);
4917 list_add(&event->list, &memcg->event_list);
4918 spin_unlock(&memcg->event_list_lock);
4930 eventfd_ctx_put(event->eventfd);
4939 static struct cftype mem_cgroup_legacy_files[] = {
4941 .name = "usage_in_bytes",
4942 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4943 .read_u64 = mem_cgroup_read_u64,
4946 .name = "max_usage_in_bytes",
4947 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4948 .write = mem_cgroup_reset,
4949 .read_u64 = mem_cgroup_read_u64,
4952 .name = "limit_in_bytes",
4953 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4954 .write = mem_cgroup_write,
4955 .read_u64 = mem_cgroup_read_u64,
4958 .name = "soft_limit_in_bytes",
4959 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4960 .write = mem_cgroup_write,
4961 .read_u64 = mem_cgroup_read_u64,
4965 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4966 .write = mem_cgroup_reset,
4967 .read_u64 = mem_cgroup_read_u64,
4971 .seq_show = memcg_stat_show,
4974 .name = "force_empty",
4975 .write = mem_cgroup_force_empty_write,
4978 .name = "use_hierarchy",
4979 .write_u64 = mem_cgroup_hierarchy_write,
4980 .read_u64 = mem_cgroup_hierarchy_read,
4983 .name = "cgroup.event_control", /* XXX: for compat */
4984 .write = memcg_write_event_control,
4985 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4988 .name = "swappiness",
4989 .read_u64 = mem_cgroup_swappiness_read,
4990 .write_u64 = mem_cgroup_swappiness_write,
4993 .name = "move_charge_at_immigrate",
4994 .read_u64 = mem_cgroup_move_charge_read,
4995 .write_u64 = mem_cgroup_move_charge_write,
4998 .name = "oom_control",
4999 .seq_show = mem_cgroup_oom_control_read,
5000 .write_u64 = mem_cgroup_oom_control_write,
5001 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5004 .name = "pressure_level",
5008 .name = "numa_stat",
5009 .seq_show = memcg_numa_stat_show,
5013 .name = "kmem.limit_in_bytes",
5014 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5015 .write = mem_cgroup_write,
5016 .read_u64 = mem_cgroup_read_u64,
5019 .name = "kmem.usage_in_bytes",
5020 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5021 .read_u64 = mem_cgroup_read_u64,
5024 .name = "kmem.failcnt",
5025 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5026 .write = mem_cgroup_reset,
5027 .read_u64 = mem_cgroup_read_u64,
5030 .name = "kmem.max_usage_in_bytes",
5031 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5032 .write = mem_cgroup_reset,
5033 .read_u64 = mem_cgroup_read_u64,
5035 #if defined(CONFIG_MEMCG_KMEM) && \
5036 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5038 .name = "kmem.slabinfo",
5039 .seq_show = memcg_slab_show,
5043 .name = "kmem.tcp.limit_in_bytes",
5044 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5045 .write = mem_cgroup_write,
5046 .read_u64 = mem_cgroup_read_u64,
5049 .name = "kmem.tcp.usage_in_bytes",
5050 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5051 .read_u64 = mem_cgroup_read_u64,
5054 .name = "kmem.tcp.failcnt",
5055 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5056 .write = mem_cgroup_reset,
5057 .read_u64 = mem_cgroup_read_u64,
5060 .name = "kmem.tcp.max_usage_in_bytes",
5061 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5062 .write = mem_cgroup_reset,
5063 .read_u64 = mem_cgroup_read_u64,
5065 { }, /* terminate */
5069 * Private memory cgroup IDR
5071 * Swap-out records and page cache shadow entries need to store memcg
5072 * references in constrained space, so we maintain an ID space that is
5073 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5074 * memory-controlled cgroups to 64k.
5076 * However, there usually are many references to the offline CSS after
5077 * the cgroup has been destroyed, such as page cache or reclaimable
5078 * slab objects, that don't need to hang on to the ID. We want to keep
5079 * those dead CSS from occupying IDs, or we might quickly exhaust the
5080 * relatively small ID space and prevent the creation of new cgroups
5081 * even when there are much fewer than 64k cgroups - possibly none.
5083 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5084 * be freed and recycled when it's no longer needed, which is usually
5085 * when the CSS is offlined.
5087 * The only exception to that are records of swapped out tmpfs/shmem
5088 * pages that need to be attributed to live ancestors on swapin. But
5089 * those references are manageable from userspace.
5092 static DEFINE_IDR(mem_cgroup_idr);
5094 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5096 if (memcg->id.id > 0) {
5097 idr_remove(&mem_cgroup_idr, memcg->id.id);
5102 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5105 refcount_add(n, &memcg->id.ref);
5108 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5110 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5111 mem_cgroup_id_remove(memcg);
5113 /* Memcg ID pins CSS */
5114 css_put(&memcg->css);
5118 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5120 mem_cgroup_id_put_many(memcg, 1);
5124 * mem_cgroup_from_id - look up a memcg from a memcg id
5125 * @id: the memcg id to look up
5127 * Caller must hold rcu_read_lock().
5129 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5131 WARN_ON_ONCE(!rcu_read_lock_held());
5132 return idr_find(&mem_cgroup_idr, id);
5135 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5137 struct mem_cgroup_per_node *pn;
5140 * This routine is called against possible nodes.
5141 * But it's BUG to call kmalloc() against offline node.
5143 * TODO: this routine can waste much memory for nodes which will
5144 * never be onlined. It's better to use memory hotplug callback
5147 if (!node_state(node, N_NORMAL_MEMORY))
5149 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5153 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5154 GFP_KERNEL_ACCOUNT);
5155 if (!pn->lruvec_stat_local) {
5160 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct lruvec_stat,
5161 GFP_KERNEL_ACCOUNT);
5162 if (!pn->lruvec_stat_cpu) {
5163 free_percpu(pn->lruvec_stat_local);
5168 lruvec_init(&pn->lruvec);
5169 pn->usage_in_excess = 0;
5170 pn->on_tree = false;
5173 memcg->nodeinfo[node] = pn;
5177 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5179 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5184 free_percpu(pn->lruvec_stat_cpu);
5185 free_percpu(pn->lruvec_stat_local);
5189 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5194 free_mem_cgroup_per_node_info(memcg, node);
5195 free_percpu(memcg->vmstats_percpu);
5196 free_percpu(memcg->vmstats_local);
5200 static void mem_cgroup_free(struct mem_cgroup *memcg)
5202 memcg_wb_domain_exit(memcg);
5204 * Flush percpu vmstats and vmevents to guarantee the value correctness
5205 * on parent's and all ancestor levels.
5207 memcg_flush_percpu_vmstats(memcg);
5208 memcg_flush_percpu_vmevents(memcg);
5209 __mem_cgroup_free(memcg);
5212 static struct mem_cgroup *mem_cgroup_alloc(void)
5214 struct mem_cgroup *memcg;
5217 int __maybe_unused i;
5218 long error = -ENOMEM;
5220 size = sizeof(struct mem_cgroup);
5221 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5223 memcg = kzalloc(size, GFP_KERNEL);
5225 return ERR_PTR(error);
5227 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5228 1, MEM_CGROUP_ID_MAX,
5230 if (memcg->id.id < 0) {
5231 error = memcg->id.id;
5235 memcg->vmstats_local = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5236 GFP_KERNEL_ACCOUNT);
5237 if (!memcg->vmstats_local)
5240 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5241 GFP_KERNEL_ACCOUNT);
5242 if (!memcg->vmstats_percpu)
5246 if (alloc_mem_cgroup_per_node_info(memcg, node))
5249 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5252 INIT_WORK(&memcg->high_work, high_work_func);
5253 INIT_LIST_HEAD(&memcg->oom_notify);
5254 mutex_init(&memcg->thresholds_lock);
5255 spin_lock_init(&memcg->move_lock);
5256 vmpressure_init(&memcg->vmpressure);
5257 INIT_LIST_HEAD(&memcg->event_list);
5258 spin_lock_init(&memcg->event_list_lock);
5259 memcg->socket_pressure = jiffies;
5260 #ifdef CONFIG_MEMCG_KMEM
5261 memcg->kmemcg_id = -1;
5262 INIT_LIST_HEAD(&memcg->objcg_list);
5264 #ifdef CONFIG_CGROUP_WRITEBACK
5265 INIT_LIST_HEAD(&memcg->cgwb_list);
5266 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5267 memcg->cgwb_frn[i].done =
5268 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5270 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5271 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5272 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5273 memcg->deferred_split_queue.split_queue_len = 0;
5275 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5278 mem_cgroup_id_remove(memcg);
5279 __mem_cgroup_free(memcg);
5280 return ERR_PTR(error);
5283 static struct cgroup_subsys_state * __ref
5284 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5286 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5287 struct mem_cgroup *memcg, *old_memcg;
5288 long error = -ENOMEM;
5290 old_memcg = set_active_memcg(parent);
5291 memcg = mem_cgroup_alloc();
5292 set_active_memcg(old_memcg);
5294 return ERR_CAST(memcg);
5296 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5297 memcg->soft_limit = PAGE_COUNTER_MAX;
5298 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5300 memcg->swappiness = mem_cgroup_swappiness(parent);
5301 memcg->oom_kill_disable = parent->oom_kill_disable;
5303 page_counter_init(&memcg->memory, &parent->memory);
5304 page_counter_init(&memcg->swap, &parent->swap);
5305 page_counter_init(&memcg->kmem, &parent->kmem);
5306 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5308 page_counter_init(&memcg->memory, NULL);
5309 page_counter_init(&memcg->swap, NULL);
5310 page_counter_init(&memcg->kmem, NULL);
5311 page_counter_init(&memcg->tcpmem, NULL);
5313 root_mem_cgroup = memcg;
5317 /* The following stuff does not apply to the root */
5318 error = memcg_online_kmem(memcg);
5322 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5323 static_branch_inc(&memcg_sockets_enabled_key);
5327 mem_cgroup_id_remove(memcg);
5328 mem_cgroup_free(memcg);
5329 return ERR_PTR(error);
5332 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5334 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5337 * A memcg must be visible for memcg_expand_shrinker_maps()
5338 * by the time the maps are allocated. So, we allocate maps
5339 * here, when for_each_mem_cgroup() can't skip it.
5341 if (memcg_alloc_shrinker_maps(memcg)) {
5342 mem_cgroup_id_remove(memcg);
5346 /* Online state pins memcg ID, memcg ID pins CSS */
5347 refcount_set(&memcg->id.ref, 1);
5352 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5354 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5355 struct mem_cgroup_event *event, *tmp;
5358 * Unregister events and notify userspace.
5359 * Notify userspace about cgroup removing only after rmdir of cgroup
5360 * directory to avoid race between userspace and kernelspace.
5362 spin_lock(&memcg->event_list_lock);
5363 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5364 list_del_init(&event->list);
5365 schedule_work(&event->remove);
5367 spin_unlock(&memcg->event_list_lock);
5369 page_counter_set_min(&memcg->memory, 0);
5370 page_counter_set_low(&memcg->memory, 0);
5372 memcg_offline_kmem(memcg);
5373 wb_memcg_offline(memcg);
5375 drain_all_stock(memcg);
5377 mem_cgroup_id_put(memcg);
5380 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5382 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5384 invalidate_reclaim_iterators(memcg);
5387 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5389 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5390 int __maybe_unused i;
5392 #ifdef CONFIG_CGROUP_WRITEBACK
5393 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5394 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5396 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5397 static_branch_dec(&memcg_sockets_enabled_key);
5399 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5400 static_branch_dec(&memcg_sockets_enabled_key);
5402 vmpressure_cleanup(&memcg->vmpressure);
5403 cancel_work_sync(&memcg->high_work);
5404 mem_cgroup_remove_from_trees(memcg);
5405 memcg_free_shrinker_maps(memcg);
5406 memcg_free_kmem(memcg);
5407 mem_cgroup_free(memcg);
5411 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5412 * @css: the target css
5414 * Reset the states of the mem_cgroup associated with @css. This is
5415 * invoked when the userland requests disabling on the default hierarchy
5416 * but the memcg is pinned through dependency. The memcg should stop
5417 * applying policies and should revert to the vanilla state as it may be
5418 * made visible again.
5420 * The current implementation only resets the essential configurations.
5421 * This needs to be expanded to cover all the visible parts.
5423 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5425 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5427 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5428 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5429 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5430 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5431 page_counter_set_min(&memcg->memory, 0);
5432 page_counter_set_low(&memcg->memory, 0);
5433 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5434 memcg->soft_limit = PAGE_COUNTER_MAX;
5435 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5436 memcg_wb_domain_size_changed(memcg);
5440 /* Handlers for move charge at task migration. */
5441 static int mem_cgroup_do_precharge(unsigned long count)
5445 /* Try a single bulk charge without reclaim first, kswapd may wake */
5446 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5448 mc.precharge += count;
5452 /* Try charges one by one with reclaim, but do not retry */
5454 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5468 enum mc_target_type {
5475 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5476 unsigned long addr, pte_t ptent)
5478 struct page *page = vm_normal_page(vma, addr, ptent);
5480 if (!page || !page_mapped(page))
5482 if (PageAnon(page)) {
5483 if (!(mc.flags & MOVE_ANON))
5486 if (!(mc.flags & MOVE_FILE))
5489 if (!get_page_unless_zero(page))
5495 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5496 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5497 pte_t ptent, swp_entry_t *entry)
5499 struct page *page = NULL;
5500 swp_entry_t ent = pte_to_swp_entry(ptent);
5502 if (!(mc.flags & MOVE_ANON))
5506 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5507 * a device and because they are not accessible by CPU they are store
5508 * as special swap entry in the CPU page table.
5510 if (is_device_private_entry(ent)) {
5511 page = device_private_entry_to_page(ent);
5513 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5514 * a refcount of 1 when free (unlike normal page)
5516 if (!page_ref_add_unless(page, 1, 1))
5521 if (non_swap_entry(ent))
5525 * Because lookup_swap_cache() updates some statistics counter,
5526 * we call find_get_page() with swapper_space directly.
5528 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5529 entry->val = ent.val;
5534 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5535 pte_t ptent, swp_entry_t *entry)
5541 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5542 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5544 if (!vma->vm_file) /* anonymous vma */
5546 if (!(mc.flags & MOVE_FILE))
5549 /* page is moved even if it's not RSS of this task(page-faulted). */
5550 /* shmem/tmpfs may report page out on swap: account for that too. */
5551 return find_get_incore_page(vma->vm_file->f_mapping,
5552 linear_page_index(vma, addr));
5556 * mem_cgroup_move_account - move account of the page
5558 * @compound: charge the page as compound or small page
5559 * @from: mem_cgroup which the page is moved from.
5560 * @to: mem_cgroup which the page is moved to. @from != @to.
5562 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5564 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5567 static int mem_cgroup_move_account(struct page *page,
5569 struct mem_cgroup *from,
5570 struct mem_cgroup *to)
5572 struct lruvec *from_vec, *to_vec;
5573 struct pglist_data *pgdat;
5574 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5577 VM_BUG_ON(from == to);
5578 VM_BUG_ON_PAGE(PageLRU(page), page);
5579 VM_BUG_ON(compound && !PageTransHuge(page));
5582 * Prevent mem_cgroup_migrate() from looking at
5583 * page's memory cgroup of its source page while we change it.
5586 if (!trylock_page(page))
5590 if (page_memcg(page) != from)
5593 pgdat = page_pgdat(page);
5594 from_vec = mem_cgroup_lruvec(from, pgdat);
5595 to_vec = mem_cgroup_lruvec(to, pgdat);
5597 lock_page_memcg(page);
5599 if (PageAnon(page)) {
5600 if (page_mapped(page)) {
5601 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5602 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5603 if (PageTransHuge(page)) {
5604 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5606 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5612 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5613 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5615 if (PageSwapBacked(page)) {
5616 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5617 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5620 if (page_mapped(page)) {
5621 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5622 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5625 if (PageDirty(page)) {
5626 struct address_space *mapping = page_mapping(page);
5628 if (mapping_can_writeback(mapping)) {
5629 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5631 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5637 if (PageWriteback(page)) {
5638 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5639 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5643 * All state has been migrated, let's switch to the new memcg.
5645 * It is safe to change page's memcg here because the page
5646 * is referenced, charged, isolated, and locked: we can't race
5647 * with (un)charging, migration, LRU putback, or anything else
5648 * that would rely on a stable page's memory cgroup.
5650 * Note that lock_page_memcg is a memcg lock, not a page lock,
5651 * to save space. As soon as we switch page's memory cgroup to a
5652 * new memcg that isn't locked, the above state can change
5653 * concurrently again. Make sure we're truly done with it.
5658 css_put(&from->css);
5660 page->memcg_data = (unsigned long)to;
5662 __unlock_page_memcg(from);
5666 local_irq_disable();
5667 mem_cgroup_charge_statistics(to, page, nr_pages);
5668 memcg_check_events(to, page);
5669 mem_cgroup_charge_statistics(from, page, -nr_pages);
5670 memcg_check_events(from, page);
5679 * get_mctgt_type - get target type of moving charge
5680 * @vma: the vma the pte to be checked belongs
5681 * @addr: the address corresponding to the pte to be checked
5682 * @ptent: the pte to be checked
5683 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5686 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5687 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5688 * move charge. if @target is not NULL, the page is stored in target->page
5689 * with extra refcnt got(Callers should handle it).
5690 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5691 * target for charge migration. if @target is not NULL, the entry is stored
5693 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5694 * (so ZONE_DEVICE page and thus not on the lru).
5695 * For now we such page is charge like a regular page would be as for all
5696 * intent and purposes it is just special memory taking the place of a
5699 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5701 * Called with pte lock held.
5704 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5705 unsigned long addr, pte_t ptent, union mc_target *target)
5707 struct page *page = NULL;
5708 enum mc_target_type ret = MC_TARGET_NONE;
5709 swp_entry_t ent = { .val = 0 };
5711 if (pte_present(ptent))
5712 page = mc_handle_present_pte(vma, addr, ptent);
5713 else if (is_swap_pte(ptent))
5714 page = mc_handle_swap_pte(vma, ptent, &ent);
5715 else if (pte_none(ptent))
5716 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5718 if (!page && !ent.val)
5722 * Do only loose check w/o serialization.
5723 * mem_cgroup_move_account() checks the page is valid or
5724 * not under LRU exclusion.
5726 if (page_memcg(page) == mc.from) {
5727 ret = MC_TARGET_PAGE;
5728 if (is_device_private_page(page))
5729 ret = MC_TARGET_DEVICE;
5731 target->page = page;
5733 if (!ret || !target)
5737 * There is a swap entry and a page doesn't exist or isn't charged.
5738 * But we cannot move a tail-page in a THP.
5740 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5741 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5742 ret = MC_TARGET_SWAP;
5749 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5751 * We don't consider PMD mapped swapping or file mapped pages because THP does
5752 * not support them for now.
5753 * Caller should make sure that pmd_trans_huge(pmd) is true.
5755 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5756 unsigned long addr, pmd_t pmd, union mc_target *target)
5758 struct page *page = NULL;
5759 enum mc_target_type ret = MC_TARGET_NONE;
5761 if (unlikely(is_swap_pmd(pmd))) {
5762 VM_BUG_ON(thp_migration_supported() &&
5763 !is_pmd_migration_entry(pmd));
5766 page = pmd_page(pmd);
5767 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5768 if (!(mc.flags & MOVE_ANON))
5770 if (page_memcg(page) == mc.from) {
5771 ret = MC_TARGET_PAGE;
5774 target->page = page;
5780 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5781 unsigned long addr, pmd_t pmd, union mc_target *target)
5783 return MC_TARGET_NONE;
5787 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5788 unsigned long addr, unsigned long end,
5789 struct mm_walk *walk)
5791 struct vm_area_struct *vma = walk->vma;
5795 ptl = pmd_trans_huge_lock(pmd, vma);
5798 * Note their can not be MC_TARGET_DEVICE for now as we do not
5799 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5800 * this might change.
5802 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5803 mc.precharge += HPAGE_PMD_NR;
5808 if (pmd_trans_unstable(pmd))
5810 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5811 for (; addr != end; pte++, addr += PAGE_SIZE)
5812 if (get_mctgt_type(vma, addr, *pte, NULL))
5813 mc.precharge++; /* increment precharge temporarily */
5814 pte_unmap_unlock(pte - 1, ptl);
5820 static const struct mm_walk_ops precharge_walk_ops = {
5821 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5824 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5826 unsigned long precharge;
5829 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5830 mmap_read_unlock(mm);
5832 precharge = mc.precharge;
5838 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5840 unsigned long precharge = mem_cgroup_count_precharge(mm);
5842 VM_BUG_ON(mc.moving_task);
5843 mc.moving_task = current;
5844 return mem_cgroup_do_precharge(precharge);
5847 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5848 static void __mem_cgroup_clear_mc(void)
5850 struct mem_cgroup *from = mc.from;
5851 struct mem_cgroup *to = mc.to;
5853 /* we must uncharge all the leftover precharges from mc.to */
5855 cancel_charge(mc.to, mc.precharge);
5859 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5860 * we must uncharge here.
5862 if (mc.moved_charge) {
5863 cancel_charge(mc.from, mc.moved_charge);
5864 mc.moved_charge = 0;
5866 /* we must fixup refcnts and charges */
5867 if (mc.moved_swap) {
5868 /* uncharge swap account from the old cgroup */
5869 if (!mem_cgroup_is_root(mc.from))
5870 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5872 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5875 * we charged both to->memory and to->memsw, so we
5876 * should uncharge to->memory.
5878 if (!mem_cgroup_is_root(mc.to))
5879 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5883 memcg_oom_recover(from);
5884 memcg_oom_recover(to);
5885 wake_up_all(&mc.waitq);
5888 static void mem_cgroup_clear_mc(void)
5890 struct mm_struct *mm = mc.mm;
5893 * we must clear moving_task before waking up waiters at the end of
5896 mc.moving_task = NULL;
5897 __mem_cgroup_clear_mc();
5898 spin_lock(&mc.lock);
5902 spin_unlock(&mc.lock);
5907 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5909 struct cgroup_subsys_state *css;
5910 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5911 struct mem_cgroup *from;
5912 struct task_struct *leader, *p;
5913 struct mm_struct *mm;
5914 unsigned long move_flags;
5917 /* charge immigration isn't supported on the default hierarchy */
5918 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5922 * Multi-process migrations only happen on the default hierarchy
5923 * where charge immigration is not used. Perform charge
5924 * immigration if @tset contains a leader and whine if there are
5928 cgroup_taskset_for_each_leader(leader, css, tset) {
5931 memcg = mem_cgroup_from_css(css);
5937 * We are now commited to this value whatever it is. Changes in this
5938 * tunable will only affect upcoming migrations, not the current one.
5939 * So we need to save it, and keep it going.
5941 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5945 from = mem_cgroup_from_task(p);
5947 VM_BUG_ON(from == memcg);
5949 mm = get_task_mm(p);
5952 /* We move charges only when we move a owner of the mm */
5953 if (mm->owner == p) {
5956 VM_BUG_ON(mc.precharge);
5957 VM_BUG_ON(mc.moved_charge);
5958 VM_BUG_ON(mc.moved_swap);
5960 spin_lock(&mc.lock);
5964 mc.flags = move_flags;
5965 spin_unlock(&mc.lock);
5966 /* We set mc.moving_task later */
5968 ret = mem_cgroup_precharge_mc(mm);
5970 mem_cgroup_clear_mc();
5977 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5980 mem_cgroup_clear_mc();
5983 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5984 unsigned long addr, unsigned long end,
5985 struct mm_walk *walk)
5988 struct vm_area_struct *vma = walk->vma;
5991 enum mc_target_type target_type;
5992 union mc_target target;
5995 ptl = pmd_trans_huge_lock(pmd, vma);
5997 if (mc.precharge < HPAGE_PMD_NR) {
6001 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6002 if (target_type == MC_TARGET_PAGE) {
6004 if (!isolate_lru_page(page)) {
6005 if (!mem_cgroup_move_account(page, true,
6007 mc.precharge -= HPAGE_PMD_NR;
6008 mc.moved_charge += HPAGE_PMD_NR;
6010 putback_lru_page(page);
6013 } else if (target_type == MC_TARGET_DEVICE) {
6015 if (!mem_cgroup_move_account(page, true,
6017 mc.precharge -= HPAGE_PMD_NR;
6018 mc.moved_charge += HPAGE_PMD_NR;
6026 if (pmd_trans_unstable(pmd))
6029 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6030 for (; addr != end; addr += PAGE_SIZE) {
6031 pte_t ptent = *(pte++);
6032 bool device = false;
6038 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6039 case MC_TARGET_DEVICE:
6042 case MC_TARGET_PAGE:
6045 * We can have a part of the split pmd here. Moving it
6046 * can be done but it would be too convoluted so simply
6047 * ignore such a partial THP and keep it in original
6048 * memcg. There should be somebody mapping the head.
6050 if (PageTransCompound(page))
6052 if (!device && isolate_lru_page(page))
6054 if (!mem_cgroup_move_account(page, false,
6057 /* we uncharge from mc.from later. */
6061 putback_lru_page(page);
6062 put: /* get_mctgt_type() gets the page */
6065 case MC_TARGET_SWAP:
6067 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6069 mem_cgroup_id_get_many(mc.to, 1);
6070 /* we fixup other refcnts and charges later. */
6078 pte_unmap_unlock(pte - 1, ptl);
6083 * We have consumed all precharges we got in can_attach().
6084 * We try charge one by one, but don't do any additional
6085 * charges to mc.to if we have failed in charge once in attach()
6088 ret = mem_cgroup_do_precharge(1);
6096 static const struct mm_walk_ops charge_walk_ops = {
6097 .pmd_entry = mem_cgroup_move_charge_pte_range,
6100 static void mem_cgroup_move_charge(void)
6102 lru_add_drain_all();
6104 * Signal lock_page_memcg() to take the memcg's move_lock
6105 * while we're moving its pages to another memcg. Then wait
6106 * for already started RCU-only updates to finish.
6108 atomic_inc(&mc.from->moving_account);
6111 if (unlikely(!mmap_read_trylock(mc.mm))) {
6113 * Someone who are holding the mmap_lock might be waiting in
6114 * waitq. So we cancel all extra charges, wake up all waiters,
6115 * and retry. Because we cancel precharges, we might not be able
6116 * to move enough charges, but moving charge is a best-effort
6117 * feature anyway, so it wouldn't be a big problem.
6119 __mem_cgroup_clear_mc();
6124 * When we have consumed all precharges and failed in doing
6125 * additional charge, the page walk just aborts.
6127 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6130 mmap_read_unlock(mc.mm);
6131 atomic_dec(&mc.from->moving_account);
6134 static void mem_cgroup_move_task(void)
6137 mem_cgroup_move_charge();
6138 mem_cgroup_clear_mc();
6141 #else /* !CONFIG_MMU */
6142 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6146 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6149 static void mem_cgroup_move_task(void)
6154 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6156 if (value == PAGE_COUNTER_MAX)
6157 seq_puts(m, "max\n");
6159 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6164 static u64 memory_current_read(struct cgroup_subsys_state *css,
6167 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6169 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6172 static int memory_min_show(struct seq_file *m, void *v)
6174 return seq_puts_memcg_tunable(m,
6175 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6178 static ssize_t memory_min_write(struct kernfs_open_file *of,
6179 char *buf, size_t nbytes, loff_t off)
6181 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6185 buf = strstrip(buf);
6186 err = page_counter_memparse(buf, "max", &min);
6190 page_counter_set_min(&memcg->memory, min);
6195 static int memory_low_show(struct seq_file *m, void *v)
6197 return seq_puts_memcg_tunable(m,
6198 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6201 static ssize_t memory_low_write(struct kernfs_open_file *of,
6202 char *buf, size_t nbytes, loff_t off)
6204 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6208 buf = strstrip(buf);
6209 err = page_counter_memparse(buf, "max", &low);
6213 page_counter_set_low(&memcg->memory, low);
6218 static int memory_high_show(struct seq_file *m, void *v)
6220 return seq_puts_memcg_tunable(m,
6221 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6224 static ssize_t memory_high_write(struct kernfs_open_file *of,
6225 char *buf, size_t nbytes, loff_t off)
6227 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6228 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6229 bool drained = false;
6233 buf = strstrip(buf);
6234 err = page_counter_memparse(buf, "max", &high);
6239 unsigned long nr_pages = page_counter_read(&memcg->memory);
6240 unsigned long reclaimed;
6242 if (nr_pages <= high)
6245 if (signal_pending(current))
6249 drain_all_stock(memcg);
6254 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6257 if (!reclaimed && !nr_retries--)
6261 page_counter_set_high(&memcg->memory, high);
6263 memcg_wb_domain_size_changed(memcg);
6268 static int memory_max_show(struct seq_file *m, void *v)
6270 return seq_puts_memcg_tunable(m,
6271 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6274 static ssize_t memory_max_write(struct kernfs_open_file *of,
6275 char *buf, size_t nbytes, loff_t off)
6277 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6278 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6279 bool drained = false;
6283 buf = strstrip(buf);
6284 err = page_counter_memparse(buf, "max", &max);
6288 xchg(&memcg->memory.max, max);
6291 unsigned long nr_pages = page_counter_read(&memcg->memory);
6293 if (nr_pages <= max)
6296 if (signal_pending(current))
6300 drain_all_stock(memcg);
6306 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6312 memcg_memory_event(memcg, MEMCG_OOM);
6313 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6317 memcg_wb_domain_size_changed(memcg);
6321 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6323 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6324 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6325 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6326 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6327 seq_printf(m, "oom_kill %lu\n",
6328 atomic_long_read(&events[MEMCG_OOM_KILL]));
6331 static int memory_events_show(struct seq_file *m, void *v)
6333 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6335 __memory_events_show(m, memcg->memory_events);
6339 static int memory_events_local_show(struct seq_file *m, void *v)
6341 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6343 __memory_events_show(m, memcg->memory_events_local);
6347 static int memory_stat_show(struct seq_file *m, void *v)
6349 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6352 buf = memory_stat_format(memcg);
6361 static int memory_numa_stat_show(struct seq_file *m, void *v)
6364 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6366 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6369 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6372 seq_printf(m, "%s", memory_stats[i].name);
6373 for_each_node_state(nid, N_MEMORY) {
6375 struct lruvec *lruvec;
6377 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6378 size = lruvec_page_state(lruvec, memory_stats[i].idx);
6379 size *= memory_stats[i].ratio;
6380 seq_printf(m, " N%d=%llu", nid, size);
6389 static int memory_oom_group_show(struct seq_file *m, void *v)
6391 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6393 seq_printf(m, "%d\n", memcg->oom_group);
6398 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6399 char *buf, size_t nbytes, loff_t off)
6401 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6404 buf = strstrip(buf);
6408 ret = kstrtoint(buf, 0, &oom_group);
6412 if (oom_group != 0 && oom_group != 1)
6415 memcg->oom_group = oom_group;
6420 static struct cftype memory_files[] = {
6423 .flags = CFTYPE_NOT_ON_ROOT,
6424 .read_u64 = memory_current_read,
6428 .flags = CFTYPE_NOT_ON_ROOT,
6429 .seq_show = memory_min_show,
6430 .write = memory_min_write,
6434 .flags = CFTYPE_NOT_ON_ROOT,
6435 .seq_show = memory_low_show,
6436 .write = memory_low_write,
6440 .flags = CFTYPE_NOT_ON_ROOT,
6441 .seq_show = memory_high_show,
6442 .write = memory_high_write,
6446 .flags = CFTYPE_NOT_ON_ROOT,
6447 .seq_show = memory_max_show,
6448 .write = memory_max_write,
6452 .flags = CFTYPE_NOT_ON_ROOT,
6453 .file_offset = offsetof(struct mem_cgroup, events_file),
6454 .seq_show = memory_events_show,
6457 .name = "events.local",
6458 .flags = CFTYPE_NOT_ON_ROOT,
6459 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6460 .seq_show = memory_events_local_show,
6464 .seq_show = memory_stat_show,
6468 .name = "numa_stat",
6469 .seq_show = memory_numa_stat_show,
6473 .name = "oom.group",
6474 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6475 .seq_show = memory_oom_group_show,
6476 .write = memory_oom_group_write,
6481 struct cgroup_subsys memory_cgrp_subsys = {
6482 .css_alloc = mem_cgroup_css_alloc,
6483 .css_online = mem_cgroup_css_online,
6484 .css_offline = mem_cgroup_css_offline,
6485 .css_released = mem_cgroup_css_released,
6486 .css_free = mem_cgroup_css_free,
6487 .css_reset = mem_cgroup_css_reset,
6488 .can_attach = mem_cgroup_can_attach,
6489 .cancel_attach = mem_cgroup_cancel_attach,
6490 .post_attach = mem_cgroup_move_task,
6491 .dfl_cftypes = memory_files,
6492 .legacy_cftypes = mem_cgroup_legacy_files,
6497 * This function calculates an individual cgroup's effective
6498 * protection which is derived from its own memory.min/low, its
6499 * parent's and siblings' settings, as well as the actual memory
6500 * distribution in the tree.
6502 * The following rules apply to the effective protection values:
6504 * 1. At the first level of reclaim, effective protection is equal to
6505 * the declared protection in memory.min and memory.low.
6507 * 2. To enable safe delegation of the protection configuration, at
6508 * subsequent levels the effective protection is capped to the
6509 * parent's effective protection.
6511 * 3. To make complex and dynamic subtrees easier to configure, the
6512 * user is allowed to overcommit the declared protection at a given
6513 * level. If that is the case, the parent's effective protection is
6514 * distributed to the children in proportion to how much protection
6515 * they have declared and how much of it they are utilizing.
6517 * This makes distribution proportional, but also work-conserving:
6518 * if one cgroup claims much more protection than it uses memory,
6519 * the unused remainder is available to its siblings.
6521 * 4. Conversely, when the declared protection is undercommitted at a
6522 * given level, the distribution of the larger parental protection
6523 * budget is NOT proportional. A cgroup's protection from a sibling
6524 * is capped to its own memory.min/low setting.
6526 * 5. However, to allow protecting recursive subtrees from each other
6527 * without having to declare each individual cgroup's fixed share
6528 * of the ancestor's claim to protection, any unutilized -
6529 * "floating" - protection from up the tree is distributed in
6530 * proportion to each cgroup's *usage*. This makes the protection
6531 * neutral wrt sibling cgroups and lets them compete freely over
6532 * the shared parental protection budget, but it protects the
6533 * subtree as a whole from neighboring subtrees.
6535 * Note that 4. and 5. are not in conflict: 4. is about protecting
6536 * against immediate siblings whereas 5. is about protecting against
6537 * neighboring subtrees.
6539 static unsigned long effective_protection(unsigned long usage,
6540 unsigned long parent_usage,
6541 unsigned long setting,
6542 unsigned long parent_effective,
6543 unsigned long siblings_protected)
6545 unsigned long protected;
6548 protected = min(usage, setting);
6550 * If all cgroups at this level combined claim and use more
6551 * protection then what the parent affords them, distribute
6552 * shares in proportion to utilization.
6554 * We are using actual utilization rather than the statically
6555 * claimed protection in order to be work-conserving: claimed
6556 * but unused protection is available to siblings that would
6557 * otherwise get a smaller chunk than what they claimed.
6559 if (siblings_protected > parent_effective)
6560 return protected * parent_effective / siblings_protected;
6563 * Ok, utilized protection of all children is within what the
6564 * parent affords them, so we know whatever this child claims
6565 * and utilizes is effectively protected.
6567 * If there is unprotected usage beyond this value, reclaim
6568 * will apply pressure in proportion to that amount.
6570 * If there is unutilized protection, the cgroup will be fully
6571 * shielded from reclaim, but we do return a smaller value for
6572 * protection than what the group could enjoy in theory. This
6573 * is okay. With the overcommit distribution above, effective
6574 * protection is always dependent on how memory is actually
6575 * consumed among the siblings anyway.
6580 * If the children aren't claiming (all of) the protection
6581 * afforded to them by the parent, distribute the remainder in
6582 * proportion to the (unprotected) memory of each cgroup. That
6583 * way, cgroups that aren't explicitly prioritized wrt each
6584 * other compete freely over the allowance, but they are
6585 * collectively protected from neighboring trees.
6587 * We're using unprotected memory for the weight so that if
6588 * some cgroups DO claim explicit protection, we don't protect
6589 * the same bytes twice.
6591 * Check both usage and parent_usage against the respective
6592 * protected values. One should imply the other, but they
6593 * aren't read atomically - make sure the division is sane.
6595 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6597 if (parent_effective > siblings_protected &&
6598 parent_usage > siblings_protected &&
6599 usage > protected) {
6600 unsigned long unclaimed;
6602 unclaimed = parent_effective - siblings_protected;
6603 unclaimed *= usage - protected;
6604 unclaimed /= parent_usage - siblings_protected;
6613 * mem_cgroup_protected - check if memory consumption is in the normal range
6614 * @root: the top ancestor of the sub-tree being checked
6615 * @memcg: the memory cgroup to check
6617 * WARNING: This function is not stateless! It can only be used as part
6618 * of a top-down tree iteration, not for isolated queries.
6620 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6621 struct mem_cgroup *memcg)
6623 unsigned long usage, parent_usage;
6624 struct mem_cgroup *parent;
6626 if (mem_cgroup_disabled())
6630 root = root_mem_cgroup;
6633 * Effective values of the reclaim targets are ignored so they
6634 * can be stale. Have a look at mem_cgroup_protection for more
6636 * TODO: calculation should be more robust so that we do not need
6637 * that special casing.
6642 usage = page_counter_read(&memcg->memory);
6646 parent = parent_mem_cgroup(memcg);
6647 /* No parent means a non-hierarchical mode on v1 memcg */
6651 if (parent == root) {
6652 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6653 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6657 parent_usage = page_counter_read(&parent->memory);
6659 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6660 READ_ONCE(memcg->memory.min),
6661 READ_ONCE(parent->memory.emin),
6662 atomic_long_read(&parent->memory.children_min_usage)));
6664 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6665 READ_ONCE(memcg->memory.low),
6666 READ_ONCE(parent->memory.elow),
6667 atomic_long_read(&parent->memory.children_low_usage)));
6671 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6672 * @page: page to charge
6673 * @mm: mm context of the victim
6674 * @gfp_mask: reclaim mode
6676 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6677 * pages according to @gfp_mask if necessary.
6679 * Returns 0 on success. Otherwise, an error code is returned.
6681 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6683 unsigned int nr_pages = thp_nr_pages(page);
6684 struct mem_cgroup *memcg = NULL;
6687 if (mem_cgroup_disabled())
6690 if (PageSwapCache(page)) {
6691 swp_entry_t ent = { .val = page_private(page), };
6695 * Every swap fault against a single page tries to charge the
6696 * page, bail as early as possible. shmem_unuse() encounters
6697 * already charged pages, too. page and memcg binding is
6698 * protected by the page lock, which serializes swap cache
6699 * removal, which in turn serializes uncharging.
6701 VM_BUG_ON_PAGE(!PageLocked(page), page);
6702 if (page_memcg(compound_head(page)))
6705 id = lookup_swap_cgroup_id(ent);
6707 memcg = mem_cgroup_from_id(id);
6708 if (memcg && !css_tryget_online(&memcg->css))
6714 memcg = get_mem_cgroup_from_mm(mm);
6716 ret = try_charge(memcg, gfp_mask, nr_pages);
6720 css_get(&memcg->css);
6721 commit_charge(page, memcg);
6723 local_irq_disable();
6724 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6725 memcg_check_events(memcg, page);
6728 if (PageSwapCache(page)) {
6729 swp_entry_t entry = { .val = page_private(page) };
6731 * The swap entry might not get freed for a long time,
6732 * let's not wait for it. The page already received a
6733 * memory+swap charge, drop the swap entry duplicate.
6735 mem_cgroup_uncharge_swap(entry, nr_pages);
6739 css_put(&memcg->css);
6744 struct uncharge_gather {
6745 struct mem_cgroup *memcg;
6746 unsigned long nr_pages;
6747 unsigned long pgpgout;
6748 unsigned long nr_kmem;
6749 struct page *dummy_page;
6752 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6754 memset(ug, 0, sizeof(*ug));
6757 static void uncharge_batch(const struct uncharge_gather *ug)
6759 unsigned long flags;
6761 if (!mem_cgroup_is_root(ug->memcg)) {
6762 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6763 if (do_memsw_account())
6764 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6765 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6766 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6767 memcg_oom_recover(ug->memcg);
6770 local_irq_save(flags);
6771 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6772 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6773 memcg_check_events(ug->memcg, ug->dummy_page);
6774 local_irq_restore(flags);
6776 /* drop reference from uncharge_page */
6777 css_put(&ug->memcg->css);
6780 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6782 unsigned long nr_pages;
6784 VM_BUG_ON_PAGE(PageLRU(page), page);
6786 if (!page_memcg(page))
6790 * Nobody should be changing or seriously looking at
6791 * page_memcg(page) at this point, we have fully
6792 * exclusive access to the page.
6795 if (ug->memcg != page_memcg(page)) {
6798 uncharge_gather_clear(ug);
6800 ug->memcg = page_memcg(page);
6802 /* pairs with css_put in uncharge_batch */
6803 css_get(&ug->memcg->css);
6806 nr_pages = compound_nr(page);
6807 ug->nr_pages += nr_pages;
6809 if (PageMemcgKmem(page))
6810 ug->nr_kmem += nr_pages;
6814 ug->dummy_page = page;
6815 page->memcg_data = 0;
6816 css_put(&ug->memcg->css);
6819 static void uncharge_list(struct list_head *page_list)
6821 struct uncharge_gather ug;
6822 struct list_head *next;
6824 uncharge_gather_clear(&ug);
6827 * Note that the list can be a single page->lru; hence the
6828 * do-while loop instead of a simple list_for_each_entry().
6830 next = page_list->next;
6834 page = list_entry(next, struct page, lru);
6835 next = page->lru.next;
6837 uncharge_page(page, &ug);
6838 } while (next != page_list);
6841 uncharge_batch(&ug);
6845 * mem_cgroup_uncharge - uncharge a page
6846 * @page: page to uncharge
6848 * Uncharge a page previously charged with mem_cgroup_charge().
6850 void mem_cgroup_uncharge(struct page *page)
6852 struct uncharge_gather ug;
6854 if (mem_cgroup_disabled())
6857 /* Don't touch page->lru of any random page, pre-check: */
6858 if (!page_memcg(page))
6861 uncharge_gather_clear(&ug);
6862 uncharge_page(page, &ug);
6863 uncharge_batch(&ug);
6867 * mem_cgroup_uncharge_list - uncharge a list of page
6868 * @page_list: list of pages to uncharge
6870 * Uncharge a list of pages previously charged with
6871 * mem_cgroup_charge().
6873 void mem_cgroup_uncharge_list(struct list_head *page_list)
6875 if (mem_cgroup_disabled())
6878 if (!list_empty(page_list))
6879 uncharge_list(page_list);
6883 * mem_cgroup_migrate - charge a page's replacement
6884 * @oldpage: currently circulating page
6885 * @newpage: replacement page
6887 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6888 * be uncharged upon free.
6890 * Both pages must be locked, @newpage->mapping must be set up.
6892 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6894 struct mem_cgroup *memcg;
6895 unsigned int nr_pages;
6896 unsigned long flags;
6898 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6899 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6900 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6901 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6904 if (mem_cgroup_disabled())
6907 /* Page cache replacement: new page already charged? */
6908 if (page_memcg(newpage))
6911 memcg = page_memcg(oldpage);
6915 /* Force-charge the new page. The old one will be freed soon */
6916 nr_pages = thp_nr_pages(newpage);
6918 page_counter_charge(&memcg->memory, nr_pages);
6919 if (do_memsw_account())
6920 page_counter_charge(&memcg->memsw, nr_pages);
6922 css_get(&memcg->css);
6923 commit_charge(newpage, memcg);
6925 local_irq_save(flags);
6926 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6927 memcg_check_events(memcg, newpage);
6928 local_irq_restore(flags);
6931 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6932 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6934 void mem_cgroup_sk_alloc(struct sock *sk)
6936 struct mem_cgroup *memcg;
6938 if (!mem_cgroup_sockets_enabled)
6941 /* Do not associate the sock with unrelated interrupted task's memcg. */
6946 memcg = mem_cgroup_from_task(current);
6947 if (memcg == root_mem_cgroup)
6949 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6951 if (css_tryget(&memcg->css))
6952 sk->sk_memcg = memcg;
6957 void mem_cgroup_sk_free(struct sock *sk)
6960 css_put(&sk->sk_memcg->css);
6964 * mem_cgroup_charge_skmem - charge socket memory
6965 * @memcg: memcg to charge
6966 * @nr_pages: number of pages to charge
6968 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6969 * @memcg's configured limit, %false if the charge had to be forced.
6971 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6973 gfp_t gfp_mask = GFP_KERNEL;
6975 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6976 struct page_counter *fail;
6978 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6979 memcg->tcpmem_pressure = 0;
6982 page_counter_charge(&memcg->tcpmem, nr_pages);
6983 memcg->tcpmem_pressure = 1;
6987 /* Don't block in the packet receive path */
6989 gfp_mask = GFP_NOWAIT;
6991 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6993 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6996 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7001 * mem_cgroup_uncharge_skmem - uncharge socket memory
7002 * @memcg: memcg to uncharge
7003 * @nr_pages: number of pages to uncharge
7005 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7007 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7008 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7012 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7014 refill_stock(memcg, nr_pages);
7017 static int __init cgroup_memory(char *s)
7021 while ((token = strsep(&s, ",")) != NULL) {
7024 if (!strcmp(token, "nosocket"))
7025 cgroup_memory_nosocket = true;
7026 if (!strcmp(token, "nokmem"))
7027 cgroup_memory_nokmem = true;
7031 __setup("cgroup.memory=", cgroup_memory);
7034 * subsys_initcall() for memory controller.
7036 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7037 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7038 * basically everything that doesn't depend on a specific mem_cgroup structure
7039 * should be initialized from here.
7041 static int __init mem_cgroup_init(void)
7045 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7046 memcg_hotplug_cpu_dead);
7048 for_each_possible_cpu(cpu)
7049 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7052 for_each_node(node) {
7053 struct mem_cgroup_tree_per_node *rtpn;
7055 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7056 node_online(node) ? node : NUMA_NO_NODE);
7058 rtpn->rb_root = RB_ROOT;
7059 rtpn->rb_rightmost = NULL;
7060 spin_lock_init(&rtpn->lock);
7061 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7066 subsys_initcall(mem_cgroup_init);
7068 #ifdef CONFIG_MEMCG_SWAP
7069 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7071 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7073 * The root cgroup cannot be destroyed, so it's refcount must
7076 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7080 memcg = parent_mem_cgroup(memcg);
7082 memcg = root_mem_cgroup;
7088 * mem_cgroup_swapout - transfer a memsw charge to swap
7089 * @page: page whose memsw charge to transfer
7090 * @entry: swap entry to move the charge to
7092 * Transfer the memsw charge of @page to @entry.
7094 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7096 struct mem_cgroup *memcg, *swap_memcg;
7097 unsigned int nr_entries;
7098 unsigned short oldid;
7100 VM_BUG_ON_PAGE(PageLRU(page), page);
7101 VM_BUG_ON_PAGE(page_count(page), page);
7103 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7106 memcg = page_memcg(page);
7108 /* Readahead page, never charged */
7113 * In case the memcg owning these pages has been offlined and doesn't
7114 * have an ID allocated to it anymore, charge the closest online
7115 * ancestor for the swap instead and transfer the memory+swap charge.
7117 swap_memcg = mem_cgroup_id_get_online(memcg);
7118 nr_entries = thp_nr_pages(page);
7119 /* Get references for the tail pages, too */
7121 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7122 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7124 VM_BUG_ON_PAGE(oldid, page);
7125 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7127 page->memcg_data = 0;
7129 if (!mem_cgroup_is_root(memcg))
7130 page_counter_uncharge(&memcg->memory, nr_entries);
7132 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7133 if (!mem_cgroup_is_root(swap_memcg))
7134 page_counter_charge(&swap_memcg->memsw, nr_entries);
7135 page_counter_uncharge(&memcg->memsw, nr_entries);
7139 * Interrupts should be disabled here because the caller holds the
7140 * i_pages lock which is taken with interrupts-off. It is
7141 * important here to have the interrupts disabled because it is the
7142 * only synchronisation we have for updating the per-CPU variables.
7144 VM_BUG_ON(!irqs_disabled());
7145 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7146 memcg_check_events(memcg, page);
7148 css_put(&memcg->css);
7152 * mem_cgroup_try_charge_swap - try charging swap space for a page
7153 * @page: page being added to swap
7154 * @entry: swap entry to charge
7156 * Try to charge @page's memcg for the swap space at @entry.
7158 * Returns 0 on success, -ENOMEM on failure.
7160 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7162 unsigned int nr_pages = thp_nr_pages(page);
7163 struct page_counter *counter;
7164 struct mem_cgroup *memcg;
7165 unsigned short oldid;
7167 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7170 memcg = page_memcg(page);
7172 /* Readahead page, never charged */
7177 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7181 memcg = mem_cgroup_id_get_online(memcg);
7183 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7184 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7185 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7186 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7187 mem_cgroup_id_put(memcg);
7191 /* Get references for the tail pages, too */
7193 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7194 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7195 VM_BUG_ON_PAGE(oldid, page);
7196 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7202 * mem_cgroup_uncharge_swap - uncharge swap space
7203 * @entry: swap entry to uncharge
7204 * @nr_pages: the amount of swap space to uncharge
7206 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7208 struct mem_cgroup *memcg;
7211 id = swap_cgroup_record(entry, 0, nr_pages);
7213 memcg = mem_cgroup_from_id(id);
7215 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7216 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7217 page_counter_uncharge(&memcg->swap, nr_pages);
7219 page_counter_uncharge(&memcg->memsw, nr_pages);
7221 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7222 mem_cgroup_id_put_many(memcg, nr_pages);
7227 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7229 long nr_swap_pages = get_nr_swap_pages();
7231 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7232 return nr_swap_pages;
7233 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7234 nr_swap_pages = min_t(long, nr_swap_pages,
7235 READ_ONCE(memcg->swap.max) -
7236 page_counter_read(&memcg->swap));
7237 return nr_swap_pages;
7240 bool mem_cgroup_swap_full(struct page *page)
7242 struct mem_cgroup *memcg;
7244 VM_BUG_ON_PAGE(!PageLocked(page), page);
7248 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7251 memcg = page_memcg(page);
7255 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7256 unsigned long usage = page_counter_read(&memcg->swap);
7258 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7259 usage * 2 >= READ_ONCE(memcg->swap.max))
7266 static int __init setup_swap_account(char *s)
7268 if (!strcmp(s, "1"))
7269 cgroup_memory_noswap = false;
7270 else if (!strcmp(s, "0"))
7271 cgroup_memory_noswap = true;
7274 __setup("swapaccount=", setup_swap_account);
7276 static u64 swap_current_read(struct cgroup_subsys_state *css,
7279 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7281 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7284 static int swap_high_show(struct seq_file *m, void *v)
7286 return seq_puts_memcg_tunable(m,
7287 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7290 static ssize_t swap_high_write(struct kernfs_open_file *of,
7291 char *buf, size_t nbytes, loff_t off)
7293 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7297 buf = strstrip(buf);
7298 err = page_counter_memparse(buf, "max", &high);
7302 page_counter_set_high(&memcg->swap, high);
7307 static int swap_max_show(struct seq_file *m, void *v)
7309 return seq_puts_memcg_tunable(m,
7310 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7313 static ssize_t swap_max_write(struct kernfs_open_file *of,
7314 char *buf, size_t nbytes, loff_t off)
7316 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7320 buf = strstrip(buf);
7321 err = page_counter_memparse(buf, "max", &max);
7325 xchg(&memcg->swap.max, max);
7330 static int swap_events_show(struct seq_file *m, void *v)
7332 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7334 seq_printf(m, "high %lu\n",
7335 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7336 seq_printf(m, "max %lu\n",
7337 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7338 seq_printf(m, "fail %lu\n",
7339 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7344 static struct cftype swap_files[] = {
7346 .name = "swap.current",
7347 .flags = CFTYPE_NOT_ON_ROOT,
7348 .read_u64 = swap_current_read,
7351 .name = "swap.high",
7352 .flags = CFTYPE_NOT_ON_ROOT,
7353 .seq_show = swap_high_show,
7354 .write = swap_high_write,
7358 .flags = CFTYPE_NOT_ON_ROOT,
7359 .seq_show = swap_max_show,
7360 .write = swap_max_write,
7363 .name = "swap.events",
7364 .flags = CFTYPE_NOT_ON_ROOT,
7365 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7366 .seq_show = swap_events_show,
7371 static struct cftype memsw_files[] = {
7373 .name = "memsw.usage_in_bytes",
7374 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7375 .read_u64 = mem_cgroup_read_u64,
7378 .name = "memsw.max_usage_in_bytes",
7379 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7380 .write = mem_cgroup_reset,
7381 .read_u64 = mem_cgroup_read_u64,
7384 .name = "memsw.limit_in_bytes",
7385 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7386 .write = mem_cgroup_write,
7387 .read_u64 = mem_cgroup_read_u64,
7390 .name = "memsw.failcnt",
7391 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7392 .write = mem_cgroup_reset,
7393 .read_u64 = mem_cgroup_read_u64,
7395 { }, /* terminate */
7399 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7400 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7401 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7402 * boot parameter. This may result in premature OOPS inside
7403 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7405 static int __init mem_cgroup_swap_init(void)
7407 /* No memory control -> no swap control */
7408 if (mem_cgroup_disabled())
7409 cgroup_memory_noswap = true;
7411 if (cgroup_memory_noswap)
7414 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7415 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7419 core_initcall(mem_cgroup_swap_init);
7421 #endif /* CONFIG_MEMCG_SWAP */