1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
25 #include <linux/page_counter.h>
26 #include <linux/memcontrol.h>
27 #include <linux/cgroup.h>
28 #include <linux/pagewalk.h>
29 #include <linux/sched/mm.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/hugetlb.h>
32 #include <linux/pagemap.h>
33 #include <linux/vm_event_item.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
59 #include <linux/tracehook.h>
60 #include <linux/psi.h>
61 #include <linux/seq_buf.h>
67 #include <linux/uaccess.h>
69 #include <trace/events/vmscan.h>
71 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
72 EXPORT_SYMBOL(memory_cgrp_subsys);
74 struct mem_cgroup *root_mem_cgroup __read_mostly;
76 /* Socket memory accounting disabled? */
77 static bool cgroup_memory_nosocket;
79 /* Kernel memory accounting disabled? */
80 static bool cgroup_memory_nokmem;
82 /* Whether the swap controller is active */
83 #ifdef CONFIG_MEMCG_SWAP
84 bool cgroup_memory_noswap __read_mostly;
86 #define cgroup_memory_noswap 1
89 #ifdef CONFIG_CGROUP_WRITEBACK
90 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
93 /* Whether legacy memory+swap accounting is active */
94 static bool do_memsw_account(void)
96 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
99 #define THRESHOLDS_EVENTS_TARGET 128
100 #define SOFTLIMIT_EVENTS_TARGET 1024
103 * Cgroups above their limits are maintained in a RB-Tree, independent of
104 * their hierarchy representation
107 struct mem_cgroup_tree_per_node {
108 struct rb_root rb_root;
109 struct rb_node *rb_rightmost;
113 struct mem_cgroup_tree {
114 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
117 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
120 struct mem_cgroup_eventfd_list {
121 struct list_head list;
122 struct eventfd_ctx *eventfd;
126 * cgroup_event represents events which userspace want to receive.
128 struct mem_cgroup_event {
130 * memcg which the event belongs to.
132 struct mem_cgroup *memcg;
134 * eventfd to signal userspace about the event.
136 struct eventfd_ctx *eventfd;
138 * Each of these stored in a list by the cgroup.
140 struct list_head list;
142 * register_event() callback will be used to add new userspace
143 * waiter for changes related to this event. Use eventfd_signal()
144 * on eventfd to send notification to userspace.
146 int (*register_event)(struct mem_cgroup *memcg,
147 struct eventfd_ctx *eventfd, const char *args);
149 * unregister_event() callback will be called when userspace closes
150 * the eventfd or on cgroup removing. This callback must be set,
151 * if you want provide notification functionality.
153 void (*unregister_event)(struct mem_cgroup *memcg,
154 struct eventfd_ctx *eventfd);
156 * All fields below needed to unregister event when
157 * userspace closes eventfd.
160 wait_queue_head_t *wqh;
161 wait_queue_entry_t wait;
162 struct work_struct remove;
165 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
166 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
168 /* Stuffs for move charges at task migration. */
170 * Types of charges to be moved.
172 #define MOVE_ANON 0x1U
173 #define MOVE_FILE 0x2U
174 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
176 /* "mc" and its members are protected by cgroup_mutex */
177 static struct move_charge_struct {
178 spinlock_t lock; /* for from, to */
179 struct mm_struct *mm;
180 struct mem_cgroup *from;
181 struct mem_cgroup *to;
183 unsigned long precharge;
184 unsigned long moved_charge;
185 unsigned long moved_swap;
186 struct task_struct *moving_task; /* a task moving charges */
187 wait_queue_head_t waitq; /* a waitq for other context */
189 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
190 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
194 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
195 * limit reclaim to prevent infinite loops, if they ever occur.
197 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
198 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
201 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
202 MEM_CGROUP_CHARGE_TYPE_ANON,
203 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
204 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
208 /* for encoding cft->private value on file */
217 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
218 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
219 #define MEMFILE_ATTR(val) ((val) & 0xffff)
220 /* Used for OOM nofiier */
221 #define OOM_CONTROL (0)
224 * Iteration constructs for visiting all cgroups (under a tree). If
225 * loops are exited prematurely (break), mem_cgroup_iter_break() must
226 * be used for reference counting.
228 #define for_each_mem_cgroup_tree(iter, root) \
229 for (iter = mem_cgroup_iter(root, NULL, NULL); \
231 iter = mem_cgroup_iter(root, iter, NULL))
233 #define for_each_mem_cgroup(iter) \
234 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
236 iter = mem_cgroup_iter(NULL, iter, NULL))
238 static inline bool should_force_charge(void)
240 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
241 (current->flags & PF_EXITING);
244 /* Some nice accessors for the vmpressure. */
245 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
248 memcg = root_mem_cgroup;
249 return &memcg->vmpressure;
252 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
254 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
257 #ifdef CONFIG_MEMCG_KMEM
258 extern spinlock_t css_set_lock;
260 static void obj_cgroup_release(struct percpu_ref *ref)
262 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
263 struct mem_cgroup *memcg;
264 unsigned int nr_bytes;
265 unsigned int nr_pages;
269 * At this point all allocated objects are freed, and
270 * objcg->nr_charged_bytes can't have an arbitrary byte value.
271 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
273 * The following sequence can lead to it:
274 * 1) CPU0: objcg == stock->cached_objcg
275 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
276 * PAGE_SIZE bytes are charged
277 * 3) CPU1: a process from another memcg is allocating something,
278 * the stock if flushed,
279 * objcg->nr_charged_bytes = PAGE_SIZE - 92
280 * 5) CPU0: we do release this object,
281 * 92 bytes are added to stock->nr_bytes
282 * 6) CPU0: stock is flushed,
283 * 92 bytes are added to objcg->nr_charged_bytes
285 * In the result, nr_charged_bytes == PAGE_SIZE.
286 * This page will be uncharged in obj_cgroup_release().
288 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
289 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
290 nr_pages = nr_bytes >> PAGE_SHIFT;
292 spin_lock_irqsave(&css_set_lock, flags);
293 memcg = obj_cgroup_memcg(objcg);
295 __memcg_kmem_uncharge(memcg, nr_pages);
296 list_del(&objcg->list);
297 mem_cgroup_put(memcg);
298 spin_unlock_irqrestore(&css_set_lock, flags);
300 percpu_ref_exit(ref);
301 kfree_rcu(objcg, rcu);
304 static struct obj_cgroup *obj_cgroup_alloc(void)
306 struct obj_cgroup *objcg;
309 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
313 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
319 INIT_LIST_HEAD(&objcg->list);
323 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
324 struct mem_cgroup *parent)
326 struct obj_cgroup *objcg, *iter;
328 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
330 spin_lock_irq(&css_set_lock);
332 /* Move active objcg to the parent's list */
333 xchg(&objcg->memcg, parent);
334 css_get(&parent->css);
335 list_add(&objcg->list, &parent->objcg_list);
337 /* Move already reparented objcgs to the parent's list */
338 list_for_each_entry(iter, &memcg->objcg_list, list) {
339 css_get(&parent->css);
340 xchg(&iter->memcg, parent);
341 css_put(&memcg->css);
343 list_splice(&memcg->objcg_list, &parent->objcg_list);
345 spin_unlock_irq(&css_set_lock);
347 percpu_ref_kill(&objcg->refcnt);
351 * This will be used as a shrinker list's index.
352 * The main reason for not using cgroup id for this:
353 * this works better in sparse environments, where we have a lot of memcgs,
354 * but only a few kmem-limited. Or also, if we have, for instance, 200
355 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
356 * 200 entry array for that.
358 * The current size of the caches array is stored in memcg_nr_cache_ids. It
359 * will double each time we have to increase it.
361 static DEFINE_IDA(memcg_cache_ida);
362 int memcg_nr_cache_ids;
364 /* Protects memcg_nr_cache_ids */
365 static DECLARE_RWSEM(memcg_cache_ids_sem);
367 void memcg_get_cache_ids(void)
369 down_read(&memcg_cache_ids_sem);
372 void memcg_put_cache_ids(void)
374 up_read(&memcg_cache_ids_sem);
378 * MIN_SIZE is different than 1, because we would like to avoid going through
379 * the alloc/free process all the time. In a small machine, 4 kmem-limited
380 * cgroups is a reasonable guess. In the future, it could be a parameter or
381 * tunable, but that is strictly not necessary.
383 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
384 * this constant directly from cgroup, but it is understandable that this is
385 * better kept as an internal representation in cgroup.c. In any case, the
386 * cgrp_id space is not getting any smaller, and we don't have to necessarily
387 * increase ours as well if it increases.
389 #define MEMCG_CACHES_MIN_SIZE 4
390 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
393 * A lot of the calls to the cache allocation functions are expected to be
394 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
395 * conditional to this static branch, we'll have to allow modules that does
396 * kmem_cache_alloc and the such to see this symbol as well
398 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
399 EXPORT_SYMBOL(memcg_kmem_enabled_key);
402 static int memcg_shrinker_map_size;
403 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
405 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
407 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
410 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
411 int size, int old_size)
413 struct memcg_shrinker_map *new, *old;
416 lockdep_assert_held(&memcg_shrinker_map_mutex);
419 old = rcu_dereference_protected(
420 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
421 /* Not yet online memcg */
425 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
429 /* Set all old bits, clear all new bits */
430 memset(new->map, (int)0xff, old_size);
431 memset((void *)new->map + old_size, 0, size - old_size);
433 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
434 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
440 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
442 struct mem_cgroup_per_node *pn;
443 struct memcg_shrinker_map *map;
446 if (mem_cgroup_is_root(memcg))
450 pn = mem_cgroup_nodeinfo(memcg, nid);
451 map = rcu_dereference_protected(pn->shrinker_map, true);
454 rcu_assign_pointer(pn->shrinker_map, NULL);
458 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
460 struct memcg_shrinker_map *map;
461 int nid, size, ret = 0;
463 if (mem_cgroup_is_root(memcg))
466 mutex_lock(&memcg_shrinker_map_mutex);
467 size = memcg_shrinker_map_size;
469 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
471 memcg_free_shrinker_maps(memcg);
475 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
477 mutex_unlock(&memcg_shrinker_map_mutex);
482 int memcg_expand_shrinker_maps(int new_id)
484 int size, old_size, ret = 0;
485 struct mem_cgroup *memcg;
487 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
488 old_size = memcg_shrinker_map_size;
489 if (size <= old_size)
492 mutex_lock(&memcg_shrinker_map_mutex);
493 if (!root_mem_cgroup)
496 for_each_mem_cgroup(memcg) {
497 if (mem_cgroup_is_root(memcg))
499 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
501 mem_cgroup_iter_break(NULL, memcg);
507 memcg_shrinker_map_size = size;
508 mutex_unlock(&memcg_shrinker_map_mutex);
512 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
514 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
515 struct memcg_shrinker_map *map;
518 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
519 /* Pairs with smp mb in shrink_slab() */
520 smp_mb__before_atomic();
521 set_bit(shrinker_id, map->map);
527 * mem_cgroup_css_from_page - css of the memcg associated with a page
528 * @page: page of interest
530 * If memcg is bound to the default hierarchy, css of the memcg associated
531 * with @page is returned. The returned css remains associated with @page
532 * until it is released.
534 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
537 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
539 struct mem_cgroup *memcg;
541 memcg = page->mem_cgroup;
543 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
544 memcg = root_mem_cgroup;
550 * page_cgroup_ino - return inode number of the memcg a page is charged to
553 * Look up the closest online ancestor of the memory cgroup @page is charged to
554 * and return its inode number or 0 if @page is not charged to any cgroup. It
555 * is safe to call this function without holding a reference to @page.
557 * Note, this function is inherently racy, because there is nothing to prevent
558 * the cgroup inode from getting torn down and potentially reallocated a moment
559 * after page_cgroup_ino() returns, so it only should be used by callers that
560 * do not care (such as procfs interfaces).
562 ino_t page_cgroup_ino(struct page *page)
564 struct mem_cgroup *memcg;
565 unsigned long ino = 0;
568 memcg = page->mem_cgroup;
571 * The lowest bit set means that memcg isn't a valid
572 * memcg pointer, but a obj_cgroups pointer.
573 * In this case the page is shared and doesn't belong
574 * to any specific memory cgroup.
576 if ((unsigned long) memcg & 0x1UL)
579 while (memcg && !(memcg->css.flags & CSS_ONLINE))
580 memcg = parent_mem_cgroup(memcg);
582 ino = cgroup_ino(memcg->css.cgroup);
587 static struct mem_cgroup_per_node *
588 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
590 int nid = page_to_nid(page);
592 return memcg->nodeinfo[nid];
595 static struct mem_cgroup_tree_per_node *
596 soft_limit_tree_node(int nid)
598 return soft_limit_tree.rb_tree_per_node[nid];
601 static struct mem_cgroup_tree_per_node *
602 soft_limit_tree_from_page(struct page *page)
604 int nid = page_to_nid(page);
606 return soft_limit_tree.rb_tree_per_node[nid];
609 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
610 struct mem_cgroup_tree_per_node *mctz,
611 unsigned long new_usage_in_excess)
613 struct rb_node **p = &mctz->rb_root.rb_node;
614 struct rb_node *parent = NULL;
615 struct mem_cgroup_per_node *mz_node;
616 bool rightmost = true;
621 mz->usage_in_excess = new_usage_in_excess;
622 if (!mz->usage_in_excess)
626 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
628 if (mz->usage_in_excess < mz_node->usage_in_excess) {
634 * We can't avoid mem cgroups that are over their soft
635 * limit by the same amount
637 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
642 mctz->rb_rightmost = &mz->tree_node;
644 rb_link_node(&mz->tree_node, parent, p);
645 rb_insert_color(&mz->tree_node, &mctz->rb_root);
649 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
650 struct mem_cgroup_tree_per_node *mctz)
655 if (&mz->tree_node == mctz->rb_rightmost)
656 mctz->rb_rightmost = rb_prev(&mz->tree_node);
658 rb_erase(&mz->tree_node, &mctz->rb_root);
662 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
663 struct mem_cgroup_tree_per_node *mctz)
667 spin_lock_irqsave(&mctz->lock, flags);
668 __mem_cgroup_remove_exceeded(mz, mctz);
669 spin_unlock_irqrestore(&mctz->lock, flags);
672 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
674 unsigned long nr_pages = page_counter_read(&memcg->memory);
675 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
676 unsigned long excess = 0;
678 if (nr_pages > soft_limit)
679 excess = nr_pages - soft_limit;
684 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
686 unsigned long excess;
687 struct mem_cgroup_per_node *mz;
688 struct mem_cgroup_tree_per_node *mctz;
690 mctz = soft_limit_tree_from_page(page);
694 * Necessary to update all ancestors when hierarchy is used.
695 * because their event counter is not touched.
697 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
698 mz = mem_cgroup_page_nodeinfo(memcg, page);
699 excess = soft_limit_excess(memcg);
701 * We have to update the tree if mz is on RB-tree or
702 * mem is over its softlimit.
704 if (excess || mz->on_tree) {
707 spin_lock_irqsave(&mctz->lock, flags);
708 /* if on-tree, remove it */
710 __mem_cgroup_remove_exceeded(mz, mctz);
712 * Insert again. mz->usage_in_excess will be updated.
713 * If excess is 0, no tree ops.
715 __mem_cgroup_insert_exceeded(mz, mctz, excess);
716 spin_unlock_irqrestore(&mctz->lock, flags);
721 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
723 struct mem_cgroup_tree_per_node *mctz;
724 struct mem_cgroup_per_node *mz;
728 mz = mem_cgroup_nodeinfo(memcg, nid);
729 mctz = soft_limit_tree_node(nid);
731 mem_cgroup_remove_exceeded(mz, mctz);
735 static struct mem_cgroup_per_node *
736 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
738 struct mem_cgroup_per_node *mz;
742 if (!mctz->rb_rightmost)
743 goto done; /* Nothing to reclaim from */
745 mz = rb_entry(mctz->rb_rightmost,
746 struct mem_cgroup_per_node, tree_node);
748 * Remove the node now but someone else can add it back,
749 * we will to add it back at the end of reclaim to its correct
750 * position in the tree.
752 __mem_cgroup_remove_exceeded(mz, mctz);
753 if (!soft_limit_excess(mz->memcg) ||
754 !css_tryget(&mz->memcg->css))
760 static struct mem_cgroup_per_node *
761 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
763 struct mem_cgroup_per_node *mz;
765 spin_lock_irq(&mctz->lock);
766 mz = __mem_cgroup_largest_soft_limit_node(mctz);
767 spin_unlock_irq(&mctz->lock);
772 * __mod_memcg_state - update cgroup memory statistics
773 * @memcg: the memory cgroup
774 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
775 * @val: delta to add to the counter, can be negative
777 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
779 long x, threshold = MEMCG_CHARGE_BATCH;
781 if (mem_cgroup_disabled())
784 if (memcg_stat_item_in_bytes(idx))
785 threshold <<= PAGE_SHIFT;
787 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
788 if (unlikely(abs(x) > threshold)) {
789 struct mem_cgroup *mi;
792 * Batch local counters to keep them in sync with
793 * the hierarchical ones.
795 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
796 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
797 atomic_long_add(x, &mi->vmstats[idx]);
800 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
803 static struct mem_cgroup_per_node *
804 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
806 struct mem_cgroup *parent;
808 parent = parent_mem_cgroup(pn->memcg);
811 return mem_cgroup_nodeinfo(parent, nid);
814 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
817 struct mem_cgroup_per_node *pn;
818 struct mem_cgroup *memcg;
819 long x, threshold = MEMCG_CHARGE_BATCH;
821 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
825 __mod_memcg_state(memcg, idx, val);
828 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
830 if (vmstat_item_in_bytes(idx))
831 threshold <<= PAGE_SHIFT;
833 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
834 if (unlikely(abs(x) > threshold)) {
835 pg_data_t *pgdat = lruvec_pgdat(lruvec);
836 struct mem_cgroup_per_node *pi;
838 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
839 atomic_long_add(x, &pi->lruvec_stat[idx]);
842 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
846 * __mod_lruvec_state - update lruvec memory statistics
847 * @lruvec: the lruvec
848 * @idx: the stat item
849 * @val: delta to add to the counter, can be negative
851 * The lruvec is the intersection of the NUMA node and a cgroup. This
852 * function updates the all three counters that are affected by a
853 * change of state at this level: per-node, per-cgroup, per-lruvec.
855 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
859 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
861 /* Update memcg and lruvec */
862 if (!mem_cgroup_disabled())
863 __mod_memcg_lruvec_state(lruvec, idx, val);
866 void __mod_lruvec_slab_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);
875 /* Untracked pages have no memcg, no lruvec. Update only the node */
876 if (!memcg || memcg == root_mem_cgroup) {
877 __mod_node_page_state(pgdat, idx, val);
879 lruvec = mem_cgroup_lruvec(memcg, pgdat);
880 __mod_lruvec_state(lruvec, idx, val);
885 void mod_memcg_obj_state(void *p, int idx, int val)
887 struct mem_cgroup *memcg;
890 memcg = mem_cgroup_from_obj(p);
892 mod_memcg_state(memcg, idx, val);
897 * __count_memcg_events - account VM events in a cgroup
898 * @memcg: the memory cgroup
899 * @idx: the event item
900 * @count: the number of events that occured
902 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
907 if (mem_cgroup_disabled())
910 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
911 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
912 struct mem_cgroup *mi;
915 * Batch local counters to keep them in sync with
916 * the hierarchical ones.
918 __this_cpu_add(memcg->vmstats_local->events[idx], x);
919 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
920 atomic_long_add(x, &mi->vmevents[idx]);
923 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
926 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
928 return atomic_long_read(&memcg->vmevents[event]);
931 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
936 for_each_possible_cpu(cpu)
937 x += per_cpu(memcg->vmstats_local->events[event], cpu);
941 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
945 /* pagein of a big page is an event. So, ignore page size */
947 __count_memcg_events(memcg, PGPGIN, 1);
949 __count_memcg_events(memcg, PGPGOUT, 1);
950 nr_pages = -nr_pages; /* for event */
953 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
956 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
957 enum mem_cgroup_events_target target)
959 unsigned long val, next;
961 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
962 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
963 /* from time_after() in jiffies.h */
964 if ((long)(next - val) < 0) {
966 case MEM_CGROUP_TARGET_THRESH:
967 next = val + THRESHOLDS_EVENTS_TARGET;
969 case MEM_CGROUP_TARGET_SOFTLIMIT:
970 next = val + SOFTLIMIT_EVENTS_TARGET;
975 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
982 * Check events in order.
985 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
987 /* threshold event is triggered in finer grain than soft limit */
988 if (unlikely(mem_cgroup_event_ratelimit(memcg,
989 MEM_CGROUP_TARGET_THRESH))) {
992 do_softlimit = mem_cgroup_event_ratelimit(memcg,
993 MEM_CGROUP_TARGET_SOFTLIMIT);
994 mem_cgroup_threshold(memcg);
995 if (unlikely(do_softlimit))
996 mem_cgroup_update_tree(memcg, page);
1000 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1003 * mm_update_next_owner() may clear mm->owner to NULL
1004 * if it races with swapoff, page migration, etc.
1005 * So this can be called with p == NULL.
1010 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1012 EXPORT_SYMBOL(mem_cgroup_from_task);
1015 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1016 * @mm: mm from which memcg should be extracted. It can be NULL.
1018 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1019 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1022 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1024 struct mem_cgroup *memcg;
1026 if (mem_cgroup_disabled())
1032 * Page cache insertions can happen withou an
1033 * actual mm context, e.g. during disk probing
1034 * on boot, loopback IO, acct() writes etc.
1037 memcg = root_mem_cgroup;
1039 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1040 if (unlikely(!memcg))
1041 memcg = root_mem_cgroup;
1043 } while (!css_tryget(&memcg->css));
1047 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1050 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1051 * @page: page from which memcg should be extracted.
1053 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1054 * root_mem_cgroup is returned.
1056 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
1058 struct mem_cgroup *memcg = page->mem_cgroup;
1060 if (mem_cgroup_disabled())
1064 /* Page should not get uncharged and freed memcg under us. */
1065 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
1066 memcg = root_mem_cgroup;
1070 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1073 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
1075 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1077 if (unlikely(current->active_memcg)) {
1078 struct mem_cgroup *memcg;
1081 /* current->active_memcg must hold a ref. */
1082 if (WARN_ON_ONCE(!css_tryget(¤t->active_memcg->css)))
1083 memcg = root_mem_cgroup;
1085 memcg = current->active_memcg;
1089 return get_mem_cgroup_from_mm(current->mm);
1093 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1094 * @root: hierarchy root
1095 * @prev: previously returned memcg, NULL on first invocation
1096 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1098 * Returns references to children of the hierarchy below @root, or
1099 * @root itself, or %NULL after a full round-trip.
1101 * Caller must pass the return value in @prev on subsequent
1102 * invocations for reference counting, or use mem_cgroup_iter_break()
1103 * to cancel a hierarchy walk before the round-trip is complete.
1105 * Reclaimers can specify a node and a priority level in @reclaim to
1106 * divide up the memcgs in the hierarchy among all concurrent
1107 * reclaimers operating on the same node and priority.
1109 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1110 struct mem_cgroup *prev,
1111 struct mem_cgroup_reclaim_cookie *reclaim)
1113 struct mem_cgroup_reclaim_iter *iter;
1114 struct cgroup_subsys_state *css = NULL;
1115 struct mem_cgroup *memcg = NULL;
1116 struct mem_cgroup *pos = NULL;
1118 if (mem_cgroup_disabled())
1122 root = root_mem_cgroup;
1124 if (prev && !reclaim)
1127 if (!root->use_hierarchy && root != root_mem_cgroup) {
1136 struct mem_cgroup_per_node *mz;
1138 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1141 if (prev && reclaim->generation != iter->generation)
1145 pos = READ_ONCE(iter->position);
1146 if (!pos || css_tryget(&pos->css))
1149 * css reference reached zero, so iter->position will
1150 * be cleared by ->css_released. However, we should not
1151 * rely on this happening soon, because ->css_released
1152 * is called from a work queue, and by busy-waiting we
1153 * might block it. So we clear iter->position right
1156 (void)cmpxchg(&iter->position, pos, NULL);
1164 css = css_next_descendant_pre(css, &root->css);
1167 * Reclaimers share the hierarchy walk, and a
1168 * new one might jump in right at the end of
1169 * the hierarchy - make sure they see at least
1170 * one group and restart from the beginning.
1178 * Verify the css and acquire a reference. The root
1179 * is provided by the caller, so we know it's alive
1180 * and kicking, and don't take an extra reference.
1182 memcg = mem_cgroup_from_css(css);
1184 if (css == &root->css)
1187 if (css_tryget(css))
1195 * The position could have already been updated by a competing
1196 * thread, so check that the value hasn't changed since we read
1197 * it to avoid reclaiming from the same cgroup twice.
1199 (void)cmpxchg(&iter->position, pos, memcg);
1207 reclaim->generation = iter->generation;
1213 if (prev && prev != root)
1214 css_put(&prev->css);
1220 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1221 * @root: hierarchy root
1222 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1224 void mem_cgroup_iter_break(struct mem_cgroup *root,
1225 struct mem_cgroup *prev)
1228 root = root_mem_cgroup;
1229 if (prev && prev != root)
1230 css_put(&prev->css);
1233 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1234 struct mem_cgroup *dead_memcg)
1236 struct mem_cgroup_reclaim_iter *iter;
1237 struct mem_cgroup_per_node *mz;
1240 for_each_node(nid) {
1241 mz = mem_cgroup_nodeinfo(from, nid);
1243 cmpxchg(&iter->position, dead_memcg, NULL);
1247 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1249 struct mem_cgroup *memcg = dead_memcg;
1250 struct mem_cgroup *last;
1253 __invalidate_reclaim_iterators(memcg, dead_memcg);
1255 } while ((memcg = parent_mem_cgroup(memcg)));
1258 * When cgruop1 non-hierarchy mode is used,
1259 * parent_mem_cgroup() does not walk all the way up to the
1260 * cgroup root (root_mem_cgroup). So we have to handle
1261 * dead_memcg from cgroup root separately.
1263 if (last != root_mem_cgroup)
1264 __invalidate_reclaim_iterators(root_mem_cgroup,
1269 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1270 * @memcg: hierarchy root
1271 * @fn: function to call for each task
1272 * @arg: argument passed to @fn
1274 * This function iterates over tasks attached to @memcg or to any of its
1275 * descendants and calls @fn for each task. If @fn returns a non-zero
1276 * value, the function breaks the iteration loop and returns the value.
1277 * Otherwise, it will iterate over all tasks and return 0.
1279 * This function must not be called for the root memory cgroup.
1281 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1282 int (*fn)(struct task_struct *, void *), void *arg)
1284 struct mem_cgroup *iter;
1287 BUG_ON(memcg == root_mem_cgroup);
1289 for_each_mem_cgroup_tree(iter, memcg) {
1290 struct css_task_iter it;
1291 struct task_struct *task;
1293 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1294 while (!ret && (task = css_task_iter_next(&it)))
1295 ret = fn(task, arg);
1296 css_task_iter_end(&it);
1298 mem_cgroup_iter_break(memcg, iter);
1306 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1308 * @pgdat: pgdat of the page
1310 * This function relies on page->mem_cgroup being stable - see the
1311 * access rules in commit_charge().
1313 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1315 struct mem_cgroup_per_node *mz;
1316 struct mem_cgroup *memcg;
1317 struct lruvec *lruvec;
1319 if (mem_cgroup_disabled()) {
1320 lruvec = &pgdat->__lruvec;
1324 memcg = page->mem_cgroup;
1326 * Swapcache readahead pages are added to the LRU - and
1327 * possibly migrated - before they are charged.
1330 memcg = root_mem_cgroup;
1332 mz = mem_cgroup_page_nodeinfo(memcg, page);
1333 lruvec = &mz->lruvec;
1336 * Since a node can be onlined after the mem_cgroup was created,
1337 * we have to be prepared to initialize lruvec->zone here;
1338 * and if offlined then reonlined, we need to reinitialize it.
1340 if (unlikely(lruvec->pgdat != pgdat))
1341 lruvec->pgdat = pgdat;
1346 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1347 * @lruvec: mem_cgroup per zone lru vector
1348 * @lru: index of lru list the page is sitting on
1349 * @zid: zone id of the accounted pages
1350 * @nr_pages: positive when adding or negative when removing
1352 * This function must be called under lru_lock, just before a page is added
1353 * to or just after a page is removed from an lru list (that ordering being
1354 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1356 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1357 int zid, int nr_pages)
1359 struct mem_cgroup_per_node *mz;
1360 unsigned long *lru_size;
1363 if (mem_cgroup_disabled())
1366 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1367 lru_size = &mz->lru_zone_size[zid][lru];
1370 *lru_size += nr_pages;
1373 if (WARN_ONCE(size < 0,
1374 "%s(%p, %d, %d): lru_size %ld\n",
1375 __func__, lruvec, lru, nr_pages, size)) {
1381 *lru_size += nr_pages;
1385 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1386 * @memcg: the memory cgroup
1388 * Returns the maximum amount of memory @mem can be charged with, in
1391 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1393 unsigned long margin = 0;
1394 unsigned long count;
1395 unsigned long limit;
1397 count = page_counter_read(&memcg->memory);
1398 limit = READ_ONCE(memcg->memory.max);
1400 margin = limit - count;
1402 if (do_memsw_account()) {
1403 count = page_counter_read(&memcg->memsw);
1404 limit = READ_ONCE(memcg->memsw.max);
1406 margin = min(margin, limit - count);
1415 * A routine for checking "mem" is under move_account() or not.
1417 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1418 * moving cgroups. This is for waiting at high-memory pressure
1421 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1423 struct mem_cgroup *from;
1424 struct mem_cgroup *to;
1427 * Unlike task_move routines, we access mc.to, mc.from not under
1428 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1430 spin_lock(&mc.lock);
1436 ret = mem_cgroup_is_descendant(from, memcg) ||
1437 mem_cgroup_is_descendant(to, memcg);
1439 spin_unlock(&mc.lock);
1443 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1445 if (mc.moving_task && current != mc.moving_task) {
1446 if (mem_cgroup_under_move(memcg)) {
1448 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1449 /* moving charge context might have finished. */
1452 finish_wait(&mc.waitq, &wait);
1459 static char *memory_stat_format(struct mem_cgroup *memcg)
1464 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1469 * Provide statistics on the state of the memory subsystem as
1470 * well as cumulative event counters that show past behavior.
1472 * This list is ordered following a combination of these gradients:
1473 * 1) generic big picture -> specifics and details
1474 * 2) reflecting userspace activity -> reflecting kernel heuristics
1476 * Current memory state:
1479 seq_buf_printf(&s, "anon %llu\n",
1480 (u64)memcg_page_state(memcg, NR_ANON_MAPPED) *
1482 seq_buf_printf(&s, "file %llu\n",
1483 (u64)memcg_page_state(memcg, NR_FILE_PAGES) *
1485 seq_buf_printf(&s, "kernel_stack %llu\n",
1486 (u64)memcg_page_state(memcg, NR_KERNEL_STACK_KB) *
1488 seq_buf_printf(&s, "slab %llu\n",
1489 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B) +
1490 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B)));
1491 seq_buf_printf(&s, "percpu %llu\n",
1492 (u64)memcg_page_state(memcg, MEMCG_PERCPU_B));
1493 seq_buf_printf(&s, "sock %llu\n",
1494 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1497 seq_buf_printf(&s, "shmem %llu\n",
1498 (u64)memcg_page_state(memcg, NR_SHMEM) *
1500 seq_buf_printf(&s, "file_mapped %llu\n",
1501 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1503 seq_buf_printf(&s, "file_dirty %llu\n",
1504 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1506 seq_buf_printf(&s, "file_writeback %llu\n",
1507 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1510 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1511 seq_buf_printf(&s, "anon_thp %llu\n",
1512 (u64)memcg_page_state(memcg, NR_ANON_THPS) *
1516 for (i = 0; i < NR_LRU_LISTS; i++)
1517 seq_buf_printf(&s, "%s %llu\n", lru_list_name(i),
1518 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1521 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1522 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B));
1523 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1524 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B));
1526 /* Accumulated memory events */
1528 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1529 memcg_events(memcg, PGFAULT));
1530 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1531 memcg_events(memcg, PGMAJFAULT));
1533 seq_buf_printf(&s, "workingset_refault_anon %lu\n",
1534 memcg_page_state(memcg, WORKINGSET_REFAULT_ANON));
1535 seq_buf_printf(&s, "workingset_refault_file %lu\n",
1536 memcg_page_state(memcg, WORKINGSET_REFAULT_FILE));
1537 seq_buf_printf(&s, "workingset_activate_anon %lu\n",
1538 memcg_page_state(memcg, WORKINGSET_ACTIVATE_ANON));
1539 seq_buf_printf(&s, "workingset_activate_file %lu\n",
1540 memcg_page_state(memcg, WORKINGSET_ACTIVATE_FILE));
1541 seq_buf_printf(&s, "workingset_restore_anon %lu\n",
1542 memcg_page_state(memcg, WORKINGSET_RESTORE_ANON));
1543 seq_buf_printf(&s, "workingset_restore_file %lu\n",
1544 memcg_page_state(memcg, WORKINGSET_RESTORE_FILE));
1545 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1546 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1548 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1549 memcg_events(memcg, PGREFILL));
1550 seq_buf_printf(&s, "pgscan %lu\n",
1551 memcg_events(memcg, PGSCAN_KSWAPD) +
1552 memcg_events(memcg, PGSCAN_DIRECT));
1553 seq_buf_printf(&s, "pgsteal %lu\n",
1554 memcg_events(memcg, PGSTEAL_KSWAPD) +
1555 memcg_events(memcg, PGSTEAL_DIRECT));
1556 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1557 memcg_events(memcg, PGACTIVATE));
1558 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1559 memcg_events(memcg, PGDEACTIVATE));
1560 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1561 memcg_events(memcg, PGLAZYFREE));
1562 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1563 memcg_events(memcg, PGLAZYFREED));
1565 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1566 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1567 memcg_events(memcg, THP_FAULT_ALLOC));
1568 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1569 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1570 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1572 /* The above should easily fit into one page */
1573 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1578 #define K(x) ((x) << (PAGE_SHIFT-10))
1580 * mem_cgroup_print_oom_context: Print OOM information relevant to
1581 * memory controller.
1582 * @memcg: The memory cgroup that went over limit
1583 * @p: Task that is going to be killed
1585 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1588 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1593 pr_cont(",oom_memcg=");
1594 pr_cont_cgroup_path(memcg->css.cgroup);
1596 pr_cont(",global_oom");
1598 pr_cont(",task_memcg=");
1599 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1605 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1606 * memory controller.
1607 * @memcg: The memory cgroup that went over limit
1609 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1613 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1614 K((u64)page_counter_read(&memcg->memory)),
1615 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1616 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1617 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1618 K((u64)page_counter_read(&memcg->swap)),
1619 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1621 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1622 K((u64)page_counter_read(&memcg->memsw)),
1623 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1624 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1625 K((u64)page_counter_read(&memcg->kmem)),
1626 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1629 pr_info("Memory cgroup stats for ");
1630 pr_cont_cgroup_path(memcg->css.cgroup);
1632 buf = memory_stat_format(memcg);
1640 * Return the memory (and swap, if configured) limit for a memcg.
1642 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1646 max = READ_ONCE(memcg->memory.max);
1647 if (mem_cgroup_swappiness(memcg)) {
1648 unsigned long memsw_max;
1649 unsigned long swap_max;
1651 memsw_max = memcg->memsw.max;
1652 swap_max = READ_ONCE(memcg->swap.max);
1653 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1654 max = min(max + swap_max, memsw_max);
1659 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1661 return page_counter_read(&memcg->memory);
1664 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1667 struct oom_control oc = {
1671 .gfp_mask = gfp_mask,
1676 if (mutex_lock_killable(&oom_lock))
1679 if (mem_cgroup_margin(memcg) >= (1 << order))
1683 * A few threads which were not waiting at mutex_lock_killable() can
1684 * fail to bail out. Therefore, check again after holding oom_lock.
1686 ret = should_force_charge() || out_of_memory(&oc);
1689 mutex_unlock(&oom_lock);
1693 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1696 unsigned long *total_scanned)
1698 struct mem_cgroup *victim = NULL;
1701 unsigned long excess;
1702 unsigned long nr_scanned;
1703 struct mem_cgroup_reclaim_cookie reclaim = {
1707 excess = soft_limit_excess(root_memcg);
1710 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1715 * If we have not been able to reclaim
1716 * anything, it might because there are
1717 * no reclaimable pages under this hierarchy
1722 * We want to do more targeted reclaim.
1723 * excess >> 2 is not to excessive so as to
1724 * reclaim too much, nor too less that we keep
1725 * coming back to reclaim from this cgroup
1727 if (total >= (excess >> 2) ||
1728 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1733 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1734 pgdat, &nr_scanned);
1735 *total_scanned += nr_scanned;
1736 if (!soft_limit_excess(root_memcg))
1739 mem_cgroup_iter_break(root_memcg, victim);
1743 #ifdef CONFIG_LOCKDEP
1744 static struct lockdep_map memcg_oom_lock_dep_map = {
1745 .name = "memcg_oom_lock",
1749 static DEFINE_SPINLOCK(memcg_oom_lock);
1752 * Check OOM-Killer is already running under our hierarchy.
1753 * If someone is running, return false.
1755 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1757 struct mem_cgroup *iter, *failed = NULL;
1759 spin_lock(&memcg_oom_lock);
1761 for_each_mem_cgroup_tree(iter, memcg) {
1762 if (iter->oom_lock) {
1764 * this subtree of our hierarchy is already locked
1765 * so we cannot give a lock.
1768 mem_cgroup_iter_break(memcg, iter);
1771 iter->oom_lock = true;
1776 * OK, we failed to lock the whole subtree so we have
1777 * to clean up what we set up to the failing subtree
1779 for_each_mem_cgroup_tree(iter, memcg) {
1780 if (iter == failed) {
1781 mem_cgroup_iter_break(memcg, iter);
1784 iter->oom_lock = false;
1787 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1789 spin_unlock(&memcg_oom_lock);
1794 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1796 struct mem_cgroup *iter;
1798 spin_lock(&memcg_oom_lock);
1799 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1800 for_each_mem_cgroup_tree(iter, memcg)
1801 iter->oom_lock = false;
1802 spin_unlock(&memcg_oom_lock);
1805 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1807 struct mem_cgroup *iter;
1809 spin_lock(&memcg_oom_lock);
1810 for_each_mem_cgroup_tree(iter, memcg)
1812 spin_unlock(&memcg_oom_lock);
1815 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1817 struct mem_cgroup *iter;
1820 * When a new child is created while the hierarchy is under oom,
1821 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1823 spin_lock(&memcg_oom_lock);
1824 for_each_mem_cgroup_tree(iter, memcg)
1825 if (iter->under_oom > 0)
1827 spin_unlock(&memcg_oom_lock);
1830 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1832 struct oom_wait_info {
1833 struct mem_cgroup *memcg;
1834 wait_queue_entry_t wait;
1837 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1838 unsigned mode, int sync, void *arg)
1840 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1841 struct mem_cgroup *oom_wait_memcg;
1842 struct oom_wait_info *oom_wait_info;
1844 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1845 oom_wait_memcg = oom_wait_info->memcg;
1847 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1848 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1850 return autoremove_wake_function(wait, mode, sync, arg);
1853 static void memcg_oom_recover(struct mem_cgroup *memcg)
1856 * For the following lockless ->under_oom test, the only required
1857 * guarantee is that it must see the state asserted by an OOM when
1858 * this function is called as a result of userland actions
1859 * triggered by the notification of the OOM. This is trivially
1860 * achieved by invoking mem_cgroup_mark_under_oom() before
1861 * triggering notification.
1863 if (memcg && memcg->under_oom)
1864 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1874 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1876 enum oom_status ret;
1879 if (order > PAGE_ALLOC_COSTLY_ORDER)
1882 memcg_memory_event(memcg, MEMCG_OOM);
1885 * We are in the middle of the charge context here, so we
1886 * don't want to block when potentially sitting on a callstack
1887 * that holds all kinds of filesystem and mm locks.
1889 * cgroup1 allows disabling the OOM killer and waiting for outside
1890 * handling until the charge can succeed; remember the context and put
1891 * the task to sleep at the end of the page fault when all locks are
1894 * On the other hand, in-kernel OOM killer allows for an async victim
1895 * memory reclaim (oom_reaper) and that means that we are not solely
1896 * relying on the oom victim to make a forward progress and we can
1897 * invoke the oom killer here.
1899 * Please note that mem_cgroup_out_of_memory might fail to find a
1900 * victim and then we have to bail out from the charge path.
1902 if (memcg->oom_kill_disable) {
1903 if (!current->in_user_fault)
1905 css_get(&memcg->css);
1906 current->memcg_in_oom = memcg;
1907 current->memcg_oom_gfp_mask = mask;
1908 current->memcg_oom_order = order;
1913 mem_cgroup_mark_under_oom(memcg);
1915 locked = mem_cgroup_oom_trylock(memcg);
1918 mem_cgroup_oom_notify(memcg);
1920 mem_cgroup_unmark_under_oom(memcg);
1921 if (mem_cgroup_out_of_memory(memcg, mask, order))
1927 mem_cgroup_oom_unlock(memcg);
1933 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1934 * @handle: actually kill/wait or just clean up the OOM state
1936 * This has to be called at the end of a page fault if the memcg OOM
1937 * handler was enabled.
1939 * Memcg supports userspace OOM handling where failed allocations must
1940 * sleep on a waitqueue until the userspace task resolves the
1941 * situation. Sleeping directly in the charge context with all kinds
1942 * of locks held is not a good idea, instead we remember an OOM state
1943 * in the task and mem_cgroup_oom_synchronize() has to be called at
1944 * the end of the page fault to complete the OOM handling.
1946 * Returns %true if an ongoing memcg OOM situation was detected and
1947 * completed, %false otherwise.
1949 bool mem_cgroup_oom_synchronize(bool handle)
1951 struct mem_cgroup *memcg = current->memcg_in_oom;
1952 struct oom_wait_info owait;
1955 /* OOM is global, do not handle */
1962 owait.memcg = memcg;
1963 owait.wait.flags = 0;
1964 owait.wait.func = memcg_oom_wake_function;
1965 owait.wait.private = current;
1966 INIT_LIST_HEAD(&owait.wait.entry);
1968 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1969 mem_cgroup_mark_under_oom(memcg);
1971 locked = mem_cgroup_oom_trylock(memcg);
1974 mem_cgroup_oom_notify(memcg);
1976 if (locked && !memcg->oom_kill_disable) {
1977 mem_cgroup_unmark_under_oom(memcg);
1978 finish_wait(&memcg_oom_waitq, &owait.wait);
1979 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1980 current->memcg_oom_order);
1983 mem_cgroup_unmark_under_oom(memcg);
1984 finish_wait(&memcg_oom_waitq, &owait.wait);
1988 mem_cgroup_oom_unlock(memcg);
1990 * There is no guarantee that an OOM-lock contender
1991 * sees the wakeups triggered by the OOM kill
1992 * uncharges. Wake any sleepers explicitely.
1994 memcg_oom_recover(memcg);
1997 current->memcg_in_oom = NULL;
1998 css_put(&memcg->css);
2003 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2004 * @victim: task to be killed by the OOM killer
2005 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2007 * Returns a pointer to a memory cgroup, which has to be cleaned up
2008 * by killing all belonging OOM-killable tasks.
2010 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2012 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2013 struct mem_cgroup *oom_domain)
2015 struct mem_cgroup *oom_group = NULL;
2016 struct mem_cgroup *memcg;
2018 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2022 oom_domain = root_mem_cgroup;
2026 memcg = mem_cgroup_from_task(victim);
2027 if (memcg == root_mem_cgroup)
2031 * If the victim task has been asynchronously moved to a different
2032 * memory cgroup, we might end up killing tasks outside oom_domain.
2033 * In this case it's better to ignore memory.group.oom.
2035 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2039 * Traverse the memory cgroup hierarchy from the victim task's
2040 * cgroup up to the OOMing cgroup (or root) to find the
2041 * highest-level memory cgroup with oom.group set.
2043 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2044 if (memcg->oom_group)
2047 if (memcg == oom_domain)
2052 css_get(&oom_group->css);
2059 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2061 pr_info("Tasks in ");
2062 pr_cont_cgroup_path(memcg->css.cgroup);
2063 pr_cont(" are going to be killed due to memory.oom.group set\n");
2067 * lock_page_memcg - lock a page->mem_cgroup binding
2070 * This function protects unlocked LRU pages from being moved to
2073 * It ensures lifetime of the returned memcg. Caller is responsible
2074 * for the lifetime of the page; __unlock_page_memcg() is available
2075 * when @page might get freed inside the locked section.
2077 struct mem_cgroup *lock_page_memcg(struct page *page)
2079 struct page *head = compound_head(page); /* rmap on tail pages */
2080 struct mem_cgroup *memcg;
2081 unsigned long flags;
2084 * The RCU lock is held throughout the transaction. The fast
2085 * path can get away without acquiring the memcg->move_lock
2086 * because page moving starts with an RCU grace period.
2088 * The RCU lock also protects the memcg from being freed when
2089 * the page state that is going to change is the only thing
2090 * preventing the page itself from being freed. E.g. writeback
2091 * doesn't hold a page reference and relies on PG_writeback to
2092 * keep off truncation, migration and so forth.
2096 if (mem_cgroup_disabled())
2099 memcg = head->mem_cgroup;
2100 if (unlikely(!memcg))
2103 if (atomic_read(&memcg->moving_account) <= 0)
2106 spin_lock_irqsave(&memcg->move_lock, flags);
2107 if (memcg != head->mem_cgroup) {
2108 spin_unlock_irqrestore(&memcg->move_lock, flags);
2113 * When charge migration first begins, we can have locked and
2114 * unlocked page stat updates happening concurrently. Track
2115 * the task who has the lock for unlock_page_memcg().
2117 memcg->move_lock_task = current;
2118 memcg->move_lock_flags = flags;
2122 EXPORT_SYMBOL(lock_page_memcg);
2125 * __unlock_page_memcg - unlock and unpin a memcg
2128 * Unlock and unpin a memcg returned by lock_page_memcg().
2130 void __unlock_page_memcg(struct mem_cgroup *memcg)
2132 if (memcg && memcg->move_lock_task == current) {
2133 unsigned long flags = memcg->move_lock_flags;
2135 memcg->move_lock_task = NULL;
2136 memcg->move_lock_flags = 0;
2138 spin_unlock_irqrestore(&memcg->move_lock, flags);
2145 * unlock_page_memcg - unlock a page->mem_cgroup binding
2148 void unlock_page_memcg(struct page *page)
2150 struct page *head = compound_head(page);
2152 __unlock_page_memcg(head->mem_cgroup);
2154 EXPORT_SYMBOL(unlock_page_memcg);
2156 struct memcg_stock_pcp {
2157 struct mem_cgroup *cached; /* this never be root cgroup */
2158 unsigned int nr_pages;
2160 #ifdef CONFIG_MEMCG_KMEM
2161 struct obj_cgroup *cached_objcg;
2162 unsigned int nr_bytes;
2165 struct work_struct work;
2166 unsigned long flags;
2167 #define FLUSHING_CACHED_CHARGE 0
2169 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2170 static DEFINE_MUTEX(percpu_charge_mutex);
2172 #ifdef CONFIG_MEMCG_KMEM
2173 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2174 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2175 struct mem_cgroup *root_memcg);
2178 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2181 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2182 struct mem_cgroup *root_memcg)
2189 * consume_stock: Try to consume stocked charge on this cpu.
2190 * @memcg: memcg to consume from.
2191 * @nr_pages: how many pages to charge.
2193 * The charges will only happen if @memcg matches the current cpu's memcg
2194 * stock, and at least @nr_pages are available in that stock. Failure to
2195 * service an allocation will refill the stock.
2197 * returns true if successful, false otherwise.
2199 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2201 struct memcg_stock_pcp *stock;
2202 unsigned long flags;
2205 if (nr_pages > MEMCG_CHARGE_BATCH)
2208 local_irq_save(flags);
2210 stock = this_cpu_ptr(&memcg_stock);
2211 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2212 stock->nr_pages -= nr_pages;
2216 local_irq_restore(flags);
2222 * Returns stocks cached in percpu and reset cached information.
2224 static void drain_stock(struct memcg_stock_pcp *stock)
2226 struct mem_cgroup *old = stock->cached;
2231 if (stock->nr_pages) {
2232 page_counter_uncharge(&old->memory, stock->nr_pages);
2233 if (do_memsw_account())
2234 page_counter_uncharge(&old->memsw, stock->nr_pages);
2235 stock->nr_pages = 0;
2239 stock->cached = NULL;
2242 static void drain_local_stock(struct work_struct *dummy)
2244 struct memcg_stock_pcp *stock;
2245 unsigned long flags;
2248 * The only protection from memory hotplug vs. drain_stock races is
2249 * that we always operate on local CPU stock here with IRQ disabled
2251 local_irq_save(flags);
2253 stock = this_cpu_ptr(&memcg_stock);
2254 drain_obj_stock(stock);
2256 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2258 local_irq_restore(flags);
2262 * Cache charges(val) to local per_cpu area.
2263 * This will be consumed by consume_stock() function, later.
2265 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2267 struct memcg_stock_pcp *stock;
2268 unsigned long flags;
2270 local_irq_save(flags);
2272 stock = this_cpu_ptr(&memcg_stock);
2273 if (stock->cached != memcg) { /* reset if necessary */
2275 css_get(&memcg->css);
2276 stock->cached = memcg;
2278 stock->nr_pages += nr_pages;
2280 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2283 local_irq_restore(flags);
2287 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2288 * of the hierarchy under it.
2290 static void drain_all_stock(struct mem_cgroup *root_memcg)
2294 /* If someone's already draining, avoid adding running more workers. */
2295 if (!mutex_trylock(&percpu_charge_mutex))
2298 * Notify other cpus that system-wide "drain" is running
2299 * We do not care about races with the cpu hotplug because cpu down
2300 * as well as workers from this path always operate on the local
2301 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2304 for_each_online_cpu(cpu) {
2305 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2306 struct mem_cgroup *memcg;
2310 memcg = stock->cached;
2311 if (memcg && stock->nr_pages &&
2312 mem_cgroup_is_descendant(memcg, root_memcg))
2314 if (obj_stock_flush_required(stock, root_memcg))
2319 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2321 drain_local_stock(&stock->work);
2323 schedule_work_on(cpu, &stock->work);
2327 mutex_unlock(&percpu_charge_mutex);
2330 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2332 struct memcg_stock_pcp *stock;
2333 struct mem_cgroup *memcg, *mi;
2335 stock = &per_cpu(memcg_stock, cpu);
2338 for_each_mem_cgroup(memcg) {
2341 for (i = 0; i < MEMCG_NR_STAT; i++) {
2345 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2347 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2348 atomic_long_add(x, &memcg->vmstats[i]);
2350 if (i >= NR_VM_NODE_STAT_ITEMS)
2353 for_each_node(nid) {
2354 struct mem_cgroup_per_node *pn;
2356 pn = mem_cgroup_nodeinfo(memcg, nid);
2357 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2360 atomic_long_add(x, &pn->lruvec_stat[i]);
2361 } while ((pn = parent_nodeinfo(pn, nid)));
2365 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2368 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2370 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2371 atomic_long_add(x, &memcg->vmevents[i]);
2378 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2379 unsigned int nr_pages,
2382 unsigned long nr_reclaimed = 0;
2385 unsigned long pflags;
2387 if (page_counter_read(&memcg->memory) <=
2388 READ_ONCE(memcg->memory.high))
2391 memcg_memory_event(memcg, MEMCG_HIGH);
2393 psi_memstall_enter(&pflags);
2394 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2396 psi_memstall_leave(&pflags);
2397 } while ((memcg = parent_mem_cgroup(memcg)) &&
2398 !mem_cgroup_is_root(memcg));
2400 return nr_reclaimed;
2403 static void high_work_func(struct work_struct *work)
2405 struct mem_cgroup *memcg;
2407 memcg = container_of(work, struct mem_cgroup, high_work);
2408 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2412 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2413 * enough to still cause a significant slowdown in most cases, while still
2414 * allowing diagnostics and tracing to proceed without becoming stuck.
2416 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2419 * When calculating the delay, we use these either side of the exponentiation to
2420 * maintain precision and scale to a reasonable number of jiffies (see the table
2423 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2424 * overage ratio to a delay.
2425 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2426 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2427 * to produce a reasonable delay curve.
2429 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2430 * reasonable delay curve compared to precision-adjusted overage, not
2431 * penalising heavily at first, but still making sure that growth beyond the
2432 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2433 * example, with a high of 100 megabytes:
2435 * +-------+------------------------+
2436 * | usage | time to allocate in ms |
2437 * +-------+------------------------+
2459 * +-------+------------------------+
2461 #define MEMCG_DELAY_PRECISION_SHIFT 20
2462 #define MEMCG_DELAY_SCALING_SHIFT 14
2464 static u64 calculate_overage(unsigned long usage, unsigned long high)
2472 * Prevent division by 0 in overage calculation by acting as if
2473 * it was a threshold of 1 page
2475 high = max(high, 1UL);
2477 overage = usage - high;
2478 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2479 return div64_u64(overage, high);
2482 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2484 u64 overage, max_overage = 0;
2487 overage = calculate_overage(page_counter_read(&memcg->memory),
2488 READ_ONCE(memcg->memory.high));
2489 max_overage = max(overage, max_overage);
2490 } while ((memcg = parent_mem_cgroup(memcg)) &&
2491 !mem_cgroup_is_root(memcg));
2496 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2498 u64 overage, max_overage = 0;
2501 overage = calculate_overage(page_counter_read(&memcg->swap),
2502 READ_ONCE(memcg->swap.high));
2504 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2505 max_overage = max(overage, max_overage);
2506 } while ((memcg = parent_mem_cgroup(memcg)) &&
2507 !mem_cgroup_is_root(memcg));
2513 * Get the number of jiffies that we should penalise a mischievous cgroup which
2514 * is exceeding its memory.high by checking both it and its ancestors.
2516 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2517 unsigned int nr_pages,
2520 unsigned long penalty_jiffies;
2526 * We use overage compared to memory.high to calculate the number of
2527 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2528 * fairly lenient on small overages, and increasingly harsh when the
2529 * memcg in question makes it clear that it has no intention of stopping
2530 * its crazy behaviour, so we exponentially increase the delay based on
2533 penalty_jiffies = max_overage * max_overage * HZ;
2534 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2535 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2538 * Factor in the task's own contribution to the overage, such that four
2539 * N-sized allocations are throttled approximately the same as one
2540 * 4N-sized allocation.
2542 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2543 * larger the current charge patch is than that.
2545 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2549 * Scheduled by try_charge() to be executed from the userland return path
2550 * and reclaims memory over the high limit.
2552 void mem_cgroup_handle_over_high(void)
2554 unsigned long penalty_jiffies;
2555 unsigned long pflags;
2556 unsigned long nr_reclaimed;
2557 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2558 int nr_retries = MAX_RECLAIM_RETRIES;
2559 struct mem_cgroup *memcg;
2560 bool in_retry = false;
2562 if (likely(!nr_pages))
2565 memcg = get_mem_cgroup_from_mm(current->mm);
2566 current->memcg_nr_pages_over_high = 0;
2570 * The allocating task should reclaim at least the batch size, but for
2571 * subsequent retries we only want to do what's necessary to prevent oom
2572 * or breaching resource isolation.
2574 * This is distinct from memory.max or page allocator behaviour because
2575 * memory.high is currently batched, whereas memory.max and the page
2576 * allocator run every time an allocation is made.
2578 nr_reclaimed = reclaim_high(memcg,
2579 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2583 * memory.high is breached and reclaim is unable to keep up. Throttle
2584 * allocators proactively to slow down excessive growth.
2586 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2587 mem_find_max_overage(memcg));
2589 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2590 swap_find_max_overage(memcg));
2593 * Clamp the max delay per usermode return so as to still keep the
2594 * application moving forwards and also permit diagnostics, albeit
2597 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2600 * Don't sleep if the amount of jiffies this memcg owes us is so low
2601 * that it's not even worth doing, in an attempt to be nice to those who
2602 * go only a small amount over their memory.high value and maybe haven't
2603 * been aggressively reclaimed enough yet.
2605 if (penalty_jiffies <= HZ / 100)
2609 * If reclaim is making forward progress but we're still over
2610 * memory.high, we want to encourage that rather than doing allocator
2613 if (nr_reclaimed || nr_retries--) {
2619 * If we exit early, we're guaranteed to die (since
2620 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2621 * need to account for any ill-begotten jiffies to pay them off later.
2623 psi_memstall_enter(&pflags);
2624 schedule_timeout_killable(penalty_jiffies);
2625 psi_memstall_leave(&pflags);
2628 css_put(&memcg->css);
2631 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2632 unsigned int nr_pages)
2634 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2635 int nr_retries = MAX_RECLAIM_RETRIES;
2636 struct mem_cgroup *mem_over_limit;
2637 struct page_counter *counter;
2638 enum oom_status oom_status;
2639 unsigned long nr_reclaimed;
2640 bool may_swap = true;
2641 bool drained = false;
2642 unsigned long pflags;
2644 if (mem_cgroup_is_root(memcg))
2647 if (consume_stock(memcg, nr_pages))
2650 if (!do_memsw_account() ||
2651 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2652 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2654 if (do_memsw_account())
2655 page_counter_uncharge(&memcg->memsw, batch);
2656 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2658 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2662 if (batch > nr_pages) {
2668 * Memcg doesn't have a dedicated reserve for atomic
2669 * allocations. But like the global atomic pool, we need to
2670 * put the burden of reclaim on regular allocation requests
2671 * and let these go through as privileged allocations.
2673 if (gfp_mask & __GFP_ATOMIC)
2677 * Unlike in global OOM situations, memcg is not in a physical
2678 * memory shortage. Allow dying and OOM-killed tasks to
2679 * bypass the last charges so that they can exit quickly and
2680 * free their memory.
2682 if (unlikely(should_force_charge()))
2686 * Prevent unbounded recursion when reclaim operations need to
2687 * allocate memory. This might exceed the limits temporarily,
2688 * but we prefer facilitating memory reclaim and getting back
2689 * under the limit over triggering OOM kills in these cases.
2691 if (unlikely(current->flags & PF_MEMALLOC))
2694 if (unlikely(task_in_memcg_oom(current)))
2697 if (!gfpflags_allow_blocking(gfp_mask))
2700 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2702 psi_memstall_enter(&pflags);
2703 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2704 gfp_mask, may_swap);
2705 psi_memstall_leave(&pflags);
2707 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2711 drain_all_stock(mem_over_limit);
2716 if (gfp_mask & __GFP_NORETRY)
2719 * Even though the limit is exceeded at this point, reclaim
2720 * may have been able to free some pages. Retry the charge
2721 * before killing the task.
2723 * Only for regular pages, though: huge pages are rather
2724 * unlikely to succeed so close to the limit, and we fall back
2725 * to regular pages anyway in case of failure.
2727 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2730 * At task move, charge accounts can be doubly counted. So, it's
2731 * better to wait until the end of task_move if something is going on.
2733 if (mem_cgroup_wait_acct_move(mem_over_limit))
2739 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2742 if (gfp_mask & __GFP_NOFAIL)
2745 if (fatal_signal_pending(current))
2749 * keep retrying as long as the memcg oom killer is able to make
2750 * a forward progress or bypass the charge if the oom killer
2751 * couldn't make any progress.
2753 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2754 get_order(nr_pages * PAGE_SIZE));
2755 switch (oom_status) {
2757 nr_retries = MAX_RECLAIM_RETRIES;
2765 if (!(gfp_mask & __GFP_NOFAIL))
2769 * The allocation either can't fail or will lead to more memory
2770 * being freed very soon. Allow memory usage go over the limit
2771 * temporarily by force charging it.
2773 page_counter_charge(&memcg->memory, nr_pages);
2774 if (do_memsw_account())
2775 page_counter_charge(&memcg->memsw, nr_pages);
2780 if (batch > nr_pages)
2781 refill_stock(memcg, batch - nr_pages);
2784 * If the hierarchy is above the normal consumption range, schedule
2785 * reclaim on returning to userland. We can perform reclaim here
2786 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2787 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2788 * not recorded as it most likely matches current's and won't
2789 * change in the meantime. As high limit is checked again before
2790 * reclaim, the cost of mismatch is negligible.
2793 bool mem_high, swap_high;
2795 mem_high = page_counter_read(&memcg->memory) >
2796 READ_ONCE(memcg->memory.high);
2797 swap_high = page_counter_read(&memcg->swap) >
2798 READ_ONCE(memcg->swap.high);
2800 /* Don't bother a random interrupted task */
2801 if (in_interrupt()) {
2803 schedule_work(&memcg->high_work);
2809 if (mem_high || swap_high) {
2811 * The allocating tasks in this cgroup will need to do
2812 * reclaim or be throttled to prevent further growth
2813 * of the memory or swap footprints.
2815 * Target some best-effort fairness between the tasks,
2816 * and distribute reclaim work and delay penalties
2817 * based on how much each task is actually allocating.
2819 current->memcg_nr_pages_over_high += batch;
2820 set_notify_resume(current);
2823 } while ((memcg = parent_mem_cgroup(memcg)));
2828 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2829 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2831 if (mem_cgroup_is_root(memcg))
2834 page_counter_uncharge(&memcg->memory, nr_pages);
2835 if (do_memsw_account())
2836 page_counter_uncharge(&memcg->memsw, nr_pages);
2840 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2842 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2844 * Any of the following ensures page->mem_cgroup stability:
2848 * - lock_page_memcg()
2849 * - exclusive reference
2851 page->mem_cgroup = memcg;
2854 #ifdef CONFIG_MEMCG_KMEM
2855 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2858 unsigned int objects = objs_per_slab_page(s, page);
2861 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2866 if (cmpxchg(&page->obj_cgroups, NULL,
2867 (struct obj_cgroup **) ((unsigned long)vec | 0x1UL)))
2870 kmemleak_not_leak(vec);
2876 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2878 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2879 * cgroup_mutex, etc.
2881 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2885 if (mem_cgroup_disabled())
2888 page = virt_to_head_page(p);
2891 * If page->mem_cgroup is set, it's either a simple mem_cgroup pointer
2892 * or a pointer to obj_cgroup vector. In the latter case the lowest
2893 * bit of the pointer is set.
2894 * The page->mem_cgroup pointer can be asynchronously changed
2895 * from NULL to (obj_cgroup_vec | 0x1UL), but can't be changed
2896 * from a valid memcg pointer to objcg vector or back.
2898 if (!page->mem_cgroup)
2902 * Slab objects are accounted individually, not per-page.
2903 * Memcg membership data for each individual object is saved in
2904 * the page->obj_cgroups.
2906 if (page_has_obj_cgroups(page)) {
2907 struct obj_cgroup *objcg;
2910 off = obj_to_index(page->slab_cache, page, p);
2911 objcg = page_obj_cgroups(page)[off];
2913 return obj_cgroup_memcg(objcg);
2918 /* All other pages use page->mem_cgroup */
2919 return page->mem_cgroup;
2922 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2924 struct obj_cgroup *objcg = NULL;
2925 struct mem_cgroup *memcg;
2927 if (unlikely(!current->mm && !current->active_memcg))
2931 if (unlikely(current->active_memcg))
2932 memcg = rcu_dereference(current->active_memcg);
2934 memcg = mem_cgroup_from_task(current);
2936 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2937 objcg = rcu_dereference(memcg->objcg);
2938 if (objcg && obj_cgroup_tryget(objcg))
2946 static int memcg_alloc_cache_id(void)
2951 id = ida_simple_get(&memcg_cache_ida,
2952 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2956 if (id < memcg_nr_cache_ids)
2960 * There's no space for the new id in memcg_caches arrays,
2961 * so we have to grow them.
2963 down_write(&memcg_cache_ids_sem);
2965 size = 2 * (id + 1);
2966 if (size < MEMCG_CACHES_MIN_SIZE)
2967 size = MEMCG_CACHES_MIN_SIZE;
2968 else if (size > MEMCG_CACHES_MAX_SIZE)
2969 size = MEMCG_CACHES_MAX_SIZE;
2971 err = memcg_update_all_list_lrus(size);
2973 memcg_nr_cache_ids = size;
2975 up_write(&memcg_cache_ids_sem);
2978 ida_simple_remove(&memcg_cache_ida, id);
2984 static void memcg_free_cache_id(int id)
2986 ida_simple_remove(&memcg_cache_ida, id);
2990 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
2991 * @memcg: memory cgroup to charge
2992 * @gfp: reclaim mode
2993 * @nr_pages: number of pages to charge
2995 * Returns 0 on success, an error code on failure.
2997 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
2998 unsigned int nr_pages)
3000 struct page_counter *counter;
3003 ret = try_charge(memcg, gfp, nr_pages);
3007 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3008 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3011 * Enforce __GFP_NOFAIL allocation because callers are not
3012 * prepared to see failures and likely do not have any failure
3015 if (gfp & __GFP_NOFAIL) {
3016 page_counter_charge(&memcg->kmem, nr_pages);
3019 cancel_charge(memcg, nr_pages);
3026 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3027 * @memcg: memcg to uncharge
3028 * @nr_pages: number of pages to uncharge
3030 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3032 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3033 page_counter_uncharge(&memcg->kmem, nr_pages);
3035 page_counter_uncharge(&memcg->memory, nr_pages);
3036 if (do_memsw_account())
3037 page_counter_uncharge(&memcg->memsw, nr_pages);
3041 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3042 * @page: page to charge
3043 * @gfp: reclaim mode
3044 * @order: allocation order
3046 * Returns 0 on success, an error code on failure.
3048 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3050 struct mem_cgroup *memcg;
3053 if (memcg_kmem_bypass())
3056 memcg = get_mem_cgroup_from_current();
3057 if (!mem_cgroup_is_root(memcg)) {
3058 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3060 page->mem_cgroup = memcg;
3061 __SetPageKmemcg(page);
3065 css_put(&memcg->css);
3070 * __memcg_kmem_uncharge_page: uncharge a kmem page
3071 * @page: page to uncharge
3072 * @order: allocation order
3074 void __memcg_kmem_uncharge_page(struct page *page, int order)
3076 struct mem_cgroup *memcg = page->mem_cgroup;
3077 unsigned int nr_pages = 1 << order;
3082 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3083 __memcg_kmem_uncharge(memcg, nr_pages);
3084 page->mem_cgroup = NULL;
3085 css_put(&memcg->css);
3087 /* slab pages do not have PageKmemcg flag set */
3088 if (PageKmemcg(page))
3089 __ClearPageKmemcg(page);
3092 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3094 struct memcg_stock_pcp *stock;
3095 unsigned long flags;
3098 local_irq_save(flags);
3100 stock = this_cpu_ptr(&memcg_stock);
3101 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3102 stock->nr_bytes -= nr_bytes;
3106 local_irq_restore(flags);
3111 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3113 struct obj_cgroup *old = stock->cached_objcg;
3118 if (stock->nr_bytes) {
3119 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3120 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3124 __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
3129 * The leftover is flushed to the centralized per-memcg value.
3130 * On the next attempt to refill obj stock it will be moved
3131 * to a per-cpu stock (probably, on an other CPU), see
3132 * refill_obj_stock().
3134 * How often it's flushed is a trade-off between the memory
3135 * limit enforcement accuracy and potential CPU contention,
3136 * so it might be changed in the future.
3138 atomic_add(nr_bytes, &old->nr_charged_bytes);
3139 stock->nr_bytes = 0;
3142 obj_cgroup_put(old);
3143 stock->cached_objcg = NULL;
3146 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3147 struct mem_cgroup *root_memcg)
3149 struct mem_cgroup *memcg;
3151 if (stock->cached_objcg) {
3152 memcg = obj_cgroup_memcg(stock->cached_objcg);
3153 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3160 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3162 struct memcg_stock_pcp *stock;
3163 unsigned long flags;
3165 local_irq_save(flags);
3167 stock = this_cpu_ptr(&memcg_stock);
3168 if (stock->cached_objcg != objcg) { /* reset if necessary */
3169 drain_obj_stock(stock);
3170 obj_cgroup_get(objcg);
3171 stock->cached_objcg = objcg;
3172 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3174 stock->nr_bytes += nr_bytes;
3176 if (stock->nr_bytes > PAGE_SIZE)
3177 drain_obj_stock(stock);
3179 local_irq_restore(flags);
3182 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3184 struct mem_cgroup *memcg;
3185 unsigned int nr_pages, nr_bytes;
3188 if (consume_obj_stock(objcg, size))
3192 * In theory, memcg->nr_charged_bytes can have enough
3193 * pre-charged bytes to satisfy the allocation. However,
3194 * flushing memcg->nr_charged_bytes requires two atomic
3195 * operations, and memcg->nr_charged_bytes can't be big,
3196 * so it's better to ignore it and try grab some new pages.
3197 * memcg->nr_charged_bytes will be flushed in
3198 * refill_obj_stock(), called from this function or
3199 * independently later.
3202 memcg = obj_cgroup_memcg(objcg);
3203 css_get(&memcg->css);
3206 nr_pages = size >> PAGE_SHIFT;
3207 nr_bytes = size & (PAGE_SIZE - 1);
3212 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3213 if (!ret && nr_bytes)
3214 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3216 css_put(&memcg->css);
3220 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3222 refill_obj_stock(objcg, size);
3225 #endif /* CONFIG_MEMCG_KMEM */
3227 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3230 * Because tail pages are not marked as "used", set it. We're under
3231 * pgdat->lru_lock and migration entries setup in all page mappings.
3233 void mem_cgroup_split_huge_fixup(struct page *head)
3235 struct mem_cgroup *memcg = head->mem_cgroup;
3238 if (mem_cgroup_disabled())
3241 for (i = 1; i < HPAGE_PMD_NR; i++) {
3242 css_get(&memcg->css);
3243 head[i].mem_cgroup = memcg;
3246 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3248 #ifdef CONFIG_MEMCG_SWAP
3250 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3251 * @entry: swap entry to be moved
3252 * @from: mem_cgroup which the entry is moved from
3253 * @to: mem_cgroup which the entry is moved to
3255 * It succeeds only when the swap_cgroup's record for this entry is the same
3256 * as the mem_cgroup's id of @from.
3258 * Returns 0 on success, -EINVAL on failure.
3260 * The caller must have charged to @to, IOW, called page_counter_charge() about
3261 * both res and memsw, and called css_get().
3263 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3264 struct mem_cgroup *from, struct mem_cgroup *to)
3266 unsigned short old_id, new_id;
3268 old_id = mem_cgroup_id(from);
3269 new_id = mem_cgroup_id(to);
3271 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3272 mod_memcg_state(from, MEMCG_SWAP, -1);
3273 mod_memcg_state(to, MEMCG_SWAP, 1);
3279 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3280 struct mem_cgroup *from, struct mem_cgroup *to)
3286 static DEFINE_MUTEX(memcg_max_mutex);
3288 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3289 unsigned long max, bool memsw)
3291 bool enlarge = false;
3292 bool drained = false;
3294 bool limits_invariant;
3295 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3298 if (signal_pending(current)) {
3303 mutex_lock(&memcg_max_mutex);
3305 * Make sure that the new limit (memsw or memory limit) doesn't
3306 * break our basic invariant rule memory.max <= memsw.max.
3308 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3309 max <= memcg->memsw.max;
3310 if (!limits_invariant) {
3311 mutex_unlock(&memcg_max_mutex);
3315 if (max > counter->max)
3317 ret = page_counter_set_max(counter, max);
3318 mutex_unlock(&memcg_max_mutex);
3324 drain_all_stock(memcg);
3329 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3330 GFP_KERNEL, !memsw)) {
3336 if (!ret && enlarge)
3337 memcg_oom_recover(memcg);
3342 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3344 unsigned long *total_scanned)
3346 unsigned long nr_reclaimed = 0;
3347 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3348 unsigned long reclaimed;
3350 struct mem_cgroup_tree_per_node *mctz;
3351 unsigned long excess;
3352 unsigned long nr_scanned;
3357 mctz = soft_limit_tree_node(pgdat->node_id);
3360 * Do not even bother to check the largest node if the root
3361 * is empty. Do it lockless to prevent lock bouncing. Races
3362 * are acceptable as soft limit is best effort anyway.
3364 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3368 * This loop can run a while, specially if mem_cgroup's continuously
3369 * keep exceeding their soft limit and putting the system under
3376 mz = mem_cgroup_largest_soft_limit_node(mctz);
3381 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3382 gfp_mask, &nr_scanned);
3383 nr_reclaimed += reclaimed;
3384 *total_scanned += nr_scanned;
3385 spin_lock_irq(&mctz->lock);
3386 __mem_cgroup_remove_exceeded(mz, mctz);
3389 * If we failed to reclaim anything from this memory cgroup
3390 * it is time to move on to the next cgroup
3394 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3396 excess = soft_limit_excess(mz->memcg);
3398 * One school of thought says that we should not add
3399 * back the node to the tree if reclaim returns 0.
3400 * But our reclaim could return 0, simply because due
3401 * to priority we are exposing a smaller subset of
3402 * memory to reclaim from. Consider this as a longer
3405 /* If excess == 0, no tree ops */
3406 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3407 spin_unlock_irq(&mctz->lock);
3408 css_put(&mz->memcg->css);
3411 * Could not reclaim anything and there are no more
3412 * mem cgroups to try or we seem to be looping without
3413 * reclaiming anything.
3415 if (!nr_reclaimed &&
3417 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3419 } while (!nr_reclaimed);
3421 css_put(&next_mz->memcg->css);
3422 return nr_reclaimed;
3426 * Test whether @memcg has children, dead or alive. Note that this
3427 * function doesn't care whether @memcg has use_hierarchy enabled and
3428 * returns %true if there are child csses according to the cgroup
3429 * hierarchy. Testing use_hierarchy is the caller's responsibility.
3431 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3436 ret = css_next_child(NULL, &memcg->css);
3442 * Reclaims as many pages from the given memcg as possible.
3444 * Caller is responsible for holding css reference for memcg.
3446 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3448 int nr_retries = MAX_RECLAIM_RETRIES;
3450 /* we call try-to-free pages for make this cgroup empty */
3451 lru_add_drain_all();
3453 drain_all_stock(memcg);
3455 /* try to free all pages in this cgroup */
3456 while (nr_retries && page_counter_read(&memcg->memory)) {
3459 if (signal_pending(current))
3462 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3466 /* maybe some writeback is necessary */
3467 congestion_wait(BLK_RW_ASYNC, HZ/10);
3475 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3476 char *buf, size_t nbytes,
3479 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3481 if (mem_cgroup_is_root(memcg))
3483 return mem_cgroup_force_empty(memcg) ?: nbytes;
3486 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3489 return mem_cgroup_from_css(css)->use_hierarchy;
3492 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3493 struct cftype *cft, u64 val)
3496 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3497 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3499 if (memcg->use_hierarchy == val)
3503 * If parent's use_hierarchy is set, we can't make any modifications
3504 * in the child subtrees. If it is unset, then the change can
3505 * occur, provided the current cgroup has no children.
3507 * For the root cgroup, parent_mem is NULL, we allow value to be
3508 * set if there are no children.
3510 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3511 (val == 1 || val == 0)) {
3512 if (!memcg_has_children(memcg))
3513 memcg->use_hierarchy = val;
3522 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3526 if (mem_cgroup_is_root(memcg)) {
3527 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3528 memcg_page_state(memcg, NR_ANON_MAPPED);
3530 val += memcg_page_state(memcg, MEMCG_SWAP);
3533 val = page_counter_read(&memcg->memory);
3535 val = page_counter_read(&memcg->memsw);
3548 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3551 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3552 struct page_counter *counter;
3554 switch (MEMFILE_TYPE(cft->private)) {
3556 counter = &memcg->memory;
3559 counter = &memcg->memsw;
3562 counter = &memcg->kmem;
3565 counter = &memcg->tcpmem;
3571 switch (MEMFILE_ATTR(cft->private)) {
3573 if (counter == &memcg->memory)
3574 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3575 if (counter == &memcg->memsw)
3576 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3577 return (u64)page_counter_read(counter) * PAGE_SIZE;
3579 return (u64)counter->max * PAGE_SIZE;
3581 return (u64)counter->watermark * PAGE_SIZE;
3583 return counter->failcnt;
3584 case RES_SOFT_LIMIT:
3585 return (u64)memcg->soft_limit * PAGE_SIZE;
3591 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3593 unsigned long stat[MEMCG_NR_STAT] = {0};
3594 struct mem_cgroup *mi;
3597 for_each_online_cpu(cpu)
3598 for (i = 0; i < MEMCG_NR_STAT; i++)
3599 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3601 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3602 for (i = 0; i < MEMCG_NR_STAT; i++)
3603 atomic_long_add(stat[i], &mi->vmstats[i]);
3605 for_each_node(node) {
3606 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3607 struct mem_cgroup_per_node *pi;
3609 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3612 for_each_online_cpu(cpu)
3613 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3615 pn->lruvec_stat_cpu->count[i], cpu);
3617 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3618 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3619 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3623 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3625 unsigned long events[NR_VM_EVENT_ITEMS];
3626 struct mem_cgroup *mi;
3629 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3632 for_each_online_cpu(cpu)
3633 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3634 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3637 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3638 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3639 atomic_long_add(events[i], &mi->vmevents[i]);
3642 #ifdef CONFIG_MEMCG_KMEM
3643 static int memcg_online_kmem(struct mem_cgroup *memcg)
3645 struct obj_cgroup *objcg;
3648 if (cgroup_memory_nokmem)
3651 BUG_ON(memcg->kmemcg_id >= 0);
3652 BUG_ON(memcg->kmem_state);
3654 memcg_id = memcg_alloc_cache_id();
3658 objcg = obj_cgroup_alloc();
3660 memcg_free_cache_id(memcg_id);
3663 objcg->memcg = memcg;
3664 rcu_assign_pointer(memcg->objcg, objcg);
3666 static_branch_enable(&memcg_kmem_enabled_key);
3669 * A memory cgroup is considered kmem-online as soon as it gets
3670 * kmemcg_id. Setting the id after enabling static branching will
3671 * guarantee no one starts accounting before all call sites are
3674 memcg->kmemcg_id = memcg_id;
3675 memcg->kmem_state = KMEM_ONLINE;
3680 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3682 struct cgroup_subsys_state *css;
3683 struct mem_cgroup *parent, *child;
3686 if (memcg->kmem_state != KMEM_ONLINE)
3689 memcg->kmem_state = KMEM_ALLOCATED;
3691 parent = parent_mem_cgroup(memcg);
3693 parent = root_mem_cgroup;
3695 memcg_reparent_objcgs(memcg, parent);
3697 kmemcg_id = memcg->kmemcg_id;
3698 BUG_ON(kmemcg_id < 0);
3701 * Change kmemcg_id of this cgroup and all its descendants to the
3702 * parent's id, and then move all entries from this cgroup's list_lrus
3703 * to ones of the parent. After we have finished, all list_lrus
3704 * corresponding to this cgroup are guaranteed to remain empty. The
3705 * ordering is imposed by list_lru_node->lock taken by
3706 * memcg_drain_all_list_lrus().
3708 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3709 css_for_each_descendant_pre(css, &memcg->css) {
3710 child = mem_cgroup_from_css(css);
3711 BUG_ON(child->kmemcg_id != kmemcg_id);
3712 child->kmemcg_id = parent->kmemcg_id;
3713 if (!memcg->use_hierarchy)
3718 memcg_drain_all_list_lrus(kmemcg_id, parent);
3720 memcg_free_cache_id(kmemcg_id);
3723 static void memcg_free_kmem(struct mem_cgroup *memcg)
3725 /* css_alloc() failed, offlining didn't happen */
3726 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3727 memcg_offline_kmem(memcg);
3730 static int memcg_online_kmem(struct mem_cgroup *memcg)
3734 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3737 static void memcg_free_kmem(struct mem_cgroup *memcg)
3740 #endif /* CONFIG_MEMCG_KMEM */
3742 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3747 mutex_lock(&memcg_max_mutex);
3748 ret = page_counter_set_max(&memcg->kmem, max);
3749 mutex_unlock(&memcg_max_mutex);
3753 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3757 mutex_lock(&memcg_max_mutex);
3759 ret = page_counter_set_max(&memcg->tcpmem, max);
3763 if (!memcg->tcpmem_active) {
3765 * The active flag needs to be written after the static_key
3766 * update. This is what guarantees that the socket activation
3767 * function is the last one to run. See mem_cgroup_sk_alloc()
3768 * for details, and note that we don't mark any socket as
3769 * belonging to this memcg until that flag is up.
3771 * We need to do this, because static_keys will span multiple
3772 * sites, but we can't control their order. If we mark a socket
3773 * as accounted, but the accounting functions are not patched in
3774 * yet, we'll lose accounting.
3776 * We never race with the readers in mem_cgroup_sk_alloc(),
3777 * because when this value change, the code to process it is not
3780 static_branch_inc(&memcg_sockets_enabled_key);
3781 memcg->tcpmem_active = true;
3784 mutex_unlock(&memcg_max_mutex);
3789 * The user of this function is...
3792 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3793 char *buf, size_t nbytes, loff_t off)
3795 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3796 unsigned long nr_pages;
3799 buf = strstrip(buf);
3800 ret = page_counter_memparse(buf, "-1", &nr_pages);
3804 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3806 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3810 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3812 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3815 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3818 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3819 "Please report your usecase to linux-mm@kvack.org if you "
3820 "depend on this functionality.\n");
3821 ret = memcg_update_kmem_max(memcg, nr_pages);
3824 ret = memcg_update_tcp_max(memcg, nr_pages);
3828 case RES_SOFT_LIMIT:
3829 memcg->soft_limit = nr_pages;
3833 return ret ?: nbytes;
3836 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3837 size_t nbytes, loff_t off)
3839 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3840 struct page_counter *counter;
3842 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3844 counter = &memcg->memory;
3847 counter = &memcg->memsw;
3850 counter = &memcg->kmem;
3853 counter = &memcg->tcpmem;
3859 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3861 page_counter_reset_watermark(counter);
3864 counter->failcnt = 0;
3873 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3876 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3880 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3881 struct cftype *cft, u64 val)
3883 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3885 if (val & ~MOVE_MASK)
3889 * No kind of locking is needed in here, because ->can_attach() will
3890 * check this value once in the beginning of the process, and then carry
3891 * on with stale data. This means that changes to this value will only
3892 * affect task migrations starting after the change.
3894 memcg->move_charge_at_immigrate = val;
3898 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3899 struct cftype *cft, u64 val)
3907 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3908 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3909 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3911 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3912 int nid, unsigned int lru_mask, bool tree)
3914 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3915 unsigned long nr = 0;
3918 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3921 if (!(BIT(lru) & lru_mask))
3924 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3926 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3931 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3932 unsigned int lru_mask,
3935 unsigned long nr = 0;
3939 if (!(BIT(lru) & lru_mask))
3942 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3944 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3949 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3953 unsigned int lru_mask;
3956 static const struct numa_stat stats[] = {
3957 { "total", LRU_ALL },
3958 { "file", LRU_ALL_FILE },
3959 { "anon", LRU_ALL_ANON },
3960 { "unevictable", BIT(LRU_UNEVICTABLE) },
3962 const struct numa_stat *stat;
3964 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3966 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3967 seq_printf(m, "%s=%lu", stat->name,
3968 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3970 for_each_node_state(nid, N_MEMORY)
3971 seq_printf(m, " N%d=%lu", nid,
3972 mem_cgroup_node_nr_lru_pages(memcg, nid,
3973 stat->lru_mask, false));
3977 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3979 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3980 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3982 for_each_node_state(nid, N_MEMORY)
3983 seq_printf(m, " N%d=%lu", nid,
3984 mem_cgroup_node_nr_lru_pages(memcg, nid,
3985 stat->lru_mask, true));
3991 #endif /* CONFIG_NUMA */
3993 static const unsigned int memcg1_stats[] = {
3996 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4006 static const char *const memcg1_stat_names[] = {
4009 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4019 /* Universal VM events cgroup1 shows, original sort order */
4020 static const unsigned int memcg1_events[] = {
4027 static int memcg_stat_show(struct seq_file *m, void *v)
4029 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4030 unsigned long memory, memsw;
4031 struct mem_cgroup *mi;
4034 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4036 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4039 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4041 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4042 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4043 if (memcg1_stats[i] == NR_ANON_THPS)
4046 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4049 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4050 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4051 memcg_events_local(memcg, memcg1_events[i]));
4053 for (i = 0; i < NR_LRU_LISTS; i++)
4054 seq_printf(m, "%s %lu\n", lru_list_name(i),
4055 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4058 /* Hierarchical information */
4059 memory = memsw = PAGE_COUNTER_MAX;
4060 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4061 memory = min(memory, READ_ONCE(mi->memory.max));
4062 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4064 seq_printf(m, "hierarchical_memory_limit %llu\n",
4065 (u64)memory * PAGE_SIZE);
4066 if (do_memsw_account())
4067 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4068 (u64)memsw * PAGE_SIZE);
4070 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4071 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4073 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4074 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
4078 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4079 seq_printf(m, "total_%s %llu\n",
4080 vm_event_name(memcg1_events[i]),
4081 (u64)memcg_events(memcg, memcg1_events[i]));
4083 for (i = 0; i < NR_LRU_LISTS; i++)
4084 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4085 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4088 #ifdef CONFIG_DEBUG_VM
4091 struct mem_cgroup_per_node *mz;
4092 unsigned long anon_cost = 0;
4093 unsigned long file_cost = 0;
4095 for_each_online_pgdat(pgdat) {
4096 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4098 anon_cost += mz->lruvec.anon_cost;
4099 file_cost += mz->lruvec.file_cost;
4101 seq_printf(m, "anon_cost %lu\n", anon_cost);
4102 seq_printf(m, "file_cost %lu\n", file_cost);
4109 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4112 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4114 return mem_cgroup_swappiness(memcg);
4117 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4118 struct cftype *cft, u64 val)
4120 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4126 memcg->swappiness = val;
4128 vm_swappiness = val;
4133 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4135 struct mem_cgroup_threshold_ary *t;
4136 unsigned long usage;
4141 t = rcu_dereference(memcg->thresholds.primary);
4143 t = rcu_dereference(memcg->memsw_thresholds.primary);
4148 usage = mem_cgroup_usage(memcg, swap);
4151 * current_threshold points to threshold just below or equal to usage.
4152 * If it's not true, a threshold was crossed after last
4153 * call of __mem_cgroup_threshold().
4155 i = t->current_threshold;
4158 * Iterate backward over array of thresholds starting from
4159 * current_threshold and check if a threshold is crossed.
4160 * If none of thresholds below usage is crossed, we read
4161 * only one element of the array here.
4163 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4164 eventfd_signal(t->entries[i].eventfd, 1);
4166 /* i = current_threshold + 1 */
4170 * Iterate forward over array of thresholds starting from
4171 * current_threshold+1 and check if a threshold is crossed.
4172 * If none of thresholds above usage is crossed, we read
4173 * only one element of the array here.
4175 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4176 eventfd_signal(t->entries[i].eventfd, 1);
4178 /* Update current_threshold */
4179 t->current_threshold = i - 1;
4184 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4187 __mem_cgroup_threshold(memcg, false);
4188 if (do_memsw_account())
4189 __mem_cgroup_threshold(memcg, true);
4191 memcg = parent_mem_cgroup(memcg);
4195 static int compare_thresholds(const void *a, const void *b)
4197 const struct mem_cgroup_threshold *_a = a;
4198 const struct mem_cgroup_threshold *_b = b;
4200 if (_a->threshold > _b->threshold)
4203 if (_a->threshold < _b->threshold)
4209 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4211 struct mem_cgroup_eventfd_list *ev;
4213 spin_lock(&memcg_oom_lock);
4215 list_for_each_entry(ev, &memcg->oom_notify, list)
4216 eventfd_signal(ev->eventfd, 1);
4218 spin_unlock(&memcg_oom_lock);
4222 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4224 struct mem_cgroup *iter;
4226 for_each_mem_cgroup_tree(iter, memcg)
4227 mem_cgroup_oom_notify_cb(iter);
4230 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4231 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4233 struct mem_cgroup_thresholds *thresholds;
4234 struct mem_cgroup_threshold_ary *new;
4235 unsigned long threshold;
4236 unsigned long usage;
4239 ret = page_counter_memparse(args, "-1", &threshold);
4243 mutex_lock(&memcg->thresholds_lock);
4246 thresholds = &memcg->thresholds;
4247 usage = mem_cgroup_usage(memcg, false);
4248 } else if (type == _MEMSWAP) {
4249 thresholds = &memcg->memsw_thresholds;
4250 usage = mem_cgroup_usage(memcg, true);
4254 /* Check if a threshold crossed before adding a new one */
4255 if (thresholds->primary)
4256 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4258 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4260 /* Allocate memory for new array of thresholds */
4261 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4268 /* Copy thresholds (if any) to new array */
4269 if (thresholds->primary)
4270 memcpy(new->entries, thresholds->primary->entries,
4271 flex_array_size(new, entries, size - 1));
4273 /* Add new threshold */
4274 new->entries[size - 1].eventfd = eventfd;
4275 new->entries[size - 1].threshold = threshold;
4277 /* Sort thresholds. Registering of new threshold isn't time-critical */
4278 sort(new->entries, size, sizeof(*new->entries),
4279 compare_thresholds, NULL);
4281 /* Find current threshold */
4282 new->current_threshold = -1;
4283 for (i = 0; i < size; i++) {
4284 if (new->entries[i].threshold <= usage) {
4286 * new->current_threshold will not be used until
4287 * rcu_assign_pointer(), so it's safe to increment
4290 ++new->current_threshold;
4295 /* Free old spare buffer and save old primary buffer as spare */
4296 kfree(thresholds->spare);
4297 thresholds->spare = thresholds->primary;
4299 rcu_assign_pointer(thresholds->primary, new);
4301 /* To be sure that nobody uses thresholds */
4305 mutex_unlock(&memcg->thresholds_lock);
4310 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4311 struct eventfd_ctx *eventfd, const char *args)
4313 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4316 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4317 struct eventfd_ctx *eventfd, const char *args)
4319 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4322 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4323 struct eventfd_ctx *eventfd, enum res_type type)
4325 struct mem_cgroup_thresholds *thresholds;
4326 struct mem_cgroup_threshold_ary *new;
4327 unsigned long usage;
4328 int i, j, size, entries;
4330 mutex_lock(&memcg->thresholds_lock);
4333 thresholds = &memcg->thresholds;
4334 usage = mem_cgroup_usage(memcg, false);
4335 } else if (type == _MEMSWAP) {
4336 thresholds = &memcg->memsw_thresholds;
4337 usage = mem_cgroup_usage(memcg, true);
4341 if (!thresholds->primary)
4344 /* Check if a threshold crossed before removing */
4345 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4347 /* Calculate new number of threshold */
4349 for (i = 0; i < thresholds->primary->size; i++) {
4350 if (thresholds->primary->entries[i].eventfd != eventfd)
4356 new = thresholds->spare;
4358 /* If no items related to eventfd have been cleared, nothing to do */
4362 /* Set thresholds array to NULL if we don't have thresholds */
4371 /* Copy thresholds and find current threshold */
4372 new->current_threshold = -1;
4373 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4374 if (thresholds->primary->entries[i].eventfd == eventfd)
4377 new->entries[j] = thresholds->primary->entries[i];
4378 if (new->entries[j].threshold <= usage) {
4380 * new->current_threshold will not be used
4381 * until rcu_assign_pointer(), so it's safe to increment
4384 ++new->current_threshold;
4390 /* Swap primary and spare array */
4391 thresholds->spare = thresholds->primary;
4393 rcu_assign_pointer(thresholds->primary, new);
4395 /* To be sure that nobody uses thresholds */
4398 /* If all events are unregistered, free the spare array */
4400 kfree(thresholds->spare);
4401 thresholds->spare = NULL;
4404 mutex_unlock(&memcg->thresholds_lock);
4407 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4408 struct eventfd_ctx *eventfd)
4410 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4413 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4414 struct eventfd_ctx *eventfd)
4416 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4419 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4420 struct eventfd_ctx *eventfd, const char *args)
4422 struct mem_cgroup_eventfd_list *event;
4424 event = kmalloc(sizeof(*event), GFP_KERNEL);
4428 spin_lock(&memcg_oom_lock);
4430 event->eventfd = eventfd;
4431 list_add(&event->list, &memcg->oom_notify);
4433 /* already in OOM ? */
4434 if (memcg->under_oom)
4435 eventfd_signal(eventfd, 1);
4436 spin_unlock(&memcg_oom_lock);
4441 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4442 struct eventfd_ctx *eventfd)
4444 struct mem_cgroup_eventfd_list *ev, *tmp;
4446 spin_lock(&memcg_oom_lock);
4448 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4449 if (ev->eventfd == eventfd) {
4450 list_del(&ev->list);
4455 spin_unlock(&memcg_oom_lock);
4458 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4460 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4462 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4463 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4464 seq_printf(sf, "oom_kill %lu\n",
4465 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4469 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4470 struct cftype *cft, u64 val)
4472 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4474 /* cannot set to root cgroup and only 0 and 1 are allowed */
4475 if (!css->parent || !((val == 0) || (val == 1)))
4478 memcg->oom_kill_disable = val;
4480 memcg_oom_recover(memcg);
4485 #ifdef CONFIG_CGROUP_WRITEBACK
4487 #include <trace/events/writeback.h>
4489 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4491 return wb_domain_init(&memcg->cgwb_domain, gfp);
4494 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4496 wb_domain_exit(&memcg->cgwb_domain);
4499 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4501 wb_domain_size_changed(&memcg->cgwb_domain);
4504 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4506 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4508 if (!memcg->css.parent)
4511 return &memcg->cgwb_domain;
4515 * idx can be of type enum memcg_stat_item or node_stat_item.
4516 * Keep in sync with memcg_exact_page().
4518 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4520 long x = atomic_long_read(&memcg->vmstats[idx]);
4523 for_each_online_cpu(cpu)
4524 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4531 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4532 * @wb: bdi_writeback in question
4533 * @pfilepages: out parameter for number of file pages
4534 * @pheadroom: out parameter for number of allocatable pages according to memcg
4535 * @pdirty: out parameter for number of dirty pages
4536 * @pwriteback: out parameter for number of pages under writeback
4538 * Determine the numbers of file, headroom, dirty, and writeback pages in
4539 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4540 * is a bit more involved.
4542 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4543 * headroom is calculated as the lowest headroom of itself and the
4544 * ancestors. Note that this doesn't consider the actual amount of
4545 * available memory in the system. The caller should further cap
4546 * *@pheadroom accordingly.
4548 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4549 unsigned long *pheadroom, unsigned long *pdirty,
4550 unsigned long *pwriteback)
4552 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4553 struct mem_cgroup *parent;
4555 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4557 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4558 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4559 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4560 *pheadroom = PAGE_COUNTER_MAX;
4562 while ((parent = parent_mem_cgroup(memcg))) {
4563 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4564 READ_ONCE(memcg->memory.high));
4565 unsigned long used = page_counter_read(&memcg->memory);
4567 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4573 * Foreign dirty flushing
4575 * There's an inherent mismatch between memcg and writeback. The former
4576 * trackes ownership per-page while the latter per-inode. This was a
4577 * deliberate design decision because honoring per-page ownership in the
4578 * writeback path is complicated, may lead to higher CPU and IO overheads
4579 * and deemed unnecessary given that write-sharing an inode across
4580 * different cgroups isn't a common use-case.
4582 * Combined with inode majority-writer ownership switching, this works well
4583 * enough in most cases but there are some pathological cases. For
4584 * example, let's say there are two cgroups A and B which keep writing to
4585 * different but confined parts of the same inode. B owns the inode and
4586 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4587 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4588 * triggering background writeback. A will be slowed down without a way to
4589 * make writeback of the dirty pages happen.
4591 * Conditions like the above can lead to a cgroup getting repatedly and
4592 * severely throttled after making some progress after each
4593 * dirty_expire_interval while the underyling IO device is almost
4596 * Solving this problem completely requires matching the ownership tracking
4597 * granularities between memcg and writeback in either direction. However,
4598 * the more egregious behaviors can be avoided by simply remembering the
4599 * most recent foreign dirtying events and initiating remote flushes on
4600 * them when local writeback isn't enough to keep the memory clean enough.
4602 * The following two functions implement such mechanism. When a foreign
4603 * page - a page whose memcg and writeback ownerships don't match - is
4604 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4605 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4606 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4607 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4608 * foreign bdi_writebacks which haven't expired. Both the numbers of
4609 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4610 * limited to MEMCG_CGWB_FRN_CNT.
4612 * The mechanism only remembers IDs and doesn't hold any object references.
4613 * As being wrong occasionally doesn't matter, updates and accesses to the
4614 * records are lockless and racy.
4616 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4617 struct bdi_writeback *wb)
4619 struct mem_cgroup *memcg = page->mem_cgroup;
4620 struct memcg_cgwb_frn *frn;
4621 u64 now = get_jiffies_64();
4622 u64 oldest_at = now;
4626 trace_track_foreign_dirty(page, wb);
4629 * Pick the slot to use. If there is already a slot for @wb, keep
4630 * using it. If not replace the oldest one which isn't being
4633 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4634 frn = &memcg->cgwb_frn[i];
4635 if (frn->bdi_id == wb->bdi->id &&
4636 frn->memcg_id == wb->memcg_css->id)
4638 if (time_before64(frn->at, oldest_at) &&
4639 atomic_read(&frn->done.cnt) == 1) {
4641 oldest_at = frn->at;
4645 if (i < MEMCG_CGWB_FRN_CNT) {
4647 * Re-using an existing one. Update timestamp lazily to
4648 * avoid making the cacheline hot. We want them to be
4649 * reasonably up-to-date and significantly shorter than
4650 * dirty_expire_interval as that's what expires the record.
4651 * Use the shorter of 1s and dirty_expire_interval / 8.
4653 unsigned long update_intv =
4654 min_t(unsigned long, HZ,
4655 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4657 if (time_before64(frn->at, now - update_intv))
4659 } else if (oldest >= 0) {
4660 /* replace the oldest free one */
4661 frn = &memcg->cgwb_frn[oldest];
4662 frn->bdi_id = wb->bdi->id;
4663 frn->memcg_id = wb->memcg_css->id;
4668 /* issue foreign writeback flushes for recorded foreign dirtying events */
4669 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4671 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4672 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4673 u64 now = jiffies_64;
4676 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4677 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4680 * If the record is older than dirty_expire_interval,
4681 * writeback on it has already started. No need to kick it
4682 * off again. Also, don't start a new one if there's
4683 * already one in flight.
4685 if (time_after64(frn->at, now - intv) &&
4686 atomic_read(&frn->done.cnt) == 1) {
4688 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4689 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4690 WB_REASON_FOREIGN_FLUSH,
4696 #else /* CONFIG_CGROUP_WRITEBACK */
4698 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4703 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4707 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4711 #endif /* CONFIG_CGROUP_WRITEBACK */
4714 * DO NOT USE IN NEW FILES.
4716 * "cgroup.event_control" implementation.
4718 * This is way over-engineered. It tries to support fully configurable
4719 * events for each user. Such level of flexibility is completely
4720 * unnecessary especially in the light of the planned unified hierarchy.
4722 * Please deprecate this and replace with something simpler if at all
4727 * Unregister event and free resources.
4729 * Gets called from workqueue.
4731 static void memcg_event_remove(struct work_struct *work)
4733 struct mem_cgroup_event *event =
4734 container_of(work, struct mem_cgroup_event, remove);
4735 struct mem_cgroup *memcg = event->memcg;
4737 remove_wait_queue(event->wqh, &event->wait);
4739 event->unregister_event(memcg, event->eventfd);
4741 /* Notify userspace the event is going away. */
4742 eventfd_signal(event->eventfd, 1);
4744 eventfd_ctx_put(event->eventfd);
4746 css_put(&memcg->css);
4750 * Gets called on EPOLLHUP on eventfd when user closes it.
4752 * Called with wqh->lock held and interrupts disabled.
4754 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4755 int sync, void *key)
4757 struct mem_cgroup_event *event =
4758 container_of(wait, struct mem_cgroup_event, wait);
4759 struct mem_cgroup *memcg = event->memcg;
4760 __poll_t flags = key_to_poll(key);
4762 if (flags & EPOLLHUP) {
4764 * If the event has been detached at cgroup removal, we
4765 * can simply return knowing the other side will cleanup
4768 * We can't race against event freeing since the other
4769 * side will require wqh->lock via remove_wait_queue(),
4772 spin_lock(&memcg->event_list_lock);
4773 if (!list_empty(&event->list)) {
4774 list_del_init(&event->list);
4776 * We are in atomic context, but cgroup_event_remove()
4777 * may sleep, so we have to call it in workqueue.
4779 schedule_work(&event->remove);
4781 spin_unlock(&memcg->event_list_lock);
4787 static void memcg_event_ptable_queue_proc(struct file *file,
4788 wait_queue_head_t *wqh, poll_table *pt)
4790 struct mem_cgroup_event *event =
4791 container_of(pt, struct mem_cgroup_event, pt);
4794 add_wait_queue(wqh, &event->wait);
4798 * DO NOT USE IN NEW FILES.
4800 * Parse input and register new cgroup event handler.
4802 * Input must be in format '<event_fd> <control_fd> <args>'.
4803 * Interpretation of args is defined by control file implementation.
4805 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4806 char *buf, size_t nbytes, loff_t off)
4808 struct cgroup_subsys_state *css = of_css(of);
4809 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4810 struct mem_cgroup_event *event;
4811 struct cgroup_subsys_state *cfile_css;
4812 unsigned int efd, cfd;
4819 buf = strstrip(buf);
4821 efd = simple_strtoul(buf, &endp, 10);
4826 cfd = simple_strtoul(buf, &endp, 10);
4827 if ((*endp != ' ') && (*endp != '\0'))
4831 event = kzalloc(sizeof(*event), GFP_KERNEL);
4835 event->memcg = memcg;
4836 INIT_LIST_HEAD(&event->list);
4837 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4838 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4839 INIT_WORK(&event->remove, memcg_event_remove);
4847 event->eventfd = eventfd_ctx_fileget(efile.file);
4848 if (IS_ERR(event->eventfd)) {
4849 ret = PTR_ERR(event->eventfd);
4856 goto out_put_eventfd;
4859 /* the process need read permission on control file */
4860 /* AV: shouldn't we check that it's been opened for read instead? */
4861 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4866 * Determine the event callbacks and set them in @event. This used
4867 * to be done via struct cftype but cgroup core no longer knows
4868 * about these events. The following is crude but the whole thing
4869 * is for compatibility anyway.
4871 * DO NOT ADD NEW FILES.
4873 name = cfile.file->f_path.dentry->d_name.name;
4875 if (!strcmp(name, "memory.usage_in_bytes")) {
4876 event->register_event = mem_cgroup_usage_register_event;
4877 event->unregister_event = mem_cgroup_usage_unregister_event;
4878 } else if (!strcmp(name, "memory.oom_control")) {
4879 event->register_event = mem_cgroup_oom_register_event;
4880 event->unregister_event = mem_cgroup_oom_unregister_event;
4881 } else if (!strcmp(name, "memory.pressure_level")) {
4882 event->register_event = vmpressure_register_event;
4883 event->unregister_event = vmpressure_unregister_event;
4884 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4885 event->register_event = memsw_cgroup_usage_register_event;
4886 event->unregister_event = memsw_cgroup_usage_unregister_event;
4893 * Verify @cfile should belong to @css. Also, remaining events are
4894 * automatically removed on cgroup destruction but the removal is
4895 * asynchronous, so take an extra ref on @css.
4897 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4898 &memory_cgrp_subsys);
4900 if (IS_ERR(cfile_css))
4902 if (cfile_css != css) {
4907 ret = event->register_event(memcg, event->eventfd, buf);
4911 vfs_poll(efile.file, &event->pt);
4913 spin_lock(&memcg->event_list_lock);
4914 list_add(&event->list, &memcg->event_list);
4915 spin_unlock(&memcg->event_list_lock);
4927 eventfd_ctx_put(event->eventfd);
4936 static struct cftype mem_cgroup_legacy_files[] = {
4938 .name = "usage_in_bytes",
4939 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4940 .read_u64 = mem_cgroup_read_u64,
4943 .name = "max_usage_in_bytes",
4944 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4945 .write = mem_cgroup_reset,
4946 .read_u64 = mem_cgroup_read_u64,
4949 .name = "limit_in_bytes",
4950 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4951 .write = mem_cgroup_write,
4952 .read_u64 = mem_cgroup_read_u64,
4955 .name = "soft_limit_in_bytes",
4956 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4957 .write = mem_cgroup_write,
4958 .read_u64 = mem_cgroup_read_u64,
4962 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4963 .write = mem_cgroup_reset,
4964 .read_u64 = mem_cgroup_read_u64,
4968 .seq_show = memcg_stat_show,
4971 .name = "force_empty",
4972 .write = mem_cgroup_force_empty_write,
4975 .name = "use_hierarchy",
4976 .write_u64 = mem_cgroup_hierarchy_write,
4977 .read_u64 = mem_cgroup_hierarchy_read,
4980 .name = "cgroup.event_control", /* XXX: for compat */
4981 .write = memcg_write_event_control,
4982 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4985 .name = "swappiness",
4986 .read_u64 = mem_cgroup_swappiness_read,
4987 .write_u64 = mem_cgroup_swappiness_write,
4990 .name = "move_charge_at_immigrate",
4991 .read_u64 = mem_cgroup_move_charge_read,
4992 .write_u64 = mem_cgroup_move_charge_write,
4995 .name = "oom_control",
4996 .seq_show = mem_cgroup_oom_control_read,
4997 .write_u64 = mem_cgroup_oom_control_write,
4998 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5001 .name = "pressure_level",
5005 .name = "numa_stat",
5006 .seq_show = memcg_numa_stat_show,
5010 .name = "kmem.limit_in_bytes",
5011 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5012 .write = mem_cgroup_write,
5013 .read_u64 = mem_cgroup_read_u64,
5016 .name = "kmem.usage_in_bytes",
5017 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5018 .read_u64 = mem_cgroup_read_u64,
5021 .name = "kmem.failcnt",
5022 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5023 .write = mem_cgroup_reset,
5024 .read_u64 = mem_cgroup_read_u64,
5027 .name = "kmem.max_usage_in_bytes",
5028 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5029 .write = mem_cgroup_reset,
5030 .read_u64 = mem_cgroup_read_u64,
5032 #if defined(CONFIG_MEMCG_KMEM) && \
5033 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5035 .name = "kmem.slabinfo",
5036 .seq_show = memcg_slab_show,
5040 .name = "kmem.tcp.limit_in_bytes",
5041 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5042 .write = mem_cgroup_write,
5043 .read_u64 = mem_cgroup_read_u64,
5046 .name = "kmem.tcp.usage_in_bytes",
5047 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5048 .read_u64 = mem_cgroup_read_u64,
5051 .name = "kmem.tcp.failcnt",
5052 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5053 .write = mem_cgroup_reset,
5054 .read_u64 = mem_cgroup_read_u64,
5057 .name = "kmem.tcp.max_usage_in_bytes",
5058 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5059 .write = mem_cgroup_reset,
5060 .read_u64 = mem_cgroup_read_u64,
5062 { }, /* terminate */
5066 * Private memory cgroup IDR
5068 * Swap-out records and page cache shadow entries need to store memcg
5069 * references in constrained space, so we maintain an ID space that is
5070 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5071 * memory-controlled cgroups to 64k.
5073 * However, there usually are many references to the offline CSS after
5074 * the cgroup has been destroyed, such as page cache or reclaimable
5075 * slab objects, that don't need to hang on to the ID. We want to keep
5076 * those dead CSS from occupying IDs, or we might quickly exhaust the
5077 * relatively small ID space and prevent the creation of new cgroups
5078 * even when there are much fewer than 64k cgroups - possibly none.
5080 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5081 * be freed and recycled when it's no longer needed, which is usually
5082 * when the CSS is offlined.
5084 * The only exception to that are records of swapped out tmpfs/shmem
5085 * pages that need to be attributed to live ancestors on swapin. But
5086 * those references are manageable from userspace.
5089 static DEFINE_IDR(mem_cgroup_idr);
5091 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5093 if (memcg->id.id > 0) {
5094 idr_remove(&mem_cgroup_idr, memcg->id.id);
5099 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5102 refcount_add(n, &memcg->id.ref);
5105 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5107 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5108 mem_cgroup_id_remove(memcg);
5110 /* Memcg ID pins CSS */
5111 css_put(&memcg->css);
5115 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5117 mem_cgroup_id_put_many(memcg, 1);
5121 * mem_cgroup_from_id - look up a memcg from a memcg id
5122 * @id: the memcg id to look up
5124 * Caller must hold rcu_read_lock().
5126 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5128 WARN_ON_ONCE(!rcu_read_lock_held());
5129 return idr_find(&mem_cgroup_idr, id);
5132 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5134 struct mem_cgroup_per_node *pn;
5137 * This routine is called against possible nodes.
5138 * But it's BUG to call kmalloc() against offline node.
5140 * TODO: this routine can waste much memory for nodes which will
5141 * never be onlined. It's better to use memory hotplug callback
5144 if (!node_state(node, N_NORMAL_MEMORY))
5146 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5150 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5151 GFP_KERNEL_ACCOUNT);
5152 if (!pn->lruvec_stat_local) {
5157 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct lruvec_stat,
5158 GFP_KERNEL_ACCOUNT);
5159 if (!pn->lruvec_stat_cpu) {
5160 free_percpu(pn->lruvec_stat_local);
5165 lruvec_init(&pn->lruvec);
5166 pn->usage_in_excess = 0;
5167 pn->on_tree = false;
5170 memcg->nodeinfo[node] = pn;
5174 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5176 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5181 free_percpu(pn->lruvec_stat_cpu);
5182 free_percpu(pn->lruvec_stat_local);
5186 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5191 free_mem_cgroup_per_node_info(memcg, node);
5192 free_percpu(memcg->vmstats_percpu);
5193 free_percpu(memcg->vmstats_local);
5197 static void mem_cgroup_free(struct mem_cgroup *memcg)
5199 memcg_wb_domain_exit(memcg);
5201 * Flush percpu vmstats and vmevents to guarantee the value correctness
5202 * on parent's and all ancestor levels.
5204 memcg_flush_percpu_vmstats(memcg);
5205 memcg_flush_percpu_vmevents(memcg);
5206 __mem_cgroup_free(memcg);
5209 static struct mem_cgroup *mem_cgroup_alloc(void)
5211 struct mem_cgroup *memcg;
5214 int __maybe_unused i;
5215 long error = -ENOMEM;
5217 size = sizeof(struct mem_cgroup);
5218 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5220 memcg = kzalloc(size, GFP_KERNEL);
5222 return ERR_PTR(error);
5224 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5225 1, MEM_CGROUP_ID_MAX,
5227 if (memcg->id.id < 0) {
5228 error = memcg->id.id;
5232 memcg->vmstats_local = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5233 GFP_KERNEL_ACCOUNT);
5234 if (!memcg->vmstats_local)
5237 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5238 GFP_KERNEL_ACCOUNT);
5239 if (!memcg->vmstats_percpu)
5243 if (alloc_mem_cgroup_per_node_info(memcg, node))
5246 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5249 INIT_WORK(&memcg->high_work, high_work_func);
5250 INIT_LIST_HEAD(&memcg->oom_notify);
5251 mutex_init(&memcg->thresholds_lock);
5252 spin_lock_init(&memcg->move_lock);
5253 vmpressure_init(&memcg->vmpressure);
5254 INIT_LIST_HEAD(&memcg->event_list);
5255 spin_lock_init(&memcg->event_list_lock);
5256 memcg->socket_pressure = jiffies;
5257 #ifdef CONFIG_MEMCG_KMEM
5258 memcg->kmemcg_id = -1;
5259 INIT_LIST_HEAD(&memcg->objcg_list);
5261 #ifdef CONFIG_CGROUP_WRITEBACK
5262 INIT_LIST_HEAD(&memcg->cgwb_list);
5263 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5264 memcg->cgwb_frn[i].done =
5265 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5267 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5268 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5269 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5270 memcg->deferred_split_queue.split_queue_len = 0;
5272 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5275 mem_cgroup_id_remove(memcg);
5276 __mem_cgroup_free(memcg);
5277 return ERR_PTR(error);
5280 static struct cgroup_subsys_state * __ref
5281 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5283 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5284 struct mem_cgroup *memcg;
5285 long error = -ENOMEM;
5287 memalloc_use_memcg(parent);
5288 memcg = mem_cgroup_alloc();
5289 memalloc_unuse_memcg();
5291 return ERR_CAST(memcg);
5293 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5294 memcg->soft_limit = PAGE_COUNTER_MAX;
5295 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5297 memcg->swappiness = mem_cgroup_swappiness(parent);
5298 memcg->oom_kill_disable = parent->oom_kill_disable;
5300 if (parent && parent->use_hierarchy) {
5301 memcg->use_hierarchy = true;
5302 page_counter_init(&memcg->memory, &parent->memory);
5303 page_counter_init(&memcg->swap, &parent->swap);
5304 page_counter_init(&memcg->memsw, &parent->memsw);
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->memsw, NULL);
5311 page_counter_init(&memcg->kmem, NULL);
5312 page_counter_init(&memcg->tcpmem, NULL);
5314 * Deeper hierachy with use_hierarchy == false doesn't make
5315 * much sense so let cgroup subsystem know about this
5316 * unfortunate state in our controller.
5318 if (parent != root_mem_cgroup)
5319 memory_cgrp_subsys.broken_hierarchy = true;
5322 /* The following stuff does not apply to the root */
5324 root_mem_cgroup = memcg;
5328 error = memcg_online_kmem(memcg);
5332 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5333 static_branch_inc(&memcg_sockets_enabled_key);
5337 mem_cgroup_id_remove(memcg);
5338 mem_cgroup_free(memcg);
5339 return ERR_PTR(error);
5342 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5344 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5347 * A memcg must be visible for memcg_expand_shrinker_maps()
5348 * by the time the maps are allocated. So, we allocate maps
5349 * here, when for_each_mem_cgroup() can't skip it.
5351 if (memcg_alloc_shrinker_maps(memcg)) {
5352 mem_cgroup_id_remove(memcg);
5356 /* Online state pins memcg ID, memcg ID pins CSS */
5357 refcount_set(&memcg->id.ref, 1);
5362 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5364 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5365 struct mem_cgroup_event *event, *tmp;
5368 * Unregister events and notify userspace.
5369 * Notify userspace about cgroup removing only after rmdir of cgroup
5370 * directory to avoid race between userspace and kernelspace.
5372 spin_lock(&memcg->event_list_lock);
5373 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5374 list_del_init(&event->list);
5375 schedule_work(&event->remove);
5377 spin_unlock(&memcg->event_list_lock);
5379 page_counter_set_min(&memcg->memory, 0);
5380 page_counter_set_low(&memcg->memory, 0);
5382 memcg_offline_kmem(memcg);
5383 wb_memcg_offline(memcg);
5385 drain_all_stock(memcg);
5387 mem_cgroup_id_put(memcg);
5390 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5392 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5394 invalidate_reclaim_iterators(memcg);
5397 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5399 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5400 int __maybe_unused i;
5402 #ifdef CONFIG_CGROUP_WRITEBACK
5403 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5404 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5406 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5407 static_branch_dec(&memcg_sockets_enabled_key);
5409 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5410 static_branch_dec(&memcg_sockets_enabled_key);
5412 vmpressure_cleanup(&memcg->vmpressure);
5413 cancel_work_sync(&memcg->high_work);
5414 mem_cgroup_remove_from_trees(memcg);
5415 memcg_free_shrinker_maps(memcg);
5416 memcg_free_kmem(memcg);
5417 mem_cgroup_free(memcg);
5421 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5422 * @css: the target css
5424 * Reset the states of the mem_cgroup associated with @css. This is
5425 * invoked when the userland requests disabling on the default hierarchy
5426 * but the memcg is pinned through dependency. The memcg should stop
5427 * applying policies and should revert to the vanilla state as it may be
5428 * made visible again.
5430 * The current implementation only resets the essential configurations.
5431 * This needs to be expanded to cover all the visible parts.
5433 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5435 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5437 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5438 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5439 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
5440 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5441 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5442 page_counter_set_min(&memcg->memory, 0);
5443 page_counter_set_low(&memcg->memory, 0);
5444 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5445 memcg->soft_limit = PAGE_COUNTER_MAX;
5446 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5447 memcg_wb_domain_size_changed(memcg);
5451 /* Handlers for move charge at task migration. */
5452 static int mem_cgroup_do_precharge(unsigned long count)
5456 /* Try a single bulk charge without reclaim first, kswapd may wake */
5457 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5459 mc.precharge += count;
5463 /* Try charges one by one with reclaim, but do not retry */
5465 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5479 enum mc_target_type {
5486 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5487 unsigned long addr, pte_t ptent)
5489 struct page *page = vm_normal_page(vma, addr, ptent);
5491 if (!page || !page_mapped(page))
5493 if (PageAnon(page)) {
5494 if (!(mc.flags & MOVE_ANON))
5497 if (!(mc.flags & MOVE_FILE))
5500 if (!get_page_unless_zero(page))
5506 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5507 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5508 pte_t ptent, swp_entry_t *entry)
5510 struct page *page = NULL;
5511 swp_entry_t ent = pte_to_swp_entry(ptent);
5513 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
5517 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5518 * a device and because they are not accessible by CPU they are store
5519 * as special swap entry in the CPU page table.
5521 if (is_device_private_entry(ent)) {
5522 page = device_private_entry_to_page(ent);
5524 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5525 * a refcount of 1 when free (unlike normal page)
5527 if (!page_ref_add_unless(page, 1, 1))
5533 * Because lookup_swap_cache() updates some statistics counter,
5534 * we call find_get_page() with swapper_space directly.
5536 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5537 entry->val = ent.val;
5542 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5543 pte_t ptent, swp_entry_t *entry)
5549 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5550 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5552 if (!vma->vm_file) /* anonymous vma */
5554 if (!(mc.flags & MOVE_FILE))
5557 /* page is moved even if it's not RSS of this task(page-faulted). */
5558 /* shmem/tmpfs may report page out on swap: account for that too. */
5559 return find_get_incore_page(vma->vm_file->f_mapping,
5560 linear_page_index(vma, addr));
5564 * mem_cgroup_move_account - move account of the page
5566 * @compound: charge the page as compound or small page
5567 * @from: mem_cgroup which the page is moved from.
5568 * @to: mem_cgroup which the page is moved to. @from != @to.
5570 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5572 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5575 static int mem_cgroup_move_account(struct page *page,
5577 struct mem_cgroup *from,
5578 struct mem_cgroup *to)
5580 struct lruvec *from_vec, *to_vec;
5581 struct pglist_data *pgdat;
5582 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5585 VM_BUG_ON(from == to);
5586 VM_BUG_ON_PAGE(PageLRU(page), page);
5587 VM_BUG_ON(compound && !PageTransHuge(page));
5590 * Prevent mem_cgroup_migrate() from looking at
5591 * page->mem_cgroup of its source page while we change it.
5594 if (!trylock_page(page))
5598 if (page->mem_cgroup != from)
5601 pgdat = page_pgdat(page);
5602 from_vec = mem_cgroup_lruvec(from, pgdat);
5603 to_vec = mem_cgroup_lruvec(to, pgdat);
5605 lock_page_memcg(page);
5607 if (PageAnon(page)) {
5608 if (page_mapped(page)) {
5609 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5610 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5611 if (PageTransHuge(page)) {
5612 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5614 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5620 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5621 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5623 if (PageSwapBacked(page)) {
5624 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5625 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5628 if (page_mapped(page)) {
5629 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5630 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5633 if (PageDirty(page)) {
5634 struct address_space *mapping = page_mapping(page);
5636 if (mapping_can_writeback(mapping)) {
5637 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5639 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5645 if (PageWriteback(page)) {
5646 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5647 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5651 * All state has been migrated, let's switch to the new memcg.
5653 * It is safe to change page->mem_cgroup here because the page
5654 * is referenced, charged, isolated, and locked: we can't race
5655 * with (un)charging, migration, LRU putback, or anything else
5656 * that would rely on a stable page->mem_cgroup.
5658 * Note that lock_page_memcg is a memcg lock, not a page lock,
5659 * to save space. As soon as we switch page->mem_cgroup to a
5660 * new memcg that isn't locked, the above state can change
5661 * concurrently again. Make sure we're truly done with it.
5666 css_put(&from->css);
5668 page->mem_cgroup = to;
5670 __unlock_page_memcg(from);
5674 local_irq_disable();
5675 mem_cgroup_charge_statistics(to, page, nr_pages);
5676 memcg_check_events(to, page);
5677 mem_cgroup_charge_statistics(from, page, -nr_pages);
5678 memcg_check_events(from, page);
5687 * get_mctgt_type - get target type of moving charge
5688 * @vma: the vma the pte to be checked belongs
5689 * @addr: the address corresponding to the pte to be checked
5690 * @ptent: the pte to be checked
5691 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5694 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5695 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5696 * move charge. if @target is not NULL, the page is stored in target->page
5697 * with extra refcnt got(Callers should handle it).
5698 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5699 * target for charge migration. if @target is not NULL, the entry is stored
5701 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5702 * (so ZONE_DEVICE page and thus not on the lru).
5703 * For now we such page is charge like a regular page would be as for all
5704 * intent and purposes it is just special memory taking the place of a
5707 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5709 * Called with pte lock held.
5712 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5713 unsigned long addr, pte_t ptent, union mc_target *target)
5715 struct page *page = NULL;
5716 enum mc_target_type ret = MC_TARGET_NONE;
5717 swp_entry_t ent = { .val = 0 };
5719 if (pte_present(ptent))
5720 page = mc_handle_present_pte(vma, addr, ptent);
5721 else if (is_swap_pte(ptent))
5722 page = mc_handle_swap_pte(vma, ptent, &ent);
5723 else if (pte_none(ptent))
5724 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5726 if (!page && !ent.val)
5730 * Do only loose check w/o serialization.
5731 * mem_cgroup_move_account() checks the page is valid or
5732 * not under LRU exclusion.
5734 if (page->mem_cgroup == mc.from) {
5735 ret = MC_TARGET_PAGE;
5736 if (is_device_private_page(page))
5737 ret = MC_TARGET_DEVICE;
5739 target->page = page;
5741 if (!ret || !target)
5745 * There is a swap entry and a page doesn't exist or isn't charged.
5746 * But we cannot move a tail-page in a THP.
5748 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5749 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5750 ret = MC_TARGET_SWAP;
5757 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5759 * We don't consider PMD mapped swapping or file mapped pages because THP does
5760 * not support them for now.
5761 * Caller should make sure that pmd_trans_huge(pmd) is true.
5763 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5764 unsigned long addr, pmd_t pmd, union mc_target *target)
5766 struct page *page = NULL;
5767 enum mc_target_type ret = MC_TARGET_NONE;
5769 if (unlikely(is_swap_pmd(pmd))) {
5770 VM_BUG_ON(thp_migration_supported() &&
5771 !is_pmd_migration_entry(pmd));
5774 page = pmd_page(pmd);
5775 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5776 if (!(mc.flags & MOVE_ANON))
5778 if (page->mem_cgroup == mc.from) {
5779 ret = MC_TARGET_PAGE;
5782 target->page = page;
5788 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5789 unsigned long addr, pmd_t pmd, union mc_target *target)
5791 return MC_TARGET_NONE;
5795 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5796 unsigned long addr, unsigned long end,
5797 struct mm_walk *walk)
5799 struct vm_area_struct *vma = walk->vma;
5803 ptl = pmd_trans_huge_lock(pmd, vma);
5806 * Note their can not be MC_TARGET_DEVICE for now as we do not
5807 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5808 * this might change.
5810 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5811 mc.precharge += HPAGE_PMD_NR;
5816 if (pmd_trans_unstable(pmd))
5818 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5819 for (; addr != end; pte++, addr += PAGE_SIZE)
5820 if (get_mctgt_type(vma, addr, *pte, NULL))
5821 mc.precharge++; /* increment precharge temporarily */
5822 pte_unmap_unlock(pte - 1, ptl);
5828 static const struct mm_walk_ops precharge_walk_ops = {
5829 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5832 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5834 unsigned long precharge;
5837 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5838 mmap_read_unlock(mm);
5840 precharge = mc.precharge;
5846 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5848 unsigned long precharge = mem_cgroup_count_precharge(mm);
5850 VM_BUG_ON(mc.moving_task);
5851 mc.moving_task = current;
5852 return mem_cgroup_do_precharge(precharge);
5855 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5856 static void __mem_cgroup_clear_mc(void)
5858 struct mem_cgroup *from = mc.from;
5859 struct mem_cgroup *to = mc.to;
5861 /* we must uncharge all the leftover precharges from mc.to */
5863 cancel_charge(mc.to, mc.precharge);
5867 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5868 * we must uncharge here.
5870 if (mc.moved_charge) {
5871 cancel_charge(mc.from, mc.moved_charge);
5872 mc.moved_charge = 0;
5874 /* we must fixup refcnts and charges */
5875 if (mc.moved_swap) {
5876 /* uncharge swap account from the old cgroup */
5877 if (!mem_cgroup_is_root(mc.from))
5878 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5880 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5883 * we charged both to->memory and to->memsw, so we
5884 * should uncharge to->memory.
5886 if (!mem_cgroup_is_root(mc.to))
5887 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5891 memcg_oom_recover(from);
5892 memcg_oom_recover(to);
5893 wake_up_all(&mc.waitq);
5896 static void mem_cgroup_clear_mc(void)
5898 struct mm_struct *mm = mc.mm;
5901 * we must clear moving_task before waking up waiters at the end of
5904 mc.moving_task = NULL;
5905 __mem_cgroup_clear_mc();
5906 spin_lock(&mc.lock);
5910 spin_unlock(&mc.lock);
5915 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5917 struct cgroup_subsys_state *css;
5918 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5919 struct mem_cgroup *from;
5920 struct task_struct *leader, *p;
5921 struct mm_struct *mm;
5922 unsigned long move_flags;
5925 /* charge immigration isn't supported on the default hierarchy */
5926 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5930 * Multi-process migrations only happen on the default hierarchy
5931 * where charge immigration is not used. Perform charge
5932 * immigration if @tset contains a leader and whine if there are
5936 cgroup_taskset_for_each_leader(leader, css, tset) {
5939 memcg = mem_cgroup_from_css(css);
5945 * We are now commited to this value whatever it is. Changes in this
5946 * tunable will only affect upcoming migrations, not the current one.
5947 * So we need to save it, and keep it going.
5949 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5953 from = mem_cgroup_from_task(p);
5955 VM_BUG_ON(from == memcg);
5957 mm = get_task_mm(p);
5960 /* We move charges only when we move a owner of the mm */
5961 if (mm->owner == p) {
5964 VM_BUG_ON(mc.precharge);
5965 VM_BUG_ON(mc.moved_charge);
5966 VM_BUG_ON(mc.moved_swap);
5968 spin_lock(&mc.lock);
5972 mc.flags = move_flags;
5973 spin_unlock(&mc.lock);
5974 /* We set mc.moving_task later */
5976 ret = mem_cgroup_precharge_mc(mm);
5978 mem_cgroup_clear_mc();
5985 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5988 mem_cgroup_clear_mc();
5991 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5992 unsigned long addr, unsigned long end,
5993 struct mm_walk *walk)
5996 struct vm_area_struct *vma = walk->vma;
5999 enum mc_target_type target_type;
6000 union mc_target target;
6003 ptl = pmd_trans_huge_lock(pmd, vma);
6005 if (mc.precharge < HPAGE_PMD_NR) {
6009 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6010 if (target_type == MC_TARGET_PAGE) {
6012 if (!isolate_lru_page(page)) {
6013 if (!mem_cgroup_move_account(page, true,
6015 mc.precharge -= HPAGE_PMD_NR;
6016 mc.moved_charge += HPAGE_PMD_NR;
6018 putback_lru_page(page);
6021 } else if (target_type == MC_TARGET_DEVICE) {
6023 if (!mem_cgroup_move_account(page, true,
6025 mc.precharge -= HPAGE_PMD_NR;
6026 mc.moved_charge += HPAGE_PMD_NR;
6034 if (pmd_trans_unstable(pmd))
6037 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6038 for (; addr != end; addr += PAGE_SIZE) {
6039 pte_t ptent = *(pte++);
6040 bool device = false;
6046 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6047 case MC_TARGET_DEVICE:
6050 case MC_TARGET_PAGE:
6053 * We can have a part of the split pmd here. Moving it
6054 * can be done but it would be too convoluted so simply
6055 * ignore such a partial THP and keep it in original
6056 * memcg. There should be somebody mapping the head.
6058 if (PageTransCompound(page))
6060 if (!device && isolate_lru_page(page))
6062 if (!mem_cgroup_move_account(page, false,
6065 /* we uncharge from mc.from later. */
6069 putback_lru_page(page);
6070 put: /* get_mctgt_type() gets the page */
6073 case MC_TARGET_SWAP:
6075 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6077 mem_cgroup_id_get_many(mc.to, 1);
6078 /* we fixup other refcnts and charges later. */
6086 pte_unmap_unlock(pte - 1, ptl);
6091 * We have consumed all precharges we got in can_attach().
6092 * We try charge one by one, but don't do any additional
6093 * charges to mc.to if we have failed in charge once in attach()
6096 ret = mem_cgroup_do_precharge(1);
6104 static const struct mm_walk_ops charge_walk_ops = {
6105 .pmd_entry = mem_cgroup_move_charge_pte_range,
6108 static void mem_cgroup_move_charge(void)
6110 lru_add_drain_all();
6112 * Signal lock_page_memcg() to take the memcg's move_lock
6113 * while we're moving its pages to another memcg. Then wait
6114 * for already started RCU-only updates to finish.
6116 atomic_inc(&mc.from->moving_account);
6119 if (unlikely(!mmap_read_trylock(mc.mm))) {
6121 * Someone who are holding the mmap_lock might be waiting in
6122 * waitq. So we cancel all extra charges, wake up all waiters,
6123 * and retry. Because we cancel precharges, we might not be able
6124 * to move enough charges, but moving charge is a best-effort
6125 * feature anyway, so it wouldn't be a big problem.
6127 __mem_cgroup_clear_mc();
6132 * When we have consumed all precharges and failed in doing
6133 * additional charge, the page walk just aborts.
6135 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6138 mmap_read_unlock(mc.mm);
6139 atomic_dec(&mc.from->moving_account);
6142 static void mem_cgroup_move_task(void)
6145 mem_cgroup_move_charge();
6146 mem_cgroup_clear_mc();
6149 #else /* !CONFIG_MMU */
6150 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6154 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6157 static void mem_cgroup_move_task(void)
6163 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6164 * to verify whether we're attached to the default hierarchy on each mount
6167 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6170 * use_hierarchy is forced on the default hierarchy. cgroup core
6171 * guarantees that @root doesn't have any children, so turning it
6172 * on for the root memcg is enough.
6174 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6175 root_mem_cgroup->use_hierarchy = true;
6177 root_mem_cgroup->use_hierarchy = false;
6180 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6182 if (value == PAGE_COUNTER_MAX)
6183 seq_puts(m, "max\n");
6185 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6190 static u64 memory_current_read(struct cgroup_subsys_state *css,
6193 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6195 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6198 static int memory_min_show(struct seq_file *m, void *v)
6200 return seq_puts_memcg_tunable(m,
6201 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6204 static ssize_t memory_min_write(struct kernfs_open_file *of,
6205 char *buf, size_t nbytes, loff_t off)
6207 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6211 buf = strstrip(buf);
6212 err = page_counter_memparse(buf, "max", &min);
6216 page_counter_set_min(&memcg->memory, min);
6221 static int memory_low_show(struct seq_file *m, void *v)
6223 return seq_puts_memcg_tunable(m,
6224 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6227 static ssize_t memory_low_write(struct kernfs_open_file *of,
6228 char *buf, size_t nbytes, loff_t off)
6230 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6234 buf = strstrip(buf);
6235 err = page_counter_memparse(buf, "max", &low);
6239 page_counter_set_low(&memcg->memory, low);
6244 static int memory_high_show(struct seq_file *m, void *v)
6246 return seq_puts_memcg_tunable(m,
6247 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6250 static ssize_t memory_high_write(struct kernfs_open_file *of,
6251 char *buf, size_t nbytes, loff_t off)
6253 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6254 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6255 bool drained = false;
6259 buf = strstrip(buf);
6260 err = page_counter_memparse(buf, "max", &high);
6265 unsigned long nr_pages = page_counter_read(&memcg->memory);
6266 unsigned long reclaimed;
6268 if (nr_pages <= high)
6271 if (signal_pending(current))
6275 drain_all_stock(memcg);
6280 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6283 if (!reclaimed && !nr_retries--)
6287 page_counter_set_high(&memcg->memory, high);
6289 memcg_wb_domain_size_changed(memcg);
6294 static int memory_max_show(struct seq_file *m, void *v)
6296 return seq_puts_memcg_tunable(m,
6297 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6300 static ssize_t memory_max_write(struct kernfs_open_file *of,
6301 char *buf, size_t nbytes, loff_t off)
6303 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6304 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6305 bool drained = false;
6309 buf = strstrip(buf);
6310 err = page_counter_memparse(buf, "max", &max);
6314 xchg(&memcg->memory.max, max);
6317 unsigned long nr_pages = page_counter_read(&memcg->memory);
6319 if (nr_pages <= max)
6322 if (signal_pending(current))
6326 drain_all_stock(memcg);
6332 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6338 memcg_memory_event(memcg, MEMCG_OOM);
6339 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6343 memcg_wb_domain_size_changed(memcg);
6347 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6349 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6350 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6351 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6352 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6353 seq_printf(m, "oom_kill %lu\n",
6354 atomic_long_read(&events[MEMCG_OOM_KILL]));
6357 static int memory_events_show(struct seq_file *m, void *v)
6359 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6361 __memory_events_show(m, memcg->memory_events);
6365 static int memory_events_local_show(struct seq_file *m, void *v)
6367 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6369 __memory_events_show(m, memcg->memory_events_local);
6373 static int memory_stat_show(struct seq_file *m, void *v)
6375 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6378 buf = memory_stat_format(memcg);
6386 static int memory_oom_group_show(struct seq_file *m, void *v)
6388 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6390 seq_printf(m, "%d\n", memcg->oom_group);
6395 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6396 char *buf, size_t nbytes, loff_t off)
6398 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6401 buf = strstrip(buf);
6405 ret = kstrtoint(buf, 0, &oom_group);
6409 if (oom_group != 0 && oom_group != 1)
6412 memcg->oom_group = oom_group;
6417 static struct cftype memory_files[] = {
6420 .flags = CFTYPE_NOT_ON_ROOT,
6421 .read_u64 = memory_current_read,
6425 .flags = CFTYPE_NOT_ON_ROOT,
6426 .seq_show = memory_min_show,
6427 .write = memory_min_write,
6431 .flags = CFTYPE_NOT_ON_ROOT,
6432 .seq_show = memory_low_show,
6433 .write = memory_low_write,
6437 .flags = CFTYPE_NOT_ON_ROOT,
6438 .seq_show = memory_high_show,
6439 .write = memory_high_write,
6443 .flags = CFTYPE_NOT_ON_ROOT,
6444 .seq_show = memory_max_show,
6445 .write = memory_max_write,
6449 .flags = CFTYPE_NOT_ON_ROOT,
6450 .file_offset = offsetof(struct mem_cgroup, events_file),
6451 .seq_show = memory_events_show,
6454 .name = "events.local",
6455 .flags = CFTYPE_NOT_ON_ROOT,
6456 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6457 .seq_show = memory_events_local_show,
6461 .seq_show = memory_stat_show,
6464 .name = "oom.group",
6465 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6466 .seq_show = memory_oom_group_show,
6467 .write = memory_oom_group_write,
6472 struct cgroup_subsys memory_cgrp_subsys = {
6473 .css_alloc = mem_cgroup_css_alloc,
6474 .css_online = mem_cgroup_css_online,
6475 .css_offline = mem_cgroup_css_offline,
6476 .css_released = mem_cgroup_css_released,
6477 .css_free = mem_cgroup_css_free,
6478 .css_reset = mem_cgroup_css_reset,
6479 .can_attach = mem_cgroup_can_attach,
6480 .cancel_attach = mem_cgroup_cancel_attach,
6481 .post_attach = mem_cgroup_move_task,
6482 .bind = mem_cgroup_bind,
6483 .dfl_cftypes = memory_files,
6484 .legacy_cftypes = mem_cgroup_legacy_files,
6489 * This function calculates an individual cgroup's effective
6490 * protection which is derived from its own memory.min/low, its
6491 * parent's and siblings' settings, as well as the actual memory
6492 * distribution in the tree.
6494 * The following rules apply to the effective protection values:
6496 * 1. At the first level of reclaim, effective protection is equal to
6497 * the declared protection in memory.min and memory.low.
6499 * 2. To enable safe delegation of the protection configuration, at
6500 * subsequent levels the effective protection is capped to the
6501 * parent's effective protection.
6503 * 3. To make complex and dynamic subtrees easier to configure, the
6504 * user is allowed to overcommit the declared protection at a given
6505 * level. If that is the case, the parent's effective protection is
6506 * distributed to the children in proportion to how much protection
6507 * they have declared and how much of it they are utilizing.
6509 * This makes distribution proportional, but also work-conserving:
6510 * if one cgroup claims much more protection than it uses memory,
6511 * the unused remainder is available to its siblings.
6513 * 4. Conversely, when the declared protection is undercommitted at a
6514 * given level, the distribution of the larger parental protection
6515 * budget is NOT proportional. A cgroup's protection from a sibling
6516 * is capped to its own memory.min/low setting.
6518 * 5. However, to allow protecting recursive subtrees from each other
6519 * without having to declare each individual cgroup's fixed share
6520 * of the ancestor's claim to protection, any unutilized -
6521 * "floating" - protection from up the tree is distributed in
6522 * proportion to each cgroup's *usage*. This makes the protection
6523 * neutral wrt sibling cgroups and lets them compete freely over
6524 * the shared parental protection budget, but it protects the
6525 * subtree as a whole from neighboring subtrees.
6527 * Note that 4. and 5. are not in conflict: 4. is about protecting
6528 * against immediate siblings whereas 5. is about protecting against
6529 * neighboring subtrees.
6531 static unsigned long effective_protection(unsigned long usage,
6532 unsigned long parent_usage,
6533 unsigned long setting,
6534 unsigned long parent_effective,
6535 unsigned long siblings_protected)
6537 unsigned long protected;
6540 protected = min(usage, setting);
6542 * If all cgroups at this level combined claim and use more
6543 * protection then what the parent affords them, distribute
6544 * shares in proportion to utilization.
6546 * We are using actual utilization rather than the statically
6547 * claimed protection in order to be work-conserving: claimed
6548 * but unused protection is available to siblings that would
6549 * otherwise get a smaller chunk than what they claimed.
6551 if (siblings_protected > parent_effective)
6552 return protected * parent_effective / siblings_protected;
6555 * Ok, utilized protection of all children is within what the
6556 * parent affords them, so we know whatever this child claims
6557 * and utilizes is effectively protected.
6559 * If there is unprotected usage beyond this value, reclaim
6560 * will apply pressure in proportion to that amount.
6562 * If there is unutilized protection, the cgroup will be fully
6563 * shielded from reclaim, but we do return a smaller value for
6564 * protection than what the group could enjoy in theory. This
6565 * is okay. With the overcommit distribution above, effective
6566 * protection is always dependent on how memory is actually
6567 * consumed among the siblings anyway.
6572 * If the children aren't claiming (all of) the protection
6573 * afforded to them by the parent, distribute the remainder in
6574 * proportion to the (unprotected) memory of each cgroup. That
6575 * way, cgroups that aren't explicitly prioritized wrt each
6576 * other compete freely over the allowance, but they are
6577 * collectively protected from neighboring trees.
6579 * We're using unprotected memory for the weight so that if
6580 * some cgroups DO claim explicit protection, we don't protect
6581 * the same bytes twice.
6583 * Check both usage and parent_usage against the respective
6584 * protected values. One should imply the other, but they
6585 * aren't read atomically - make sure the division is sane.
6587 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6589 if (parent_effective > siblings_protected &&
6590 parent_usage > siblings_protected &&
6591 usage > protected) {
6592 unsigned long unclaimed;
6594 unclaimed = parent_effective - siblings_protected;
6595 unclaimed *= usage - protected;
6596 unclaimed /= parent_usage - siblings_protected;
6605 * mem_cgroup_protected - check if memory consumption is in the normal range
6606 * @root: the top ancestor of the sub-tree being checked
6607 * @memcg: the memory cgroup to check
6609 * WARNING: This function is not stateless! It can only be used as part
6610 * of a top-down tree iteration, not for isolated queries.
6612 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6613 struct mem_cgroup *memcg)
6615 unsigned long usage, parent_usage;
6616 struct mem_cgroup *parent;
6618 if (mem_cgroup_disabled())
6622 root = root_mem_cgroup;
6625 * Effective values of the reclaim targets are ignored so they
6626 * can be stale. Have a look at mem_cgroup_protection for more
6628 * TODO: calculation should be more robust so that we do not need
6629 * that special casing.
6634 usage = page_counter_read(&memcg->memory);
6638 parent = parent_mem_cgroup(memcg);
6639 /* No parent means a non-hierarchical mode on v1 memcg */
6643 if (parent == root) {
6644 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6645 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6649 parent_usage = page_counter_read(&parent->memory);
6651 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6652 READ_ONCE(memcg->memory.min),
6653 READ_ONCE(parent->memory.emin),
6654 atomic_long_read(&parent->memory.children_min_usage)));
6656 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6657 READ_ONCE(memcg->memory.low),
6658 READ_ONCE(parent->memory.elow),
6659 atomic_long_read(&parent->memory.children_low_usage)));
6663 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6664 * @page: page to charge
6665 * @mm: mm context of the victim
6666 * @gfp_mask: reclaim mode
6668 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6669 * pages according to @gfp_mask if necessary.
6671 * Returns 0 on success. Otherwise, an error code is returned.
6673 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6675 unsigned int nr_pages = thp_nr_pages(page);
6676 struct mem_cgroup *memcg = NULL;
6679 if (mem_cgroup_disabled())
6682 if (PageSwapCache(page)) {
6683 swp_entry_t ent = { .val = page_private(page), };
6687 * Every swap fault against a single page tries to charge the
6688 * page, bail as early as possible. shmem_unuse() encounters
6689 * already charged pages, too. page->mem_cgroup is protected
6690 * by the page lock, which serializes swap cache removal, which
6691 * in turn serializes uncharging.
6693 VM_BUG_ON_PAGE(!PageLocked(page), page);
6694 if (compound_head(page)->mem_cgroup)
6697 id = lookup_swap_cgroup_id(ent);
6699 memcg = mem_cgroup_from_id(id);
6700 if (memcg && !css_tryget_online(&memcg->css))
6706 memcg = get_mem_cgroup_from_mm(mm);
6708 ret = try_charge(memcg, gfp_mask, nr_pages);
6712 css_get(&memcg->css);
6713 commit_charge(page, memcg);
6715 local_irq_disable();
6716 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6717 memcg_check_events(memcg, page);
6720 if (PageSwapCache(page)) {
6721 swp_entry_t entry = { .val = page_private(page) };
6723 * The swap entry might not get freed for a long time,
6724 * let's not wait for it. The page already received a
6725 * memory+swap charge, drop the swap entry duplicate.
6727 mem_cgroup_uncharge_swap(entry, nr_pages);
6731 css_put(&memcg->css);
6736 struct uncharge_gather {
6737 struct mem_cgroup *memcg;
6738 unsigned long nr_pages;
6739 unsigned long pgpgout;
6740 unsigned long nr_kmem;
6741 struct page *dummy_page;
6744 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6746 memset(ug, 0, sizeof(*ug));
6749 static void uncharge_batch(const struct uncharge_gather *ug)
6751 unsigned long flags;
6753 if (!mem_cgroup_is_root(ug->memcg)) {
6754 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6755 if (do_memsw_account())
6756 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6757 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6758 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6759 memcg_oom_recover(ug->memcg);
6762 local_irq_save(flags);
6763 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6764 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6765 memcg_check_events(ug->memcg, ug->dummy_page);
6766 local_irq_restore(flags);
6768 /* drop reference from uncharge_page */
6769 css_put(&ug->memcg->css);
6772 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6774 unsigned long nr_pages;
6776 VM_BUG_ON_PAGE(PageLRU(page), page);
6778 if (!page->mem_cgroup)
6782 * Nobody should be changing or seriously looking at
6783 * page->mem_cgroup at this point, we have fully
6784 * exclusive access to the page.
6787 if (ug->memcg != page->mem_cgroup) {
6790 uncharge_gather_clear(ug);
6792 ug->memcg = page->mem_cgroup;
6794 /* pairs with css_put in uncharge_batch */
6795 css_get(&ug->memcg->css);
6798 nr_pages = compound_nr(page);
6799 ug->nr_pages += nr_pages;
6801 if (!PageKmemcg(page)) {
6804 ug->nr_kmem += nr_pages;
6805 __ClearPageKmemcg(page);
6808 ug->dummy_page = page;
6809 page->mem_cgroup = NULL;
6810 css_put(&ug->memcg->css);
6813 static void uncharge_list(struct list_head *page_list)
6815 struct uncharge_gather ug;
6816 struct list_head *next;
6818 uncharge_gather_clear(&ug);
6821 * Note that the list can be a single page->lru; hence the
6822 * do-while loop instead of a simple list_for_each_entry().
6824 next = page_list->next;
6828 page = list_entry(next, struct page, lru);
6829 next = page->lru.next;
6831 uncharge_page(page, &ug);
6832 } while (next != page_list);
6835 uncharge_batch(&ug);
6839 * mem_cgroup_uncharge - uncharge a page
6840 * @page: page to uncharge
6842 * Uncharge a page previously charged with mem_cgroup_charge().
6844 void mem_cgroup_uncharge(struct page *page)
6846 struct uncharge_gather ug;
6848 if (mem_cgroup_disabled())
6851 /* Don't touch page->lru of any random page, pre-check: */
6852 if (!page->mem_cgroup)
6855 uncharge_gather_clear(&ug);
6856 uncharge_page(page, &ug);
6857 uncharge_batch(&ug);
6861 * mem_cgroup_uncharge_list - uncharge a list of page
6862 * @page_list: list of pages to uncharge
6864 * Uncharge a list of pages previously charged with
6865 * mem_cgroup_charge().
6867 void mem_cgroup_uncharge_list(struct list_head *page_list)
6869 if (mem_cgroup_disabled())
6872 if (!list_empty(page_list))
6873 uncharge_list(page_list);
6877 * mem_cgroup_migrate - charge a page's replacement
6878 * @oldpage: currently circulating page
6879 * @newpage: replacement page
6881 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6882 * be uncharged upon free.
6884 * Both pages must be locked, @newpage->mapping must be set up.
6886 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6888 struct mem_cgroup *memcg;
6889 unsigned int nr_pages;
6890 unsigned long flags;
6892 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6893 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6894 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6895 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6898 if (mem_cgroup_disabled())
6901 /* Page cache replacement: new page already charged? */
6902 if (newpage->mem_cgroup)
6905 /* Swapcache readahead pages can get replaced before being charged */
6906 memcg = oldpage->mem_cgroup;
6910 /* Force-charge the new page. The old one will be freed soon */
6911 nr_pages = thp_nr_pages(newpage);
6913 page_counter_charge(&memcg->memory, nr_pages);
6914 if (do_memsw_account())
6915 page_counter_charge(&memcg->memsw, nr_pages);
6917 css_get(&memcg->css);
6918 commit_charge(newpage, memcg);
6920 local_irq_save(flags);
6921 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6922 memcg_check_events(memcg, newpage);
6923 local_irq_restore(flags);
6926 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6927 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6929 void mem_cgroup_sk_alloc(struct sock *sk)
6931 struct mem_cgroup *memcg;
6933 if (!mem_cgroup_sockets_enabled)
6936 /* Do not associate the sock with unrelated interrupted task's memcg. */
6941 memcg = mem_cgroup_from_task(current);
6942 if (memcg == root_mem_cgroup)
6944 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6946 if (css_tryget(&memcg->css))
6947 sk->sk_memcg = memcg;
6952 void mem_cgroup_sk_free(struct sock *sk)
6955 css_put(&sk->sk_memcg->css);
6959 * mem_cgroup_charge_skmem - charge socket memory
6960 * @memcg: memcg to charge
6961 * @nr_pages: number of pages to charge
6963 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6964 * @memcg's configured limit, %false if the charge had to be forced.
6966 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6968 gfp_t gfp_mask = GFP_KERNEL;
6970 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6971 struct page_counter *fail;
6973 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6974 memcg->tcpmem_pressure = 0;
6977 page_counter_charge(&memcg->tcpmem, nr_pages);
6978 memcg->tcpmem_pressure = 1;
6982 /* Don't block in the packet receive path */
6984 gfp_mask = GFP_NOWAIT;
6986 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6988 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6991 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6996 * mem_cgroup_uncharge_skmem - uncharge socket memory
6997 * @memcg: memcg to uncharge
6998 * @nr_pages: number of pages to uncharge
7000 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7002 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7003 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7007 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7009 refill_stock(memcg, nr_pages);
7012 static int __init cgroup_memory(char *s)
7016 while ((token = strsep(&s, ",")) != NULL) {
7019 if (!strcmp(token, "nosocket"))
7020 cgroup_memory_nosocket = true;
7021 if (!strcmp(token, "nokmem"))
7022 cgroup_memory_nokmem = true;
7026 __setup("cgroup.memory=", cgroup_memory);
7029 * subsys_initcall() for memory controller.
7031 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7032 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7033 * basically everything that doesn't depend on a specific mem_cgroup structure
7034 * should be initialized from here.
7036 static int __init mem_cgroup_init(void)
7040 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7041 memcg_hotplug_cpu_dead);
7043 for_each_possible_cpu(cpu)
7044 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7047 for_each_node(node) {
7048 struct mem_cgroup_tree_per_node *rtpn;
7050 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7051 node_online(node) ? node : NUMA_NO_NODE);
7053 rtpn->rb_root = RB_ROOT;
7054 rtpn->rb_rightmost = NULL;
7055 spin_lock_init(&rtpn->lock);
7056 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7061 subsys_initcall(mem_cgroup_init);
7063 #ifdef CONFIG_MEMCG_SWAP
7064 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7066 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7068 * The root cgroup cannot be destroyed, so it's refcount must
7071 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7075 memcg = parent_mem_cgroup(memcg);
7077 memcg = root_mem_cgroup;
7083 * mem_cgroup_swapout - transfer a memsw charge to swap
7084 * @page: page whose memsw charge to transfer
7085 * @entry: swap entry to move the charge to
7087 * Transfer the memsw charge of @page to @entry.
7089 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7091 struct mem_cgroup *memcg, *swap_memcg;
7092 unsigned int nr_entries;
7093 unsigned short oldid;
7095 VM_BUG_ON_PAGE(PageLRU(page), page);
7096 VM_BUG_ON_PAGE(page_count(page), page);
7098 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7101 memcg = page->mem_cgroup;
7103 /* Readahead page, never charged */
7108 * In case the memcg owning these pages has been offlined and doesn't
7109 * have an ID allocated to it anymore, charge the closest online
7110 * ancestor for the swap instead and transfer the memory+swap charge.
7112 swap_memcg = mem_cgroup_id_get_online(memcg);
7113 nr_entries = thp_nr_pages(page);
7114 /* Get references for the tail pages, too */
7116 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7117 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7119 VM_BUG_ON_PAGE(oldid, page);
7120 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7122 page->mem_cgroup = NULL;
7124 if (!mem_cgroup_is_root(memcg))
7125 page_counter_uncharge(&memcg->memory, nr_entries);
7127 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7128 if (!mem_cgroup_is_root(swap_memcg))
7129 page_counter_charge(&swap_memcg->memsw, nr_entries);
7130 page_counter_uncharge(&memcg->memsw, nr_entries);
7134 * Interrupts should be disabled here because the caller holds the
7135 * i_pages lock which is taken with interrupts-off. It is
7136 * important here to have the interrupts disabled because it is the
7137 * only synchronisation we have for updating the per-CPU variables.
7139 VM_BUG_ON(!irqs_disabled());
7140 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7141 memcg_check_events(memcg, page);
7143 css_put(&memcg->css);
7147 * mem_cgroup_try_charge_swap - try charging swap space for a page
7148 * @page: page being added to swap
7149 * @entry: swap entry to charge
7151 * Try to charge @page's memcg for the swap space at @entry.
7153 * Returns 0 on success, -ENOMEM on failure.
7155 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7157 unsigned int nr_pages = thp_nr_pages(page);
7158 struct page_counter *counter;
7159 struct mem_cgroup *memcg;
7160 unsigned short oldid;
7162 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7165 memcg = page->mem_cgroup;
7167 /* Readahead page, never charged */
7172 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7176 memcg = mem_cgroup_id_get_online(memcg);
7178 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7179 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7180 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7181 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7182 mem_cgroup_id_put(memcg);
7186 /* Get references for the tail pages, too */
7188 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7189 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7190 VM_BUG_ON_PAGE(oldid, page);
7191 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7197 * mem_cgroup_uncharge_swap - uncharge swap space
7198 * @entry: swap entry to uncharge
7199 * @nr_pages: the amount of swap space to uncharge
7201 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7203 struct mem_cgroup *memcg;
7206 id = swap_cgroup_record(entry, 0, nr_pages);
7208 memcg = mem_cgroup_from_id(id);
7210 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7211 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7212 page_counter_uncharge(&memcg->swap, nr_pages);
7214 page_counter_uncharge(&memcg->memsw, nr_pages);
7216 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7217 mem_cgroup_id_put_many(memcg, nr_pages);
7222 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7224 long nr_swap_pages = get_nr_swap_pages();
7226 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7227 return nr_swap_pages;
7228 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7229 nr_swap_pages = min_t(long, nr_swap_pages,
7230 READ_ONCE(memcg->swap.max) -
7231 page_counter_read(&memcg->swap));
7232 return nr_swap_pages;
7235 bool mem_cgroup_swap_full(struct page *page)
7237 struct mem_cgroup *memcg;
7239 VM_BUG_ON_PAGE(!PageLocked(page), page);
7243 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7246 memcg = page->mem_cgroup;
7250 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7251 unsigned long usage = page_counter_read(&memcg->swap);
7253 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7254 usage * 2 >= READ_ONCE(memcg->swap.max))
7261 static int __init setup_swap_account(char *s)
7263 if (!strcmp(s, "1"))
7264 cgroup_memory_noswap = 0;
7265 else if (!strcmp(s, "0"))
7266 cgroup_memory_noswap = 1;
7269 __setup("swapaccount=", setup_swap_account);
7271 static u64 swap_current_read(struct cgroup_subsys_state *css,
7274 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7276 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7279 static int swap_high_show(struct seq_file *m, void *v)
7281 return seq_puts_memcg_tunable(m,
7282 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7285 static ssize_t swap_high_write(struct kernfs_open_file *of,
7286 char *buf, size_t nbytes, loff_t off)
7288 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7292 buf = strstrip(buf);
7293 err = page_counter_memparse(buf, "max", &high);
7297 page_counter_set_high(&memcg->swap, high);
7302 static int swap_max_show(struct seq_file *m, void *v)
7304 return seq_puts_memcg_tunable(m,
7305 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7308 static ssize_t swap_max_write(struct kernfs_open_file *of,
7309 char *buf, size_t nbytes, loff_t off)
7311 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7315 buf = strstrip(buf);
7316 err = page_counter_memparse(buf, "max", &max);
7320 xchg(&memcg->swap.max, max);
7325 static int swap_events_show(struct seq_file *m, void *v)
7327 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7329 seq_printf(m, "high %lu\n",
7330 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7331 seq_printf(m, "max %lu\n",
7332 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7333 seq_printf(m, "fail %lu\n",
7334 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7339 static struct cftype swap_files[] = {
7341 .name = "swap.current",
7342 .flags = CFTYPE_NOT_ON_ROOT,
7343 .read_u64 = swap_current_read,
7346 .name = "swap.high",
7347 .flags = CFTYPE_NOT_ON_ROOT,
7348 .seq_show = swap_high_show,
7349 .write = swap_high_write,
7353 .flags = CFTYPE_NOT_ON_ROOT,
7354 .seq_show = swap_max_show,
7355 .write = swap_max_write,
7358 .name = "swap.events",
7359 .flags = CFTYPE_NOT_ON_ROOT,
7360 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7361 .seq_show = swap_events_show,
7366 static struct cftype memsw_files[] = {
7368 .name = "memsw.usage_in_bytes",
7369 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7370 .read_u64 = mem_cgroup_read_u64,
7373 .name = "memsw.max_usage_in_bytes",
7374 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7375 .write = mem_cgroup_reset,
7376 .read_u64 = mem_cgroup_read_u64,
7379 .name = "memsw.limit_in_bytes",
7380 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7381 .write = mem_cgroup_write,
7382 .read_u64 = mem_cgroup_read_u64,
7385 .name = "memsw.failcnt",
7386 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7387 .write = mem_cgroup_reset,
7388 .read_u64 = mem_cgroup_read_u64,
7390 { }, /* terminate */
7394 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7395 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7396 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7397 * boot parameter. This may result in premature OOPS inside
7398 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7400 static int __init mem_cgroup_swap_init(void)
7402 /* No memory control -> no swap control */
7403 if (mem_cgroup_disabled())
7404 cgroup_memory_noswap = true;
7406 if (cgroup_memory_noswap)
7409 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7410 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7414 core_initcall(mem_cgroup_swap_init);
7416 #endif /* CONFIG_MEMCG_SWAP */