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 #define MEM_CGROUP_RECLAIM_RETRIES 5
78 /* Socket memory accounting disabled? */
79 static bool cgroup_memory_nosocket;
81 /* Kernel memory accounting disabled? */
82 static bool cgroup_memory_nokmem;
84 /* Whether the swap controller is active */
85 #ifdef CONFIG_MEMCG_SWAP
86 bool cgroup_memory_noswap __read_mostly;
88 #define cgroup_memory_noswap 1
91 #ifdef CONFIG_CGROUP_WRITEBACK
92 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
101 #define THRESHOLDS_EVENTS_TARGET 128
102 #define SOFTLIMIT_EVENTS_TARGET 1024
105 * Cgroups above their limits are maintained in a RB-Tree, independent of
106 * their hierarchy representation
109 struct mem_cgroup_tree_per_node {
110 struct rb_root rb_root;
111 struct rb_node *rb_rightmost;
115 struct mem_cgroup_tree {
116 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
119 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
122 struct mem_cgroup_eventfd_list {
123 struct list_head list;
124 struct eventfd_ctx *eventfd;
128 * cgroup_event represents events which userspace want to receive.
130 struct mem_cgroup_event {
132 * memcg which the event belongs to.
134 struct mem_cgroup *memcg;
136 * eventfd to signal userspace about the event.
138 struct eventfd_ctx *eventfd;
140 * Each of these stored in a list by the cgroup.
142 struct list_head list;
144 * register_event() callback will be used to add new userspace
145 * waiter for changes related to this event. Use eventfd_signal()
146 * on eventfd to send notification to userspace.
148 int (*register_event)(struct mem_cgroup *memcg,
149 struct eventfd_ctx *eventfd, const char *args);
151 * unregister_event() callback will be called when userspace closes
152 * the eventfd or on cgroup removing. This callback must be set,
153 * if you want provide notification functionality.
155 void (*unregister_event)(struct mem_cgroup *memcg,
156 struct eventfd_ctx *eventfd);
158 * All fields below needed to unregister event when
159 * userspace closes eventfd.
162 wait_queue_head_t *wqh;
163 wait_queue_entry_t wait;
164 struct work_struct remove;
167 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
168 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
170 /* Stuffs for move charges at task migration. */
172 * Types of charges to be moved.
174 #define MOVE_ANON 0x1U
175 #define MOVE_FILE 0x2U
176 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
178 /* "mc" and its members are protected by cgroup_mutex */
179 static struct move_charge_struct {
180 spinlock_t lock; /* for from, to */
181 struct mm_struct *mm;
182 struct mem_cgroup *from;
183 struct mem_cgroup *to;
185 unsigned long precharge;
186 unsigned long moved_charge;
187 unsigned long moved_swap;
188 struct task_struct *moving_task; /* a task moving charges */
189 wait_queue_head_t waitq; /* a waitq for other context */
191 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
192 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
196 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
197 * limit reclaim to prevent infinite loops, if they ever occur.
199 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
200 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
203 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
204 MEM_CGROUP_CHARGE_TYPE_ANON,
205 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
206 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
210 /* for encoding cft->private value on file */
219 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
220 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
221 #define MEMFILE_ATTR(val) ((val) & 0xffff)
222 /* Used for OOM nofiier */
223 #define OOM_CONTROL (0)
226 * Iteration constructs for visiting all cgroups (under a tree). If
227 * loops are exited prematurely (break), mem_cgroup_iter_break() must
228 * be used for reference counting.
230 #define for_each_mem_cgroup_tree(iter, root) \
231 for (iter = mem_cgroup_iter(root, NULL, NULL); \
233 iter = mem_cgroup_iter(root, iter, NULL))
235 #define for_each_mem_cgroup(iter) \
236 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
238 iter = mem_cgroup_iter(NULL, iter, NULL))
240 static inline bool should_force_charge(void)
242 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
243 (current->flags & PF_EXITING);
246 /* Some nice accessors for the vmpressure. */
247 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
250 memcg = root_mem_cgroup;
251 return &memcg->vmpressure;
254 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
256 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
259 #ifdef CONFIG_MEMCG_KMEM
260 extern spinlock_t css_set_lock;
262 static void obj_cgroup_release(struct percpu_ref *ref)
264 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
265 struct mem_cgroup *memcg;
266 unsigned int nr_bytes;
267 unsigned int nr_pages;
271 * At this point all allocated objects are freed, and
272 * objcg->nr_charged_bytes can't have an arbitrary byte value.
273 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
275 * The following sequence can lead to it:
276 * 1) CPU0: objcg == stock->cached_objcg
277 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
278 * PAGE_SIZE bytes are charged
279 * 3) CPU1: a process from another memcg is allocating something,
280 * the stock if flushed,
281 * objcg->nr_charged_bytes = PAGE_SIZE - 92
282 * 5) CPU0: we do release this object,
283 * 92 bytes are added to stock->nr_bytes
284 * 6) CPU0: stock is flushed,
285 * 92 bytes are added to objcg->nr_charged_bytes
287 * In the result, nr_charged_bytes == PAGE_SIZE.
288 * This page will be uncharged in obj_cgroup_release().
290 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
291 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
292 nr_pages = nr_bytes >> PAGE_SHIFT;
294 spin_lock_irqsave(&css_set_lock, flags);
295 memcg = obj_cgroup_memcg(objcg);
297 __memcg_kmem_uncharge(memcg, nr_pages);
298 list_del(&objcg->list);
299 mem_cgroup_put(memcg);
300 spin_unlock_irqrestore(&css_set_lock, flags);
302 percpu_ref_exit(ref);
303 kfree_rcu(objcg, rcu);
306 static struct obj_cgroup *obj_cgroup_alloc(void)
308 struct obj_cgroup *objcg;
311 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
315 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
321 INIT_LIST_HEAD(&objcg->list);
325 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
326 struct mem_cgroup *parent)
328 struct obj_cgroup *objcg, *iter;
330 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
332 spin_lock_irq(&css_set_lock);
334 /* Move active objcg to the parent's list */
335 xchg(&objcg->memcg, parent);
336 css_get(&parent->css);
337 list_add(&objcg->list, &parent->objcg_list);
339 /* Move already reparented objcgs to the parent's list */
340 list_for_each_entry(iter, &memcg->objcg_list, list) {
341 css_get(&parent->css);
342 xchg(&iter->memcg, parent);
343 css_put(&memcg->css);
345 list_splice(&memcg->objcg_list, &parent->objcg_list);
347 spin_unlock_irq(&css_set_lock);
349 percpu_ref_kill(&objcg->refcnt);
353 * This will be used as a shrinker list's index.
354 * The main reason for not using cgroup id for this:
355 * this works better in sparse environments, where we have a lot of memcgs,
356 * but only a few kmem-limited. Or also, if we have, for instance, 200
357 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
358 * 200 entry array for that.
360 * The current size of the caches array is stored in memcg_nr_cache_ids. It
361 * will double each time we have to increase it.
363 static DEFINE_IDA(memcg_cache_ida);
364 int memcg_nr_cache_ids;
366 /* Protects memcg_nr_cache_ids */
367 static DECLARE_RWSEM(memcg_cache_ids_sem);
369 void memcg_get_cache_ids(void)
371 down_read(&memcg_cache_ids_sem);
374 void memcg_put_cache_ids(void)
376 up_read(&memcg_cache_ids_sem);
380 * MIN_SIZE is different than 1, because we would like to avoid going through
381 * the alloc/free process all the time. In a small machine, 4 kmem-limited
382 * cgroups is a reasonable guess. In the future, it could be a parameter or
383 * tunable, but that is strictly not necessary.
385 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
386 * this constant directly from cgroup, but it is understandable that this is
387 * better kept as an internal representation in cgroup.c. In any case, the
388 * cgrp_id space is not getting any smaller, and we don't have to necessarily
389 * increase ours as well if it increases.
391 #define MEMCG_CACHES_MIN_SIZE 4
392 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
395 * A lot of the calls to the cache allocation functions are expected to be
396 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
397 * conditional to this static branch, we'll have to allow modules that does
398 * kmem_cache_alloc and the such to see this symbol as well
400 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
401 EXPORT_SYMBOL(memcg_kmem_enabled_key);
404 static int memcg_shrinker_map_size;
405 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
407 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
409 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
412 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
413 int size, int old_size)
415 struct memcg_shrinker_map *new, *old;
418 lockdep_assert_held(&memcg_shrinker_map_mutex);
421 old = rcu_dereference_protected(
422 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
423 /* Not yet online memcg */
427 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
431 /* Set all old bits, clear all new bits */
432 memset(new->map, (int)0xff, old_size);
433 memset((void *)new->map + old_size, 0, size - old_size);
435 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
436 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
442 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
444 struct mem_cgroup_per_node *pn;
445 struct memcg_shrinker_map *map;
448 if (mem_cgroup_is_root(memcg))
452 pn = mem_cgroup_nodeinfo(memcg, nid);
453 map = rcu_dereference_protected(pn->shrinker_map, true);
456 rcu_assign_pointer(pn->shrinker_map, NULL);
460 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
462 struct memcg_shrinker_map *map;
463 int nid, size, ret = 0;
465 if (mem_cgroup_is_root(memcg))
468 mutex_lock(&memcg_shrinker_map_mutex);
469 size = memcg_shrinker_map_size;
471 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
473 memcg_free_shrinker_maps(memcg);
477 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
479 mutex_unlock(&memcg_shrinker_map_mutex);
484 int memcg_expand_shrinker_maps(int new_id)
486 int size, old_size, ret = 0;
487 struct mem_cgroup *memcg;
489 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
490 old_size = memcg_shrinker_map_size;
491 if (size <= old_size)
494 mutex_lock(&memcg_shrinker_map_mutex);
495 if (!root_mem_cgroup)
498 for_each_mem_cgroup(memcg) {
499 if (mem_cgroup_is_root(memcg))
501 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
503 mem_cgroup_iter_break(NULL, memcg);
509 memcg_shrinker_map_size = size;
510 mutex_unlock(&memcg_shrinker_map_mutex);
514 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
516 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
517 struct memcg_shrinker_map *map;
520 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
521 /* Pairs with smp mb in shrink_slab() */
522 smp_mb__before_atomic();
523 set_bit(shrinker_id, map->map);
529 * mem_cgroup_css_from_page - css of the memcg associated with a page
530 * @page: page of interest
532 * If memcg is bound to the default hierarchy, css of the memcg associated
533 * with @page is returned. The returned css remains associated with @page
534 * until it is released.
536 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
539 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
541 struct mem_cgroup *memcg;
543 memcg = page->mem_cgroup;
545 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
546 memcg = root_mem_cgroup;
552 * page_cgroup_ino - return inode number of the memcg a page is charged to
555 * Look up the closest online ancestor of the memory cgroup @page is charged to
556 * and return its inode number or 0 if @page is not charged to any cgroup. It
557 * is safe to call this function without holding a reference to @page.
559 * Note, this function is inherently racy, because there is nothing to prevent
560 * the cgroup inode from getting torn down and potentially reallocated a moment
561 * after page_cgroup_ino() returns, so it only should be used by callers that
562 * do not care (such as procfs interfaces).
564 ino_t page_cgroup_ino(struct page *page)
566 struct mem_cgroup *memcg;
567 unsigned long ino = 0;
570 memcg = page->mem_cgroup;
573 * The lowest bit set means that memcg isn't a valid
574 * memcg pointer, but a obj_cgroups pointer.
575 * In this case the page is shared and doesn't belong
576 * to any specific memory cgroup.
578 if ((unsigned long) memcg & 0x1UL)
581 while (memcg && !(memcg->css.flags & CSS_ONLINE))
582 memcg = parent_mem_cgroup(memcg);
584 ino = cgroup_ino(memcg->css.cgroup);
589 static struct mem_cgroup_per_node *
590 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
592 int nid = page_to_nid(page);
594 return memcg->nodeinfo[nid];
597 static struct mem_cgroup_tree_per_node *
598 soft_limit_tree_node(int nid)
600 return soft_limit_tree.rb_tree_per_node[nid];
603 static struct mem_cgroup_tree_per_node *
604 soft_limit_tree_from_page(struct page *page)
606 int nid = page_to_nid(page);
608 return soft_limit_tree.rb_tree_per_node[nid];
611 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
612 struct mem_cgroup_tree_per_node *mctz,
613 unsigned long new_usage_in_excess)
615 struct rb_node **p = &mctz->rb_root.rb_node;
616 struct rb_node *parent = NULL;
617 struct mem_cgroup_per_node *mz_node;
618 bool rightmost = true;
623 mz->usage_in_excess = new_usage_in_excess;
624 if (!mz->usage_in_excess)
628 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
630 if (mz->usage_in_excess < mz_node->usage_in_excess) {
636 * We can't avoid mem cgroups that are over their soft
637 * limit by the same amount
639 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
644 mctz->rb_rightmost = &mz->tree_node;
646 rb_link_node(&mz->tree_node, parent, p);
647 rb_insert_color(&mz->tree_node, &mctz->rb_root);
651 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
652 struct mem_cgroup_tree_per_node *mctz)
657 if (&mz->tree_node == mctz->rb_rightmost)
658 mctz->rb_rightmost = rb_prev(&mz->tree_node);
660 rb_erase(&mz->tree_node, &mctz->rb_root);
664 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
665 struct mem_cgroup_tree_per_node *mctz)
669 spin_lock_irqsave(&mctz->lock, flags);
670 __mem_cgroup_remove_exceeded(mz, mctz);
671 spin_unlock_irqrestore(&mctz->lock, flags);
674 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
676 unsigned long nr_pages = page_counter_read(&memcg->memory);
677 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
678 unsigned long excess = 0;
680 if (nr_pages > soft_limit)
681 excess = nr_pages - soft_limit;
686 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
688 unsigned long excess;
689 struct mem_cgroup_per_node *mz;
690 struct mem_cgroup_tree_per_node *mctz;
692 mctz = soft_limit_tree_from_page(page);
696 * Necessary to update all ancestors when hierarchy is used.
697 * because their event counter is not touched.
699 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
700 mz = mem_cgroup_page_nodeinfo(memcg, page);
701 excess = soft_limit_excess(memcg);
703 * We have to update the tree if mz is on RB-tree or
704 * mem is over its softlimit.
706 if (excess || mz->on_tree) {
709 spin_lock_irqsave(&mctz->lock, flags);
710 /* if on-tree, remove it */
712 __mem_cgroup_remove_exceeded(mz, mctz);
714 * Insert again. mz->usage_in_excess will be updated.
715 * If excess is 0, no tree ops.
717 __mem_cgroup_insert_exceeded(mz, mctz, excess);
718 spin_unlock_irqrestore(&mctz->lock, flags);
723 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
725 struct mem_cgroup_tree_per_node *mctz;
726 struct mem_cgroup_per_node *mz;
730 mz = mem_cgroup_nodeinfo(memcg, nid);
731 mctz = soft_limit_tree_node(nid);
733 mem_cgroup_remove_exceeded(mz, mctz);
737 static struct mem_cgroup_per_node *
738 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
740 struct mem_cgroup_per_node *mz;
744 if (!mctz->rb_rightmost)
745 goto done; /* Nothing to reclaim from */
747 mz = rb_entry(mctz->rb_rightmost,
748 struct mem_cgroup_per_node, tree_node);
750 * Remove the node now but someone else can add it back,
751 * we will to add it back at the end of reclaim to its correct
752 * position in the tree.
754 __mem_cgroup_remove_exceeded(mz, mctz);
755 if (!soft_limit_excess(mz->memcg) ||
756 !css_tryget(&mz->memcg->css))
762 static struct mem_cgroup_per_node *
763 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
765 struct mem_cgroup_per_node *mz;
767 spin_lock_irq(&mctz->lock);
768 mz = __mem_cgroup_largest_soft_limit_node(mctz);
769 spin_unlock_irq(&mctz->lock);
774 * __mod_memcg_state - update cgroup memory statistics
775 * @memcg: the memory cgroup
776 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
777 * @val: delta to add to the counter, can be negative
779 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
781 long x, threshold = MEMCG_CHARGE_BATCH;
783 if (mem_cgroup_disabled())
786 if (vmstat_item_in_bytes(idx))
787 threshold <<= PAGE_SHIFT;
789 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
790 if (unlikely(abs(x) > threshold)) {
791 struct mem_cgroup *mi;
794 * Batch local counters to keep them in sync with
795 * the hierarchical ones.
797 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
798 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
799 atomic_long_add(x, &mi->vmstats[idx]);
802 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
805 static struct mem_cgroup_per_node *
806 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
808 struct mem_cgroup *parent;
810 parent = parent_mem_cgroup(pn->memcg);
813 return mem_cgroup_nodeinfo(parent, nid);
816 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
819 struct mem_cgroup_per_node *pn;
820 struct mem_cgroup *memcg;
821 long x, threshold = MEMCG_CHARGE_BATCH;
823 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
827 __mod_memcg_state(memcg, idx, val);
830 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
832 if (vmstat_item_in_bytes(idx))
833 threshold <<= PAGE_SHIFT;
835 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
836 if (unlikely(abs(x) > threshold)) {
837 pg_data_t *pgdat = lruvec_pgdat(lruvec);
838 struct mem_cgroup_per_node *pi;
840 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
841 atomic_long_add(x, &pi->lruvec_stat[idx]);
844 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
848 * __mod_lruvec_state - update lruvec memory statistics
849 * @lruvec: the lruvec
850 * @idx: the stat item
851 * @val: delta to add to the counter, can be negative
853 * The lruvec is the intersection of the NUMA node and a cgroup. This
854 * function updates the all three counters that are affected by a
855 * change of state at this level: per-node, per-cgroup, per-lruvec.
857 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
861 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
863 /* Update memcg and lruvec */
864 if (!mem_cgroup_disabled())
865 __mod_memcg_lruvec_state(lruvec, idx, val);
868 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
870 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
871 struct mem_cgroup *memcg;
872 struct lruvec *lruvec;
875 memcg = mem_cgroup_from_obj(p);
877 /* Untracked pages have no memcg, no lruvec. Update only the node */
878 if (!memcg || memcg == root_mem_cgroup) {
879 __mod_node_page_state(pgdat, idx, val);
881 lruvec = mem_cgroup_lruvec(memcg, pgdat);
882 __mod_lruvec_state(lruvec, idx, val);
887 void mod_memcg_obj_state(void *p, int idx, int val)
889 struct mem_cgroup *memcg;
892 memcg = mem_cgroup_from_obj(p);
894 mod_memcg_state(memcg, idx, val);
899 * __count_memcg_events - account VM events in a cgroup
900 * @memcg: the memory cgroup
901 * @idx: the event item
902 * @count: the number of events that occured
904 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
909 if (mem_cgroup_disabled())
912 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
913 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
914 struct mem_cgroup *mi;
917 * Batch local counters to keep them in sync with
918 * the hierarchical ones.
920 __this_cpu_add(memcg->vmstats_local->events[idx], x);
921 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
922 atomic_long_add(x, &mi->vmevents[idx]);
925 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
928 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
930 return atomic_long_read(&memcg->vmevents[event]);
933 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
938 for_each_possible_cpu(cpu)
939 x += per_cpu(memcg->vmstats_local->events[event], cpu);
943 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
947 /* pagein of a big page is an event. So, ignore page size */
949 __count_memcg_events(memcg, PGPGIN, 1);
951 __count_memcg_events(memcg, PGPGOUT, 1);
952 nr_pages = -nr_pages; /* for event */
955 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
958 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
959 enum mem_cgroup_events_target target)
961 unsigned long val, next;
963 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
964 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
965 /* from time_after() in jiffies.h */
966 if ((long)(next - val) < 0) {
968 case MEM_CGROUP_TARGET_THRESH:
969 next = val + THRESHOLDS_EVENTS_TARGET;
971 case MEM_CGROUP_TARGET_SOFTLIMIT:
972 next = val + SOFTLIMIT_EVENTS_TARGET;
977 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
984 * Check events in order.
987 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
989 /* threshold event is triggered in finer grain than soft limit */
990 if (unlikely(mem_cgroup_event_ratelimit(memcg,
991 MEM_CGROUP_TARGET_THRESH))) {
994 do_softlimit = mem_cgroup_event_ratelimit(memcg,
995 MEM_CGROUP_TARGET_SOFTLIMIT);
996 mem_cgroup_threshold(memcg);
997 if (unlikely(do_softlimit))
998 mem_cgroup_update_tree(memcg, page);
1002 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1005 * mm_update_next_owner() may clear mm->owner to NULL
1006 * if it races with swapoff, page migration, etc.
1007 * So this can be called with p == NULL.
1012 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1014 EXPORT_SYMBOL(mem_cgroup_from_task);
1017 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1018 * @mm: mm from which memcg should be extracted. It can be NULL.
1020 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1021 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1024 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1026 struct mem_cgroup *memcg;
1028 if (mem_cgroup_disabled())
1034 * Page cache insertions can happen withou an
1035 * actual mm context, e.g. during disk probing
1036 * on boot, loopback IO, acct() writes etc.
1039 memcg = root_mem_cgroup;
1041 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1042 if (unlikely(!memcg))
1043 memcg = root_mem_cgroup;
1045 } while (!css_tryget(&memcg->css));
1049 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1052 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1053 * @page: page from which memcg should be extracted.
1055 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1056 * root_mem_cgroup is returned.
1058 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
1060 struct mem_cgroup *memcg = page->mem_cgroup;
1062 if (mem_cgroup_disabled())
1066 /* Page should not get uncharged and freed memcg under us. */
1067 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
1068 memcg = root_mem_cgroup;
1072 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1075 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
1077 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1079 if (unlikely(current->active_memcg)) {
1080 struct mem_cgroup *memcg;
1083 /* current->active_memcg must hold a ref. */
1084 if (WARN_ON_ONCE(!css_tryget(¤t->active_memcg->css)))
1085 memcg = root_mem_cgroup;
1087 memcg = current->active_memcg;
1091 return get_mem_cgroup_from_mm(current->mm);
1095 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1096 * @root: hierarchy root
1097 * @prev: previously returned memcg, NULL on first invocation
1098 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1100 * Returns references to children of the hierarchy below @root, or
1101 * @root itself, or %NULL after a full round-trip.
1103 * Caller must pass the return value in @prev on subsequent
1104 * invocations for reference counting, or use mem_cgroup_iter_break()
1105 * to cancel a hierarchy walk before the round-trip is complete.
1107 * Reclaimers can specify a node and a priority level in @reclaim to
1108 * divide up the memcgs in the hierarchy among all concurrent
1109 * reclaimers operating on the same node and priority.
1111 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1112 struct mem_cgroup *prev,
1113 struct mem_cgroup_reclaim_cookie *reclaim)
1115 struct mem_cgroup_reclaim_iter *iter;
1116 struct cgroup_subsys_state *css = NULL;
1117 struct mem_cgroup *memcg = NULL;
1118 struct mem_cgroup *pos = NULL;
1120 if (mem_cgroup_disabled())
1124 root = root_mem_cgroup;
1126 if (prev && !reclaim)
1129 if (!root->use_hierarchy && root != root_mem_cgroup) {
1138 struct mem_cgroup_per_node *mz;
1140 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1143 if (prev && reclaim->generation != iter->generation)
1147 pos = READ_ONCE(iter->position);
1148 if (!pos || css_tryget(&pos->css))
1151 * css reference reached zero, so iter->position will
1152 * be cleared by ->css_released. However, we should not
1153 * rely on this happening soon, because ->css_released
1154 * is called from a work queue, and by busy-waiting we
1155 * might block it. So we clear iter->position right
1158 (void)cmpxchg(&iter->position, pos, NULL);
1166 css = css_next_descendant_pre(css, &root->css);
1169 * Reclaimers share the hierarchy walk, and a
1170 * new one might jump in right at the end of
1171 * the hierarchy - make sure they see at least
1172 * one group and restart from the beginning.
1180 * Verify the css and acquire a reference. The root
1181 * is provided by the caller, so we know it's alive
1182 * and kicking, and don't take an extra reference.
1184 memcg = mem_cgroup_from_css(css);
1186 if (css == &root->css)
1189 if (css_tryget(css))
1197 * The position could have already been updated by a competing
1198 * thread, so check that the value hasn't changed since we read
1199 * it to avoid reclaiming from the same cgroup twice.
1201 (void)cmpxchg(&iter->position, pos, memcg);
1209 reclaim->generation = iter->generation;
1215 if (prev && prev != root)
1216 css_put(&prev->css);
1222 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1223 * @root: hierarchy root
1224 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1226 void mem_cgroup_iter_break(struct mem_cgroup *root,
1227 struct mem_cgroup *prev)
1230 root = root_mem_cgroup;
1231 if (prev && prev != root)
1232 css_put(&prev->css);
1235 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1236 struct mem_cgroup *dead_memcg)
1238 struct mem_cgroup_reclaim_iter *iter;
1239 struct mem_cgroup_per_node *mz;
1242 for_each_node(nid) {
1243 mz = mem_cgroup_nodeinfo(from, nid);
1245 cmpxchg(&iter->position, dead_memcg, NULL);
1249 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1251 struct mem_cgroup *memcg = dead_memcg;
1252 struct mem_cgroup *last;
1255 __invalidate_reclaim_iterators(memcg, dead_memcg);
1257 } while ((memcg = parent_mem_cgroup(memcg)));
1260 * When cgruop1 non-hierarchy mode is used,
1261 * parent_mem_cgroup() does not walk all the way up to the
1262 * cgroup root (root_mem_cgroup). So we have to handle
1263 * dead_memcg from cgroup root separately.
1265 if (last != root_mem_cgroup)
1266 __invalidate_reclaim_iterators(root_mem_cgroup,
1271 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1272 * @memcg: hierarchy root
1273 * @fn: function to call for each task
1274 * @arg: argument passed to @fn
1276 * This function iterates over tasks attached to @memcg or to any of its
1277 * descendants and calls @fn for each task. If @fn returns a non-zero
1278 * value, the function breaks the iteration loop and returns the value.
1279 * Otherwise, it will iterate over all tasks and return 0.
1281 * This function must not be called for the root memory cgroup.
1283 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1284 int (*fn)(struct task_struct *, void *), void *arg)
1286 struct mem_cgroup *iter;
1289 BUG_ON(memcg == root_mem_cgroup);
1291 for_each_mem_cgroup_tree(iter, memcg) {
1292 struct css_task_iter it;
1293 struct task_struct *task;
1295 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1296 while (!ret && (task = css_task_iter_next(&it)))
1297 ret = fn(task, arg);
1298 css_task_iter_end(&it);
1300 mem_cgroup_iter_break(memcg, iter);
1308 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1310 * @pgdat: pgdat of the page
1312 * This function relies on page->mem_cgroup being stable - see the
1313 * access rules in commit_charge().
1315 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1317 struct mem_cgroup_per_node *mz;
1318 struct mem_cgroup *memcg;
1319 struct lruvec *lruvec;
1321 if (mem_cgroup_disabled()) {
1322 lruvec = &pgdat->__lruvec;
1326 memcg = page->mem_cgroup;
1328 * Swapcache readahead pages are added to the LRU - and
1329 * possibly migrated - before they are charged.
1332 memcg = root_mem_cgroup;
1334 mz = mem_cgroup_page_nodeinfo(memcg, page);
1335 lruvec = &mz->lruvec;
1338 * Since a node can be onlined after the mem_cgroup was created,
1339 * we have to be prepared to initialize lruvec->zone here;
1340 * and if offlined then reonlined, we need to reinitialize it.
1342 if (unlikely(lruvec->pgdat != pgdat))
1343 lruvec->pgdat = pgdat;
1348 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1349 * @lruvec: mem_cgroup per zone lru vector
1350 * @lru: index of lru list the page is sitting on
1351 * @zid: zone id of the accounted pages
1352 * @nr_pages: positive when adding or negative when removing
1354 * This function must be called under lru_lock, just before a page is added
1355 * to or just after a page is removed from an lru list (that ordering being
1356 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1358 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1359 int zid, int nr_pages)
1361 struct mem_cgroup_per_node *mz;
1362 unsigned long *lru_size;
1365 if (mem_cgroup_disabled())
1368 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1369 lru_size = &mz->lru_zone_size[zid][lru];
1372 *lru_size += nr_pages;
1375 if (WARN_ONCE(size < 0,
1376 "%s(%p, %d, %d): lru_size %ld\n",
1377 __func__, lruvec, lru, nr_pages, size)) {
1383 *lru_size += nr_pages;
1387 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1388 * @memcg: the memory cgroup
1390 * Returns the maximum amount of memory @mem can be charged with, in
1393 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1395 unsigned long margin = 0;
1396 unsigned long count;
1397 unsigned long limit;
1399 count = page_counter_read(&memcg->memory);
1400 limit = READ_ONCE(memcg->memory.max);
1402 margin = limit - count;
1404 if (do_memsw_account()) {
1405 count = page_counter_read(&memcg->memsw);
1406 limit = READ_ONCE(memcg->memsw.max);
1408 margin = min(margin, limit - count);
1417 * A routine for checking "mem" is under move_account() or not.
1419 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1420 * moving cgroups. This is for waiting at high-memory pressure
1423 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1425 struct mem_cgroup *from;
1426 struct mem_cgroup *to;
1429 * Unlike task_move routines, we access mc.to, mc.from not under
1430 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1432 spin_lock(&mc.lock);
1438 ret = mem_cgroup_is_descendant(from, memcg) ||
1439 mem_cgroup_is_descendant(to, memcg);
1441 spin_unlock(&mc.lock);
1445 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1447 if (mc.moving_task && current != mc.moving_task) {
1448 if (mem_cgroup_under_move(memcg)) {
1450 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1451 /* moving charge context might have finished. */
1454 finish_wait(&mc.waitq, &wait);
1461 static char *memory_stat_format(struct mem_cgroup *memcg)
1466 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1471 * Provide statistics on the state of the memory subsystem as
1472 * well as cumulative event counters that show past behavior.
1474 * This list is ordered following a combination of these gradients:
1475 * 1) generic big picture -> specifics and details
1476 * 2) reflecting userspace activity -> reflecting kernel heuristics
1478 * Current memory state:
1481 seq_buf_printf(&s, "anon %llu\n",
1482 (u64)memcg_page_state(memcg, NR_ANON_MAPPED) *
1484 seq_buf_printf(&s, "file %llu\n",
1485 (u64)memcg_page_state(memcg, NR_FILE_PAGES) *
1487 seq_buf_printf(&s, "kernel_stack %llu\n",
1488 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) *
1490 seq_buf_printf(&s, "slab %llu\n",
1491 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B) +
1492 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_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 %lu\n",
1534 memcg_page_state(memcg, WORKINGSET_REFAULT));
1535 seq_buf_printf(&s, "workingset_activate %lu\n",
1536 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1537 seq_buf_printf(&s, "workingset_restore %lu\n",
1538 memcg_page_state(memcg, WORKINGSET_RESTORE));
1539 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1540 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1542 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1543 memcg_events(memcg, PGREFILL));
1544 seq_buf_printf(&s, "pgscan %lu\n",
1545 memcg_events(memcg, PGSCAN_KSWAPD) +
1546 memcg_events(memcg, PGSCAN_DIRECT));
1547 seq_buf_printf(&s, "pgsteal %lu\n",
1548 memcg_events(memcg, PGSTEAL_KSWAPD) +
1549 memcg_events(memcg, PGSTEAL_DIRECT));
1550 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1551 memcg_events(memcg, PGACTIVATE));
1552 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1553 memcg_events(memcg, PGDEACTIVATE));
1554 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1555 memcg_events(memcg, PGLAZYFREE));
1556 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1557 memcg_events(memcg, PGLAZYFREED));
1559 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1560 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1561 memcg_events(memcg, THP_FAULT_ALLOC));
1562 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1563 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1564 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1566 /* The above should easily fit into one page */
1567 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1572 #define K(x) ((x) << (PAGE_SHIFT-10))
1574 * mem_cgroup_print_oom_context: Print OOM information relevant to
1575 * memory controller.
1576 * @memcg: The memory cgroup that went over limit
1577 * @p: Task that is going to be killed
1579 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1582 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1587 pr_cont(",oom_memcg=");
1588 pr_cont_cgroup_path(memcg->css.cgroup);
1590 pr_cont(",global_oom");
1592 pr_cont(",task_memcg=");
1593 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1599 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1600 * memory controller.
1601 * @memcg: The memory cgroup that went over limit
1603 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1607 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1608 K((u64)page_counter_read(&memcg->memory)),
1609 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1610 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1611 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1612 K((u64)page_counter_read(&memcg->swap)),
1613 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1615 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1616 K((u64)page_counter_read(&memcg->memsw)),
1617 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1618 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1619 K((u64)page_counter_read(&memcg->kmem)),
1620 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1623 pr_info("Memory cgroup stats for ");
1624 pr_cont_cgroup_path(memcg->css.cgroup);
1626 buf = memory_stat_format(memcg);
1634 * Return the memory (and swap, if configured) limit for a memcg.
1636 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1640 max = READ_ONCE(memcg->memory.max);
1641 if (mem_cgroup_swappiness(memcg)) {
1642 unsigned long memsw_max;
1643 unsigned long swap_max;
1645 memsw_max = memcg->memsw.max;
1646 swap_max = READ_ONCE(memcg->swap.max);
1647 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1648 max = min(max + swap_max, memsw_max);
1653 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1655 return page_counter_read(&memcg->memory);
1658 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1661 struct oom_control oc = {
1665 .gfp_mask = gfp_mask,
1670 if (mutex_lock_killable(&oom_lock))
1673 * A few threads which were not waiting at mutex_lock_killable() can
1674 * fail to bail out. Therefore, check again after holding oom_lock.
1676 ret = should_force_charge() || out_of_memory(&oc);
1677 mutex_unlock(&oom_lock);
1681 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1684 unsigned long *total_scanned)
1686 struct mem_cgroup *victim = NULL;
1689 unsigned long excess;
1690 unsigned long nr_scanned;
1691 struct mem_cgroup_reclaim_cookie reclaim = {
1695 excess = soft_limit_excess(root_memcg);
1698 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1703 * If we have not been able to reclaim
1704 * anything, it might because there are
1705 * no reclaimable pages under this hierarchy
1710 * We want to do more targeted reclaim.
1711 * excess >> 2 is not to excessive so as to
1712 * reclaim too much, nor too less that we keep
1713 * coming back to reclaim from this cgroup
1715 if (total >= (excess >> 2) ||
1716 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1721 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1722 pgdat, &nr_scanned);
1723 *total_scanned += nr_scanned;
1724 if (!soft_limit_excess(root_memcg))
1727 mem_cgroup_iter_break(root_memcg, victim);
1731 #ifdef CONFIG_LOCKDEP
1732 static struct lockdep_map memcg_oom_lock_dep_map = {
1733 .name = "memcg_oom_lock",
1737 static DEFINE_SPINLOCK(memcg_oom_lock);
1740 * Check OOM-Killer is already running under our hierarchy.
1741 * If someone is running, return false.
1743 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1745 struct mem_cgroup *iter, *failed = NULL;
1747 spin_lock(&memcg_oom_lock);
1749 for_each_mem_cgroup_tree(iter, memcg) {
1750 if (iter->oom_lock) {
1752 * this subtree of our hierarchy is already locked
1753 * so we cannot give a lock.
1756 mem_cgroup_iter_break(memcg, iter);
1759 iter->oom_lock = true;
1764 * OK, we failed to lock the whole subtree so we have
1765 * to clean up what we set up to the failing subtree
1767 for_each_mem_cgroup_tree(iter, memcg) {
1768 if (iter == failed) {
1769 mem_cgroup_iter_break(memcg, iter);
1772 iter->oom_lock = false;
1775 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1777 spin_unlock(&memcg_oom_lock);
1782 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1784 struct mem_cgroup *iter;
1786 spin_lock(&memcg_oom_lock);
1787 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1788 for_each_mem_cgroup_tree(iter, memcg)
1789 iter->oom_lock = false;
1790 spin_unlock(&memcg_oom_lock);
1793 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1795 struct mem_cgroup *iter;
1797 spin_lock(&memcg_oom_lock);
1798 for_each_mem_cgroup_tree(iter, memcg)
1800 spin_unlock(&memcg_oom_lock);
1803 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1805 struct mem_cgroup *iter;
1808 * When a new child is created while the hierarchy is under oom,
1809 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1811 spin_lock(&memcg_oom_lock);
1812 for_each_mem_cgroup_tree(iter, memcg)
1813 if (iter->under_oom > 0)
1815 spin_unlock(&memcg_oom_lock);
1818 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1820 struct oom_wait_info {
1821 struct mem_cgroup *memcg;
1822 wait_queue_entry_t wait;
1825 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1826 unsigned mode, int sync, void *arg)
1828 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1829 struct mem_cgroup *oom_wait_memcg;
1830 struct oom_wait_info *oom_wait_info;
1832 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1833 oom_wait_memcg = oom_wait_info->memcg;
1835 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1836 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1838 return autoremove_wake_function(wait, mode, sync, arg);
1841 static void memcg_oom_recover(struct mem_cgroup *memcg)
1844 * For the following lockless ->under_oom test, the only required
1845 * guarantee is that it must see the state asserted by an OOM when
1846 * this function is called as a result of userland actions
1847 * triggered by the notification of the OOM. This is trivially
1848 * achieved by invoking mem_cgroup_mark_under_oom() before
1849 * triggering notification.
1851 if (memcg && memcg->under_oom)
1852 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1862 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1864 enum oom_status ret;
1867 if (order > PAGE_ALLOC_COSTLY_ORDER)
1870 memcg_memory_event(memcg, MEMCG_OOM);
1873 * We are in the middle of the charge context here, so we
1874 * don't want to block when potentially sitting on a callstack
1875 * that holds all kinds of filesystem and mm locks.
1877 * cgroup1 allows disabling the OOM killer and waiting for outside
1878 * handling until the charge can succeed; remember the context and put
1879 * the task to sleep at the end of the page fault when all locks are
1882 * On the other hand, in-kernel OOM killer allows for an async victim
1883 * memory reclaim (oom_reaper) and that means that we are not solely
1884 * relying on the oom victim to make a forward progress and we can
1885 * invoke the oom killer here.
1887 * Please note that mem_cgroup_out_of_memory might fail to find a
1888 * victim and then we have to bail out from the charge path.
1890 if (memcg->oom_kill_disable) {
1891 if (!current->in_user_fault)
1893 css_get(&memcg->css);
1894 current->memcg_in_oom = memcg;
1895 current->memcg_oom_gfp_mask = mask;
1896 current->memcg_oom_order = order;
1901 mem_cgroup_mark_under_oom(memcg);
1903 locked = mem_cgroup_oom_trylock(memcg);
1906 mem_cgroup_oom_notify(memcg);
1908 mem_cgroup_unmark_under_oom(memcg);
1909 if (mem_cgroup_out_of_memory(memcg, mask, order))
1915 mem_cgroup_oom_unlock(memcg);
1921 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1922 * @handle: actually kill/wait or just clean up the OOM state
1924 * This has to be called at the end of a page fault if the memcg OOM
1925 * handler was enabled.
1927 * Memcg supports userspace OOM handling where failed allocations must
1928 * sleep on a waitqueue until the userspace task resolves the
1929 * situation. Sleeping directly in the charge context with all kinds
1930 * of locks held is not a good idea, instead we remember an OOM state
1931 * in the task and mem_cgroup_oom_synchronize() has to be called at
1932 * the end of the page fault to complete the OOM handling.
1934 * Returns %true if an ongoing memcg OOM situation was detected and
1935 * completed, %false otherwise.
1937 bool mem_cgroup_oom_synchronize(bool handle)
1939 struct mem_cgroup *memcg = current->memcg_in_oom;
1940 struct oom_wait_info owait;
1943 /* OOM is global, do not handle */
1950 owait.memcg = memcg;
1951 owait.wait.flags = 0;
1952 owait.wait.func = memcg_oom_wake_function;
1953 owait.wait.private = current;
1954 INIT_LIST_HEAD(&owait.wait.entry);
1956 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1957 mem_cgroup_mark_under_oom(memcg);
1959 locked = mem_cgroup_oom_trylock(memcg);
1962 mem_cgroup_oom_notify(memcg);
1964 if (locked && !memcg->oom_kill_disable) {
1965 mem_cgroup_unmark_under_oom(memcg);
1966 finish_wait(&memcg_oom_waitq, &owait.wait);
1967 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1968 current->memcg_oom_order);
1971 mem_cgroup_unmark_under_oom(memcg);
1972 finish_wait(&memcg_oom_waitq, &owait.wait);
1976 mem_cgroup_oom_unlock(memcg);
1978 * There is no guarantee that an OOM-lock contender
1979 * sees the wakeups triggered by the OOM kill
1980 * uncharges. Wake any sleepers explicitely.
1982 memcg_oom_recover(memcg);
1985 current->memcg_in_oom = NULL;
1986 css_put(&memcg->css);
1991 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1992 * @victim: task to be killed by the OOM killer
1993 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1995 * Returns a pointer to a memory cgroup, which has to be cleaned up
1996 * by killing all belonging OOM-killable tasks.
1998 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2000 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2001 struct mem_cgroup *oom_domain)
2003 struct mem_cgroup *oom_group = NULL;
2004 struct mem_cgroup *memcg;
2006 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2010 oom_domain = root_mem_cgroup;
2014 memcg = mem_cgroup_from_task(victim);
2015 if (memcg == root_mem_cgroup)
2019 * If the victim task has been asynchronously moved to a different
2020 * memory cgroup, we might end up killing tasks outside oom_domain.
2021 * In this case it's better to ignore memory.group.oom.
2023 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2027 * Traverse the memory cgroup hierarchy from the victim task's
2028 * cgroup up to the OOMing cgroup (or root) to find the
2029 * highest-level memory cgroup with oom.group set.
2031 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2032 if (memcg->oom_group)
2035 if (memcg == oom_domain)
2040 css_get(&oom_group->css);
2047 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2049 pr_info("Tasks in ");
2050 pr_cont_cgroup_path(memcg->css.cgroup);
2051 pr_cont(" are going to be killed due to memory.oom.group set\n");
2055 * lock_page_memcg - lock a page->mem_cgroup binding
2058 * This function protects unlocked LRU pages from being moved to
2061 * It ensures lifetime of the returned memcg. Caller is responsible
2062 * for the lifetime of the page; __unlock_page_memcg() is available
2063 * when @page might get freed inside the locked section.
2065 struct mem_cgroup *lock_page_memcg(struct page *page)
2067 struct page *head = compound_head(page); /* rmap on tail pages */
2068 struct mem_cgroup *memcg;
2069 unsigned long flags;
2072 * The RCU lock is held throughout the transaction. The fast
2073 * path can get away without acquiring the memcg->move_lock
2074 * because page moving starts with an RCU grace period.
2076 * The RCU lock also protects the memcg from being freed when
2077 * the page state that is going to change is the only thing
2078 * preventing the page itself from being freed. E.g. writeback
2079 * doesn't hold a page reference and relies on PG_writeback to
2080 * keep off truncation, migration and so forth.
2084 if (mem_cgroup_disabled())
2087 memcg = head->mem_cgroup;
2088 if (unlikely(!memcg))
2091 if (atomic_read(&memcg->moving_account) <= 0)
2094 spin_lock_irqsave(&memcg->move_lock, flags);
2095 if (memcg != head->mem_cgroup) {
2096 spin_unlock_irqrestore(&memcg->move_lock, flags);
2101 * When charge migration first begins, we can have locked and
2102 * unlocked page stat updates happening concurrently. Track
2103 * the task who has the lock for unlock_page_memcg().
2105 memcg->move_lock_task = current;
2106 memcg->move_lock_flags = flags;
2110 EXPORT_SYMBOL(lock_page_memcg);
2113 * __unlock_page_memcg - unlock and unpin a memcg
2116 * Unlock and unpin a memcg returned by lock_page_memcg().
2118 void __unlock_page_memcg(struct mem_cgroup *memcg)
2120 if (memcg && memcg->move_lock_task == current) {
2121 unsigned long flags = memcg->move_lock_flags;
2123 memcg->move_lock_task = NULL;
2124 memcg->move_lock_flags = 0;
2126 spin_unlock_irqrestore(&memcg->move_lock, flags);
2133 * unlock_page_memcg - unlock a page->mem_cgroup binding
2136 void unlock_page_memcg(struct page *page)
2138 struct page *head = compound_head(page);
2140 __unlock_page_memcg(head->mem_cgroup);
2142 EXPORT_SYMBOL(unlock_page_memcg);
2144 struct memcg_stock_pcp {
2145 struct mem_cgroup *cached; /* this never be root cgroup */
2146 unsigned int nr_pages;
2148 #ifdef CONFIG_MEMCG_KMEM
2149 struct obj_cgroup *cached_objcg;
2150 unsigned int nr_bytes;
2153 struct work_struct work;
2154 unsigned long flags;
2155 #define FLUSHING_CACHED_CHARGE 0
2157 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2158 static DEFINE_MUTEX(percpu_charge_mutex);
2160 #ifdef CONFIG_MEMCG_KMEM
2161 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2162 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2163 struct mem_cgroup *root_memcg);
2166 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2169 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2170 struct mem_cgroup *root_memcg)
2177 * consume_stock: Try to consume stocked charge on this cpu.
2178 * @memcg: memcg to consume from.
2179 * @nr_pages: how many pages to charge.
2181 * The charges will only happen if @memcg matches the current cpu's memcg
2182 * stock, and at least @nr_pages are available in that stock. Failure to
2183 * service an allocation will refill the stock.
2185 * returns true if successful, false otherwise.
2187 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2189 struct memcg_stock_pcp *stock;
2190 unsigned long flags;
2193 if (nr_pages > MEMCG_CHARGE_BATCH)
2196 local_irq_save(flags);
2198 stock = this_cpu_ptr(&memcg_stock);
2199 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2200 stock->nr_pages -= nr_pages;
2204 local_irq_restore(flags);
2210 * Returns stocks cached in percpu and reset cached information.
2212 static void drain_stock(struct memcg_stock_pcp *stock)
2214 struct mem_cgroup *old = stock->cached;
2219 if (stock->nr_pages) {
2220 page_counter_uncharge(&old->memory, stock->nr_pages);
2221 if (do_memsw_account())
2222 page_counter_uncharge(&old->memsw, stock->nr_pages);
2223 stock->nr_pages = 0;
2227 stock->cached = NULL;
2230 static void drain_local_stock(struct work_struct *dummy)
2232 struct memcg_stock_pcp *stock;
2233 unsigned long flags;
2236 * The only protection from memory hotplug vs. drain_stock races is
2237 * that we always operate on local CPU stock here with IRQ disabled
2239 local_irq_save(flags);
2241 stock = this_cpu_ptr(&memcg_stock);
2242 drain_obj_stock(stock);
2244 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2246 local_irq_restore(flags);
2250 * Cache charges(val) to local per_cpu area.
2251 * This will be consumed by consume_stock() function, later.
2253 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2255 struct memcg_stock_pcp *stock;
2256 unsigned long flags;
2258 local_irq_save(flags);
2260 stock = this_cpu_ptr(&memcg_stock);
2261 if (stock->cached != memcg) { /* reset if necessary */
2263 css_get(&memcg->css);
2264 stock->cached = memcg;
2266 stock->nr_pages += nr_pages;
2268 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2271 local_irq_restore(flags);
2275 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2276 * of the hierarchy under it.
2278 static void drain_all_stock(struct mem_cgroup *root_memcg)
2282 /* If someone's already draining, avoid adding running more workers. */
2283 if (!mutex_trylock(&percpu_charge_mutex))
2286 * Notify other cpus that system-wide "drain" is running
2287 * We do not care about races with the cpu hotplug because cpu down
2288 * as well as workers from this path always operate on the local
2289 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2292 for_each_online_cpu(cpu) {
2293 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2294 struct mem_cgroup *memcg;
2298 memcg = stock->cached;
2299 if (memcg && stock->nr_pages &&
2300 mem_cgroup_is_descendant(memcg, root_memcg))
2302 if (obj_stock_flush_required(stock, root_memcg))
2307 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2309 drain_local_stock(&stock->work);
2311 schedule_work_on(cpu, &stock->work);
2315 mutex_unlock(&percpu_charge_mutex);
2318 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2320 struct memcg_stock_pcp *stock;
2321 struct mem_cgroup *memcg, *mi;
2323 stock = &per_cpu(memcg_stock, cpu);
2326 for_each_mem_cgroup(memcg) {
2329 for (i = 0; i < MEMCG_NR_STAT; i++) {
2333 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2335 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2336 atomic_long_add(x, &memcg->vmstats[i]);
2338 if (i >= NR_VM_NODE_STAT_ITEMS)
2341 for_each_node(nid) {
2342 struct mem_cgroup_per_node *pn;
2344 pn = mem_cgroup_nodeinfo(memcg, nid);
2345 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2348 atomic_long_add(x, &pn->lruvec_stat[i]);
2349 } while ((pn = parent_nodeinfo(pn, nid)));
2353 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2356 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2358 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2359 atomic_long_add(x, &memcg->vmevents[i]);
2366 static void reclaim_high(struct mem_cgroup *memcg,
2367 unsigned int nr_pages,
2371 if (page_counter_read(&memcg->memory) <=
2372 READ_ONCE(memcg->memory.high))
2374 memcg_memory_event(memcg, MEMCG_HIGH);
2375 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2376 } while ((memcg = parent_mem_cgroup(memcg)) &&
2377 !mem_cgroup_is_root(memcg));
2380 static void high_work_func(struct work_struct *work)
2382 struct mem_cgroup *memcg;
2384 memcg = container_of(work, struct mem_cgroup, high_work);
2385 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2389 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2390 * enough to still cause a significant slowdown in most cases, while still
2391 * allowing diagnostics and tracing to proceed without becoming stuck.
2393 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2396 * When calculating the delay, we use these either side of the exponentiation to
2397 * maintain precision and scale to a reasonable number of jiffies (see the table
2400 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2401 * overage ratio to a delay.
2402 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2403 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2404 * to produce a reasonable delay curve.
2406 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2407 * reasonable delay curve compared to precision-adjusted overage, not
2408 * penalising heavily at first, but still making sure that growth beyond the
2409 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2410 * example, with a high of 100 megabytes:
2412 * +-------+------------------------+
2413 * | usage | time to allocate in ms |
2414 * +-------+------------------------+
2436 * +-------+------------------------+
2438 #define MEMCG_DELAY_PRECISION_SHIFT 20
2439 #define MEMCG_DELAY_SCALING_SHIFT 14
2441 static u64 calculate_overage(unsigned long usage, unsigned long high)
2449 * Prevent division by 0 in overage calculation by acting as if
2450 * it was a threshold of 1 page
2452 high = max(high, 1UL);
2454 overage = usage - high;
2455 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2456 return div64_u64(overage, high);
2459 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2461 u64 overage, max_overage = 0;
2464 overage = calculate_overage(page_counter_read(&memcg->memory),
2465 READ_ONCE(memcg->memory.high));
2466 max_overage = max(overage, max_overage);
2467 } while ((memcg = parent_mem_cgroup(memcg)) &&
2468 !mem_cgroup_is_root(memcg));
2473 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2475 u64 overage, max_overage = 0;
2478 overage = calculate_overage(page_counter_read(&memcg->swap),
2479 READ_ONCE(memcg->swap.high));
2481 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2482 max_overage = max(overage, max_overage);
2483 } while ((memcg = parent_mem_cgroup(memcg)) &&
2484 !mem_cgroup_is_root(memcg));
2490 * Get the number of jiffies that we should penalise a mischievous cgroup which
2491 * is exceeding its memory.high by checking both it and its ancestors.
2493 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2494 unsigned int nr_pages,
2497 unsigned long penalty_jiffies;
2503 * We use overage compared to memory.high to calculate the number of
2504 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2505 * fairly lenient on small overages, and increasingly harsh when the
2506 * memcg in question makes it clear that it has no intention of stopping
2507 * its crazy behaviour, so we exponentially increase the delay based on
2510 penalty_jiffies = max_overage * max_overage * HZ;
2511 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2512 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2515 * Factor in the task's own contribution to the overage, such that four
2516 * N-sized allocations are throttled approximately the same as one
2517 * 4N-sized allocation.
2519 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2520 * larger the current charge patch is than that.
2522 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2526 * Scheduled by try_charge() to be executed from the userland return path
2527 * and reclaims memory over the high limit.
2529 void mem_cgroup_handle_over_high(void)
2531 unsigned long penalty_jiffies;
2532 unsigned long pflags;
2533 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2534 struct mem_cgroup *memcg;
2536 if (likely(!nr_pages))
2539 memcg = get_mem_cgroup_from_mm(current->mm);
2540 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2541 current->memcg_nr_pages_over_high = 0;
2544 * memory.high is breached and reclaim is unable to keep up. Throttle
2545 * allocators proactively to slow down excessive growth.
2547 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2548 mem_find_max_overage(memcg));
2550 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2551 swap_find_max_overage(memcg));
2554 * Clamp the max delay per usermode return so as to still keep the
2555 * application moving forwards and also permit diagnostics, albeit
2558 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2561 * Don't sleep if the amount of jiffies this memcg owes us is so low
2562 * that it's not even worth doing, in an attempt to be nice to those who
2563 * go only a small amount over their memory.high value and maybe haven't
2564 * been aggressively reclaimed enough yet.
2566 if (penalty_jiffies <= HZ / 100)
2570 * If we exit early, we're guaranteed to die (since
2571 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2572 * need to account for any ill-begotten jiffies to pay them off later.
2574 psi_memstall_enter(&pflags);
2575 schedule_timeout_killable(penalty_jiffies);
2576 psi_memstall_leave(&pflags);
2579 css_put(&memcg->css);
2582 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2583 unsigned int nr_pages)
2585 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2586 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2587 struct mem_cgroup *mem_over_limit;
2588 struct page_counter *counter;
2589 unsigned long nr_reclaimed;
2590 bool may_swap = true;
2591 bool drained = false;
2592 enum oom_status oom_status;
2594 if (mem_cgroup_is_root(memcg))
2597 if (consume_stock(memcg, nr_pages))
2600 if (!do_memsw_account() ||
2601 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2602 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2604 if (do_memsw_account())
2605 page_counter_uncharge(&memcg->memsw, batch);
2606 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2608 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2612 if (batch > nr_pages) {
2618 * Memcg doesn't have a dedicated reserve for atomic
2619 * allocations. But like the global atomic pool, we need to
2620 * put the burden of reclaim on regular allocation requests
2621 * and let these go through as privileged allocations.
2623 if (gfp_mask & __GFP_ATOMIC)
2627 * Unlike in global OOM situations, memcg is not in a physical
2628 * memory shortage. Allow dying and OOM-killed tasks to
2629 * bypass the last charges so that they can exit quickly and
2630 * free their memory.
2632 if (unlikely(should_force_charge()))
2636 * Prevent unbounded recursion when reclaim operations need to
2637 * allocate memory. This might exceed the limits temporarily,
2638 * but we prefer facilitating memory reclaim and getting back
2639 * under the limit over triggering OOM kills in these cases.
2641 if (unlikely(current->flags & PF_MEMALLOC))
2644 if (unlikely(task_in_memcg_oom(current)))
2647 if (!gfpflags_allow_blocking(gfp_mask))
2650 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2652 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2653 gfp_mask, may_swap);
2655 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2659 drain_all_stock(mem_over_limit);
2664 if (gfp_mask & __GFP_NORETRY)
2667 * Even though the limit is exceeded at this point, reclaim
2668 * may have been able to free some pages. Retry the charge
2669 * before killing the task.
2671 * Only for regular pages, though: huge pages are rather
2672 * unlikely to succeed so close to the limit, and we fall back
2673 * to regular pages anyway in case of failure.
2675 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2678 * At task move, charge accounts can be doubly counted. So, it's
2679 * better to wait until the end of task_move if something is going on.
2681 if (mem_cgroup_wait_acct_move(mem_over_limit))
2687 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2690 if (gfp_mask & __GFP_NOFAIL)
2693 if (fatal_signal_pending(current))
2697 * keep retrying as long as the memcg oom killer is able to make
2698 * a forward progress or bypass the charge if the oom killer
2699 * couldn't make any progress.
2701 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2702 get_order(nr_pages * PAGE_SIZE));
2703 switch (oom_status) {
2705 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2713 if (!(gfp_mask & __GFP_NOFAIL))
2717 * The allocation either can't fail or will lead to more memory
2718 * being freed very soon. Allow memory usage go over the limit
2719 * temporarily by force charging it.
2721 page_counter_charge(&memcg->memory, nr_pages);
2722 if (do_memsw_account())
2723 page_counter_charge(&memcg->memsw, nr_pages);
2728 if (batch > nr_pages)
2729 refill_stock(memcg, batch - nr_pages);
2732 * If the hierarchy is above the normal consumption range, schedule
2733 * reclaim on returning to userland. We can perform reclaim here
2734 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2735 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2736 * not recorded as it most likely matches current's and won't
2737 * change in the meantime. As high limit is checked again before
2738 * reclaim, the cost of mismatch is negligible.
2741 bool mem_high, swap_high;
2743 mem_high = page_counter_read(&memcg->memory) >
2744 READ_ONCE(memcg->memory.high);
2745 swap_high = page_counter_read(&memcg->swap) >
2746 READ_ONCE(memcg->swap.high);
2748 /* Don't bother a random interrupted task */
2749 if (in_interrupt()) {
2751 schedule_work(&memcg->high_work);
2757 if (mem_high || swap_high) {
2759 * The allocating tasks in this cgroup will need to do
2760 * reclaim or be throttled to prevent further growth
2761 * of the memory or swap footprints.
2763 * Target some best-effort fairness between the tasks,
2764 * and distribute reclaim work and delay penalties
2765 * based on how much each task is actually allocating.
2767 current->memcg_nr_pages_over_high += batch;
2768 set_notify_resume(current);
2771 } while ((memcg = parent_mem_cgroup(memcg)));
2776 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2777 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2779 if (mem_cgroup_is_root(memcg))
2782 page_counter_uncharge(&memcg->memory, nr_pages);
2783 if (do_memsw_account())
2784 page_counter_uncharge(&memcg->memsw, nr_pages);
2788 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2790 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2792 * Any of the following ensures page->mem_cgroup stability:
2796 * - lock_page_memcg()
2797 * - exclusive reference
2799 page->mem_cgroup = memcg;
2802 #ifdef CONFIG_MEMCG_KMEM
2804 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2806 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2807 * cgroup_mutex, etc.
2809 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2813 if (mem_cgroup_disabled())
2816 page = virt_to_head_page(p);
2819 * Slab objects are accounted individually, not per-page.
2820 * Memcg membership data for each individual object is saved in
2821 * the page->obj_cgroups.
2823 if (page_has_obj_cgroups(page)) {
2824 struct obj_cgroup *objcg;
2827 off = obj_to_index(page->slab_cache, page, p);
2828 objcg = page_obj_cgroups(page)[off];
2829 return obj_cgroup_memcg(objcg);
2832 /* All other pages use page->mem_cgroup */
2833 return page->mem_cgroup;
2836 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2838 struct obj_cgroup *objcg = NULL;
2839 struct mem_cgroup *memcg;
2841 if (unlikely(!current->mm && !current->active_memcg))
2845 if (unlikely(current->active_memcg))
2846 memcg = rcu_dereference(current->active_memcg);
2848 memcg = mem_cgroup_from_task(current);
2850 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2851 objcg = rcu_dereference(memcg->objcg);
2852 if (objcg && obj_cgroup_tryget(objcg))
2860 static int memcg_alloc_cache_id(void)
2865 id = ida_simple_get(&memcg_cache_ida,
2866 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2870 if (id < memcg_nr_cache_ids)
2874 * There's no space for the new id in memcg_caches arrays,
2875 * so we have to grow them.
2877 down_write(&memcg_cache_ids_sem);
2879 size = 2 * (id + 1);
2880 if (size < MEMCG_CACHES_MIN_SIZE)
2881 size = MEMCG_CACHES_MIN_SIZE;
2882 else if (size > MEMCG_CACHES_MAX_SIZE)
2883 size = MEMCG_CACHES_MAX_SIZE;
2885 err = memcg_update_all_list_lrus(size);
2887 memcg_nr_cache_ids = size;
2889 up_write(&memcg_cache_ids_sem);
2892 ida_simple_remove(&memcg_cache_ida, id);
2898 static void memcg_free_cache_id(int id)
2900 ida_simple_remove(&memcg_cache_ida, id);
2904 * memcg_kmem_get_cache: select memcg or root cache for allocation
2905 * @cachep: the original global kmem cache
2907 * Return the kmem_cache we're supposed to use for a slab allocation.
2909 * If the cache does not exist yet, if we are the first user of it, we
2910 * create it asynchronously in a workqueue and let the current allocation
2911 * go through with the original cache.
2913 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2915 struct kmem_cache *memcg_cachep;
2917 memcg_cachep = READ_ONCE(cachep->memcg_params.memcg_cache);
2918 if (unlikely(!memcg_cachep)) {
2919 queue_work(system_wq, &cachep->memcg_params.work);
2923 return memcg_cachep;
2927 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
2928 * @memcg: memory cgroup to charge
2929 * @gfp: reclaim mode
2930 * @nr_pages: number of pages to charge
2932 * Returns 0 on success, an error code on failure.
2934 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
2935 unsigned int nr_pages)
2937 struct page_counter *counter;
2940 ret = try_charge(memcg, gfp, nr_pages);
2944 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2945 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2948 * Enforce __GFP_NOFAIL allocation because callers are not
2949 * prepared to see failures and likely do not have any failure
2952 if (gfp & __GFP_NOFAIL) {
2953 page_counter_charge(&memcg->kmem, nr_pages);
2956 cancel_charge(memcg, nr_pages);
2963 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
2964 * @memcg: memcg to uncharge
2965 * @nr_pages: number of pages to uncharge
2967 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
2969 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2970 page_counter_uncharge(&memcg->kmem, nr_pages);
2972 page_counter_uncharge(&memcg->memory, nr_pages);
2973 if (do_memsw_account())
2974 page_counter_uncharge(&memcg->memsw, nr_pages);
2978 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
2979 * @page: page to charge
2980 * @gfp: reclaim mode
2981 * @order: allocation order
2983 * Returns 0 on success, an error code on failure.
2985 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
2987 struct mem_cgroup *memcg;
2990 if (memcg_kmem_bypass())
2993 memcg = get_mem_cgroup_from_current();
2994 if (!mem_cgroup_is_root(memcg)) {
2995 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
2997 page->mem_cgroup = memcg;
2998 __SetPageKmemcg(page);
3002 css_put(&memcg->css);
3007 * __memcg_kmem_uncharge_page: uncharge a kmem page
3008 * @page: page to uncharge
3009 * @order: allocation order
3011 void __memcg_kmem_uncharge_page(struct page *page, int order)
3013 struct mem_cgroup *memcg = page->mem_cgroup;
3014 unsigned int nr_pages = 1 << order;
3019 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3020 __memcg_kmem_uncharge(memcg, nr_pages);
3021 page->mem_cgroup = NULL;
3022 css_put(&memcg->css);
3024 /* slab pages do not have PageKmemcg flag set */
3025 if (PageKmemcg(page))
3026 __ClearPageKmemcg(page);
3029 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3031 struct memcg_stock_pcp *stock;
3032 unsigned long flags;
3035 local_irq_save(flags);
3037 stock = this_cpu_ptr(&memcg_stock);
3038 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3039 stock->nr_bytes -= nr_bytes;
3043 local_irq_restore(flags);
3048 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3050 struct obj_cgroup *old = stock->cached_objcg;
3055 if (stock->nr_bytes) {
3056 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3057 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3061 __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
3066 * The leftover is flushed to the centralized per-memcg value.
3067 * On the next attempt to refill obj stock it will be moved
3068 * to a per-cpu stock (probably, on an other CPU), see
3069 * refill_obj_stock().
3071 * How often it's flushed is a trade-off between the memory
3072 * limit enforcement accuracy and potential CPU contention,
3073 * so it might be changed in the future.
3075 atomic_add(nr_bytes, &old->nr_charged_bytes);
3076 stock->nr_bytes = 0;
3079 obj_cgroup_put(old);
3080 stock->cached_objcg = NULL;
3083 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3084 struct mem_cgroup *root_memcg)
3086 struct mem_cgroup *memcg;
3088 if (stock->cached_objcg) {
3089 memcg = obj_cgroup_memcg(stock->cached_objcg);
3090 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3097 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3099 struct memcg_stock_pcp *stock;
3100 unsigned long flags;
3102 local_irq_save(flags);
3104 stock = this_cpu_ptr(&memcg_stock);
3105 if (stock->cached_objcg != objcg) { /* reset if necessary */
3106 drain_obj_stock(stock);
3107 obj_cgroup_get(objcg);
3108 stock->cached_objcg = objcg;
3109 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3111 stock->nr_bytes += nr_bytes;
3113 if (stock->nr_bytes > PAGE_SIZE)
3114 drain_obj_stock(stock);
3116 local_irq_restore(flags);
3119 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3121 struct mem_cgroup *memcg;
3122 unsigned int nr_pages, nr_bytes;
3125 if (consume_obj_stock(objcg, size))
3129 * In theory, memcg->nr_charged_bytes can have enough
3130 * pre-charged bytes to satisfy the allocation. However,
3131 * flushing memcg->nr_charged_bytes requires two atomic
3132 * operations, and memcg->nr_charged_bytes can't be big,
3133 * so it's better to ignore it and try grab some new pages.
3134 * memcg->nr_charged_bytes will be flushed in
3135 * refill_obj_stock(), called from this function or
3136 * independently later.
3139 memcg = obj_cgroup_memcg(objcg);
3140 css_get(&memcg->css);
3143 nr_pages = size >> PAGE_SHIFT;
3144 nr_bytes = size & (PAGE_SIZE - 1);
3149 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3150 if (!ret && nr_bytes)
3151 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3153 css_put(&memcg->css);
3157 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3159 refill_obj_stock(objcg, size);
3162 #endif /* CONFIG_MEMCG_KMEM */
3164 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3167 * Because tail pages are not marked as "used", set it. We're under
3168 * pgdat->lru_lock and migration entries setup in all page mappings.
3170 void mem_cgroup_split_huge_fixup(struct page *head)
3172 struct mem_cgroup *memcg = head->mem_cgroup;
3175 if (mem_cgroup_disabled())
3178 for (i = 1; i < HPAGE_PMD_NR; i++) {
3179 css_get(&memcg->css);
3180 head[i].mem_cgroup = memcg;
3183 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3185 #ifdef CONFIG_MEMCG_SWAP
3187 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3188 * @entry: swap entry to be moved
3189 * @from: mem_cgroup which the entry is moved from
3190 * @to: mem_cgroup which the entry is moved to
3192 * It succeeds only when the swap_cgroup's record for this entry is the same
3193 * as the mem_cgroup's id of @from.
3195 * Returns 0 on success, -EINVAL on failure.
3197 * The caller must have charged to @to, IOW, called page_counter_charge() about
3198 * both res and memsw, and called css_get().
3200 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3201 struct mem_cgroup *from, struct mem_cgroup *to)
3203 unsigned short old_id, new_id;
3205 old_id = mem_cgroup_id(from);
3206 new_id = mem_cgroup_id(to);
3208 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3209 mod_memcg_state(from, MEMCG_SWAP, -1);
3210 mod_memcg_state(to, MEMCG_SWAP, 1);
3216 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3217 struct mem_cgroup *from, struct mem_cgroup *to)
3223 static DEFINE_MUTEX(memcg_max_mutex);
3225 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3226 unsigned long max, bool memsw)
3228 bool enlarge = false;
3229 bool drained = false;
3231 bool limits_invariant;
3232 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3235 if (signal_pending(current)) {
3240 mutex_lock(&memcg_max_mutex);
3242 * Make sure that the new limit (memsw or memory limit) doesn't
3243 * break our basic invariant rule memory.max <= memsw.max.
3245 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3246 max <= memcg->memsw.max;
3247 if (!limits_invariant) {
3248 mutex_unlock(&memcg_max_mutex);
3252 if (max > counter->max)
3254 ret = page_counter_set_max(counter, max);
3255 mutex_unlock(&memcg_max_mutex);
3261 drain_all_stock(memcg);
3266 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3267 GFP_KERNEL, !memsw)) {
3273 if (!ret && enlarge)
3274 memcg_oom_recover(memcg);
3279 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3281 unsigned long *total_scanned)
3283 unsigned long nr_reclaimed = 0;
3284 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3285 unsigned long reclaimed;
3287 struct mem_cgroup_tree_per_node *mctz;
3288 unsigned long excess;
3289 unsigned long nr_scanned;
3294 mctz = soft_limit_tree_node(pgdat->node_id);
3297 * Do not even bother to check the largest node if the root
3298 * is empty. Do it lockless to prevent lock bouncing. Races
3299 * are acceptable as soft limit is best effort anyway.
3301 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3305 * This loop can run a while, specially if mem_cgroup's continuously
3306 * keep exceeding their soft limit and putting the system under
3313 mz = mem_cgroup_largest_soft_limit_node(mctz);
3318 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3319 gfp_mask, &nr_scanned);
3320 nr_reclaimed += reclaimed;
3321 *total_scanned += nr_scanned;
3322 spin_lock_irq(&mctz->lock);
3323 __mem_cgroup_remove_exceeded(mz, mctz);
3326 * If we failed to reclaim anything from this memory cgroup
3327 * it is time to move on to the next cgroup
3331 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3333 excess = soft_limit_excess(mz->memcg);
3335 * One school of thought says that we should not add
3336 * back the node to the tree if reclaim returns 0.
3337 * But our reclaim could return 0, simply because due
3338 * to priority we are exposing a smaller subset of
3339 * memory to reclaim from. Consider this as a longer
3342 /* If excess == 0, no tree ops */
3343 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3344 spin_unlock_irq(&mctz->lock);
3345 css_put(&mz->memcg->css);
3348 * Could not reclaim anything and there are no more
3349 * mem cgroups to try or we seem to be looping without
3350 * reclaiming anything.
3352 if (!nr_reclaimed &&
3354 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3356 } while (!nr_reclaimed);
3358 css_put(&next_mz->memcg->css);
3359 return nr_reclaimed;
3363 * Test whether @memcg has children, dead or alive. Note that this
3364 * function doesn't care whether @memcg has use_hierarchy enabled and
3365 * returns %true if there are child csses according to the cgroup
3366 * hierarchy. Testing use_hierarchy is the caller's responsibility.
3368 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3373 ret = css_next_child(NULL, &memcg->css);
3379 * Reclaims as many pages from the given memcg as possible.
3381 * Caller is responsible for holding css reference for memcg.
3383 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3385 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3387 /* we call try-to-free pages for make this cgroup empty */
3388 lru_add_drain_all();
3390 drain_all_stock(memcg);
3392 /* try to free all pages in this cgroup */
3393 while (nr_retries && page_counter_read(&memcg->memory)) {
3396 if (signal_pending(current))
3399 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3403 /* maybe some writeback is necessary */
3404 congestion_wait(BLK_RW_ASYNC, HZ/10);
3412 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3413 char *buf, size_t nbytes,
3416 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3418 if (mem_cgroup_is_root(memcg))
3420 return mem_cgroup_force_empty(memcg) ?: nbytes;
3423 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3426 return mem_cgroup_from_css(css)->use_hierarchy;
3429 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3430 struct cftype *cft, u64 val)
3433 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3434 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3436 if (memcg->use_hierarchy == val)
3440 * If parent's use_hierarchy is set, we can't make any modifications
3441 * in the child subtrees. If it is unset, then the change can
3442 * occur, provided the current cgroup has no children.
3444 * For the root cgroup, parent_mem is NULL, we allow value to be
3445 * set if there are no children.
3447 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3448 (val == 1 || val == 0)) {
3449 if (!memcg_has_children(memcg))
3450 memcg->use_hierarchy = val;
3459 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3463 if (mem_cgroup_is_root(memcg)) {
3464 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3465 memcg_page_state(memcg, NR_ANON_MAPPED);
3467 val += memcg_page_state(memcg, MEMCG_SWAP);
3470 val = page_counter_read(&memcg->memory);
3472 val = page_counter_read(&memcg->memsw);
3485 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3488 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3489 struct page_counter *counter;
3491 switch (MEMFILE_TYPE(cft->private)) {
3493 counter = &memcg->memory;
3496 counter = &memcg->memsw;
3499 counter = &memcg->kmem;
3502 counter = &memcg->tcpmem;
3508 switch (MEMFILE_ATTR(cft->private)) {
3510 if (counter == &memcg->memory)
3511 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3512 if (counter == &memcg->memsw)
3513 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3514 return (u64)page_counter_read(counter) * PAGE_SIZE;
3516 return (u64)counter->max * PAGE_SIZE;
3518 return (u64)counter->watermark * PAGE_SIZE;
3520 return counter->failcnt;
3521 case RES_SOFT_LIMIT:
3522 return (u64)memcg->soft_limit * PAGE_SIZE;
3528 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3530 unsigned long stat[MEMCG_NR_STAT] = {0};
3531 struct mem_cgroup *mi;
3534 for_each_online_cpu(cpu)
3535 for (i = 0; i < MEMCG_NR_STAT; i++)
3536 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3538 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3539 for (i = 0; i < MEMCG_NR_STAT; i++)
3540 atomic_long_add(stat[i], &mi->vmstats[i]);
3542 for_each_node(node) {
3543 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3544 struct mem_cgroup_per_node *pi;
3546 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3549 for_each_online_cpu(cpu)
3550 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3552 pn->lruvec_stat_cpu->count[i], cpu);
3554 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3555 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3556 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3560 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3562 unsigned long events[NR_VM_EVENT_ITEMS];
3563 struct mem_cgroup *mi;
3566 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3569 for_each_online_cpu(cpu)
3570 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3571 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3574 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3575 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3576 atomic_long_add(events[i], &mi->vmevents[i]);
3579 #ifdef CONFIG_MEMCG_KMEM
3580 static int memcg_online_kmem(struct mem_cgroup *memcg)
3582 struct obj_cgroup *objcg;
3585 if (cgroup_memory_nokmem)
3588 BUG_ON(memcg->kmemcg_id >= 0);
3589 BUG_ON(memcg->kmem_state);
3591 memcg_id = memcg_alloc_cache_id();
3595 objcg = obj_cgroup_alloc();
3597 memcg_free_cache_id(memcg_id);
3600 objcg->memcg = memcg;
3601 rcu_assign_pointer(memcg->objcg, objcg);
3603 static_branch_enable(&memcg_kmem_enabled_key);
3606 * A memory cgroup is considered kmem-online as soon as it gets
3607 * kmemcg_id. Setting the id after enabling static branching will
3608 * guarantee no one starts accounting before all call sites are
3611 memcg->kmemcg_id = memcg_id;
3612 memcg->kmem_state = KMEM_ONLINE;
3617 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3619 struct cgroup_subsys_state *css;
3620 struct mem_cgroup *parent, *child;
3623 if (memcg->kmem_state != KMEM_ONLINE)
3626 memcg->kmem_state = KMEM_ALLOCATED;
3628 parent = parent_mem_cgroup(memcg);
3630 parent = root_mem_cgroup;
3632 memcg_reparent_objcgs(memcg, parent);
3634 kmemcg_id = memcg->kmemcg_id;
3635 BUG_ON(kmemcg_id < 0);
3638 * Change kmemcg_id of this cgroup and all its descendants to the
3639 * parent's id, and then move all entries from this cgroup's list_lrus
3640 * to ones of the parent. After we have finished, all list_lrus
3641 * corresponding to this cgroup are guaranteed to remain empty. The
3642 * ordering is imposed by list_lru_node->lock taken by
3643 * memcg_drain_all_list_lrus().
3645 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3646 css_for_each_descendant_pre(css, &memcg->css) {
3647 child = mem_cgroup_from_css(css);
3648 BUG_ON(child->kmemcg_id != kmemcg_id);
3649 child->kmemcg_id = parent->kmemcg_id;
3650 if (!memcg->use_hierarchy)
3655 memcg_drain_all_list_lrus(kmemcg_id, parent);
3657 memcg_free_cache_id(kmemcg_id);
3660 static void memcg_free_kmem(struct mem_cgroup *memcg)
3662 /* css_alloc() failed, offlining didn't happen */
3663 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3664 memcg_offline_kmem(memcg);
3667 static int memcg_online_kmem(struct mem_cgroup *memcg)
3671 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3674 static void memcg_free_kmem(struct mem_cgroup *memcg)
3677 #endif /* CONFIG_MEMCG_KMEM */
3679 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3684 mutex_lock(&memcg_max_mutex);
3685 ret = page_counter_set_max(&memcg->kmem, max);
3686 mutex_unlock(&memcg_max_mutex);
3690 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3694 mutex_lock(&memcg_max_mutex);
3696 ret = page_counter_set_max(&memcg->tcpmem, max);
3700 if (!memcg->tcpmem_active) {
3702 * The active flag needs to be written after the static_key
3703 * update. This is what guarantees that the socket activation
3704 * function is the last one to run. See mem_cgroup_sk_alloc()
3705 * for details, and note that we don't mark any socket as
3706 * belonging to this memcg until that flag is up.
3708 * We need to do this, because static_keys will span multiple
3709 * sites, but we can't control their order. If we mark a socket
3710 * as accounted, but the accounting functions are not patched in
3711 * yet, we'll lose accounting.
3713 * We never race with the readers in mem_cgroup_sk_alloc(),
3714 * because when this value change, the code to process it is not
3717 static_branch_inc(&memcg_sockets_enabled_key);
3718 memcg->tcpmem_active = true;
3721 mutex_unlock(&memcg_max_mutex);
3726 * The user of this function is...
3729 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3730 char *buf, size_t nbytes, loff_t off)
3732 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3733 unsigned long nr_pages;
3736 buf = strstrip(buf);
3737 ret = page_counter_memparse(buf, "-1", &nr_pages);
3741 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3743 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3747 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3749 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3752 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3755 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3756 "Please report your usecase to linux-mm@kvack.org if you "
3757 "depend on this functionality.\n");
3758 ret = memcg_update_kmem_max(memcg, nr_pages);
3761 ret = memcg_update_tcp_max(memcg, nr_pages);
3765 case RES_SOFT_LIMIT:
3766 memcg->soft_limit = nr_pages;
3770 return ret ?: nbytes;
3773 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3774 size_t nbytes, loff_t off)
3776 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3777 struct page_counter *counter;
3779 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3781 counter = &memcg->memory;
3784 counter = &memcg->memsw;
3787 counter = &memcg->kmem;
3790 counter = &memcg->tcpmem;
3796 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3798 page_counter_reset_watermark(counter);
3801 counter->failcnt = 0;
3810 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3813 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3817 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3818 struct cftype *cft, u64 val)
3820 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3822 if (val & ~MOVE_MASK)
3826 * No kind of locking is needed in here, because ->can_attach() will
3827 * check this value once in the beginning of the process, and then carry
3828 * on with stale data. This means that changes to this value will only
3829 * affect task migrations starting after the change.
3831 memcg->move_charge_at_immigrate = val;
3835 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3836 struct cftype *cft, u64 val)
3844 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3845 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3846 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3848 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3849 int nid, unsigned int lru_mask, bool tree)
3851 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3852 unsigned long nr = 0;
3855 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3858 if (!(BIT(lru) & lru_mask))
3861 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3863 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3868 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3869 unsigned int lru_mask,
3872 unsigned long nr = 0;
3876 if (!(BIT(lru) & lru_mask))
3879 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3881 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3886 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3890 unsigned int lru_mask;
3893 static const struct numa_stat stats[] = {
3894 { "total", LRU_ALL },
3895 { "file", LRU_ALL_FILE },
3896 { "anon", LRU_ALL_ANON },
3897 { "unevictable", BIT(LRU_UNEVICTABLE) },
3899 const struct numa_stat *stat;
3901 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3903 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3904 seq_printf(m, "%s=%lu", stat->name,
3905 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3907 for_each_node_state(nid, N_MEMORY)
3908 seq_printf(m, " N%d=%lu", nid,
3909 mem_cgroup_node_nr_lru_pages(memcg, nid,
3910 stat->lru_mask, false));
3914 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3916 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3917 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3919 for_each_node_state(nid, N_MEMORY)
3920 seq_printf(m, " N%d=%lu", nid,
3921 mem_cgroup_node_nr_lru_pages(memcg, nid,
3922 stat->lru_mask, true));
3928 #endif /* CONFIG_NUMA */
3930 static const unsigned int memcg1_stats[] = {
3933 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3943 static const char *const memcg1_stat_names[] = {
3946 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3956 /* Universal VM events cgroup1 shows, original sort order */
3957 static const unsigned int memcg1_events[] = {
3964 static int memcg_stat_show(struct seq_file *m, void *v)
3966 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3967 unsigned long memory, memsw;
3968 struct mem_cgroup *mi;
3971 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3973 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3976 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3978 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
3979 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3980 if (memcg1_stats[i] == NR_ANON_THPS)
3983 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
3986 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3987 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
3988 memcg_events_local(memcg, memcg1_events[i]));
3990 for (i = 0; i < NR_LRU_LISTS; i++)
3991 seq_printf(m, "%s %lu\n", lru_list_name(i),
3992 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3995 /* Hierarchical information */
3996 memory = memsw = PAGE_COUNTER_MAX;
3997 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3998 memory = min(memory, READ_ONCE(mi->memory.max));
3999 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4001 seq_printf(m, "hierarchical_memory_limit %llu\n",
4002 (u64)memory * PAGE_SIZE);
4003 if (do_memsw_account())
4004 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4005 (u64)memsw * PAGE_SIZE);
4007 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4008 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4010 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4011 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
4015 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4016 seq_printf(m, "total_%s %llu\n",
4017 vm_event_name(memcg1_events[i]),
4018 (u64)memcg_events(memcg, memcg1_events[i]));
4020 for (i = 0; i < NR_LRU_LISTS; i++)
4021 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4022 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4025 #ifdef CONFIG_DEBUG_VM
4028 struct mem_cgroup_per_node *mz;
4029 unsigned long anon_cost = 0;
4030 unsigned long file_cost = 0;
4032 for_each_online_pgdat(pgdat) {
4033 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4035 anon_cost += mz->lruvec.anon_cost;
4036 file_cost += mz->lruvec.file_cost;
4038 seq_printf(m, "anon_cost %lu\n", anon_cost);
4039 seq_printf(m, "file_cost %lu\n", file_cost);
4046 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4049 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4051 return mem_cgroup_swappiness(memcg);
4054 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4055 struct cftype *cft, u64 val)
4057 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4063 memcg->swappiness = val;
4065 vm_swappiness = val;
4070 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4072 struct mem_cgroup_threshold_ary *t;
4073 unsigned long usage;
4078 t = rcu_dereference(memcg->thresholds.primary);
4080 t = rcu_dereference(memcg->memsw_thresholds.primary);
4085 usage = mem_cgroup_usage(memcg, swap);
4088 * current_threshold points to threshold just below or equal to usage.
4089 * If it's not true, a threshold was crossed after last
4090 * call of __mem_cgroup_threshold().
4092 i = t->current_threshold;
4095 * Iterate backward over array of thresholds starting from
4096 * current_threshold and check if a threshold is crossed.
4097 * If none of thresholds below usage is crossed, we read
4098 * only one element of the array here.
4100 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4101 eventfd_signal(t->entries[i].eventfd, 1);
4103 /* i = current_threshold + 1 */
4107 * Iterate forward over array of thresholds starting from
4108 * current_threshold+1 and check if a threshold is crossed.
4109 * If none of thresholds above usage is crossed, we read
4110 * only one element of the array here.
4112 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4113 eventfd_signal(t->entries[i].eventfd, 1);
4115 /* Update current_threshold */
4116 t->current_threshold = i - 1;
4121 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4124 __mem_cgroup_threshold(memcg, false);
4125 if (do_memsw_account())
4126 __mem_cgroup_threshold(memcg, true);
4128 memcg = parent_mem_cgroup(memcg);
4132 static int compare_thresholds(const void *a, const void *b)
4134 const struct mem_cgroup_threshold *_a = a;
4135 const struct mem_cgroup_threshold *_b = b;
4137 if (_a->threshold > _b->threshold)
4140 if (_a->threshold < _b->threshold)
4146 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4148 struct mem_cgroup_eventfd_list *ev;
4150 spin_lock(&memcg_oom_lock);
4152 list_for_each_entry(ev, &memcg->oom_notify, list)
4153 eventfd_signal(ev->eventfd, 1);
4155 spin_unlock(&memcg_oom_lock);
4159 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4161 struct mem_cgroup *iter;
4163 for_each_mem_cgroup_tree(iter, memcg)
4164 mem_cgroup_oom_notify_cb(iter);
4167 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4168 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4170 struct mem_cgroup_thresholds *thresholds;
4171 struct mem_cgroup_threshold_ary *new;
4172 unsigned long threshold;
4173 unsigned long usage;
4176 ret = page_counter_memparse(args, "-1", &threshold);
4180 mutex_lock(&memcg->thresholds_lock);
4183 thresholds = &memcg->thresholds;
4184 usage = mem_cgroup_usage(memcg, false);
4185 } else if (type == _MEMSWAP) {
4186 thresholds = &memcg->memsw_thresholds;
4187 usage = mem_cgroup_usage(memcg, true);
4191 /* Check if a threshold crossed before adding a new one */
4192 if (thresholds->primary)
4193 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4195 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4197 /* Allocate memory for new array of thresholds */
4198 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4205 /* Copy thresholds (if any) to new array */
4206 if (thresholds->primary) {
4207 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4208 sizeof(struct mem_cgroup_threshold));
4211 /* Add new threshold */
4212 new->entries[size - 1].eventfd = eventfd;
4213 new->entries[size - 1].threshold = threshold;
4215 /* Sort thresholds. Registering of new threshold isn't time-critical */
4216 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4217 compare_thresholds, NULL);
4219 /* Find current threshold */
4220 new->current_threshold = -1;
4221 for (i = 0; i < size; i++) {
4222 if (new->entries[i].threshold <= usage) {
4224 * new->current_threshold will not be used until
4225 * rcu_assign_pointer(), so it's safe to increment
4228 ++new->current_threshold;
4233 /* Free old spare buffer and save old primary buffer as spare */
4234 kfree(thresholds->spare);
4235 thresholds->spare = thresholds->primary;
4237 rcu_assign_pointer(thresholds->primary, new);
4239 /* To be sure that nobody uses thresholds */
4243 mutex_unlock(&memcg->thresholds_lock);
4248 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4249 struct eventfd_ctx *eventfd, const char *args)
4251 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4254 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4255 struct eventfd_ctx *eventfd, const char *args)
4257 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4260 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4261 struct eventfd_ctx *eventfd, enum res_type type)
4263 struct mem_cgroup_thresholds *thresholds;
4264 struct mem_cgroup_threshold_ary *new;
4265 unsigned long usage;
4266 int i, j, size, entries;
4268 mutex_lock(&memcg->thresholds_lock);
4271 thresholds = &memcg->thresholds;
4272 usage = mem_cgroup_usage(memcg, false);
4273 } else if (type == _MEMSWAP) {
4274 thresholds = &memcg->memsw_thresholds;
4275 usage = mem_cgroup_usage(memcg, true);
4279 if (!thresholds->primary)
4282 /* Check if a threshold crossed before removing */
4283 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4285 /* Calculate new number of threshold */
4287 for (i = 0; i < thresholds->primary->size; i++) {
4288 if (thresholds->primary->entries[i].eventfd != eventfd)
4294 new = thresholds->spare;
4296 /* If no items related to eventfd have been cleared, nothing to do */
4300 /* Set thresholds array to NULL if we don't have thresholds */
4309 /* Copy thresholds and find current threshold */
4310 new->current_threshold = -1;
4311 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4312 if (thresholds->primary->entries[i].eventfd == eventfd)
4315 new->entries[j] = thresholds->primary->entries[i];
4316 if (new->entries[j].threshold <= usage) {
4318 * new->current_threshold will not be used
4319 * until rcu_assign_pointer(), so it's safe to increment
4322 ++new->current_threshold;
4328 /* Swap primary and spare array */
4329 thresholds->spare = thresholds->primary;
4331 rcu_assign_pointer(thresholds->primary, new);
4333 /* To be sure that nobody uses thresholds */
4336 /* If all events are unregistered, free the spare array */
4338 kfree(thresholds->spare);
4339 thresholds->spare = NULL;
4342 mutex_unlock(&memcg->thresholds_lock);
4345 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4346 struct eventfd_ctx *eventfd)
4348 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4351 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4352 struct eventfd_ctx *eventfd)
4354 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4357 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4358 struct eventfd_ctx *eventfd, const char *args)
4360 struct mem_cgroup_eventfd_list *event;
4362 event = kmalloc(sizeof(*event), GFP_KERNEL);
4366 spin_lock(&memcg_oom_lock);
4368 event->eventfd = eventfd;
4369 list_add(&event->list, &memcg->oom_notify);
4371 /* already in OOM ? */
4372 if (memcg->under_oom)
4373 eventfd_signal(eventfd, 1);
4374 spin_unlock(&memcg_oom_lock);
4379 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4380 struct eventfd_ctx *eventfd)
4382 struct mem_cgroup_eventfd_list *ev, *tmp;
4384 spin_lock(&memcg_oom_lock);
4386 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4387 if (ev->eventfd == eventfd) {
4388 list_del(&ev->list);
4393 spin_unlock(&memcg_oom_lock);
4396 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4398 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4400 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4401 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4402 seq_printf(sf, "oom_kill %lu\n",
4403 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4407 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4408 struct cftype *cft, u64 val)
4410 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4412 /* cannot set to root cgroup and only 0 and 1 are allowed */
4413 if (!css->parent || !((val == 0) || (val == 1)))
4416 memcg->oom_kill_disable = val;
4418 memcg_oom_recover(memcg);
4423 #ifdef CONFIG_CGROUP_WRITEBACK
4425 #include <trace/events/writeback.h>
4427 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4429 return wb_domain_init(&memcg->cgwb_domain, gfp);
4432 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4434 wb_domain_exit(&memcg->cgwb_domain);
4437 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4439 wb_domain_size_changed(&memcg->cgwb_domain);
4442 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4444 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4446 if (!memcg->css.parent)
4449 return &memcg->cgwb_domain;
4453 * idx can be of type enum memcg_stat_item or node_stat_item.
4454 * Keep in sync with memcg_exact_page().
4456 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4458 long x = atomic_long_read(&memcg->vmstats[idx]);
4461 for_each_online_cpu(cpu)
4462 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4469 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4470 * @wb: bdi_writeback in question
4471 * @pfilepages: out parameter for number of file pages
4472 * @pheadroom: out parameter for number of allocatable pages according to memcg
4473 * @pdirty: out parameter for number of dirty pages
4474 * @pwriteback: out parameter for number of pages under writeback
4476 * Determine the numbers of file, headroom, dirty, and writeback pages in
4477 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4478 * is a bit more involved.
4480 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4481 * headroom is calculated as the lowest headroom of itself and the
4482 * ancestors. Note that this doesn't consider the actual amount of
4483 * available memory in the system. The caller should further cap
4484 * *@pheadroom accordingly.
4486 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4487 unsigned long *pheadroom, unsigned long *pdirty,
4488 unsigned long *pwriteback)
4490 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4491 struct mem_cgroup *parent;
4493 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4495 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4496 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4497 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4498 *pheadroom = PAGE_COUNTER_MAX;
4500 while ((parent = parent_mem_cgroup(memcg))) {
4501 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4502 READ_ONCE(memcg->memory.high));
4503 unsigned long used = page_counter_read(&memcg->memory);
4505 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4511 * Foreign dirty flushing
4513 * There's an inherent mismatch between memcg and writeback. The former
4514 * trackes ownership per-page while the latter per-inode. This was a
4515 * deliberate design decision because honoring per-page ownership in the
4516 * writeback path is complicated, may lead to higher CPU and IO overheads
4517 * and deemed unnecessary given that write-sharing an inode across
4518 * different cgroups isn't a common use-case.
4520 * Combined with inode majority-writer ownership switching, this works well
4521 * enough in most cases but there are some pathological cases. For
4522 * example, let's say there are two cgroups A and B which keep writing to
4523 * different but confined parts of the same inode. B owns the inode and
4524 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4525 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4526 * triggering background writeback. A will be slowed down without a way to
4527 * make writeback of the dirty pages happen.
4529 * Conditions like the above can lead to a cgroup getting repatedly and
4530 * severely throttled after making some progress after each
4531 * dirty_expire_interval while the underyling IO device is almost
4534 * Solving this problem completely requires matching the ownership tracking
4535 * granularities between memcg and writeback in either direction. However,
4536 * the more egregious behaviors can be avoided by simply remembering the
4537 * most recent foreign dirtying events and initiating remote flushes on
4538 * them when local writeback isn't enough to keep the memory clean enough.
4540 * The following two functions implement such mechanism. When a foreign
4541 * page - a page whose memcg and writeback ownerships don't match - is
4542 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4543 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4544 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4545 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4546 * foreign bdi_writebacks which haven't expired. Both the numbers of
4547 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4548 * limited to MEMCG_CGWB_FRN_CNT.
4550 * The mechanism only remembers IDs and doesn't hold any object references.
4551 * As being wrong occasionally doesn't matter, updates and accesses to the
4552 * records are lockless and racy.
4554 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4555 struct bdi_writeback *wb)
4557 struct mem_cgroup *memcg = page->mem_cgroup;
4558 struct memcg_cgwb_frn *frn;
4559 u64 now = get_jiffies_64();
4560 u64 oldest_at = now;
4564 trace_track_foreign_dirty(page, wb);
4567 * Pick the slot to use. If there is already a slot for @wb, keep
4568 * using it. If not replace the oldest one which isn't being
4571 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4572 frn = &memcg->cgwb_frn[i];
4573 if (frn->bdi_id == wb->bdi->id &&
4574 frn->memcg_id == wb->memcg_css->id)
4576 if (time_before64(frn->at, oldest_at) &&
4577 atomic_read(&frn->done.cnt) == 1) {
4579 oldest_at = frn->at;
4583 if (i < MEMCG_CGWB_FRN_CNT) {
4585 * Re-using an existing one. Update timestamp lazily to
4586 * avoid making the cacheline hot. We want them to be
4587 * reasonably up-to-date and significantly shorter than
4588 * dirty_expire_interval as that's what expires the record.
4589 * Use the shorter of 1s and dirty_expire_interval / 8.
4591 unsigned long update_intv =
4592 min_t(unsigned long, HZ,
4593 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4595 if (time_before64(frn->at, now - update_intv))
4597 } else if (oldest >= 0) {
4598 /* replace the oldest free one */
4599 frn = &memcg->cgwb_frn[oldest];
4600 frn->bdi_id = wb->bdi->id;
4601 frn->memcg_id = wb->memcg_css->id;
4606 /* issue foreign writeback flushes for recorded foreign dirtying events */
4607 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4609 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4610 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4611 u64 now = jiffies_64;
4614 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4615 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4618 * If the record is older than dirty_expire_interval,
4619 * writeback on it has already started. No need to kick it
4620 * off again. Also, don't start a new one if there's
4621 * already one in flight.
4623 if (time_after64(frn->at, now - intv) &&
4624 atomic_read(&frn->done.cnt) == 1) {
4626 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4627 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4628 WB_REASON_FOREIGN_FLUSH,
4634 #else /* CONFIG_CGROUP_WRITEBACK */
4636 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4641 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4645 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4649 #endif /* CONFIG_CGROUP_WRITEBACK */
4652 * DO NOT USE IN NEW FILES.
4654 * "cgroup.event_control" implementation.
4656 * This is way over-engineered. It tries to support fully configurable
4657 * events for each user. Such level of flexibility is completely
4658 * unnecessary especially in the light of the planned unified hierarchy.
4660 * Please deprecate this and replace with something simpler if at all
4665 * Unregister event and free resources.
4667 * Gets called from workqueue.
4669 static void memcg_event_remove(struct work_struct *work)
4671 struct mem_cgroup_event *event =
4672 container_of(work, struct mem_cgroup_event, remove);
4673 struct mem_cgroup *memcg = event->memcg;
4675 remove_wait_queue(event->wqh, &event->wait);
4677 event->unregister_event(memcg, event->eventfd);
4679 /* Notify userspace the event is going away. */
4680 eventfd_signal(event->eventfd, 1);
4682 eventfd_ctx_put(event->eventfd);
4684 css_put(&memcg->css);
4688 * Gets called on EPOLLHUP on eventfd when user closes it.
4690 * Called with wqh->lock held and interrupts disabled.
4692 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4693 int sync, void *key)
4695 struct mem_cgroup_event *event =
4696 container_of(wait, struct mem_cgroup_event, wait);
4697 struct mem_cgroup *memcg = event->memcg;
4698 __poll_t flags = key_to_poll(key);
4700 if (flags & EPOLLHUP) {
4702 * If the event has been detached at cgroup removal, we
4703 * can simply return knowing the other side will cleanup
4706 * We can't race against event freeing since the other
4707 * side will require wqh->lock via remove_wait_queue(),
4710 spin_lock(&memcg->event_list_lock);
4711 if (!list_empty(&event->list)) {
4712 list_del_init(&event->list);
4714 * We are in atomic context, but cgroup_event_remove()
4715 * may sleep, so we have to call it in workqueue.
4717 schedule_work(&event->remove);
4719 spin_unlock(&memcg->event_list_lock);
4725 static void memcg_event_ptable_queue_proc(struct file *file,
4726 wait_queue_head_t *wqh, poll_table *pt)
4728 struct mem_cgroup_event *event =
4729 container_of(pt, struct mem_cgroup_event, pt);
4732 add_wait_queue(wqh, &event->wait);
4736 * DO NOT USE IN NEW FILES.
4738 * Parse input and register new cgroup event handler.
4740 * Input must be in format '<event_fd> <control_fd> <args>'.
4741 * Interpretation of args is defined by control file implementation.
4743 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4744 char *buf, size_t nbytes, loff_t off)
4746 struct cgroup_subsys_state *css = of_css(of);
4747 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4748 struct mem_cgroup_event *event;
4749 struct cgroup_subsys_state *cfile_css;
4750 unsigned int efd, cfd;
4757 buf = strstrip(buf);
4759 efd = simple_strtoul(buf, &endp, 10);
4764 cfd = simple_strtoul(buf, &endp, 10);
4765 if ((*endp != ' ') && (*endp != '\0'))
4769 event = kzalloc(sizeof(*event), GFP_KERNEL);
4773 event->memcg = memcg;
4774 INIT_LIST_HEAD(&event->list);
4775 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4776 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4777 INIT_WORK(&event->remove, memcg_event_remove);
4785 event->eventfd = eventfd_ctx_fileget(efile.file);
4786 if (IS_ERR(event->eventfd)) {
4787 ret = PTR_ERR(event->eventfd);
4794 goto out_put_eventfd;
4797 /* the process need read permission on control file */
4798 /* AV: shouldn't we check that it's been opened for read instead? */
4799 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4804 * Determine the event callbacks and set them in @event. This used
4805 * to be done via struct cftype but cgroup core no longer knows
4806 * about these events. The following is crude but the whole thing
4807 * is for compatibility anyway.
4809 * DO NOT ADD NEW FILES.
4811 name = cfile.file->f_path.dentry->d_name.name;
4813 if (!strcmp(name, "memory.usage_in_bytes")) {
4814 event->register_event = mem_cgroup_usage_register_event;
4815 event->unregister_event = mem_cgroup_usage_unregister_event;
4816 } else if (!strcmp(name, "memory.oom_control")) {
4817 event->register_event = mem_cgroup_oom_register_event;
4818 event->unregister_event = mem_cgroup_oom_unregister_event;
4819 } else if (!strcmp(name, "memory.pressure_level")) {
4820 event->register_event = vmpressure_register_event;
4821 event->unregister_event = vmpressure_unregister_event;
4822 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4823 event->register_event = memsw_cgroup_usage_register_event;
4824 event->unregister_event = memsw_cgroup_usage_unregister_event;
4831 * Verify @cfile should belong to @css. Also, remaining events are
4832 * automatically removed on cgroup destruction but the removal is
4833 * asynchronous, so take an extra ref on @css.
4835 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4836 &memory_cgrp_subsys);
4838 if (IS_ERR(cfile_css))
4840 if (cfile_css != css) {
4845 ret = event->register_event(memcg, event->eventfd, buf);
4849 vfs_poll(efile.file, &event->pt);
4851 spin_lock(&memcg->event_list_lock);
4852 list_add(&event->list, &memcg->event_list);
4853 spin_unlock(&memcg->event_list_lock);
4865 eventfd_ctx_put(event->eventfd);
4874 static struct cftype mem_cgroup_legacy_files[] = {
4876 .name = "usage_in_bytes",
4877 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4878 .read_u64 = mem_cgroup_read_u64,
4881 .name = "max_usage_in_bytes",
4882 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4883 .write = mem_cgroup_reset,
4884 .read_u64 = mem_cgroup_read_u64,
4887 .name = "limit_in_bytes",
4888 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4889 .write = mem_cgroup_write,
4890 .read_u64 = mem_cgroup_read_u64,
4893 .name = "soft_limit_in_bytes",
4894 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4895 .write = mem_cgroup_write,
4896 .read_u64 = mem_cgroup_read_u64,
4900 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4901 .write = mem_cgroup_reset,
4902 .read_u64 = mem_cgroup_read_u64,
4906 .seq_show = memcg_stat_show,
4909 .name = "force_empty",
4910 .write = mem_cgroup_force_empty_write,
4913 .name = "use_hierarchy",
4914 .write_u64 = mem_cgroup_hierarchy_write,
4915 .read_u64 = mem_cgroup_hierarchy_read,
4918 .name = "cgroup.event_control", /* XXX: for compat */
4919 .write = memcg_write_event_control,
4920 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4923 .name = "swappiness",
4924 .read_u64 = mem_cgroup_swappiness_read,
4925 .write_u64 = mem_cgroup_swappiness_write,
4928 .name = "move_charge_at_immigrate",
4929 .read_u64 = mem_cgroup_move_charge_read,
4930 .write_u64 = mem_cgroup_move_charge_write,
4933 .name = "oom_control",
4934 .seq_show = mem_cgroup_oom_control_read,
4935 .write_u64 = mem_cgroup_oom_control_write,
4936 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4939 .name = "pressure_level",
4943 .name = "numa_stat",
4944 .seq_show = memcg_numa_stat_show,
4948 .name = "kmem.limit_in_bytes",
4949 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4950 .write = mem_cgroup_write,
4951 .read_u64 = mem_cgroup_read_u64,
4954 .name = "kmem.usage_in_bytes",
4955 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4956 .read_u64 = mem_cgroup_read_u64,
4959 .name = "kmem.failcnt",
4960 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4961 .write = mem_cgroup_reset,
4962 .read_u64 = mem_cgroup_read_u64,
4965 .name = "kmem.max_usage_in_bytes",
4966 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4967 .write = mem_cgroup_reset,
4968 .read_u64 = mem_cgroup_read_u64,
4970 #if defined(CONFIG_MEMCG_KMEM) && \
4971 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4973 .name = "kmem.slabinfo",
4974 .seq_show = memcg_slab_show,
4978 .name = "kmem.tcp.limit_in_bytes",
4979 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4980 .write = mem_cgroup_write,
4981 .read_u64 = mem_cgroup_read_u64,
4984 .name = "kmem.tcp.usage_in_bytes",
4985 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4986 .read_u64 = mem_cgroup_read_u64,
4989 .name = "kmem.tcp.failcnt",
4990 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4991 .write = mem_cgroup_reset,
4992 .read_u64 = mem_cgroup_read_u64,
4995 .name = "kmem.tcp.max_usage_in_bytes",
4996 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4997 .write = mem_cgroup_reset,
4998 .read_u64 = mem_cgroup_read_u64,
5000 { }, /* terminate */
5004 * Private memory cgroup IDR
5006 * Swap-out records and page cache shadow entries need to store memcg
5007 * references in constrained space, so we maintain an ID space that is
5008 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5009 * memory-controlled cgroups to 64k.
5011 * However, there usually are many references to the offline CSS after
5012 * the cgroup has been destroyed, such as page cache or reclaimable
5013 * slab objects, that don't need to hang on to the ID. We want to keep
5014 * those dead CSS from occupying IDs, or we might quickly exhaust the
5015 * relatively small ID space and prevent the creation of new cgroups
5016 * even when there are much fewer than 64k cgroups - possibly none.
5018 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5019 * be freed and recycled when it's no longer needed, which is usually
5020 * when the CSS is offlined.
5022 * The only exception to that are records of swapped out tmpfs/shmem
5023 * pages that need to be attributed to live ancestors on swapin. But
5024 * those references are manageable from userspace.
5027 static DEFINE_IDR(mem_cgroup_idr);
5029 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5031 if (memcg->id.id > 0) {
5032 idr_remove(&mem_cgroup_idr, memcg->id.id);
5037 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5040 refcount_add(n, &memcg->id.ref);
5043 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5045 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5046 mem_cgroup_id_remove(memcg);
5048 /* Memcg ID pins CSS */
5049 css_put(&memcg->css);
5053 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5055 mem_cgroup_id_put_many(memcg, 1);
5059 * mem_cgroup_from_id - look up a memcg from a memcg id
5060 * @id: the memcg id to look up
5062 * Caller must hold rcu_read_lock().
5064 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5066 WARN_ON_ONCE(!rcu_read_lock_held());
5067 return idr_find(&mem_cgroup_idr, id);
5070 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5072 struct mem_cgroup_per_node *pn;
5075 * This routine is called against possible nodes.
5076 * But it's BUG to call kmalloc() against offline node.
5078 * TODO: this routine can waste much memory for nodes which will
5079 * never be onlined. It's better to use memory hotplug callback
5082 if (!node_state(node, N_NORMAL_MEMORY))
5084 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5088 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
5089 if (!pn->lruvec_stat_local) {
5094 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
5095 if (!pn->lruvec_stat_cpu) {
5096 free_percpu(pn->lruvec_stat_local);
5101 lruvec_init(&pn->lruvec);
5102 pn->usage_in_excess = 0;
5103 pn->on_tree = false;
5106 memcg->nodeinfo[node] = pn;
5110 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5112 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5117 free_percpu(pn->lruvec_stat_cpu);
5118 free_percpu(pn->lruvec_stat_local);
5122 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5127 free_mem_cgroup_per_node_info(memcg, node);
5128 free_percpu(memcg->vmstats_percpu);
5129 free_percpu(memcg->vmstats_local);
5133 static void mem_cgroup_free(struct mem_cgroup *memcg)
5135 memcg_wb_domain_exit(memcg);
5137 * Flush percpu vmstats and vmevents to guarantee the value correctness
5138 * on parent's and all ancestor levels.
5140 memcg_flush_percpu_vmstats(memcg);
5141 memcg_flush_percpu_vmevents(memcg);
5142 __mem_cgroup_free(memcg);
5145 static struct mem_cgroup *mem_cgroup_alloc(void)
5147 struct mem_cgroup *memcg;
5150 int __maybe_unused i;
5151 long error = -ENOMEM;
5153 size = sizeof(struct mem_cgroup);
5154 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5156 memcg = kzalloc(size, GFP_KERNEL);
5158 return ERR_PTR(error);
5160 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5161 1, MEM_CGROUP_ID_MAX,
5163 if (memcg->id.id < 0) {
5164 error = memcg->id.id;
5168 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
5169 if (!memcg->vmstats_local)
5172 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
5173 if (!memcg->vmstats_percpu)
5177 if (alloc_mem_cgroup_per_node_info(memcg, node))
5180 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5183 INIT_WORK(&memcg->high_work, high_work_func);
5184 INIT_LIST_HEAD(&memcg->oom_notify);
5185 mutex_init(&memcg->thresholds_lock);
5186 spin_lock_init(&memcg->move_lock);
5187 vmpressure_init(&memcg->vmpressure);
5188 INIT_LIST_HEAD(&memcg->event_list);
5189 spin_lock_init(&memcg->event_list_lock);
5190 memcg->socket_pressure = jiffies;
5191 #ifdef CONFIG_MEMCG_KMEM
5192 memcg->kmemcg_id = -1;
5193 INIT_LIST_HEAD(&memcg->objcg_list);
5195 #ifdef CONFIG_CGROUP_WRITEBACK
5196 INIT_LIST_HEAD(&memcg->cgwb_list);
5197 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5198 memcg->cgwb_frn[i].done =
5199 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5201 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5202 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5203 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5204 memcg->deferred_split_queue.split_queue_len = 0;
5206 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5209 mem_cgroup_id_remove(memcg);
5210 __mem_cgroup_free(memcg);
5211 return ERR_PTR(error);
5214 static struct cgroup_subsys_state * __ref
5215 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5217 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5218 struct mem_cgroup *memcg;
5219 long error = -ENOMEM;
5221 memcg = mem_cgroup_alloc();
5223 return ERR_CAST(memcg);
5225 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5226 memcg->soft_limit = PAGE_COUNTER_MAX;
5227 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5229 memcg->swappiness = mem_cgroup_swappiness(parent);
5230 memcg->oom_kill_disable = parent->oom_kill_disable;
5232 if (parent && parent->use_hierarchy) {
5233 memcg->use_hierarchy = true;
5234 page_counter_init(&memcg->memory, &parent->memory);
5235 page_counter_init(&memcg->swap, &parent->swap);
5236 page_counter_init(&memcg->memsw, &parent->memsw);
5237 page_counter_init(&memcg->kmem, &parent->kmem);
5238 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5240 page_counter_init(&memcg->memory, NULL);
5241 page_counter_init(&memcg->swap, NULL);
5242 page_counter_init(&memcg->memsw, NULL);
5243 page_counter_init(&memcg->kmem, NULL);
5244 page_counter_init(&memcg->tcpmem, NULL);
5246 * Deeper hierachy with use_hierarchy == false doesn't make
5247 * much sense so let cgroup subsystem know about this
5248 * unfortunate state in our controller.
5250 if (parent != root_mem_cgroup)
5251 memory_cgrp_subsys.broken_hierarchy = true;
5254 /* The following stuff does not apply to the root */
5256 root_mem_cgroup = memcg;
5260 error = memcg_online_kmem(memcg);
5264 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5265 static_branch_inc(&memcg_sockets_enabled_key);
5269 mem_cgroup_id_remove(memcg);
5270 mem_cgroup_free(memcg);
5271 return ERR_PTR(error);
5274 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5276 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5279 * A memcg must be visible for memcg_expand_shrinker_maps()
5280 * by the time the maps are allocated. So, we allocate maps
5281 * here, when for_each_mem_cgroup() can't skip it.
5283 if (memcg_alloc_shrinker_maps(memcg)) {
5284 mem_cgroup_id_remove(memcg);
5288 /* Online state pins memcg ID, memcg ID pins CSS */
5289 refcount_set(&memcg->id.ref, 1);
5294 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5296 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5297 struct mem_cgroup_event *event, *tmp;
5300 * Unregister events and notify userspace.
5301 * Notify userspace about cgroup removing only after rmdir of cgroup
5302 * directory to avoid race between userspace and kernelspace.
5304 spin_lock(&memcg->event_list_lock);
5305 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5306 list_del_init(&event->list);
5307 schedule_work(&event->remove);
5309 spin_unlock(&memcg->event_list_lock);
5311 page_counter_set_min(&memcg->memory, 0);
5312 page_counter_set_low(&memcg->memory, 0);
5314 memcg_offline_kmem(memcg);
5315 wb_memcg_offline(memcg);
5317 drain_all_stock(memcg);
5319 mem_cgroup_id_put(memcg);
5322 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5324 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5326 invalidate_reclaim_iterators(memcg);
5329 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5331 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5332 int __maybe_unused i;
5334 #ifdef CONFIG_CGROUP_WRITEBACK
5335 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5336 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5338 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5339 static_branch_dec(&memcg_sockets_enabled_key);
5341 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5342 static_branch_dec(&memcg_sockets_enabled_key);
5344 vmpressure_cleanup(&memcg->vmpressure);
5345 cancel_work_sync(&memcg->high_work);
5346 mem_cgroup_remove_from_trees(memcg);
5347 memcg_free_shrinker_maps(memcg);
5348 memcg_free_kmem(memcg);
5349 mem_cgroup_free(memcg);
5353 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5354 * @css: the target css
5356 * Reset the states of the mem_cgroup associated with @css. This is
5357 * invoked when the userland requests disabling on the default hierarchy
5358 * but the memcg is pinned through dependency. The memcg should stop
5359 * applying policies and should revert to the vanilla state as it may be
5360 * made visible again.
5362 * The current implementation only resets the essential configurations.
5363 * This needs to be expanded to cover all the visible parts.
5365 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5367 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5369 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5370 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5371 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
5372 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5373 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5374 page_counter_set_min(&memcg->memory, 0);
5375 page_counter_set_low(&memcg->memory, 0);
5376 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5377 memcg->soft_limit = PAGE_COUNTER_MAX;
5378 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5379 memcg_wb_domain_size_changed(memcg);
5383 /* Handlers for move charge at task migration. */
5384 static int mem_cgroup_do_precharge(unsigned long count)
5388 /* Try a single bulk charge without reclaim first, kswapd may wake */
5389 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5391 mc.precharge += count;
5395 /* Try charges one by one with reclaim, but do not retry */
5397 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5411 enum mc_target_type {
5418 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5419 unsigned long addr, pte_t ptent)
5421 struct page *page = vm_normal_page(vma, addr, ptent);
5423 if (!page || !page_mapped(page))
5425 if (PageAnon(page)) {
5426 if (!(mc.flags & MOVE_ANON))
5429 if (!(mc.flags & MOVE_FILE))
5432 if (!get_page_unless_zero(page))
5438 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5439 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5440 pte_t ptent, swp_entry_t *entry)
5442 struct page *page = NULL;
5443 swp_entry_t ent = pte_to_swp_entry(ptent);
5445 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
5449 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5450 * a device and because they are not accessible by CPU they are store
5451 * as special swap entry in the CPU page table.
5453 if (is_device_private_entry(ent)) {
5454 page = device_private_entry_to_page(ent);
5456 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5457 * a refcount of 1 when free (unlike normal page)
5459 if (!page_ref_add_unless(page, 1, 1))
5465 * Because lookup_swap_cache() updates some statistics counter,
5466 * we call find_get_page() with swapper_space directly.
5468 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5469 entry->val = ent.val;
5474 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5475 pte_t ptent, swp_entry_t *entry)
5481 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5482 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5484 struct page *page = NULL;
5485 struct address_space *mapping;
5488 if (!vma->vm_file) /* anonymous vma */
5490 if (!(mc.flags & MOVE_FILE))
5493 mapping = vma->vm_file->f_mapping;
5494 pgoff = linear_page_index(vma, addr);
5496 /* page is moved even if it's not RSS of this task(page-faulted). */
5498 /* shmem/tmpfs may report page out on swap: account for that too. */
5499 if (shmem_mapping(mapping)) {
5500 page = find_get_entry(mapping, pgoff);
5501 if (xa_is_value(page)) {
5502 swp_entry_t swp = radix_to_swp_entry(page);
5504 page = find_get_page(swap_address_space(swp),
5508 page = find_get_page(mapping, pgoff);
5510 page = find_get_page(mapping, pgoff);
5516 * mem_cgroup_move_account - move account of the page
5518 * @compound: charge the page as compound or small page
5519 * @from: mem_cgroup which the page is moved from.
5520 * @to: mem_cgroup which the page is moved to. @from != @to.
5522 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5524 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5527 static int mem_cgroup_move_account(struct page *page,
5529 struct mem_cgroup *from,
5530 struct mem_cgroup *to)
5532 struct lruvec *from_vec, *to_vec;
5533 struct pglist_data *pgdat;
5534 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5537 VM_BUG_ON(from == to);
5538 VM_BUG_ON_PAGE(PageLRU(page), page);
5539 VM_BUG_ON(compound && !PageTransHuge(page));
5542 * Prevent mem_cgroup_migrate() from looking at
5543 * page->mem_cgroup of its source page while we change it.
5546 if (!trylock_page(page))
5550 if (page->mem_cgroup != from)
5553 pgdat = page_pgdat(page);
5554 from_vec = mem_cgroup_lruvec(from, pgdat);
5555 to_vec = mem_cgroup_lruvec(to, pgdat);
5557 lock_page_memcg(page);
5559 if (PageAnon(page)) {
5560 if (page_mapped(page)) {
5561 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5562 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5563 if (PageTransHuge(page)) {
5564 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5566 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5572 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5573 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5575 if (PageSwapBacked(page)) {
5576 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5577 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5580 if (page_mapped(page)) {
5581 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5582 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5585 if (PageDirty(page)) {
5586 struct address_space *mapping = page_mapping(page);
5588 if (mapping_cap_account_dirty(mapping)) {
5589 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5591 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5597 if (PageWriteback(page)) {
5598 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5599 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5603 * All state has been migrated, let's switch to the new memcg.
5605 * It is safe to change page->mem_cgroup here because the page
5606 * is referenced, charged, isolated, and locked: we can't race
5607 * with (un)charging, migration, LRU putback, or anything else
5608 * that would rely on a stable page->mem_cgroup.
5610 * Note that lock_page_memcg is a memcg lock, not a page lock,
5611 * to save space. As soon as we switch page->mem_cgroup to a
5612 * new memcg that isn't locked, the above state can change
5613 * concurrently again. Make sure we're truly done with it.
5618 css_put(&from->css);
5620 page->mem_cgroup = to;
5622 __unlock_page_memcg(from);
5626 local_irq_disable();
5627 mem_cgroup_charge_statistics(to, page, nr_pages);
5628 memcg_check_events(to, page);
5629 mem_cgroup_charge_statistics(from, page, -nr_pages);
5630 memcg_check_events(from, page);
5639 * get_mctgt_type - get target type of moving charge
5640 * @vma: the vma the pte to be checked belongs
5641 * @addr: the address corresponding to the pte to be checked
5642 * @ptent: the pte to be checked
5643 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5646 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5647 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5648 * move charge. if @target is not NULL, the page is stored in target->page
5649 * with extra refcnt got(Callers should handle it).
5650 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5651 * target for charge migration. if @target is not NULL, the entry is stored
5653 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5654 * (so ZONE_DEVICE page and thus not on the lru).
5655 * For now we such page is charge like a regular page would be as for all
5656 * intent and purposes it is just special memory taking the place of a
5659 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5661 * Called with pte lock held.
5664 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5665 unsigned long addr, pte_t ptent, union mc_target *target)
5667 struct page *page = NULL;
5668 enum mc_target_type ret = MC_TARGET_NONE;
5669 swp_entry_t ent = { .val = 0 };
5671 if (pte_present(ptent))
5672 page = mc_handle_present_pte(vma, addr, ptent);
5673 else if (is_swap_pte(ptent))
5674 page = mc_handle_swap_pte(vma, ptent, &ent);
5675 else if (pte_none(ptent))
5676 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5678 if (!page && !ent.val)
5682 * Do only loose check w/o serialization.
5683 * mem_cgroup_move_account() checks the page is valid or
5684 * not under LRU exclusion.
5686 if (page->mem_cgroup == mc.from) {
5687 ret = MC_TARGET_PAGE;
5688 if (is_device_private_page(page))
5689 ret = MC_TARGET_DEVICE;
5691 target->page = page;
5693 if (!ret || !target)
5697 * There is a swap entry and a page doesn't exist or isn't charged.
5698 * But we cannot move a tail-page in a THP.
5700 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5701 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5702 ret = MC_TARGET_SWAP;
5709 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5711 * We don't consider PMD mapped swapping or file mapped pages because THP does
5712 * not support them for now.
5713 * Caller should make sure that pmd_trans_huge(pmd) is true.
5715 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5716 unsigned long addr, pmd_t pmd, union mc_target *target)
5718 struct page *page = NULL;
5719 enum mc_target_type ret = MC_TARGET_NONE;
5721 if (unlikely(is_swap_pmd(pmd))) {
5722 VM_BUG_ON(thp_migration_supported() &&
5723 !is_pmd_migration_entry(pmd));
5726 page = pmd_page(pmd);
5727 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5728 if (!(mc.flags & MOVE_ANON))
5730 if (page->mem_cgroup == mc.from) {
5731 ret = MC_TARGET_PAGE;
5734 target->page = page;
5740 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5741 unsigned long addr, pmd_t pmd, union mc_target *target)
5743 return MC_TARGET_NONE;
5747 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5748 unsigned long addr, unsigned long end,
5749 struct mm_walk *walk)
5751 struct vm_area_struct *vma = walk->vma;
5755 ptl = pmd_trans_huge_lock(pmd, vma);
5758 * Note their can not be MC_TARGET_DEVICE for now as we do not
5759 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5760 * this might change.
5762 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5763 mc.precharge += HPAGE_PMD_NR;
5768 if (pmd_trans_unstable(pmd))
5770 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5771 for (; addr != end; pte++, addr += PAGE_SIZE)
5772 if (get_mctgt_type(vma, addr, *pte, NULL))
5773 mc.precharge++; /* increment precharge temporarily */
5774 pte_unmap_unlock(pte - 1, ptl);
5780 static const struct mm_walk_ops precharge_walk_ops = {
5781 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5784 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5786 unsigned long precharge;
5789 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5790 mmap_read_unlock(mm);
5792 precharge = mc.precharge;
5798 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5800 unsigned long precharge = mem_cgroup_count_precharge(mm);
5802 VM_BUG_ON(mc.moving_task);
5803 mc.moving_task = current;
5804 return mem_cgroup_do_precharge(precharge);
5807 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5808 static void __mem_cgroup_clear_mc(void)
5810 struct mem_cgroup *from = mc.from;
5811 struct mem_cgroup *to = mc.to;
5813 /* we must uncharge all the leftover precharges from mc.to */
5815 cancel_charge(mc.to, mc.precharge);
5819 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5820 * we must uncharge here.
5822 if (mc.moved_charge) {
5823 cancel_charge(mc.from, mc.moved_charge);
5824 mc.moved_charge = 0;
5826 /* we must fixup refcnts and charges */
5827 if (mc.moved_swap) {
5828 /* uncharge swap account from the old cgroup */
5829 if (!mem_cgroup_is_root(mc.from))
5830 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5832 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5835 * we charged both to->memory and to->memsw, so we
5836 * should uncharge to->memory.
5838 if (!mem_cgroup_is_root(mc.to))
5839 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5843 memcg_oom_recover(from);
5844 memcg_oom_recover(to);
5845 wake_up_all(&mc.waitq);
5848 static void mem_cgroup_clear_mc(void)
5850 struct mm_struct *mm = mc.mm;
5853 * we must clear moving_task before waking up waiters at the end of
5856 mc.moving_task = NULL;
5857 __mem_cgroup_clear_mc();
5858 spin_lock(&mc.lock);
5862 spin_unlock(&mc.lock);
5867 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5869 struct cgroup_subsys_state *css;
5870 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5871 struct mem_cgroup *from;
5872 struct task_struct *leader, *p;
5873 struct mm_struct *mm;
5874 unsigned long move_flags;
5877 /* charge immigration isn't supported on the default hierarchy */
5878 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5882 * Multi-process migrations only happen on the default hierarchy
5883 * where charge immigration is not used. Perform charge
5884 * immigration if @tset contains a leader and whine if there are
5888 cgroup_taskset_for_each_leader(leader, css, tset) {
5891 memcg = mem_cgroup_from_css(css);
5897 * We are now commited to this value whatever it is. Changes in this
5898 * tunable will only affect upcoming migrations, not the current one.
5899 * So we need to save it, and keep it going.
5901 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5905 from = mem_cgroup_from_task(p);
5907 VM_BUG_ON(from == memcg);
5909 mm = get_task_mm(p);
5912 /* We move charges only when we move a owner of the mm */
5913 if (mm->owner == p) {
5916 VM_BUG_ON(mc.precharge);
5917 VM_BUG_ON(mc.moved_charge);
5918 VM_BUG_ON(mc.moved_swap);
5920 spin_lock(&mc.lock);
5924 mc.flags = move_flags;
5925 spin_unlock(&mc.lock);
5926 /* We set mc.moving_task later */
5928 ret = mem_cgroup_precharge_mc(mm);
5930 mem_cgroup_clear_mc();
5937 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5940 mem_cgroup_clear_mc();
5943 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5944 unsigned long addr, unsigned long end,
5945 struct mm_walk *walk)
5948 struct vm_area_struct *vma = walk->vma;
5951 enum mc_target_type target_type;
5952 union mc_target target;
5955 ptl = pmd_trans_huge_lock(pmd, vma);
5957 if (mc.precharge < HPAGE_PMD_NR) {
5961 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5962 if (target_type == MC_TARGET_PAGE) {
5964 if (!isolate_lru_page(page)) {
5965 if (!mem_cgroup_move_account(page, true,
5967 mc.precharge -= HPAGE_PMD_NR;
5968 mc.moved_charge += HPAGE_PMD_NR;
5970 putback_lru_page(page);
5973 } else if (target_type == MC_TARGET_DEVICE) {
5975 if (!mem_cgroup_move_account(page, true,
5977 mc.precharge -= HPAGE_PMD_NR;
5978 mc.moved_charge += HPAGE_PMD_NR;
5986 if (pmd_trans_unstable(pmd))
5989 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5990 for (; addr != end; addr += PAGE_SIZE) {
5991 pte_t ptent = *(pte++);
5992 bool device = false;
5998 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5999 case MC_TARGET_DEVICE:
6002 case MC_TARGET_PAGE:
6005 * We can have a part of the split pmd here. Moving it
6006 * can be done but it would be too convoluted so simply
6007 * ignore such a partial THP and keep it in original
6008 * memcg. There should be somebody mapping the head.
6010 if (PageTransCompound(page))
6012 if (!device && isolate_lru_page(page))
6014 if (!mem_cgroup_move_account(page, false,
6017 /* we uncharge from mc.from later. */
6021 putback_lru_page(page);
6022 put: /* get_mctgt_type() gets the page */
6025 case MC_TARGET_SWAP:
6027 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6029 mem_cgroup_id_get_many(mc.to, 1);
6030 /* we fixup other refcnts and charges later. */
6038 pte_unmap_unlock(pte - 1, ptl);
6043 * We have consumed all precharges we got in can_attach().
6044 * We try charge one by one, but don't do any additional
6045 * charges to mc.to if we have failed in charge once in attach()
6048 ret = mem_cgroup_do_precharge(1);
6056 static const struct mm_walk_ops charge_walk_ops = {
6057 .pmd_entry = mem_cgroup_move_charge_pte_range,
6060 static void mem_cgroup_move_charge(void)
6062 lru_add_drain_all();
6064 * Signal lock_page_memcg() to take the memcg's move_lock
6065 * while we're moving its pages to another memcg. Then wait
6066 * for already started RCU-only updates to finish.
6068 atomic_inc(&mc.from->moving_account);
6071 if (unlikely(!mmap_read_trylock(mc.mm))) {
6073 * Someone who are holding the mmap_lock might be waiting in
6074 * waitq. So we cancel all extra charges, wake up all waiters,
6075 * and retry. Because we cancel precharges, we might not be able
6076 * to move enough charges, but moving charge is a best-effort
6077 * feature anyway, so it wouldn't be a big problem.
6079 __mem_cgroup_clear_mc();
6084 * When we have consumed all precharges and failed in doing
6085 * additional charge, the page walk just aborts.
6087 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6090 mmap_read_unlock(mc.mm);
6091 atomic_dec(&mc.from->moving_account);
6094 static void mem_cgroup_move_task(void)
6097 mem_cgroup_move_charge();
6098 mem_cgroup_clear_mc();
6101 #else /* !CONFIG_MMU */
6102 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6106 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6109 static void mem_cgroup_move_task(void)
6115 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6116 * to verify whether we're attached to the default hierarchy on each mount
6119 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6122 * use_hierarchy is forced on the default hierarchy. cgroup core
6123 * guarantees that @root doesn't have any children, so turning it
6124 * on for the root memcg is enough.
6126 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6127 root_mem_cgroup->use_hierarchy = true;
6129 root_mem_cgroup->use_hierarchy = false;
6132 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6134 if (value == PAGE_COUNTER_MAX)
6135 seq_puts(m, "max\n");
6137 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6142 static u64 memory_current_read(struct cgroup_subsys_state *css,
6145 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6147 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6150 static int memory_min_show(struct seq_file *m, void *v)
6152 return seq_puts_memcg_tunable(m,
6153 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6156 static ssize_t memory_min_write(struct kernfs_open_file *of,
6157 char *buf, size_t nbytes, loff_t off)
6159 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6163 buf = strstrip(buf);
6164 err = page_counter_memparse(buf, "max", &min);
6168 page_counter_set_min(&memcg->memory, min);
6173 static int memory_low_show(struct seq_file *m, void *v)
6175 return seq_puts_memcg_tunable(m,
6176 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6179 static ssize_t memory_low_write(struct kernfs_open_file *of,
6180 char *buf, size_t nbytes, loff_t off)
6182 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6186 buf = strstrip(buf);
6187 err = page_counter_memparse(buf, "max", &low);
6191 page_counter_set_low(&memcg->memory, low);
6196 static int memory_high_show(struct seq_file *m, void *v)
6198 return seq_puts_memcg_tunable(m,
6199 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6202 static ssize_t memory_high_write(struct kernfs_open_file *of,
6203 char *buf, size_t nbytes, loff_t off)
6205 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6206 unsigned int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
6207 bool drained = false;
6211 buf = strstrip(buf);
6212 err = page_counter_memparse(buf, "max", &high);
6216 page_counter_set_high(&memcg->memory, high);
6219 unsigned long nr_pages = page_counter_read(&memcg->memory);
6220 unsigned long reclaimed;
6222 if (nr_pages <= high)
6225 if (signal_pending(current))
6229 drain_all_stock(memcg);
6234 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6237 if (!reclaimed && !nr_retries--)
6244 static int memory_max_show(struct seq_file *m, void *v)
6246 return seq_puts_memcg_tunable(m,
6247 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6250 static ssize_t memory_max_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_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
6255 bool drained = false;
6259 buf = strstrip(buf);
6260 err = page_counter_memparse(buf, "max", &max);
6264 xchg(&memcg->memory.max, max);
6267 unsigned long nr_pages = page_counter_read(&memcg->memory);
6269 if (nr_pages <= max)
6272 if (signal_pending(current))
6276 drain_all_stock(memcg);
6282 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6288 memcg_memory_event(memcg, MEMCG_OOM);
6289 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6293 memcg_wb_domain_size_changed(memcg);
6297 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6299 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6300 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6301 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6302 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6303 seq_printf(m, "oom_kill %lu\n",
6304 atomic_long_read(&events[MEMCG_OOM_KILL]));
6307 static int memory_events_show(struct seq_file *m, void *v)
6309 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6311 __memory_events_show(m, memcg->memory_events);
6315 static int memory_events_local_show(struct seq_file *m, void *v)
6317 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6319 __memory_events_show(m, memcg->memory_events_local);
6323 static int memory_stat_show(struct seq_file *m, void *v)
6325 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6328 buf = memory_stat_format(memcg);
6336 static int memory_oom_group_show(struct seq_file *m, void *v)
6338 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6340 seq_printf(m, "%d\n", memcg->oom_group);
6345 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6346 char *buf, size_t nbytes, loff_t off)
6348 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6351 buf = strstrip(buf);
6355 ret = kstrtoint(buf, 0, &oom_group);
6359 if (oom_group != 0 && oom_group != 1)
6362 memcg->oom_group = oom_group;
6367 static struct cftype memory_files[] = {
6370 .flags = CFTYPE_NOT_ON_ROOT,
6371 .read_u64 = memory_current_read,
6375 .flags = CFTYPE_NOT_ON_ROOT,
6376 .seq_show = memory_min_show,
6377 .write = memory_min_write,
6381 .flags = CFTYPE_NOT_ON_ROOT,
6382 .seq_show = memory_low_show,
6383 .write = memory_low_write,
6387 .flags = CFTYPE_NOT_ON_ROOT,
6388 .seq_show = memory_high_show,
6389 .write = memory_high_write,
6393 .flags = CFTYPE_NOT_ON_ROOT,
6394 .seq_show = memory_max_show,
6395 .write = memory_max_write,
6399 .flags = CFTYPE_NOT_ON_ROOT,
6400 .file_offset = offsetof(struct mem_cgroup, events_file),
6401 .seq_show = memory_events_show,
6404 .name = "events.local",
6405 .flags = CFTYPE_NOT_ON_ROOT,
6406 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6407 .seq_show = memory_events_local_show,
6411 .seq_show = memory_stat_show,
6414 .name = "oom.group",
6415 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6416 .seq_show = memory_oom_group_show,
6417 .write = memory_oom_group_write,
6422 struct cgroup_subsys memory_cgrp_subsys = {
6423 .css_alloc = mem_cgroup_css_alloc,
6424 .css_online = mem_cgroup_css_online,
6425 .css_offline = mem_cgroup_css_offline,
6426 .css_released = mem_cgroup_css_released,
6427 .css_free = mem_cgroup_css_free,
6428 .css_reset = mem_cgroup_css_reset,
6429 .can_attach = mem_cgroup_can_attach,
6430 .cancel_attach = mem_cgroup_cancel_attach,
6431 .post_attach = mem_cgroup_move_task,
6432 .bind = mem_cgroup_bind,
6433 .dfl_cftypes = memory_files,
6434 .legacy_cftypes = mem_cgroup_legacy_files,
6439 * This function calculates an individual cgroup's effective
6440 * protection which is derived from its own memory.min/low, its
6441 * parent's and siblings' settings, as well as the actual memory
6442 * distribution in the tree.
6444 * The following rules apply to the effective protection values:
6446 * 1. At the first level of reclaim, effective protection is equal to
6447 * the declared protection in memory.min and memory.low.
6449 * 2. To enable safe delegation of the protection configuration, at
6450 * subsequent levels the effective protection is capped to the
6451 * parent's effective protection.
6453 * 3. To make complex and dynamic subtrees easier to configure, the
6454 * user is allowed to overcommit the declared protection at a given
6455 * level. If that is the case, the parent's effective protection is
6456 * distributed to the children in proportion to how much protection
6457 * they have declared and how much of it they are utilizing.
6459 * This makes distribution proportional, but also work-conserving:
6460 * if one cgroup claims much more protection than it uses memory,
6461 * the unused remainder is available to its siblings.
6463 * 4. Conversely, when the declared protection is undercommitted at a
6464 * given level, the distribution of the larger parental protection
6465 * budget is NOT proportional. A cgroup's protection from a sibling
6466 * is capped to its own memory.min/low setting.
6468 * 5. However, to allow protecting recursive subtrees from each other
6469 * without having to declare each individual cgroup's fixed share
6470 * of the ancestor's claim to protection, any unutilized -
6471 * "floating" - protection from up the tree is distributed in
6472 * proportion to each cgroup's *usage*. This makes the protection
6473 * neutral wrt sibling cgroups and lets them compete freely over
6474 * the shared parental protection budget, but it protects the
6475 * subtree as a whole from neighboring subtrees.
6477 * Note that 4. and 5. are not in conflict: 4. is about protecting
6478 * against immediate siblings whereas 5. is about protecting against
6479 * neighboring subtrees.
6481 static unsigned long effective_protection(unsigned long usage,
6482 unsigned long parent_usage,
6483 unsigned long setting,
6484 unsigned long parent_effective,
6485 unsigned long siblings_protected)
6487 unsigned long protected;
6490 protected = min(usage, setting);
6492 * If all cgroups at this level combined claim and use more
6493 * protection then what the parent affords them, distribute
6494 * shares in proportion to utilization.
6496 * We are using actual utilization rather than the statically
6497 * claimed protection in order to be work-conserving: claimed
6498 * but unused protection is available to siblings that would
6499 * otherwise get a smaller chunk than what they claimed.
6501 if (siblings_protected > parent_effective)
6502 return protected * parent_effective / siblings_protected;
6505 * Ok, utilized protection of all children is within what the
6506 * parent affords them, so we know whatever this child claims
6507 * and utilizes is effectively protected.
6509 * If there is unprotected usage beyond this value, reclaim
6510 * will apply pressure in proportion to that amount.
6512 * If there is unutilized protection, the cgroup will be fully
6513 * shielded from reclaim, but we do return a smaller value for
6514 * protection than what the group could enjoy in theory. This
6515 * is okay. With the overcommit distribution above, effective
6516 * protection is always dependent on how memory is actually
6517 * consumed among the siblings anyway.
6522 * If the children aren't claiming (all of) the protection
6523 * afforded to them by the parent, distribute the remainder in
6524 * proportion to the (unprotected) memory of each cgroup. That
6525 * way, cgroups that aren't explicitly prioritized wrt each
6526 * other compete freely over the allowance, but they are
6527 * collectively protected from neighboring trees.
6529 * We're using unprotected memory for the weight so that if
6530 * some cgroups DO claim explicit protection, we don't protect
6531 * the same bytes twice.
6533 * Check both usage and parent_usage against the respective
6534 * protected values. One should imply the other, but they
6535 * aren't read atomically - make sure the division is sane.
6537 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6539 if (parent_effective > siblings_protected &&
6540 parent_usage > siblings_protected &&
6541 usage > protected) {
6542 unsigned long unclaimed;
6544 unclaimed = parent_effective - siblings_protected;
6545 unclaimed *= usage - protected;
6546 unclaimed /= parent_usage - siblings_protected;
6555 * mem_cgroup_protected - check if memory consumption is in the normal range
6556 * @root: the top ancestor of the sub-tree being checked
6557 * @memcg: the memory cgroup to check
6559 * WARNING: This function is not stateless! It can only be used as part
6560 * of a top-down tree iteration, not for isolated queries.
6562 * Returns one of the following:
6563 * MEMCG_PROT_NONE: cgroup memory is not protected
6564 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6565 * an unprotected supply of reclaimable memory from other cgroups.
6566 * MEMCG_PROT_MIN: cgroup memory is protected
6568 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
6569 struct mem_cgroup *memcg)
6571 unsigned long usage, parent_usage;
6572 struct mem_cgroup *parent;
6574 if (mem_cgroup_disabled())
6575 return MEMCG_PROT_NONE;
6578 root = root_mem_cgroup;
6580 return MEMCG_PROT_NONE;
6582 usage = page_counter_read(&memcg->memory);
6584 return MEMCG_PROT_NONE;
6586 parent = parent_mem_cgroup(memcg);
6587 /* No parent means a non-hierarchical mode on v1 memcg */
6589 return MEMCG_PROT_NONE;
6591 if (parent == root) {
6592 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6593 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6597 parent_usage = page_counter_read(&parent->memory);
6599 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6600 READ_ONCE(memcg->memory.min),
6601 READ_ONCE(parent->memory.emin),
6602 atomic_long_read(&parent->memory.children_min_usage)));
6604 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6605 READ_ONCE(memcg->memory.low),
6606 READ_ONCE(parent->memory.elow),
6607 atomic_long_read(&parent->memory.children_low_usage)));
6610 if (usage <= memcg->memory.emin)
6611 return MEMCG_PROT_MIN;
6612 else if (usage <= memcg->memory.elow)
6613 return MEMCG_PROT_LOW;
6615 return MEMCG_PROT_NONE;
6619 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6620 * @page: page to charge
6621 * @mm: mm context of the victim
6622 * @gfp_mask: reclaim mode
6624 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6625 * pages according to @gfp_mask if necessary.
6627 * Returns 0 on success. Otherwise, an error code is returned.
6629 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6631 unsigned int nr_pages = hpage_nr_pages(page);
6632 struct mem_cgroup *memcg = NULL;
6635 if (mem_cgroup_disabled())
6638 if (PageSwapCache(page)) {
6639 swp_entry_t ent = { .val = page_private(page), };
6643 * Every swap fault against a single page tries to charge the
6644 * page, bail as early as possible. shmem_unuse() encounters
6645 * already charged pages, too. page->mem_cgroup is protected
6646 * by the page lock, which serializes swap cache removal, which
6647 * in turn serializes uncharging.
6649 VM_BUG_ON_PAGE(!PageLocked(page), page);
6650 if (compound_head(page)->mem_cgroup)
6653 id = lookup_swap_cgroup_id(ent);
6655 memcg = mem_cgroup_from_id(id);
6656 if (memcg && !css_tryget_online(&memcg->css))
6662 memcg = get_mem_cgroup_from_mm(mm);
6664 ret = try_charge(memcg, gfp_mask, nr_pages);
6668 css_get(&memcg->css);
6669 commit_charge(page, memcg);
6671 local_irq_disable();
6672 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6673 memcg_check_events(memcg, page);
6676 if (PageSwapCache(page)) {
6677 swp_entry_t entry = { .val = page_private(page) };
6679 * The swap entry might not get freed for a long time,
6680 * let's not wait for it. The page already received a
6681 * memory+swap charge, drop the swap entry duplicate.
6683 mem_cgroup_uncharge_swap(entry, nr_pages);
6687 css_put(&memcg->css);
6692 struct uncharge_gather {
6693 struct mem_cgroup *memcg;
6694 unsigned long nr_pages;
6695 unsigned long pgpgout;
6696 unsigned long nr_kmem;
6697 struct page *dummy_page;
6700 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6702 memset(ug, 0, sizeof(*ug));
6705 static void uncharge_batch(const struct uncharge_gather *ug)
6707 unsigned long flags;
6709 if (!mem_cgroup_is_root(ug->memcg)) {
6710 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6711 if (do_memsw_account())
6712 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6713 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6714 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6715 memcg_oom_recover(ug->memcg);
6718 local_irq_save(flags);
6719 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6720 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6721 memcg_check_events(ug->memcg, ug->dummy_page);
6722 local_irq_restore(flags);
6725 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6727 unsigned long nr_pages;
6729 VM_BUG_ON_PAGE(PageLRU(page), page);
6731 if (!page->mem_cgroup)
6735 * Nobody should be changing or seriously looking at
6736 * page->mem_cgroup at this point, we have fully
6737 * exclusive access to the page.
6740 if (ug->memcg != page->mem_cgroup) {
6743 uncharge_gather_clear(ug);
6745 ug->memcg = page->mem_cgroup;
6748 nr_pages = compound_nr(page);
6749 ug->nr_pages += nr_pages;
6751 if (!PageKmemcg(page)) {
6754 ug->nr_kmem += nr_pages;
6755 __ClearPageKmemcg(page);
6758 ug->dummy_page = page;
6759 page->mem_cgroup = NULL;
6760 css_put(&ug->memcg->css);
6763 static void uncharge_list(struct list_head *page_list)
6765 struct uncharge_gather ug;
6766 struct list_head *next;
6768 uncharge_gather_clear(&ug);
6771 * Note that the list can be a single page->lru; hence the
6772 * do-while loop instead of a simple list_for_each_entry().
6774 next = page_list->next;
6778 page = list_entry(next, struct page, lru);
6779 next = page->lru.next;
6781 uncharge_page(page, &ug);
6782 } while (next != page_list);
6785 uncharge_batch(&ug);
6789 * mem_cgroup_uncharge - uncharge a page
6790 * @page: page to uncharge
6792 * Uncharge a page previously charged with mem_cgroup_charge().
6794 void mem_cgroup_uncharge(struct page *page)
6796 struct uncharge_gather ug;
6798 if (mem_cgroup_disabled())
6801 /* Don't touch page->lru of any random page, pre-check: */
6802 if (!page->mem_cgroup)
6805 uncharge_gather_clear(&ug);
6806 uncharge_page(page, &ug);
6807 uncharge_batch(&ug);
6811 * mem_cgroup_uncharge_list - uncharge a list of page
6812 * @page_list: list of pages to uncharge
6814 * Uncharge a list of pages previously charged with
6815 * mem_cgroup_charge().
6817 void mem_cgroup_uncharge_list(struct list_head *page_list)
6819 if (mem_cgroup_disabled())
6822 if (!list_empty(page_list))
6823 uncharge_list(page_list);
6827 * mem_cgroup_migrate - charge a page's replacement
6828 * @oldpage: currently circulating page
6829 * @newpage: replacement page
6831 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6832 * be uncharged upon free.
6834 * Both pages must be locked, @newpage->mapping must be set up.
6836 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6838 struct mem_cgroup *memcg;
6839 unsigned int nr_pages;
6840 unsigned long flags;
6842 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6843 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6844 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6845 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6848 if (mem_cgroup_disabled())
6851 /* Page cache replacement: new page already charged? */
6852 if (newpage->mem_cgroup)
6855 /* Swapcache readahead pages can get replaced before being charged */
6856 memcg = oldpage->mem_cgroup;
6860 /* Force-charge the new page. The old one will be freed soon */
6861 nr_pages = hpage_nr_pages(newpage);
6863 page_counter_charge(&memcg->memory, nr_pages);
6864 if (do_memsw_account())
6865 page_counter_charge(&memcg->memsw, nr_pages);
6867 css_get(&memcg->css);
6868 commit_charge(newpage, memcg);
6870 local_irq_save(flags);
6871 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6872 memcg_check_events(memcg, newpage);
6873 local_irq_restore(flags);
6876 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6877 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6879 void mem_cgroup_sk_alloc(struct sock *sk)
6881 struct mem_cgroup *memcg;
6883 if (!mem_cgroup_sockets_enabled)
6886 /* Do not associate the sock with unrelated interrupted task's memcg. */
6891 memcg = mem_cgroup_from_task(current);
6892 if (memcg == root_mem_cgroup)
6894 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6896 if (css_tryget(&memcg->css))
6897 sk->sk_memcg = memcg;
6902 void mem_cgroup_sk_free(struct sock *sk)
6905 css_put(&sk->sk_memcg->css);
6909 * mem_cgroup_charge_skmem - charge socket memory
6910 * @memcg: memcg to charge
6911 * @nr_pages: number of pages to charge
6913 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6914 * @memcg's configured limit, %false if the charge had to be forced.
6916 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6918 gfp_t gfp_mask = GFP_KERNEL;
6920 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6921 struct page_counter *fail;
6923 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6924 memcg->tcpmem_pressure = 0;
6927 page_counter_charge(&memcg->tcpmem, nr_pages);
6928 memcg->tcpmem_pressure = 1;
6932 /* Don't block in the packet receive path */
6934 gfp_mask = GFP_NOWAIT;
6936 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6938 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6941 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6946 * mem_cgroup_uncharge_skmem - uncharge socket memory
6947 * @memcg: memcg to uncharge
6948 * @nr_pages: number of pages to uncharge
6950 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6952 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6953 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6957 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6959 refill_stock(memcg, nr_pages);
6962 static int __init cgroup_memory(char *s)
6966 while ((token = strsep(&s, ",")) != NULL) {
6969 if (!strcmp(token, "nosocket"))
6970 cgroup_memory_nosocket = true;
6971 if (!strcmp(token, "nokmem"))
6972 cgroup_memory_nokmem = true;
6976 __setup("cgroup.memory=", cgroup_memory);
6979 * subsys_initcall() for memory controller.
6981 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6982 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6983 * basically everything that doesn't depend on a specific mem_cgroup structure
6984 * should be initialized from here.
6986 static int __init mem_cgroup_init(void)
6990 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6991 memcg_hotplug_cpu_dead);
6993 for_each_possible_cpu(cpu)
6994 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6997 for_each_node(node) {
6998 struct mem_cgroup_tree_per_node *rtpn;
7000 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7001 node_online(node) ? node : NUMA_NO_NODE);
7003 rtpn->rb_root = RB_ROOT;
7004 rtpn->rb_rightmost = NULL;
7005 spin_lock_init(&rtpn->lock);
7006 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7011 subsys_initcall(mem_cgroup_init);
7013 #ifdef CONFIG_MEMCG_SWAP
7014 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7016 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7018 * The root cgroup cannot be destroyed, so it's refcount must
7021 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7025 memcg = parent_mem_cgroup(memcg);
7027 memcg = root_mem_cgroup;
7033 * mem_cgroup_swapout - transfer a memsw charge to swap
7034 * @page: page whose memsw charge to transfer
7035 * @entry: swap entry to move the charge to
7037 * Transfer the memsw charge of @page to @entry.
7039 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7041 struct mem_cgroup *memcg, *swap_memcg;
7042 unsigned int nr_entries;
7043 unsigned short oldid;
7045 VM_BUG_ON_PAGE(PageLRU(page), page);
7046 VM_BUG_ON_PAGE(page_count(page), page);
7048 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7051 memcg = page->mem_cgroup;
7053 /* Readahead page, never charged */
7058 * In case the memcg owning these pages has been offlined and doesn't
7059 * have an ID allocated to it anymore, charge the closest online
7060 * ancestor for the swap instead and transfer the memory+swap charge.
7062 swap_memcg = mem_cgroup_id_get_online(memcg);
7063 nr_entries = hpage_nr_pages(page);
7064 /* Get references for the tail pages, too */
7066 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7067 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7069 VM_BUG_ON_PAGE(oldid, page);
7070 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7072 page->mem_cgroup = NULL;
7074 if (!mem_cgroup_is_root(memcg))
7075 page_counter_uncharge(&memcg->memory, nr_entries);
7077 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7078 if (!mem_cgroup_is_root(swap_memcg))
7079 page_counter_charge(&swap_memcg->memsw, nr_entries);
7080 page_counter_uncharge(&memcg->memsw, nr_entries);
7084 * Interrupts should be disabled here because the caller holds the
7085 * i_pages lock which is taken with interrupts-off. It is
7086 * important here to have the interrupts disabled because it is the
7087 * only synchronisation we have for updating the per-CPU variables.
7089 VM_BUG_ON(!irqs_disabled());
7090 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7091 memcg_check_events(memcg, page);
7093 css_put(&memcg->css);
7097 * mem_cgroup_try_charge_swap - try charging swap space for a page
7098 * @page: page being added to swap
7099 * @entry: swap entry to charge
7101 * Try to charge @page's memcg for the swap space at @entry.
7103 * Returns 0 on success, -ENOMEM on failure.
7105 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7107 unsigned int nr_pages = hpage_nr_pages(page);
7108 struct page_counter *counter;
7109 struct mem_cgroup *memcg;
7110 unsigned short oldid;
7112 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7115 memcg = page->mem_cgroup;
7117 /* Readahead page, never charged */
7122 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7126 memcg = mem_cgroup_id_get_online(memcg);
7128 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7129 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7130 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7131 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7132 mem_cgroup_id_put(memcg);
7136 /* Get references for the tail pages, too */
7138 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7139 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7140 VM_BUG_ON_PAGE(oldid, page);
7141 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7147 * mem_cgroup_uncharge_swap - uncharge swap space
7148 * @entry: swap entry to uncharge
7149 * @nr_pages: the amount of swap space to uncharge
7151 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7153 struct mem_cgroup *memcg;
7156 id = swap_cgroup_record(entry, 0, nr_pages);
7158 memcg = mem_cgroup_from_id(id);
7160 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7161 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7162 page_counter_uncharge(&memcg->swap, nr_pages);
7164 page_counter_uncharge(&memcg->memsw, nr_pages);
7166 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7167 mem_cgroup_id_put_many(memcg, nr_pages);
7172 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7174 long nr_swap_pages = get_nr_swap_pages();
7176 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7177 return nr_swap_pages;
7178 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7179 nr_swap_pages = min_t(long, nr_swap_pages,
7180 READ_ONCE(memcg->swap.max) -
7181 page_counter_read(&memcg->swap));
7182 return nr_swap_pages;
7185 bool mem_cgroup_swap_full(struct page *page)
7187 struct mem_cgroup *memcg;
7189 VM_BUG_ON_PAGE(!PageLocked(page), page);
7193 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7196 memcg = page->mem_cgroup;
7200 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7201 unsigned long usage = page_counter_read(&memcg->swap);
7203 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7204 usage * 2 >= READ_ONCE(memcg->swap.max))
7211 static int __init setup_swap_account(char *s)
7213 if (!strcmp(s, "1"))
7214 cgroup_memory_noswap = 0;
7215 else if (!strcmp(s, "0"))
7216 cgroup_memory_noswap = 1;
7219 __setup("swapaccount=", setup_swap_account);
7221 static u64 swap_current_read(struct cgroup_subsys_state *css,
7224 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7226 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7229 static int swap_high_show(struct seq_file *m, void *v)
7231 return seq_puts_memcg_tunable(m,
7232 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7235 static ssize_t swap_high_write(struct kernfs_open_file *of,
7236 char *buf, size_t nbytes, loff_t off)
7238 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7242 buf = strstrip(buf);
7243 err = page_counter_memparse(buf, "max", &high);
7247 page_counter_set_high(&memcg->swap, high);
7252 static int swap_max_show(struct seq_file *m, void *v)
7254 return seq_puts_memcg_tunable(m,
7255 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7258 static ssize_t swap_max_write(struct kernfs_open_file *of,
7259 char *buf, size_t nbytes, loff_t off)
7261 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7265 buf = strstrip(buf);
7266 err = page_counter_memparse(buf, "max", &max);
7270 xchg(&memcg->swap.max, max);
7275 static int swap_events_show(struct seq_file *m, void *v)
7277 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7279 seq_printf(m, "high %lu\n",
7280 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7281 seq_printf(m, "max %lu\n",
7282 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7283 seq_printf(m, "fail %lu\n",
7284 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7289 static struct cftype swap_files[] = {
7291 .name = "swap.current",
7292 .flags = CFTYPE_NOT_ON_ROOT,
7293 .read_u64 = swap_current_read,
7296 .name = "swap.high",
7297 .flags = CFTYPE_NOT_ON_ROOT,
7298 .seq_show = swap_high_show,
7299 .write = swap_high_write,
7303 .flags = CFTYPE_NOT_ON_ROOT,
7304 .seq_show = swap_max_show,
7305 .write = swap_max_write,
7308 .name = "swap.events",
7309 .flags = CFTYPE_NOT_ON_ROOT,
7310 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7311 .seq_show = swap_events_show,
7316 static struct cftype memsw_files[] = {
7318 .name = "memsw.usage_in_bytes",
7319 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7320 .read_u64 = mem_cgroup_read_u64,
7323 .name = "memsw.max_usage_in_bytes",
7324 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7325 .write = mem_cgroup_reset,
7326 .read_u64 = mem_cgroup_read_u64,
7329 .name = "memsw.limit_in_bytes",
7330 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7331 .write = mem_cgroup_write,
7332 .read_u64 = mem_cgroup_read_u64,
7335 .name = "memsw.failcnt",
7336 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7337 .write = mem_cgroup_reset,
7338 .read_u64 = mem_cgroup_read_u64,
7340 { }, /* terminate */
7344 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7345 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7346 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7347 * boot parameter. This may result in premature OOPS inside
7348 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7350 static int __init mem_cgroup_swap_init(void)
7352 /* No memory control -> no swap control */
7353 if (mem_cgroup_disabled())
7354 cgroup_memory_noswap = true;
7356 if (cgroup_memory_noswap)
7359 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7360 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7364 core_initcall(mem_cgroup_swap_init);
7366 #endif /* CONFIG_MEMCG_SWAP */