1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
25 #include <linux/page_counter.h>
26 #include <linux/memcontrol.h>
27 #include <linux/cgroup.h>
28 #include <linux/pagewalk.h>
29 #include <linux/sched/mm.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/hugetlb.h>
32 #include <linux/pagemap.h>
33 #include <linux/vm_event_item.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
59 #include <linux/tracehook.h>
60 #include <linux/psi.h>
61 #include <linux/seq_buf.h>
67 #include <linux/uaccess.h>
69 #include <trace/events/vmscan.h>
71 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
72 EXPORT_SYMBOL(memory_cgrp_subsys);
74 struct mem_cgroup *root_mem_cgroup __read_mostly;
76 /* Active memory cgroup to use from an interrupt context */
77 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
79 /* Socket memory accounting disabled? */
80 static bool cgroup_memory_nosocket;
82 /* Kernel memory accounting disabled? */
83 static bool cgroup_memory_nokmem;
85 /* Whether the swap controller is active */
86 #ifdef CONFIG_MEMCG_SWAP
87 bool cgroup_memory_noswap __read_mostly;
89 #define cgroup_memory_noswap 1
92 #ifdef CONFIG_CGROUP_WRITEBACK
93 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
96 /* Whether legacy memory+swap accounting is active */
97 static bool do_memsw_account(void)
99 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
102 #define THRESHOLDS_EVENTS_TARGET 128
103 #define SOFTLIMIT_EVENTS_TARGET 1024
106 * Cgroups above their limits are maintained in a RB-Tree, independent of
107 * their hierarchy representation
110 struct mem_cgroup_tree_per_node {
111 struct rb_root rb_root;
112 struct rb_node *rb_rightmost;
116 struct mem_cgroup_tree {
117 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
120 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
123 struct mem_cgroup_eventfd_list {
124 struct list_head list;
125 struct eventfd_ctx *eventfd;
129 * cgroup_event represents events which userspace want to receive.
131 struct mem_cgroup_event {
133 * memcg which the event belongs to.
135 struct mem_cgroup *memcg;
137 * eventfd to signal userspace about the event.
139 struct eventfd_ctx *eventfd;
141 * Each of these stored in a list by the cgroup.
143 struct list_head list;
145 * register_event() callback will be used to add new userspace
146 * waiter for changes related to this event. Use eventfd_signal()
147 * on eventfd to send notification to userspace.
149 int (*register_event)(struct mem_cgroup *memcg,
150 struct eventfd_ctx *eventfd, const char *args);
152 * unregister_event() callback will be called when userspace closes
153 * the eventfd or on cgroup removing. This callback must be set,
154 * if you want provide notification functionality.
156 void (*unregister_event)(struct mem_cgroup *memcg,
157 struct eventfd_ctx *eventfd);
159 * All fields below needed to unregister event when
160 * userspace closes eventfd.
163 wait_queue_head_t *wqh;
164 wait_queue_entry_t wait;
165 struct work_struct remove;
168 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
169 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
171 /* Stuffs for move charges at task migration. */
173 * Types of charges to be moved.
175 #define MOVE_ANON 0x1U
176 #define MOVE_FILE 0x2U
177 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
179 /* "mc" and its members are protected by cgroup_mutex */
180 static struct move_charge_struct {
181 spinlock_t lock; /* for from, to */
182 struct mm_struct *mm;
183 struct mem_cgroup *from;
184 struct mem_cgroup *to;
186 unsigned long precharge;
187 unsigned long moved_charge;
188 unsigned long moved_swap;
189 struct task_struct *moving_task; /* a task moving charges */
190 wait_queue_head_t waitq; /* a waitq for other context */
192 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
193 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
197 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
198 * limit reclaim to prevent infinite loops, if they ever occur.
200 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
201 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
203 /* for encoding cft->private value on file */
212 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
213 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
214 #define MEMFILE_ATTR(val) ((val) & 0xffff)
215 /* Used for OOM nofiier */
216 #define OOM_CONTROL (0)
219 * Iteration constructs for visiting all cgroups (under a tree). If
220 * loops are exited prematurely (break), mem_cgroup_iter_break() must
221 * be used for reference counting.
223 #define for_each_mem_cgroup_tree(iter, root) \
224 for (iter = mem_cgroup_iter(root, NULL, NULL); \
226 iter = mem_cgroup_iter(root, iter, NULL))
228 #define for_each_mem_cgroup(iter) \
229 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
231 iter = mem_cgroup_iter(NULL, iter, NULL))
233 static inline bool should_force_charge(void)
235 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
236 (current->flags & PF_EXITING);
239 /* Some nice accessors for the vmpressure. */
240 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
243 memcg = root_mem_cgroup;
244 return &memcg->vmpressure;
247 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
249 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
252 #ifdef CONFIG_MEMCG_KMEM
253 extern spinlock_t css_set_lock;
255 static void obj_cgroup_release(struct percpu_ref *ref)
257 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
258 struct mem_cgroup *memcg;
259 unsigned int nr_bytes;
260 unsigned int nr_pages;
264 * At this point all allocated objects are freed, and
265 * objcg->nr_charged_bytes can't have an arbitrary byte value.
266 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
268 * The following sequence can lead to it:
269 * 1) CPU0: objcg == stock->cached_objcg
270 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
271 * PAGE_SIZE bytes are charged
272 * 3) CPU1: a process from another memcg is allocating something,
273 * the stock if flushed,
274 * objcg->nr_charged_bytes = PAGE_SIZE - 92
275 * 5) CPU0: we do release this object,
276 * 92 bytes are added to stock->nr_bytes
277 * 6) CPU0: stock is flushed,
278 * 92 bytes are added to objcg->nr_charged_bytes
280 * In the result, nr_charged_bytes == PAGE_SIZE.
281 * This page will be uncharged in obj_cgroup_release().
283 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
284 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
285 nr_pages = nr_bytes >> PAGE_SHIFT;
287 spin_lock_irqsave(&css_set_lock, flags);
288 memcg = obj_cgroup_memcg(objcg);
290 __memcg_kmem_uncharge(memcg, nr_pages);
291 list_del(&objcg->list);
292 mem_cgroup_put(memcg);
293 spin_unlock_irqrestore(&css_set_lock, flags);
295 percpu_ref_exit(ref);
296 kfree_rcu(objcg, rcu);
299 static struct obj_cgroup *obj_cgroup_alloc(void)
301 struct obj_cgroup *objcg;
304 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
308 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
314 INIT_LIST_HEAD(&objcg->list);
318 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
319 struct mem_cgroup *parent)
321 struct obj_cgroup *objcg, *iter;
323 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
325 spin_lock_irq(&css_set_lock);
327 /* Move active objcg to the parent's list */
328 xchg(&objcg->memcg, parent);
329 css_get(&parent->css);
330 list_add(&objcg->list, &parent->objcg_list);
332 /* Move already reparented objcgs to the parent's list */
333 list_for_each_entry(iter, &memcg->objcg_list, list) {
334 css_get(&parent->css);
335 xchg(&iter->memcg, parent);
336 css_put(&memcg->css);
338 list_splice(&memcg->objcg_list, &parent->objcg_list);
340 spin_unlock_irq(&css_set_lock);
342 percpu_ref_kill(&objcg->refcnt);
346 * This will be used as a shrinker list's index.
347 * The main reason for not using cgroup id for this:
348 * this works better in sparse environments, where we have a lot of memcgs,
349 * but only a few kmem-limited. Or also, if we have, for instance, 200
350 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
351 * 200 entry array for that.
353 * The current size of the caches array is stored in memcg_nr_cache_ids. It
354 * will double each time we have to increase it.
356 static DEFINE_IDA(memcg_cache_ida);
357 int memcg_nr_cache_ids;
359 /* Protects memcg_nr_cache_ids */
360 static DECLARE_RWSEM(memcg_cache_ids_sem);
362 void memcg_get_cache_ids(void)
364 down_read(&memcg_cache_ids_sem);
367 void memcg_put_cache_ids(void)
369 up_read(&memcg_cache_ids_sem);
373 * MIN_SIZE is different than 1, because we would like to avoid going through
374 * the alloc/free process all the time. In a small machine, 4 kmem-limited
375 * cgroups is a reasonable guess. In the future, it could be a parameter or
376 * tunable, but that is strictly not necessary.
378 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
379 * this constant directly from cgroup, but it is understandable that this is
380 * better kept as an internal representation in cgroup.c. In any case, the
381 * cgrp_id space is not getting any smaller, and we don't have to necessarily
382 * increase ours as well if it increases.
384 #define MEMCG_CACHES_MIN_SIZE 4
385 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
388 * A lot of the calls to the cache allocation functions are expected to be
389 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
390 * conditional to this static branch, we'll have to allow modules that does
391 * kmem_cache_alloc and the such to see this symbol as well
393 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
394 EXPORT_SYMBOL(memcg_kmem_enabled_key);
397 static int memcg_shrinker_map_size;
398 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
400 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
402 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
405 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
406 int size, int old_size)
408 struct memcg_shrinker_map *new, *old;
411 lockdep_assert_held(&memcg_shrinker_map_mutex);
414 old = rcu_dereference_protected(
415 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
416 /* Not yet online memcg */
420 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
424 /* Set all old bits, clear all new bits */
425 memset(new->map, (int)0xff, old_size);
426 memset((void *)new->map + old_size, 0, size - old_size);
428 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
429 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
435 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
437 struct mem_cgroup_per_node *pn;
438 struct memcg_shrinker_map *map;
441 if (mem_cgroup_is_root(memcg))
445 pn = mem_cgroup_nodeinfo(memcg, nid);
446 map = rcu_dereference_protected(pn->shrinker_map, true);
449 rcu_assign_pointer(pn->shrinker_map, NULL);
453 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
455 struct memcg_shrinker_map *map;
456 int nid, size, ret = 0;
458 if (mem_cgroup_is_root(memcg))
461 mutex_lock(&memcg_shrinker_map_mutex);
462 size = memcg_shrinker_map_size;
464 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
466 memcg_free_shrinker_maps(memcg);
470 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
472 mutex_unlock(&memcg_shrinker_map_mutex);
477 int memcg_expand_shrinker_maps(int new_id)
479 int size, old_size, ret = 0;
480 struct mem_cgroup *memcg;
482 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
483 old_size = memcg_shrinker_map_size;
484 if (size <= old_size)
487 mutex_lock(&memcg_shrinker_map_mutex);
488 if (!root_mem_cgroup)
491 for_each_mem_cgroup(memcg) {
492 if (mem_cgroup_is_root(memcg))
494 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
496 mem_cgroup_iter_break(NULL, memcg);
502 memcg_shrinker_map_size = size;
503 mutex_unlock(&memcg_shrinker_map_mutex);
507 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
509 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
510 struct memcg_shrinker_map *map;
513 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
514 /* Pairs with smp mb in shrink_slab() */
515 smp_mb__before_atomic();
516 set_bit(shrinker_id, map->map);
522 * mem_cgroup_css_from_page - css of the memcg associated with a page
523 * @page: page of interest
525 * If memcg is bound to the default hierarchy, css of the memcg associated
526 * with @page is returned. The returned css remains associated with @page
527 * until it is released.
529 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
532 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
534 struct mem_cgroup *memcg;
536 memcg = page->mem_cgroup;
538 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
539 memcg = root_mem_cgroup;
545 * page_cgroup_ino - return inode number of the memcg a page is charged to
548 * Look up the closest online ancestor of the memory cgroup @page is charged to
549 * and return its inode number or 0 if @page is not charged to any cgroup. It
550 * is safe to call this function without holding a reference to @page.
552 * Note, this function is inherently racy, because there is nothing to prevent
553 * the cgroup inode from getting torn down and potentially reallocated a moment
554 * after page_cgroup_ino() returns, so it only should be used by callers that
555 * do not care (such as procfs interfaces).
557 ino_t page_cgroup_ino(struct page *page)
559 struct mem_cgroup *memcg;
560 unsigned long ino = 0;
563 memcg = page->mem_cgroup;
566 * The lowest bit set means that memcg isn't a valid
567 * memcg pointer, but a obj_cgroups pointer.
568 * In this case the page is shared and doesn't belong
569 * to any specific memory cgroup.
571 if ((unsigned long) memcg & 0x1UL)
574 while (memcg && !(memcg->css.flags & CSS_ONLINE))
575 memcg = parent_mem_cgroup(memcg);
577 ino = cgroup_ino(memcg->css.cgroup);
582 static struct mem_cgroup_per_node *
583 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
585 int nid = page_to_nid(page);
587 return memcg->nodeinfo[nid];
590 static struct mem_cgroup_tree_per_node *
591 soft_limit_tree_node(int nid)
593 return soft_limit_tree.rb_tree_per_node[nid];
596 static struct mem_cgroup_tree_per_node *
597 soft_limit_tree_from_page(struct page *page)
599 int nid = page_to_nid(page);
601 return soft_limit_tree.rb_tree_per_node[nid];
604 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
605 struct mem_cgroup_tree_per_node *mctz,
606 unsigned long new_usage_in_excess)
608 struct rb_node **p = &mctz->rb_root.rb_node;
609 struct rb_node *parent = NULL;
610 struct mem_cgroup_per_node *mz_node;
611 bool rightmost = true;
616 mz->usage_in_excess = new_usage_in_excess;
617 if (!mz->usage_in_excess)
621 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
623 if (mz->usage_in_excess < mz_node->usage_in_excess) {
629 * We can't avoid mem cgroups that are over their soft
630 * limit by the same amount
632 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
637 mctz->rb_rightmost = &mz->tree_node;
639 rb_link_node(&mz->tree_node, parent, p);
640 rb_insert_color(&mz->tree_node, &mctz->rb_root);
644 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
645 struct mem_cgroup_tree_per_node *mctz)
650 if (&mz->tree_node == mctz->rb_rightmost)
651 mctz->rb_rightmost = rb_prev(&mz->tree_node);
653 rb_erase(&mz->tree_node, &mctz->rb_root);
657 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
658 struct mem_cgroup_tree_per_node *mctz)
662 spin_lock_irqsave(&mctz->lock, flags);
663 __mem_cgroup_remove_exceeded(mz, mctz);
664 spin_unlock_irqrestore(&mctz->lock, flags);
667 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
669 unsigned long nr_pages = page_counter_read(&memcg->memory);
670 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
671 unsigned long excess = 0;
673 if (nr_pages > soft_limit)
674 excess = nr_pages - soft_limit;
679 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
681 unsigned long excess;
682 struct mem_cgroup_per_node *mz;
683 struct mem_cgroup_tree_per_node *mctz;
685 mctz = soft_limit_tree_from_page(page);
689 * Necessary to update all ancestors when hierarchy is used.
690 * because their event counter is not touched.
692 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
693 mz = mem_cgroup_page_nodeinfo(memcg, page);
694 excess = soft_limit_excess(memcg);
696 * We have to update the tree if mz is on RB-tree or
697 * mem is over its softlimit.
699 if (excess || mz->on_tree) {
702 spin_lock_irqsave(&mctz->lock, flags);
703 /* if on-tree, remove it */
705 __mem_cgroup_remove_exceeded(mz, mctz);
707 * Insert again. mz->usage_in_excess will be updated.
708 * If excess is 0, no tree ops.
710 __mem_cgroup_insert_exceeded(mz, mctz, excess);
711 spin_unlock_irqrestore(&mctz->lock, flags);
716 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
718 struct mem_cgroup_tree_per_node *mctz;
719 struct mem_cgroup_per_node *mz;
723 mz = mem_cgroup_nodeinfo(memcg, nid);
724 mctz = soft_limit_tree_node(nid);
726 mem_cgroup_remove_exceeded(mz, mctz);
730 static struct mem_cgroup_per_node *
731 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
733 struct mem_cgroup_per_node *mz;
737 if (!mctz->rb_rightmost)
738 goto done; /* Nothing to reclaim from */
740 mz = rb_entry(mctz->rb_rightmost,
741 struct mem_cgroup_per_node, tree_node);
743 * Remove the node now but someone else can add it back,
744 * we will to add it back at the end of reclaim to its correct
745 * position in the tree.
747 __mem_cgroup_remove_exceeded(mz, mctz);
748 if (!soft_limit_excess(mz->memcg) ||
749 !css_tryget(&mz->memcg->css))
755 static struct mem_cgroup_per_node *
756 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
758 struct mem_cgroup_per_node *mz;
760 spin_lock_irq(&mctz->lock);
761 mz = __mem_cgroup_largest_soft_limit_node(mctz);
762 spin_unlock_irq(&mctz->lock);
767 * __mod_memcg_state - update cgroup memory statistics
768 * @memcg: the memory cgroup
769 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
770 * @val: delta to add to the counter, can be negative
772 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
774 long x, threshold = MEMCG_CHARGE_BATCH;
776 if (mem_cgroup_disabled())
779 if (memcg_stat_item_in_bytes(idx))
780 threshold <<= PAGE_SHIFT;
782 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
783 if (unlikely(abs(x) > threshold)) {
784 struct mem_cgroup *mi;
787 * Batch local counters to keep them in sync with
788 * the hierarchical ones.
790 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
791 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
792 atomic_long_add(x, &mi->vmstats[idx]);
795 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
798 static struct mem_cgroup_per_node *
799 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
801 struct mem_cgroup *parent;
803 parent = parent_mem_cgroup(pn->memcg);
806 return mem_cgroup_nodeinfo(parent, nid);
809 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
812 struct mem_cgroup_per_node *pn;
813 struct mem_cgroup *memcg;
814 long x, threshold = MEMCG_CHARGE_BATCH;
816 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
820 __mod_memcg_state(memcg, idx, val);
823 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
825 if (vmstat_item_in_bytes(idx))
826 threshold <<= PAGE_SHIFT;
828 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
829 if (unlikely(abs(x) > threshold)) {
830 pg_data_t *pgdat = lruvec_pgdat(lruvec);
831 struct mem_cgroup_per_node *pi;
833 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
834 atomic_long_add(x, &pi->lruvec_stat[idx]);
837 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
841 * __mod_lruvec_state - update lruvec memory statistics
842 * @lruvec: the lruvec
843 * @idx: the stat item
844 * @val: delta to add to the counter, can be negative
846 * The lruvec is the intersection of the NUMA node and a cgroup. This
847 * function updates the all three counters that are affected by a
848 * change of state at this level: per-node, per-cgroup, per-lruvec.
850 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
854 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
856 /* Update memcg and lruvec */
857 if (!mem_cgroup_disabled())
858 __mod_memcg_lruvec_state(lruvec, idx, val);
861 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
863 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
864 struct mem_cgroup *memcg;
865 struct lruvec *lruvec;
868 memcg = mem_cgroup_from_obj(p);
871 * Untracked pages have no memcg, no lruvec. Update only the
872 * node. If we reparent the slab objects to the root memcg,
873 * when we free the slab object, we need to update the per-memcg
874 * vmstats to keep it correct for the root memcg.
877 __mod_node_page_state(pgdat, idx, val);
879 lruvec = mem_cgroup_lruvec(memcg, pgdat);
880 __mod_lruvec_state(lruvec, idx, val);
885 void mod_memcg_obj_state(void *p, int idx, int val)
887 struct mem_cgroup *memcg;
890 memcg = mem_cgroup_from_obj(p);
892 mod_memcg_state(memcg, idx, val);
897 * __count_memcg_events - account VM events in a cgroup
898 * @memcg: the memory cgroup
899 * @idx: the event item
900 * @count: the number of events that occured
902 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
907 if (mem_cgroup_disabled())
910 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
911 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
912 struct mem_cgroup *mi;
915 * Batch local counters to keep them in sync with
916 * the hierarchical ones.
918 __this_cpu_add(memcg->vmstats_local->events[idx], x);
919 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
920 atomic_long_add(x, &mi->vmevents[idx]);
923 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
926 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
928 return atomic_long_read(&memcg->vmevents[event]);
931 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
936 for_each_possible_cpu(cpu)
937 x += per_cpu(memcg->vmstats_local->events[event], cpu);
941 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
945 /* pagein of a big page is an event. So, ignore page size */
947 __count_memcg_events(memcg, PGPGIN, 1);
949 __count_memcg_events(memcg, PGPGOUT, 1);
950 nr_pages = -nr_pages; /* for event */
953 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
956 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
957 enum mem_cgroup_events_target target)
959 unsigned long val, next;
961 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
962 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
963 /* from time_after() in jiffies.h */
964 if ((long)(next - val) < 0) {
966 case MEM_CGROUP_TARGET_THRESH:
967 next = val + THRESHOLDS_EVENTS_TARGET;
969 case MEM_CGROUP_TARGET_SOFTLIMIT:
970 next = val + SOFTLIMIT_EVENTS_TARGET;
975 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
982 * Check events in order.
985 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
987 /* threshold event is triggered in finer grain than soft limit */
988 if (unlikely(mem_cgroup_event_ratelimit(memcg,
989 MEM_CGROUP_TARGET_THRESH))) {
992 do_softlimit = mem_cgroup_event_ratelimit(memcg,
993 MEM_CGROUP_TARGET_SOFTLIMIT);
994 mem_cgroup_threshold(memcg);
995 if (unlikely(do_softlimit))
996 mem_cgroup_update_tree(memcg, page);
1000 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1003 * mm_update_next_owner() may clear mm->owner to NULL
1004 * if it races with swapoff, page migration, etc.
1005 * So this can be called with p == NULL.
1010 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1012 EXPORT_SYMBOL(mem_cgroup_from_task);
1015 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1016 * @mm: mm from which memcg should be extracted. It can be NULL.
1018 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1019 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1022 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1024 struct mem_cgroup *memcg;
1026 if (mem_cgroup_disabled())
1032 * Page cache insertions can happen withou an
1033 * actual mm context, e.g. during disk probing
1034 * on boot, loopback IO, acct() writes etc.
1037 memcg = root_mem_cgroup;
1039 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1040 if (unlikely(!memcg))
1041 memcg = root_mem_cgroup;
1043 } while (!css_tryget(&memcg->css));
1047 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1050 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1051 * @page: page from which memcg should be extracted.
1053 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1054 * root_mem_cgroup is returned.
1056 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
1058 struct mem_cgroup *memcg = page->mem_cgroup;
1060 if (mem_cgroup_disabled())
1064 /* Page should not get uncharged and freed memcg under us. */
1065 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
1066 memcg = root_mem_cgroup;
1070 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1072 static __always_inline struct mem_cgroup *active_memcg(void)
1075 return this_cpu_read(int_active_memcg);
1077 return current->active_memcg;
1080 static __always_inline struct mem_cgroup *get_active_memcg(void)
1082 struct mem_cgroup *memcg;
1085 memcg = active_memcg();
1087 /* current->active_memcg must hold a ref. */
1088 if (WARN_ON_ONCE(!css_tryget(&memcg->css)))
1089 memcg = root_mem_cgroup;
1091 memcg = current->active_memcg;
1098 static __always_inline bool memcg_kmem_bypass(void)
1100 /* Allow remote memcg charging from any context. */
1101 if (unlikely(active_memcg()))
1104 /* Memcg to charge can't be determined. */
1105 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
1112 * If active memcg is set, do not fallback to current->mm->memcg.
1114 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1116 if (memcg_kmem_bypass())
1119 if (unlikely(active_memcg()))
1120 return get_active_memcg();
1122 return get_mem_cgroup_from_mm(current->mm);
1126 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1127 * @root: hierarchy root
1128 * @prev: previously returned memcg, NULL on first invocation
1129 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1131 * Returns references to children of the hierarchy below @root, or
1132 * @root itself, or %NULL after a full round-trip.
1134 * Caller must pass the return value in @prev on subsequent
1135 * invocations for reference counting, or use mem_cgroup_iter_break()
1136 * to cancel a hierarchy walk before the round-trip is complete.
1138 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1139 * in the hierarchy among all concurrent reclaimers operating on the
1142 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1143 struct mem_cgroup *prev,
1144 struct mem_cgroup_reclaim_cookie *reclaim)
1146 struct mem_cgroup_reclaim_iter *iter;
1147 struct cgroup_subsys_state *css = NULL;
1148 struct mem_cgroup *memcg = NULL;
1149 struct mem_cgroup *pos = NULL;
1151 if (mem_cgroup_disabled())
1155 root = root_mem_cgroup;
1157 if (prev && !reclaim)
1160 if (!root->use_hierarchy && root != root_mem_cgroup) {
1169 struct mem_cgroup_per_node *mz;
1171 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1174 if (prev && reclaim->generation != iter->generation)
1178 pos = READ_ONCE(iter->position);
1179 if (!pos || css_tryget(&pos->css))
1182 * css reference reached zero, so iter->position will
1183 * be cleared by ->css_released. However, we should not
1184 * rely on this happening soon, because ->css_released
1185 * is called from a work queue, and by busy-waiting we
1186 * might block it. So we clear iter->position right
1189 (void)cmpxchg(&iter->position, pos, NULL);
1197 css = css_next_descendant_pre(css, &root->css);
1200 * Reclaimers share the hierarchy walk, and a
1201 * new one might jump in right at the end of
1202 * the hierarchy - make sure they see at least
1203 * one group and restart from the beginning.
1211 * Verify the css and acquire a reference. The root
1212 * is provided by the caller, so we know it's alive
1213 * and kicking, and don't take an extra reference.
1215 memcg = mem_cgroup_from_css(css);
1217 if (css == &root->css)
1220 if (css_tryget(css))
1228 * The position could have already been updated by a competing
1229 * thread, so check that the value hasn't changed since we read
1230 * it to avoid reclaiming from the same cgroup twice.
1232 (void)cmpxchg(&iter->position, pos, memcg);
1240 reclaim->generation = iter->generation;
1246 if (prev && prev != root)
1247 css_put(&prev->css);
1253 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1254 * @root: hierarchy root
1255 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1257 void mem_cgroup_iter_break(struct mem_cgroup *root,
1258 struct mem_cgroup *prev)
1261 root = root_mem_cgroup;
1262 if (prev && prev != root)
1263 css_put(&prev->css);
1266 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1267 struct mem_cgroup *dead_memcg)
1269 struct mem_cgroup_reclaim_iter *iter;
1270 struct mem_cgroup_per_node *mz;
1273 for_each_node(nid) {
1274 mz = mem_cgroup_nodeinfo(from, nid);
1276 cmpxchg(&iter->position, dead_memcg, NULL);
1280 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1282 struct mem_cgroup *memcg = dead_memcg;
1283 struct mem_cgroup *last;
1286 __invalidate_reclaim_iterators(memcg, dead_memcg);
1288 } while ((memcg = parent_mem_cgroup(memcg)));
1291 * When cgruop1 non-hierarchy mode is used,
1292 * parent_mem_cgroup() does not walk all the way up to the
1293 * cgroup root (root_mem_cgroup). So we have to handle
1294 * dead_memcg from cgroup root separately.
1296 if (last != root_mem_cgroup)
1297 __invalidate_reclaim_iterators(root_mem_cgroup,
1302 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1303 * @memcg: hierarchy root
1304 * @fn: function to call for each task
1305 * @arg: argument passed to @fn
1307 * This function iterates over tasks attached to @memcg or to any of its
1308 * descendants and calls @fn for each task. If @fn returns a non-zero
1309 * value, the function breaks the iteration loop and returns the value.
1310 * Otherwise, it will iterate over all tasks and return 0.
1312 * This function must not be called for the root memory cgroup.
1314 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1315 int (*fn)(struct task_struct *, void *), void *arg)
1317 struct mem_cgroup *iter;
1320 BUG_ON(memcg == root_mem_cgroup);
1322 for_each_mem_cgroup_tree(iter, memcg) {
1323 struct css_task_iter it;
1324 struct task_struct *task;
1326 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1327 while (!ret && (task = css_task_iter_next(&it)))
1328 ret = fn(task, arg);
1329 css_task_iter_end(&it);
1331 mem_cgroup_iter_break(memcg, iter);
1339 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1341 * @pgdat: pgdat of the page
1343 * This function relies on page->mem_cgroup being stable - see the
1344 * access rules in commit_charge().
1346 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1348 struct mem_cgroup_per_node *mz;
1349 struct mem_cgroup *memcg;
1350 struct lruvec *lruvec;
1352 if (mem_cgroup_disabled()) {
1353 lruvec = &pgdat->__lruvec;
1357 memcg = page->mem_cgroup;
1359 * Swapcache readahead pages are added to the LRU - and
1360 * possibly migrated - before they are charged.
1363 memcg = root_mem_cgroup;
1365 mz = mem_cgroup_page_nodeinfo(memcg, page);
1366 lruvec = &mz->lruvec;
1369 * Since a node can be onlined after the mem_cgroup was created,
1370 * we have to be prepared to initialize lruvec->zone here;
1371 * and if offlined then reonlined, we need to reinitialize it.
1373 if (unlikely(lruvec->pgdat != pgdat))
1374 lruvec->pgdat = pgdat;
1379 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1380 * @lruvec: mem_cgroup per zone lru vector
1381 * @lru: index of lru list the page is sitting on
1382 * @zid: zone id of the accounted pages
1383 * @nr_pages: positive when adding or negative when removing
1385 * This function must be called under lru_lock, just before a page is added
1386 * to or just after a page is removed from an lru list (that ordering being
1387 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1389 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1390 int zid, int nr_pages)
1392 struct mem_cgroup_per_node *mz;
1393 unsigned long *lru_size;
1396 if (mem_cgroup_disabled())
1399 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1400 lru_size = &mz->lru_zone_size[zid][lru];
1403 *lru_size += nr_pages;
1406 if (WARN_ONCE(size < 0,
1407 "%s(%p, %d, %d): lru_size %ld\n",
1408 __func__, lruvec, lru, nr_pages, size)) {
1414 *lru_size += nr_pages;
1418 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1419 * @memcg: the memory cgroup
1421 * Returns the maximum amount of memory @mem can be charged with, in
1424 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1426 unsigned long margin = 0;
1427 unsigned long count;
1428 unsigned long limit;
1430 count = page_counter_read(&memcg->memory);
1431 limit = READ_ONCE(memcg->memory.max);
1433 margin = limit - count;
1435 if (do_memsw_account()) {
1436 count = page_counter_read(&memcg->memsw);
1437 limit = READ_ONCE(memcg->memsw.max);
1439 margin = min(margin, limit - count);
1448 * A routine for checking "mem" is under move_account() or not.
1450 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1451 * moving cgroups. This is for waiting at high-memory pressure
1454 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1456 struct mem_cgroup *from;
1457 struct mem_cgroup *to;
1460 * Unlike task_move routines, we access mc.to, mc.from not under
1461 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1463 spin_lock(&mc.lock);
1469 ret = mem_cgroup_is_descendant(from, memcg) ||
1470 mem_cgroup_is_descendant(to, memcg);
1472 spin_unlock(&mc.lock);
1476 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1478 if (mc.moving_task && current != mc.moving_task) {
1479 if (mem_cgroup_under_move(memcg)) {
1481 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1482 /* moving charge context might have finished. */
1485 finish_wait(&mc.waitq, &wait);
1492 struct memory_stat {
1498 static struct memory_stat memory_stats[] = {
1499 { "anon", PAGE_SIZE, NR_ANON_MAPPED },
1500 { "file", PAGE_SIZE, NR_FILE_PAGES },
1501 { "kernel_stack", 1024, NR_KERNEL_STACK_KB },
1502 { "percpu", 1, MEMCG_PERCPU_B },
1503 { "sock", PAGE_SIZE, MEMCG_SOCK },
1504 { "shmem", PAGE_SIZE, NR_SHMEM },
1505 { "file_mapped", PAGE_SIZE, NR_FILE_MAPPED },
1506 { "file_dirty", PAGE_SIZE, NR_FILE_DIRTY },
1507 { "file_writeback", PAGE_SIZE, NR_WRITEBACK },
1508 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1510 * The ratio will be initialized in memory_stats_init(). Because
1511 * on some architectures, the macro of HPAGE_PMD_SIZE is not
1512 * constant(e.g. powerpc).
1514 { "anon_thp", 0, NR_ANON_THPS },
1515 { "file_thp", 0, NR_FILE_THPS },
1516 { "shmem_thp", 0, NR_SHMEM_THPS },
1518 { "inactive_anon", PAGE_SIZE, NR_INACTIVE_ANON },
1519 { "active_anon", PAGE_SIZE, NR_ACTIVE_ANON },
1520 { "inactive_file", PAGE_SIZE, NR_INACTIVE_FILE },
1521 { "active_file", PAGE_SIZE, NR_ACTIVE_FILE },
1522 { "unevictable", PAGE_SIZE, NR_UNEVICTABLE },
1525 * Note: The slab_reclaimable and slab_unreclaimable must be
1526 * together and slab_reclaimable must be in front.
1528 { "slab_reclaimable", 1, NR_SLAB_RECLAIMABLE_B },
1529 { "slab_unreclaimable", 1, NR_SLAB_UNRECLAIMABLE_B },
1531 /* The memory events */
1532 { "workingset_refault_anon", 1, WORKINGSET_REFAULT_ANON },
1533 { "workingset_refault_file", 1, WORKINGSET_REFAULT_FILE },
1534 { "workingset_activate_anon", 1, WORKINGSET_ACTIVATE_ANON },
1535 { "workingset_activate_file", 1, WORKINGSET_ACTIVATE_FILE },
1536 { "workingset_restore_anon", 1, WORKINGSET_RESTORE_ANON },
1537 { "workingset_restore_file", 1, WORKINGSET_RESTORE_FILE },
1538 { "workingset_nodereclaim", 1, WORKINGSET_NODERECLAIM },
1541 static int __init memory_stats_init(void)
1545 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1546 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1547 if (memory_stats[i].idx == NR_ANON_THPS ||
1548 memory_stats[i].idx == NR_FILE_THPS ||
1549 memory_stats[i].idx == NR_SHMEM_THPS)
1550 memory_stats[i].ratio = HPAGE_PMD_SIZE;
1552 VM_BUG_ON(!memory_stats[i].ratio);
1553 VM_BUG_ON(memory_stats[i].idx >= MEMCG_NR_STAT);
1558 pure_initcall(memory_stats_init);
1560 static char *memory_stat_format(struct mem_cgroup *memcg)
1565 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1570 * Provide statistics on the state of the memory subsystem as
1571 * well as cumulative event counters that show past behavior.
1573 * This list is ordered following a combination of these gradients:
1574 * 1) generic big picture -> specifics and details
1575 * 2) reflecting userspace activity -> reflecting kernel heuristics
1577 * Current memory state:
1580 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1583 size = memcg_page_state(memcg, memory_stats[i].idx);
1584 size *= memory_stats[i].ratio;
1585 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1587 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1588 size = memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B) +
1589 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B);
1590 seq_buf_printf(&s, "slab %llu\n", size);
1594 /* Accumulated memory events */
1596 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1597 memcg_events(memcg, PGFAULT));
1598 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1599 memcg_events(memcg, PGMAJFAULT));
1600 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1601 memcg_events(memcg, PGREFILL));
1602 seq_buf_printf(&s, "pgscan %lu\n",
1603 memcg_events(memcg, PGSCAN_KSWAPD) +
1604 memcg_events(memcg, PGSCAN_DIRECT));
1605 seq_buf_printf(&s, "pgsteal %lu\n",
1606 memcg_events(memcg, PGSTEAL_KSWAPD) +
1607 memcg_events(memcg, PGSTEAL_DIRECT));
1608 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1609 memcg_events(memcg, PGACTIVATE));
1610 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1611 memcg_events(memcg, PGDEACTIVATE));
1612 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1613 memcg_events(memcg, PGLAZYFREE));
1614 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1615 memcg_events(memcg, PGLAZYFREED));
1617 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1618 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1619 memcg_events(memcg, THP_FAULT_ALLOC));
1620 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1621 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1622 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1624 /* The above should easily fit into one page */
1625 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1630 #define K(x) ((x) << (PAGE_SHIFT-10))
1632 * mem_cgroup_print_oom_context: Print OOM information relevant to
1633 * memory controller.
1634 * @memcg: The memory cgroup that went over limit
1635 * @p: Task that is going to be killed
1637 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1640 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1645 pr_cont(",oom_memcg=");
1646 pr_cont_cgroup_path(memcg->css.cgroup);
1648 pr_cont(",global_oom");
1650 pr_cont(",task_memcg=");
1651 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1657 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1658 * memory controller.
1659 * @memcg: The memory cgroup that went over limit
1661 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1665 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1666 K((u64)page_counter_read(&memcg->memory)),
1667 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1668 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1669 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1670 K((u64)page_counter_read(&memcg->swap)),
1671 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1673 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1674 K((u64)page_counter_read(&memcg->memsw)),
1675 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1676 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1677 K((u64)page_counter_read(&memcg->kmem)),
1678 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1681 pr_info("Memory cgroup stats for ");
1682 pr_cont_cgroup_path(memcg->css.cgroup);
1684 buf = memory_stat_format(memcg);
1692 * Return the memory (and swap, if configured) limit for a memcg.
1694 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1696 unsigned long max = READ_ONCE(memcg->memory.max);
1698 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1699 if (mem_cgroup_swappiness(memcg))
1700 max += min(READ_ONCE(memcg->swap.max),
1701 (unsigned long)total_swap_pages);
1703 if (mem_cgroup_swappiness(memcg)) {
1704 /* Calculate swap excess capacity from memsw limit */
1705 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1707 max += min(swap, (unsigned long)total_swap_pages);
1713 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1715 return page_counter_read(&memcg->memory);
1718 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1721 struct oom_control oc = {
1725 .gfp_mask = gfp_mask,
1730 if (mutex_lock_killable(&oom_lock))
1733 if (mem_cgroup_margin(memcg) >= (1 << order))
1737 * A few threads which were not waiting at mutex_lock_killable() can
1738 * fail to bail out. Therefore, check again after holding oom_lock.
1740 ret = should_force_charge() || out_of_memory(&oc);
1743 mutex_unlock(&oom_lock);
1747 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1750 unsigned long *total_scanned)
1752 struct mem_cgroup *victim = NULL;
1755 unsigned long excess;
1756 unsigned long nr_scanned;
1757 struct mem_cgroup_reclaim_cookie reclaim = {
1761 excess = soft_limit_excess(root_memcg);
1764 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1769 * If we have not been able to reclaim
1770 * anything, it might because there are
1771 * no reclaimable pages under this hierarchy
1776 * We want to do more targeted reclaim.
1777 * excess >> 2 is not to excessive so as to
1778 * reclaim too much, nor too less that we keep
1779 * coming back to reclaim from this cgroup
1781 if (total >= (excess >> 2) ||
1782 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1787 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1788 pgdat, &nr_scanned);
1789 *total_scanned += nr_scanned;
1790 if (!soft_limit_excess(root_memcg))
1793 mem_cgroup_iter_break(root_memcg, victim);
1797 #ifdef CONFIG_LOCKDEP
1798 static struct lockdep_map memcg_oom_lock_dep_map = {
1799 .name = "memcg_oom_lock",
1803 static DEFINE_SPINLOCK(memcg_oom_lock);
1806 * Check OOM-Killer is already running under our hierarchy.
1807 * If someone is running, return false.
1809 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1811 struct mem_cgroup *iter, *failed = NULL;
1813 spin_lock(&memcg_oom_lock);
1815 for_each_mem_cgroup_tree(iter, memcg) {
1816 if (iter->oom_lock) {
1818 * this subtree of our hierarchy is already locked
1819 * so we cannot give a lock.
1822 mem_cgroup_iter_break(memcg, iter);
1825 iter->oom_lock = true;
1830 * OK, we failed to lock the whole subtree so we have
1831 * to clean up what we set up to the failing subtree
1833 for_each_mem_cgroup_tree(iter, memcg) {
1834 if (iter == failed) {
1835 mem_cgroup_iter_break(memcg, iter);
1838 iter->oom_lock = false;
1841 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1843 spin_unlock(&memcg_oom_lock);
1848 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1850 struct mem_cgroup *iter;
1852 spin_lock(&memcg_oom_lock);
1853 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1854 for_each_mem_cgroup_tree(iter, memcg)
1855 iter->oom_lock = false;
1856 spin_unlock(&memcg_oom_lock);
1859 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1861 struct mem_cgroup *iter;
1863 spin_lock(&memcg_oom_lock);
1864 for_each_mem_cgroup_tree(iter, memcg)
1866 spin_unlock(&memcg_oom_lock);
1869 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1871 struct mem_cgroup *iter;
1874 * Be careful about under_oom underflows becase a child memcg
1875 * could have been added after mem_cgroup_mark_under_oom.
1877 spin_lock(&memcg_oom_lock);
1878 for_each_mem_cgroup_tree(iter, memcg)
1879 if (iter->under_oom > 0)
1881 spin_unlock(&memcg_oom_lock);
1884 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1886 struct oom_wait_info {
1887 struct mem_cgroup *memcg;
1888 wait_queue_entry_t wait;
1891 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1892 unsigned mode, int sync, void *arg)
1894 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1895 struct mem_cgroup *oom_wait_memcg;
1896 struct oom_wait_info *oom_wait_info;
1898 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1899 oom_wait_memcg = oom_wait_info->memcg;
1901 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1902 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1904 return autoremove_wake_function(wait, mode, sync, arg);
1907 static void memcg_oom_recover(struct mem_cgroup *memcg)
1910 * For the following lockless ->under_oom test, the only required
1911 * guarantee is that it must see the state asserted by an OOM when
1912 * this function is called as a result of userland actions
1913 * triggered by the notification of the OOM. This is trivially
1914 * achieved by invoking mem_cgroup_mark_under_oom() before
1915 * triggering notification.
1917 if (memcg && memcg->under_oom)
1918 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1928 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1930 enum oom_status ret;
1933 if (order > PAGE_ALLOC_COSTLY_ORDER)
1936 memcg_memory_event(memcg, MEMCG_OOM);
1939 * We are in the middle of the charge context here, so we
1940 * don't want to block when potentially sitting on a callstack
1941 * that holds all kinds of filesystem and mm locks.
1943 * cgroup1 allows disabling the OOM killer and waiting for outside
1944 * handling until the charge can succeed; remember the context and put
1945 * the task to sleep at the end of the page fault when all locks are
1948 * On the other hand, in-kernel OOM killer allows for an async victim
1949 * memory reclaim (oom_reaper) and that means that we are not solely
1950 * relying on the oom victim to make a forward progress and we can
1951 * invoke the oom killer here.
1953 * Please note that mem_cgroup_out_of_memory might fail to find a
1954 * victim and then we have to bail out from the charge path.
1956 if (memcg->oom_kill_disable) {
1957 if (!current->in_user_fault)
1959 css_get(&memcg->css);
1960 current->memcg_in_oom = memcg;
1961 current->memcg_oom_gfp_mask = mask;
1962 current->memcg_oom_order = order;
1967 mem_cgroup_mark_under_oom(memcg);
1969 locked = mem_cgroup_oom_trylock(memcg);
1972 mem_cgroup_oom_notify(memcg);
1974 mem_cgroup_unmark_under_oom(memcg);
1975 if (mem_cgroup_out_of_memory(memcg, mask, order))
1981 mem_cgroup_oom_unlock(memcg);
1987 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1988 * @handle: actually kill/wait or just clean up the OOM state
1990 * This has to be called at the end of a page fault if the memcg OOM
1991 * handler was enabled.
1993 * Memcg supports userspace OOM handling where failed allocations must
1994 * sleep on a waitqueue until the userspace task resolves the
1995 * situation. Sleeping directly in the charge context with all kinds
1996 * of locks held is not a good idea, instead we remember an OOM state
1997 * in the task and mem_cgroup_oom_synchronize() has to be called at
1998 * the end of the page fault to complete the OOM handling.
2000 * Returns %true if an ongoing memcg OOM situation was detected and
2001 * completed, %false otherwise.
2003 bool mem_cgroup_oom_synchronize(bool handle)
2005 struct mem_cgroup *memcg = current->memcg_in_oom;
2006 struct oom_wait_info owait;
2009 /* OOM is global, do not handle */
2016 owait.memcg = memcg;
2017 owait.wait.flags = 0;
2018 owait.wait.func = memcg_oom_wake_function;
2019 owait.wait.private = current;
2020 INIT_LIST_HEAD(&owait.wait.entry);
2022 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2023 mem_cgroup_mark_under_oom(memcg);
2025 locked = mem_cgroup_oom_trylock(memcg);
2028 mem_cgroup_oom_notify(memcg);
2030 if (locked && !memcg->oom_kill_disable) {
2031 mem_cgroup_unmark_under_oom(memcg);
2032 finish_wait(&memcg_oom_waitq, &owait.wait);
2033 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
2034 current->memcg_oom_order);
2037 mem_cgroup_unmark_under_oom(memcg);
2038 finish_wait(&memcg_oom_waitq, &owait.wait);
2042 mem_cgroup_oom_unlock(memcg);
2044 * There is no guarantee that an OOM-lock contender
2045 * sees the wakeups triggered by the OOM kill
2046 * uncharges. Wake any sleepers explicitely.
2048 memcg_oom_recover(memcg);
2051 current->memcg_in_oom = NULL;
2052 css_put(&memcg->css);
2057 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2058 * @victim: task to be killed by the OOM killer
2059 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2061 * Returns a pointer to a memory cgroup, which has to be cleaned up
2062 * by killing all belonging OOM-killable tasks.
2064 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2066 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2067 struct mem_cgroup *oom_domain)
2069 struct mem_cgroup *oom_group = NULL;
2070 struct mem_cgroup *memcg;
2072 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2076 oom_domain = root_mem_cgroup;
2080 memcg = mem_cgroup_from_task(victim);
2081 if (memcg == root_mem_cgroup)
2085 * If the victim task has been asynchronously moved to a different
2086 * memory cgroup, we might end up killing tasks outside oom_domain.
2087 * In this case it's better to ignore memory.group.oom.
2089 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2093 * Traverse the memory cgroup hierarchy from the victim task's
2094 * cgroup up to the OOMing cgroup (or root) to find the
2095 * highest-level memory cgroup with oom.group set.
2097 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2098 if (memcg->oom_group)
2101 if (memcg == oom_domain)
2106 css_get(&oom_group->css);
2113 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2115 pr_info("Tasks in ");
2116 pr_cont_cgroup_path(memcg->css.cgroup);
2117 pr_cont(" are going to be killed due to memory.oom.group set\n");
2121 * lock_page_memcg - lock a page->mem_cgroup binding
2124 * This function protects unlocked LRU pages from being moved to
2127 * It ensures lifetime of the returned memcg. Caller is responsible
2128 * for the lifetime of the page; __unlock_page_memcg() is available
2129 * when @page might get freed inside the locked section.
2131 struct mem_cgroup *lock_page_memcg(struct page *page)
2133 struct page *head = compound_head(page); /* rmap on tail pages */
2134 struct mem_cgroup *memcg;
2135 unsigned long flags;
2138 * The RCU lock is held throughout the transaction. The fast
2139 * path can get away without acquiring the memcg->move_lock
2140 * because page moving starts with an RCU grace period.
2142 * The RCU lock also protects the memcg from being freed when
2143 * the page state that is going to change is the only thing
2144 * preventing the page itself from being freed. E.g. writeback
2145 * doesn't hold a page reference and relies on PG_writeback to
2146 * keep off truncation, migration and so forth.
2150 if (mem_cgroup_disabled())
2153 memcg = head->mem_cgroup;
2154 if (unlikely(!memcg))
2157 if (atomic_read(&memcg->moving_account) <= 0)
2160 spin_lock_irqsave(&memcg->move_lock, flags);
2161 if (memcg != head->mem_cgroup) {
2162 spin_unlock_irqrestore(&memcg->move_lock, flags);
2167 * When charge migration first begins, we can have locked and
2168 * unlocked page stat updates happening concurrently. Track
2169 * the task who has the lock for unlock_page_memcg().
2171 memcg->move_lock_task = current;
2172 memcg->move_lock_flags = flags;
2176 EXPORT_SYMBOL(lock_page_memcg);
2179 * __unlock_page_memcg - unlock and unpin a memcg
2182 * Unlock and unpin a memcg returned by lock_page_memcg().
2184 void __unlock_page_memcg(struct mem_cgroup *memcg)
2186 if (memcg && memcg->move_lock_task == current) {
2187 unsigned long flags = memcg->move_lock_flags;
2189 memcg->move_lock_task = NULL;
2190 memcg->move_lock_flags = 0;
2192 spin_unlock_irqrestore(&memcg->move_lock, flags);
2199 * unlock_page_memcg - unlock a page->mem_cgroup binding
2202 void unlock_page_memcg(struct page *page)
2204 struct page *head = compound_head(page);
2206 __unlock_page_memcg(head->mem_cgroup);
2208 EXPORT_SYMBOL(unlock_page_memcg);
2210 struct memcg_stock_pcp {
2211 struct mem_cgroup *cached; /* this never be root cgroup */
2212 unsigned int nr_pages;
2214 #ifdef CONFIG_MEMCG_KMEM
2215 struct obj_cgroup *cached_objcg;
2216 unsigned int nr_bytes;
2219 struct work_struct work;
2220 unsigned long flags;
2221 #define FLUSHING_CACHED_CHARGE 0
2223 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2224 static DEFINE_MUTEX(percpu_charge_mutex);
2226 #ifdef CONFIG_MEMCG_KMEM
2227 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2228 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2229 struct mem_cgroup *root_memcg);
2232 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2235 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2236 struct mem_cgroup *root_memcg)
2243 * consume_stock: Try to consume stocked charge on this cpu.
2244 * @memcg: memcg to consume from.
2245 * @nr_pages: how many pages to charge.
2247 * The charges will only happen if @memcg matches the current cpu's memcg
2248 * stock, and at least @nr_pages are available in that stock. Failure to
2249 * service an allocation will refill the stock.
2251 * returns true if successful, false otherwise.
2253 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2255 struct memcg_stock_pcp *stock;
2256 unsigned long flags;
2259 if (nr_pages > MEMCG_CHARGE_BATCH)
2262 local_irq_save(flags);
2264 stock = this_cpu_ptr(&memcg_stock);
2265 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2266 stock->nr_pages -= nr_pages;
2270 local_irq_restore(flags);
2276 * Returns stocks cached in percpu and reset cached information.
2278 static void drain_stock(struct memcg_stock_pcp *stock)
2280 struct mem_cgroup *old = stock->cached;
2285 if (stock->nr_pages) {
2286 page_counter_uncharge(&old->memory, stock->nr_pages);
2287 if (do_memsw_account())
2288 page_counter_uncharge(&old->memsw, stock->nr_pages);
2289 stock->nr_pages = 0;
2293 stock->cached = NULL;
2296 static void drain_local_stock(struct work_struct *dummy)
2298 struct memcg_stock_pcp *stock;
2299 unsigned long flags;
2302 * The only protection from memory hotplug vs. drain_stock races is
2303 * that we always operate on local CPU stock here with IRQ disabled
2305 local_irq_save(flags);
2307 stock = this_cpu_ptr(&memcg_stock);
2308 drain_obj_stock(stock);
2310 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2312 local_irq_restore(flags);
2316 * Cache charges(val) to local per_cpu area.
2317 * This will be consumed by consume_stock() function, later.
2319 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2321 struct memcg_stock_pcp *stock;
2322 unsigned long flags;
2324 local_irq_save(flags);
2326 stock = this_cpu_ptr(&memcg_stock);
2327 if (stock->cached != memcg) { /* reset if necessary */
2329 css_get(&memcg->css);
2330 stock->cached = memcg;
2332 stock->nr_pages += nr_pages;
2334 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2337 local_irq_restore(flags);
2341 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2342 * of the hierarchy under it.
2344 static void drain_all_stock(struct mem_cgroup *root_memcg)
2348 /* If someone's already draining, avoid adding running more workers. */
2349 if (!mutex_trylock(&percpu_charge_mutex))
2352 * Notify other cpus that system-wide "drain" is running
2353 * We do not care about races with the cpu hotplug because cpu down
2354 * as well as workers from this path always operate on the local
2355 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2358 for_each_online_cpu(cpu) {
2359 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2360 struct mem_cgroup *memcg;
2364 memcg = stock->cached;
2365 if (memcg && stock->nr_pages &&
2366 mem_cgroup_is_descendant(memcg, root_memcg))
2368 if (obj_stock_flush_required(stock, root_memcg))
2373 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2375 drain_local_stock(&stock->work);
2377 schedule_work_on(cpu, &stock->work);
2381 mutex_unlock(&percpu_charge_mutex);
2384 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2386 struct memcg_stock_pcp *stock;
2387 struct mem_cgroup *memcg, *mi;
2389 stock = &per_cpu(memcg_stock, cpu);
2392 for_each_mem_cgroup(memcg) {
2395 for (i = 0; i < MEMCG_NR_STAT; i++) {
2399 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2401 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2402 atomic_long_add(x, &memcg->vmstats[i]);
2404 if (i >= NR_VM_NODE_STAT_ITEMS)
2407 for_each_node(nid) {
2408 struct mem_cgroup_per_node *pn;
2410 pn = mem_cgroup_nodeinfo(memcg, nid);
2411 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2414 atomic_long_add(x, &pn->lruvec_stat[i]);
2415 } while ((pn = parent_nodeinfo(pn, nid)));
2419 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2422 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2424 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2425 atomic_long_add(x, &memcg->vmevents[i]);
2432 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2433 unsigned int nr_pages,
2436 unsigned long nr_reclaimed = 0;
2439 unsigned long pflags;
2441 if (page_counter_read(&memcg->memory) <=
2442 READ_ONCE(memcg->memory.high))
2445 memcg_memory_event(memcg, MEMCG_HIGH);
2447 psi_memstall_enter(&pflags);
2448 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2450 psi_memstall_leave(&pflags);
2451 } while ((memcg = parent_mem_cgroup(memcg)) &&
2452 !mem_cgroup_is_root(memcg));
2454 return nr_reclaimed;
2457 static void high_work_func(struct work_struct *work)
2459 struct mem_cgroup *memcg;
2461 memcg = container_of(work, struct mem_cgroup, high_work);
2462 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2466 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2467 * enough to still cause a significant slowdown in most cases, while still
2468 * allowing diagnostics and tracing to proceed without becoming stuck.
2470 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2473 * When calculating the delay, we use these either side of the exponentiation to
2474 * maintain precision and scale to a reasonable number of jiffies (see the table
2477 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2478 * overage ratio to a delay.
2479 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2480 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2481 * to produce a reasonable delay curve.
2483 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2484 * reasonable delay curve compared to precision-adjusted overage, not
2485 * penalising heavily at first, but still making sure that growth beyond the
2486 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2487 * example, with a high of 100 megabytes:
2489 * +-------+------------------------+
2490 * | usage | time to allocate in ms |
2491 * +-------+------------------------+
2513 * +-------+------------------------+
2515 #define MEMCG_DELAY_PRECISION_SHIFT 20
2516 #define MEMCG_DELAY_SCALING_SHIFT 14
2518 static u64 calculate_overage(unsigned long usage, unsigned long high)
2526 * Prevent division by 0 in overage calculation by acting as if
2527 * it was a threshold of 1 page
2529 high = max(high, 1UL);
2531 overage = usage - high;
2532 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2533 return div64_u64(overage, high);
2536 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2538 u64 overage, max_overage = 0;
2541 overage = calculate_overage(page_counter_read(&memcg->memory),
2542 READ_ONCE(memcg->memory.high));
2543 max_overage = max(overage, max_overage);
2544 } while ((memcg = parent_mem_cgroup(memcg)) &&
2545 !mem_cgroup_is_root(memcg));
2550 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2552 u64 overage, max_overage = 0;
2555 overage = calculate_overage(page_counter_read(&memcg->swap),
2556 READ_ONCE(memcg->swap.high));
2558 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2559 max_overage = max(overage, max_overage);
2560 } while ((memcg = parent_mem_cgroup(memcg)) &&
2561 !mem_cgroup_is_root(memcg));
2567 * Get the number of jiffies that we should penalise a mischievous cgroup which
2568 * is exceeding its memory.high by checking both it and its ancestors.
2570 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2571 unsigned int nr_pages,
2574 unsigned long penalty_jiffies;
2580 * We use overage compared to memory.high to calculate the number of
2581 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2582 * fairly lenient on small overages, and increasingly harsh when the
2583 * memcg in question makes it clear that it has no intention of stopping
2584 * its crazy behaviour, so we exponentially increase the delay based on
2587 penalty_jiffies = max_overage * max_overage * HZ;
2588 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2589 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2592 * Factor in the task's own contribution to the overage, such that four
2593 * N-sized allocations are throttled approximately the same as one
2594 * 4N-sized allocation.
2596 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2597 * larger the current charge patch is than that.
2599 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2603 * Scheduled by try_charge() to be executed from the userland return path
2604 * and reclaims memory over the high limit.
2606 void mem_cgroup_handle_over_high(void)
2608 unsigned long penalty_jiffies;
2609 unsigned long pflags;
2610 unsigned long nr_reclaimed;
2611 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2612 int nr_retries = MAX_RECLAIM_RETRIES;
2613 struct mem_cgroup *memcg;
2614 bool in_retry = false;
2616 if (likely(!nr_pages))
2619 memcg = get_mem_cgroup_from_mm(current->mm);
2620 current->memcg_nr_pages_over_high = 0;
2624 * The allocating task should reclaim at least the batch size, but for
2625 * subsequent retries we only want to do what's necessary to prevent oom
2626 * or breaching resource isolation.
2628 * This is distinct from memory.max or page allocator behaviour because
2629 * memory.high is currently batched, whereas memory.max and the page
2630 * allocator run every time an allocation is made.
2632 nr_reclaimed = reclaim_high(memcg,
2633 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2637 * memory.high is breached and reclaim is unable to keep up. Throttle
2638 * allocators proactively to slow down excessive growth.
2640 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2641 mem_find_max_overage(memcg));
2643 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2644 swap_find_max_overage(memcg));
2647 * Clamp the max delay per usermode return so as to still keep the
2648 * application moving forwards and also permit diagnostics, albeit
2651 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2654 * Don't sleep if the amount of jiffies this memcg owes us is so low
2655 * that it's not even worth doing, in an attempt to be nice to those who
2656 * go only a small amount over their memory.high value and maybe haven't
2657 * been aggressively reclaimed enough yet.
2659 if (penalty_jiffies <= HZ / 100)
2663 * If reclaim is making forward progress but we're still over
2664 * memory.high, we want to encourage that rather than doing allocator
2667 if (nr_reclaimed || nr_retries--) {
2673 * If we exit early, we're guaranteed to die (since
2674 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2675 * need to account for any ill-begotten jiffies to pay them off later.
2677 psi_memstall_enter(&pflags);
2678 schedule_timeout_killable(penalty_jiffies);
2679 psi_memstall_leave(&pflags);
2682 css_put(&memcg->css);
2685 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2686 unsigned int nr_pages)
2688 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2689 int nr_retries = MAX_RECLAIM_RETRIES;
2690 struct mem_cgroup *mem_over_limit;
2691 struct page_counter *counter;
2692 enum oom_status oom_status;
2693 unsigned long nr_reclaimed;
2694 bool may_swap = true;
2695 bool drained = false;
2696 unsigned long pflags;
2698 if (mem_cgroup_is_root(memcg))
2701 if (consume_stock(memcg, nr_pages))
2704 if (!do_memsw_account() ||
2705 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2706 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2708 if (do_memsw_account())
2709 page_counter_uncharge(&memcg->memsw, batch);
2710 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2712 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2716 if (batch > nr_pages) {
2722 * Memcg doesn't have a dedicated reserve for atomic
2723 * allocations. But like the global atomic pool, we need to
2724 * put the burden of reclaim on regular allocation requests
2725 * and let these go through as privileged allocations.
2727 if (gfp_mask & __GFP_ATOMIC)
2731 * Unlike in global OOM situations, memcg is not in a physical
2732 * memory shortage. Allow dying and OOM-killed tasks to
2733 * bypass the last charges so that they can exit quickly and
2734 * free their memory.
2736 if (unlikely(should_force_charge()))
2740 * Prevent unbounded recursion when reclaim operations need to
2741 * allocate memory. This might exceed the limits temporarily,
2742 * but we prefer facilitating memory reclaim and getting back
2743 * under the limit over triggering OOM kills in these cases.
2745 if (unlikely(current->flags & PF_MEMALLOC))
2748 if (unlikely(task_in_memcg_oom(current)))
2751 if (!gfpflags_allow_blocking(gfp_mask))
2754 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2756 psi_memstall_enter(&pflags);
2757 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2758 gfp_mask, may_swap);
2759 psi_memstall_leave(&pflags);
2761 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2765 drain_all_stock(mem_over_limit);
2770 if (gfp_mask & __GFP_NORETRY)
2773 * Even though the limit is exceeded at this point, reclaim
2774 * may have been able to free some pages. Retry the charge
2775 * before killing the task.
2777 * Only for regular pages, though: huge pages are rather
2778 * unlikely to succeed so close to the limit, and we fall back
2779 * to regular pages anyway in case of failure.
2781 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2784 * At task move, charge accounts can be doubly counted. So, it's
2785 * better to wait until the end of task_move if something is going on.
2787 if (mem_cgroup_wait_acct_move(mem_over_limit))
2793 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2796 if (gfp_mask & __GFP_NOFAIL)
2799 if (fatal_signal_pending(current))
2803 * keep retrying as long as the memcg oom killer is able to make
2804 * a forward progress or bypass the charge if the oom killer
2805 * couldn't make any progress.
2807 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2808 get_order(nr_pages * PAGE_SIZE));
2809 switch (oom_status) {
2811 nr_retries = MAX_RECLAIM_RETRIES;
2819 if (!(gfp_mask & __GFP_NOFAIL))
2823 * The allocation either can't fail or will lead to more memory
2824 * being freed very soon. Allow memory usage go over the limit
2825 * temporarily by force charging it.
2827 page_counter_charge(&memcg->memory, nr_pages);
2828 if (do_memsw_account())
2829 page_counter_charge(&memcg->memsw, nr_pages);
2834 if (batch > nr_pages)
2835 refill_stock(memcg, batch - nr_pages);
2838 * If the hierarchy is above the normal consumption range, schedule
2839 * reclaim on returning to userland. We can perform reclaim here
2840 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2841 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2842 * not recorded as it most likely matches current's and won't
2843 * change in the meantime. As high limit is checked again before
2844 * reclaim, the cost of mismatch is negligible.
2847 bool mem_high, swap_high;
2849 mem_high = page_counter_read(&memcg->memory) >
2850 READ_ONCE(memcg->memory.high);
2851 swap_high = page_counter_read(&memcg->swap) >
2852 READ_ONCE(memcg->swap.high);
2854 /* Don't bother a random interrupted task */
2855 if (in_interrupt()) {
2857 schedule_work(&memcg->high_work);
2863 if (mem_high || swap_high) {
2865 * The allocating tasks in this cgroup will need to do
2866 * reclaim or be throttled to prevent further growth
2867 * of the memory or swap footprints.
2869 * Target some best-effort fairness between the tasks,
2870 * and distribute reclaim work and delay penalties
2871 * based on how much each task is actually allocating.
2873 current->memcg_nr_pages_over_high += batch;
2874 set_notify_resume(current);
2877 } while ((memcg = parent_mem_cgroup(memcg)));
2882 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2883 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2885 if (mem_cgroup_is_root(memcg))
2888 page_counter_uncharge(&memcg->memory, nr_pages);
2889 if (do_memsw_account())
2890 page_counter_uncharge(&memcg->memsw, nr_pages);
2894 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2896 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2898 * Any of the following ensures page->mem_cgroup stability:
2902 * - lock_page_memcg()
2903 * - exclusive reference
2905 page->mem_cgroup = memcg;
2908 #ifdef CONFIG_MEMCG_KMEM
2909 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2912 unsigned int objects = objs_per_slab_page(s, page);
2915 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2920 if (cmpxchg(&page->obj_cgroups, NULL,
2921 (struct obj_cgroup **) ((unsigned long)vec | 0x1UL)))
2924 kmemleak_not_leak(vec);
2930 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2932 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2933 * cgroup_mutex, etc.
2935 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2939 if (mem_cgroup_disabled())
2942 page = virt_to_head_page(p);
2945 * If page->mem_cgroup is set, it's either a simple mem_cgroup pointer
2946 * or a pointer to obj_cgroup vector. In the latter case the lowest
2947 * bit of the pointer is set.
2948 * The page->mem_cgroup pointer can be asynchronously changed
2949 * from NULL to (obj_cgroup_vec | 0x1UL), but can't be changed
2950 * from a valid memcg pointer to objcg vector or back.
2952 if (!page->mem_cgroup)
2956 * Slab objects are accounted individually, not per-page.
2957 * Memcg membership data for each individual object is saved in
2958 * the page->obj_cgroups.
2960 if (page_has_obj_cgroups(page)) {
2961 struct obj_cgroup *objcg;
2964 off = obj_to_index(page->slab_cache, page, p);
2965 objcg = page_obj_cgroups(page)[off];
2967 return obj_cgroup_memcg(objcg);
2972 /* All other pages use page->mem_cgroup */
2973 return page->mem_cgroup;
2976 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2978 struct obj_cgroup *objcg = NULL;
2979 struct mem_cgroup *memcg;
2981 if (memcg_kmem_bypass())
2985 if (unlikely(active_memcg()))
2986 memcg = active_memcg();
2988 memcg = mem_cgroup_from_task(current);
2990 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2991 objcg = rcu_dereference(memcg->objcg);
2992 if (objcg && obj_cgroup_tryget(objcg))
3000 static int memcg_alloc_cache_id(void)
3005 id = ida_simple_get(&memcg_cache_ida,
3006 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3010 if (id < memcg_nr_cache_ids)
3014 * There's no space for the new id in memcg_caches arrays,
3015 * so we have to grow them.
3017 down_write(&memcg_cache_ids_sem);
3019 size = 2 * (id + 1);
3020 if (size < MEMCG_CACHES_MIN_SIZE)
3021 size = MEMCG_CACHES_MIN_SIZE;
3022 else if (size > MEMCG_CACHES_MAX_SIZE)
3023 size = MEMCG_CACHES_MAX_SIZE;
3025 err = memcg_update_all_list_lrus(size);
3027 memcg_nr_cache_ids = size;
3029 up_write(&memcg_cache_ids_sem);
3032 ida_simple_remove(&memcg_cache_ida, id);
3038 static void memcg_free_cache_id(int id)
3040 ida_simple_remove(&memcg_cache_ida, id);
3044 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3045 * @memcg: memory cgroup to charge
3046 * @gfp: reclaim mode
3047 * @nr_pages: number of pages to charge
3049 * Returns 0 on success, an error code on failure.
3051 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
3052 unsigned int nr_pages)
3054 struct page_counter *counter;
3057 ret = try_charge(memcg, gfp, nr_pages);
3061 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3062 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3065 * Enforce __GFP_NOFAIL allocation because callers are not
3066 * prepared to see failures and likely do not have any failure
3069 if (gfp & __GFP_NOFAIL) {
3070 page_counter_charge(&memcg->kmem, nr_pages);
3073 cancel_charge(memcg, nr_pages);
3080 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3081 * @memcg: memcg to uncharge
3082 * @nr_pages: number of pages to uncharge
3084 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3086 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3087 page_counter_uncharge(&memcg->kmem, nr_pages);
3089 page_counter_uncharge(&memcg->memory, nr_pages);
3090 if (do_memsw_account())
3091 page_counter_uncharge(&memcg->memsw, nr_pages);
3095 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3096 * @page: page to charge
3097 * @gfp: reclaim mode
3098 * @order: allocation order
3100 * Returns 0 on success, an error code on failure.
3102 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3104 struct mem_cgroup *memcg;
3107 memcg = get_mem_cgroup_from_current();
3108 if (memcg && !mem_cgroup_is_root(memcg)) {
3109 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3111 page->mem_cgroup = memcg;
3112 __SetPageKmemcg(page);
3115 css_put(&memcg->css);
3121 * __memcg_kmem_uncharge_page: uncharge a kmem page
3122 * @page: page to uncharge
3123 * @order: allocation order
3125 void __memcg_kmem_uncharge_page(struct page *page, int order)
3127 struct mem_cgroup *memcg = page->mem_cgroup;
3128 unsigned int nr_pages = 1 << order;
3133 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3134 __memcg_kmem_uncharge(memcg, nr_pages);
3135 page->mem_cgroup = NULL;
3136 css_put(&memcg->css);
3138 /* slab pages do not have PageKmemcg flag set */
3139 if (PageKmemcg(page))
3140 __ClearPageKmemcg(page);
3143 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3145 struct memcg_stock_pcp *stock;
3146 unsigned long flags;
3149 local_irq_save(flags);
3151 stock = this_cpu_ptr(&memcg_stock);
3152 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3153 stock->nr_bytes -= nr_bytes;
3157 local_irq_restore(flags);
3162 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3164 struct obj_cgroup *old = stock->cached_objcg;
3169 if (stock->nr_bytes) {
3170 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3171 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3175 __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
3180 * The leftover is flushed to the centralized per-memcg value.
3181 * On the next attempt to refill obj stock it will be moved
3182 * to a per-cpu stock (probably, on an other CPU), see
3183 * refill_obj_stock().
3185 * How often it's flushed is a trade-off between the memory
3186 * limit enforcement accuracy and potential CPU contention,
3187 * so it might be changed in the future.
3189 atomic_add(nr_bytes, &old->nr_charged_bytes);
3190 stock->nr_bytes = 0;
3193 obj_cgroup_put(old);
3194 stock->cached_objcg = NULL;
3197 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3198 struct mem_cgroup *root_memcg)
3200 struct mem_cgroup *memcg;
3202 if (stock->cached_objcg) {
3203 memcg = obj_cgroup_memcg(stock->cached_objcg);
3204 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3211 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3213 struct memcg_stock_pcp *stock;
3214 unsigned long flags;
3216 local_irq_save(flags);
3218 stock = this_cpu_ptr(&memcg_stock);
3219 if (stock->cached_objcg != objcg) { /* reset if necessary */
3220 drain_obj_stock(stock);
3221 obj_cgroup_get(objcg);
3222 stock->cached_objcg = objcg;
3223 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3225 stock->nr_bytes += nr_bytes;
3227 if (stock->nr_bytes > PAGE_SIZE)
3228 drain_obj_stock(stock);
3230 local_irq_restore(flags);
3233 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3235 struct mem_cgroup *memcg;
3236 unsigned int nr_pages, nr_bytes;
3239 if (consume_obj_stock(objcg, size))
3243 * In theory, memcg->nr_charged_bytes can have enough
3244 * pre-charged bytes to satisfy the allocation. However,
3245 * flushing memcg->nr_charged_bytes requires two atomic
3246 * operations, and memcg->nr_charged_bytes can't be big,
3247 * so it's better to ignore it and try grab some new pages.
3248 * memcg->nr_charged_bytes will be flushed in
3249 * refill_obj_stock(), called from this function or
3250 * independently later.
3253 memcg = obj_cgroup_memcg(objcg);
3254 css_get(&memcg->css);
3257 nr_pages = size >> PAGE_SHIFT;
3258 nr_bytes = size & (PAGE_SIZE - 1);
3263 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3264 if (!ret && nr_bytes)
3265 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3267 css_put(&memcg->css);
3271 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3273 refill_obj_stock(objcg, size);
3276 #endif /* CONFIG_MEMCG_KMEM */
3278 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3281 * Because tail pages are not marked as "used", set it. We're under
3282 * pgdat->lru_lock and migration entries setup in all page mappings.
3284 void mem_cgroup_split_huge_fixup(struct page *head)
3286 struct mem_cgroup *memcg = head->mem_cgroup;
3289 if (mem_cgroup_disabled())
3292 for (i = 1; i < HPAGE_PMD_NR; i++) {
3293 css_get(&memcg->css);
3294 head[i].mem_cgroup = memcg;
3297 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3299 #ifdef CONFIG_MEMCG_SWAP
3301 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3302 * @entry: swap entry to be moved
3303 * @from: mem_cgroup which the entry is moved from
3304 * @to: mem_cgroup which the entry is moved to
3306 * It succeeds only when the swap_cgroup's record for this entry is the same
3307 * as the mem_cgroup's id of @from.
3309 * Returns 0 on success, -EINVAL on failure.
3311 * The caller must have charged to @to, IOW, called page_counter_charge() about
3312 * both res and memsw, and called css_get().
3314 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3315 struct mem_cgroup *from, struct mem_cgroup *to)
3317 unsigned short old_id, new_id;
3319 old_id = mem_cgroup_id(from);
3320 new_id = mem_cgroup_id(to);
3322 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3323 mod_memcg_state(from, MEMCG_SWAP, -1);
3324 mod_memcg_state(to, MEMCG_SWAP, 1);
3330 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3331 struct mem_cgroup *from, struct mem_cgroup *to)
3337 static DEFINE_MUTEX(memcg_max_mutex);
3339 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3340 unsigned long max, bool memsw)
3342 bool enlarge = false;
3343 bool drained = false;
3345 bool limits_invariant;
3346 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3349 if (signal_pending(current)) {
3354 mutex_lock(&memcg_max_mutex);
3356 * Make sure that the new limit (memsw or memory limit) doesn't
3357 * break our basic invariant rule memory.max <= memsw.max.
3359 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3360 max <= memcg->memsw.max;
3361 if (!limits_invariant) {
3362 mutex_unlock(&memcg_max_mutex);
3366 if (max > counter->max)
3368 ret = page_counter_set_max(counter, max);
3369 mutex_unlock(&memcg_max_mutex);
3375 drain_all_stock(memcg);
3380 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3381 GFP_KERNEL, !memsw)) {
3387 if (!ret && enlarge)
3388 memcg_oom_recover(memcg);
3393 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3395 unsigned long *total_scanned)
3397 unsigned long nr_reclaimed = 0;
3398 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3399 unsigned long reclaimed;
3401 struct mem_cgroup_tree_per_node *mctz;
3402 unsigned long excess;
3403 unsigned long nr_scanned;
3408 mctz = soft_limit_tree_node(pgdat->node_id);
3411 * Do not even bother to check the largest node if the root
3412 * is empty. Do it lockless to prevent lock bouncing. Races
3413 * are acceptable as soft limit is best effort anyway.
3415 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3419 * This loop can run a while, specially if mem_cgroup's continuously
3420 * keep exceeding their soft limit and putting the system under
3427 mz = mem_cgroup_largest_soft_limit_node(mctz);
3432 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3433 gfp_mask, &nr_scanned);
3434 nr_reclaimed += reclaimed;
3435 *total_scanned += nr_scanned;
3436 spin_lock_irq(&mctz->lock);
3437 __mem_cgroup_remove_exceeded(mz, mctz);
3440 * If we failed to reclaim anything from this memory cgroup
3441 * it is time to move on to the next cgroup
3445 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3447 excess = soft_limit_excess(mz->memcg);
3449 * One school of thought says that we should not add
3450 * back the node to the tree if reclaim returns 0.
3451 * But our reclaim could return 0, simply because due
3452 * to priority we are exposing a smaller subset of
3453 * memory to reclaim from. Consider this as a longer
3456 /* If excess == 0, no tree ops */
3457 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3458 spin_unlock_irq(&mctz->lock);
3459 css_put(&mz->memcg->css);
3462 * Could not reclaim anything and there are no more
3463 * mem cgroups to try or we seem to be looping without
3464 * reclaiming anything.
3466 if (!nr_reclaimed &&
3468 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3470 } while (!nr_reclaimed);
3472 css_put(&next_mz->memcg->css);
3473 return nr_reclaimed;
3477 * Test whether @memcg has children, dead or alive. Note that this
3478 * function doesn't care whether @memcg has use_hierarchy enabled and
3479 * returns %true if there are child csses according to the cgroup
3480 * hierarchy. Testing use_hierarchy is the caller's responsibility.
3482 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3487 ret = css_next_child(NULL, &memcg->css);
3493 * Reclaims as many pages from the given memcg as possible.
3495 * Caller is responsible for holding css reference for memcg.
3497 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3499 int nr_retries = MAX_RECLAIM_RETRIES;
3501 /* we call try-to-free pages for make this cgroup empty */
3502 lru_add_drain_all();
3504 drain_all_stock(memcg);
3506 /* try to free all pages in this cgroup */
3507 while (nr_retries && page_counter_read(&memcg->memory)) {
3510 if (signal_pending(current))
3513 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3517 /* maybe some writeback is necessary */
3518 congestion_wait(BLK_RW_ASYNC, HZ/10);
3526 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3527 char *buf, size_t nbytes,
3530 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3532 if (mem_cgroup_is_root(memcg))
3534 return mem_cgroup_force_empty(memcg) ?: nbytes;
3537 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3540 return mem_cgroup_from_css(css)->use_hierarchy;
3543 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3544 struct cftype *cft, u64 val)
3547 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3548 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3550 if (memcg->use_hierarchy == val)
3554 * If parent's use_hierarchy is set, we can't make any modifications
3555 * in the child subtrees. If it is unset, then the change can
3556 * occur, provided the current cgroup has no children.
3558 * For the root cgroup, parent_mem is NULL, we allow value to be
3559 * set if there are no children.
3561 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3562 (val == 1 || val == 0)) {
3563 if (!memcg_has_children(memcg))
3564 memcg->use_hierarchy = val;
3573 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3577 if (mem_cgroup_is_root(memcg)) {
3578 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3579 memcg_page_state(memcg, NR_ANON_MAPPED);
3581 val += memcg_page_state(memcg, MEMCG_SWAP);
3584 val = page_counter_read(&memcg->memory);
3586 val = page_counter_read(&memcg->memsw);
3599 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3602 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3603 struct page_counter *counter;
3605 switch (MEMFILE_TYPE(cft->private)) {
3607 counter = &memcg->memory;
3610 counter = &memcg->memsw;
3613 counter = &memcg->kmem;
3616 counter = &memcg->tcpmem;
3622 switch (MEMFILE_ATTR(cft->private)) {
3624 if (counter == &memcg->memory)
3625 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3626 if (counter == &memcg->memsw)
3627 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3628 return (u64)page_counter_read(counter) * PAGE_SIZE;
3630 return (u64)counter->max * PAGE_SIZE;
3632 return (u64)counter->watermark * PAGE_SIZE;
3634 return counter->failcnt;
3635 case RES_SOFT_LIMIT:
3636 return (u64)memcg->soft_limit * PAGE_SIZE;
3642 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3644 unsigned long stat[MEMCG_NR_STAT] = {0};
3645 struct mem_cgroup *mi;
3648 for_each_online_cpu(cpu)
3649 for (i = 0; i < MEMCG_NR_STAT; i++)
3650 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3652 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3653 for (i = 0; i < MEMCG_NR_STAT; i++)
3654 atomic_long_add(stat[i], &mi->vmstats[i]);
3656 for_each_node(node) {
3657 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3658 struct mem_cgroup_per_node *pi;
3660 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3663 for_each_online_cpu(cpu)
3664 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3666 pn->lruvec_stat_cpu->count[i], cpu);
3668 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3669 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3670 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3674 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3676 unsigned long events[NR_VM_EVENT_ITEMS];
3677 struct mem_cgroup *mi;
3680 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3683 for_each_online_cpu(cpu)
3684 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3685 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3688 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3689 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3690 atomic_long_add(events[i], &mi->vmevents[i]);
3693 #ifdef CONFIG_MEMCG_KMEM
3694 static int memcg_online_kmem(struct mem_cgroup *memcg)
3696 struct obj_cgroup *objcg;
3699 if (cgroup_memory_nokmem)
3702 BUG_ON(memcg->kmemcg_id >= 0);
3703 BUG_ON(memcg->kmem_state);
3705 memcg_id = memcg_alloc_cache_id();
3709 objcg = obj_cgroup_alloc();
3711 memcg_free_cache_id(memcg_id);
3714 objcg->memcg = memcg;
3715 rcu_assign_pointer(memcg->objcg, objcg);
3717 static_branch_enable(&memcg_kmem_enabled_key);
3720 * A memory cgroup is considered kmem-online as soon as it gets
3721 * kmemcg_id. Setting the id after enabling static branching will
3722 * guarantee no one starts accounting before all call sites are
3725 memcg->kmemcg_id = memcg_id;
3726 memcg->kmem_state = KMEM_ONLINE;
3731 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3733 struct cgroup_subsys_state *css;
3734 struct mem_cgroup *parent, *child;
3737 if (memcg->kmem_state != KMEM_ONLINE)
3740 memcg->kmem_state = KMEM_ALLOCATED;
3742 parent = parent_mem_cgroup(memcg);
3744 parent = root_mem_cgroup;
3746 memcg_reparent_objcgs(memcg, parent);
3748 kmemcg_id = memcg->kmemcg_id;
3749 BUG_ON(kmemcg_id < 0);
3752 * Change kmemcg_id of this cgroup and all its descendants to the
3753 * parent's id, and then move all entries from this cgroup's list_lrus
3754 * to ones of the parent. After we have finished, all list_lrus
3755 * corresponding to this cgroup are guaranteed to remain empty. The
3756 * ordering is imposed by list_lru_node->lock taken by
3757 * memcg_drain_all_list_lrus().
3759 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3760 css_for_each_descendant_pre(css, &memcg->css) {
3761 child = mem_cgroup_from_css(css);
3762 BUG_ON(child->kmemcg_id != kmemcg_id);
3763 child->kmemcg_id = parent->kmemcg_id;
3764 if (!memcg->use_hierarchy)
3769 memcg_drain_all_list_lrus(kmemcg_id, parent);
3771 memcg_free_cache_id(kmemcg_id);
3774 static void memcg_free_kmem(struct mem_cgroup *memcg)
3776 /* css_alloc() failed, offlining didn't happen */
3777 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3778 memcg_offline_kmem(memcg);
3781 static int memcg_online_kmem(struct mem_cgroup *memcg)
3785 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3788 static void memcg_free_kmem(struct mem_cgroup *memcg)
3791 #endif /* CONFIG_MEMCG_KMEM */
3793 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3798 mutex_lock(&memcg_max_mutex);
3799 ret = page_counter_set_max(&memcg->kmem, max);
3800 mutex_unlock(&memcg_max_mutex);
3804 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3808 mutex_lock(&memcg_max_mutex);
3810 ret = page_counter_set_max(&memcg->tcpmem, max);
3814 if (!memcg->tcpmem_active) {
3816 * The active flag needs to be written after the static_key
3817 * update. This is what guarantees that the socket activation
3818 * function is the last one to run. See mem_cgroup_sk_alloc()
3819 * for details, and note that we don't mark any socket as
3820 * belonging to this memcg until that flag is up.
3822 * We need to do this, because static_keys will span multiple
3823 * sites, but we can't control their order. If we mark a socket
3824 * as accounted, but the accounting functions are not patched in
3825 * yet, we'll lose accounting.
3827 * We never race with the readers in mem_cgroup_sk_alloc(),
3828 * because when this value change, the code to process it is not
3831 static_branch_inc(&memcg_sockets_enabled_key);
3832 memcg->tcpmem_active = true;
3835 mutex_unlock(&memcg_max_mutex);
3840 * The user of this function is...
3843 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3844 char *buf, size_t nbytes, loff_t off)
3846 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3847 unsigned long nr_pages;
3850 buf = strstrip(buf);
3851 ret = page_counter_memparse(buf, "-1", &nr_pages);
3855 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3857 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3861 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3863 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3866 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3869 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3870 "Please report your usecase to linux-mm@kvack.org if you "
3871 "depend on this functionality.\n");
3872 ret = memcg_update_kmem_max(memcg, nr_pages);
3875 ret = memcg_update_tcp_max(memcg, nr_pages);
3879 case RES_SOFT_LIMIT:
3880 memcg->soft_limit = nr_pages;
3884 return ret ?: nbytes;
3887 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3888 size_t nbytes, loff_t off)
3890 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3891 struct page_counter *counter;
3893 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3895 counter = &memcg->memory;
3898 counter = &memcg->memsw;
3901 counter = &memcg->kmem;
3904 counter = &memcg->tcpmem;
3910 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3912 page_counter_reset_watermark(counter);
3915 counter->failcnt = 0;
3924 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3927 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3931 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3932 struct cftype *cft, u64 val)
3934 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3936 if (val & ~MOVE_MASK)
3940 * No kind of locking is needed in here, because ->can_attach() will
3941 * check this value once in the beginning of the process, and then carry
3942 * on with stale data. This means that changes to this value will only
3943 * affect task migrations starting after the change.
3945 memcg->move_charge_at_immigrate = val;
3949 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3950 struct cftype *cft, u64 val)
3958 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3959 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3960 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3962 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3963 int nid, unsigned int lru_mask, bool tree)
3965 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3966 unsigned long nr = 0;
3969 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3972 if (!(BIT(lru) & lru_mask))
3975 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3977 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3982 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3983 unsigned int lru_mask,
3986 unsigned long nr = 0;
3990 if (!(BIT(lru) & lru_mask))
3993 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3995 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
4000 static int memcg_numa_stat_show(struct seq_file *m, void *v)
4004 unsigned int lru_mask;
4007 static const struct numa_stat stats[] = {
4008 { "total", LRU_ALL },
4009 { "file", LRU_ALL_FILE },
4010 { "anon", LRU_ALL_ANON },
4011 { "unevictable", BIT(LRU_UNEVICTABLE) },
4013 const struct numa_stat *stat;
4015 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4017 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4018 seq_printf(m, "%s=%lu", stat->name,
4019 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4021 for_each_node_state(nid, N_MEMORY)
4022 seq_printf(m, " N%d=%lu", nid,
4023 mem_cgroup_node_nr_lru_pages(memcg, nid,
4024 stat->lru_mask, false));
4028 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4030 seq_printf(m, "hierarchical_%s=%lu", stat->name,
4031 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4033 for_each_node_state(nid, N_MEMORY)
4034 seq_printf(m, " N%d=%lu", nid,
4035 mem_cgroup_node_nr_lru_pages(memcg, nid,
4036 stat->lru_mask, true));
4042 #endif /* CONFIG_NUMA */
4044 static const unsigned int memcg1_stats[] = {
4047 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4057 static const char *const memcg1_stat_names[] = {
4060 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4070 /* Universal VM events cgroup1 shows, original sort order */
4071 static const unsigned int memcg1_events[] = {
4078 static int memcg_stat_show(struct seq_file *m, void *v)
4080 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4081 unsigned long memory, memsw;
4082 struct mem_cgroup *mi;
4085 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4087 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4090 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4092 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4093 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4094 if (memcg1_stats[i] == NR_ANON_THPS)
4097 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4100 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4101 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4102 memcg_events_local(memcg, memcg1_events[i]));
4104 for (i = 0; i < NR_LRU_LISTS; i++)
4105 seq_printf(m, "%s %lu\n", lru_list_name(i),
4106 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4109 /* Hierarchical information */
4110 memory = memsw = PAGE_COUNTER_MAX;
4111 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4112 memory = min(memory, READ_ONCE(mi->memory.max));
4113 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4115 seq_printf(m, "hierarchical_memory_limit %llu\n",
4116 (u64)memory * PAGE_SIZE);
4117 if (do_memsw_account())
4118 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4119 (u64)memsw * PAGE_SIZE);
4121 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4124 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4126 nr = memcg_page_state(memcg, memcg1_stats[i]);
4127 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4128 if (memcg1_stats[i] == NR_ANON_THPS)
4131 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4132 (u64)nr * PAGE_SIZE);
4135 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4136 seq_printf(m, "total_%s %llu\n",
4137 vm_event_name(memcg1_events[i]),
4138 (u64)memcg_events(memcg, memcg1_events[i]));
4140 for (i = 0; i < NR_LRU_LISTS; i++)
4141 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4142 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4145 #ifdef CONFIG_DEBUG_VM
4148 struct mem_cgroup_per_node *mz;
4149 unsigned long anon_cost = 0;
4150 unsigned long file_cost = 0;
4152 for_each_online_pgdat(pgdat) {
4153 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4155 anon_cost += mz->lruvec.anon_cost;
4156 file_cost += mz->lruvec.file_cost;
4158 seq_printf(m, "anon_cost %lu\n", anon_cost);
4159 seq_printf(m, "file_cost %lu\n", file_cost);
4166 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4169 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4171 return mem_cgroup_swappiness(memcg);
4174 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4175 struct cftype *cft, u64 val)
4177 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4183 memcg->swappiness = val;
4185 vm_swappiness = val;
4190 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4192 struct mem_cgroup_threshold_ary *t;
4193 unsigned long usage;
4198 t = rcu_dereference(memcg->thresholds.primary);
4200 t = rcu_dereference(memcg->memsw_thresholds.primary);
4205 usage = mem_cgroup_usage(memcg, swap);
4208 * current_threshold points to threshold just below or equal to usage.
4209 * If it's not true, a threshold was crossed after last
4210 * call of __mem_cgroup_threshold().
4212 i = t->current_threshold;
4215 * Iterate backward over array of thresholds starting from
4216 * current_threshold and check if a threshold is crossed.
4217 * If none of thresholds below usage is crossed, we read
4218 * only one element of the array here.
4220 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4221 eventfd_signal(t->entries[i].eventfd, 1);
4223 /* i = current_threshold + 1 */
4227 * Iterate forward over array of thresholds starting from
4228 * current_threshold+1 and check if a threshold is crossed.
4229 * If none of thresholds above usage is crossed, we read
4230 * only one element of the array here.
4232 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4233 eventfd_signal(t->entries[i].eventfd, 1);
4235 /* Update current_threshold */
4236 t->current_threshold = i - 1;
4241 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4244 __mem_cgroup_threshold(memcg, false);
4245 if (do_memsw_account())
4246 __mem_cgroup_threshold(memcg, true);
4248 memcg = parent_mem_cgroup(memcg);
4252 static int compare_thresholds(const void *a, const void *b)
4254 const struct mem_cgroup_threshold *_a = a;
4255 const struct mem_cgroup_threshold *_b = b;
4257 if (_a->threshold > _b->threshold)
4260 if (_a->threshold < _b->threshold)
4266 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4268 struct mem_cgroup_eventfd_list *ev;
4270 spin_lock(&memcg_oom_lock);
4272 list_for_each_entry(ev, &memcg->oom_notify, list)
4273 eventfd_signal(ev->eventfd, 1);
4275 spin_unlock(&memcg_oom_lock);
4279 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4281 struct mem_cgroup *iter;
4283 for_each_mem_cgroup_tree(iter, memcg)
4284 mem_cgroup_oom_notify_cb(iter);
4287 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4288 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4290 struct mem_cgroup_thresholds *thresholds;
4291 struct mem_cgroup_threshold_ary *new;
4292 unsigned long threshold;
4293 unsigned long usage;
4296 ret = page_counter_memparse(args, "-1", &threshold);
4300 mutex_lock(&memcg->thresholds_lock);
4303 thresholds = &memcg->thresholds;
4304 usage = mem_cgroup_usage(memcg, false);
4305 } else if (type == _MEMSWAP) {
4306 thresholds = &memcg->memsw_thresholds;
4307 usage = mem_cgroup_usage(memcg, true);
4311 /* Check if a threshold crossed before adding a new one */
4312 if (thresholds->primary)
4313 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4315 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4317 /* Allocate memory for new array of thresholds */
4318 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4325 /* Copy thresholds (if any) to new array */
4326 if (thresholds->primary)
4327 memcpy(new->entries, thresholds->primary->entries,
4328 flex_array_size(new, entries, size - 1));
4330 /* Add new threshold */
4331 new->entries[size - 1].eventfd = eventfd;
4332 new->entries[size - 1].threshold = threshold;
4334 /* Sort thresholds. Registering of new threshold isn't time-critical */
4335 sort(new->entries, size, sizeof(*new->entries),
4336 compare_thresholds, NULL);
4338 /* Find current threshold */
4339 new->current_threshold = -1;
4340 for (i = 0; i < size; i++) {
4341 if (new->entries[i].threshold <= usage) {
4343 * new->current_threshold will not be used until
4344 * rcu_assign_pointer(), so it's safe to increment
4347 ++new->current_threshold;
4352 /* Free old spare buffer and save old primary buffer as spare */
4353 kfree(thresholds->spare);
4354 thresholds->spare = thresholds->primary;
4356 rcu_assign_pointer(thresholds->primary, new);
4358 /* To be sure that nobody uses thresholds */
4362 mutex_unlock(&memcg->thresholds_lock);
4367 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4368 struct eventfd_ctx *eventfd, const char *args)
4370 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4373 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4374 struct eventfd_ctx *eventfd, const char *args)
4376 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4379 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4380 struct eventfd_ctx *eventfd, enum res_type type)
4382 struct mem_cgroup_thresholds *thresholds;
4383 struct mem_cgroup_threshold_ary *new;
4384 unsigned long usage;
4385 int i, j, size, entries;
4387 mutex_lock(&memcg->thresholds_lock);
4390 thresholds = &memcg->thresholds;
4391 usage = mem_cgroup_usage(memcg, false);
4392 } else if (type == _MEMSWAP) {
4393 thresholds = &memcg->memsw_thresholds;
4394 usage = mem_cgroup_usage(memcg, true);
4398 if (!thresholds->primary)
4401 /* Check if a threshold crossed before removing */
4402 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4404 /* Calculate new number of threshold */
4406 for (i = 0; i < thresholds->primary->size; i++) {
4407 if (thresholds->primary->entries[i].eventfd != eventfd)
4413 new = thresholds->spare;
4415 /* If no items related to eventfd have been cleared, nothing to do */
4419 /* Set thresholds array to NULL if we don't have thresholds */
4428 /* Copy thresholds and find current threshold */
4429 new->current_threshold = -1;
4430 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4431 if (thresholds->primary->entries[i].eventfd == eventfd)
4434 new->entries[j] = thresholds->primary->entries[i];
4435 if (new->entries[j].threshold <= usage) {
4437 * new->current_threshold will not be used
4438 * until rcu_assign_pointer(), so it's safe to increment
4441 ++new->current_threshold;
4447 /* Swap primary and spare array */
4448 thresholds->spare = thresholds->primary;
4450 rcu_assign_pointer(thresholds->primary, new);
4452 /* To be sure that nobody uses thresholds */
4455 /* If all events are unregistered, free the spare array */
4457 kfree(thresholds->spare);
4458 thresholds->spare = NULL;
4461 mutex_unlock(&memcg->thresholds_lock);
4464 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4465 struct eventfd_ctx *eventfd)
4467 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4470 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4471 struct eventfd_ctx *eventfd)
4473 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4476 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4477 struct eventfd_ctx *eventfd, const char *args)
4479 struct mem_cgroup_eventfd_list *event;
4481 event = kmalloc(sizeof(*event), GFP_KERNEL);
4485 spin_lock(&memcg_oom_lock);
4487 event->eventfd = eventfd;
4488 list_add(&event->list, &memcg->oom_notify);
4490 /* already in OOM ? */
4491 if (memcg->under_oom)
4492 eventfd_signal(eventfd, 1);
4493 spin_unlock(&memcg_oom_lock);
4498 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4499 struct eventfd_ctx *eventfd)
4501 struct mem_cgroup_eventfd_list *ev, *tmp;
4503 spin_lock(&memcg_oom_lock);
4505 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4506 if (ev->eventfd == eventfd) {
4507 list_del(&ev->list);
4512 spin_unlock(&memcg_oom_lock);
4515 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4517 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4519 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4520 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4521 seq_printf(sf, "oom_kill %lu\n",
4522 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4526 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4527 struct cftype *cft, u64 val)
4529 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4531 /* cannot set to root cgroup and only 0 and 1 are allowed */
4532 if (!css->parent || !((val == 0) || (val == 1)))
4535 memcg->oom_kill_disable = val;
4537 memcg_oom_recover(memcg);
4542 #ifdef CONFIG_CGROUP_WRITEBACK
4544 #include <trace/events/writeback.h>
4546 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4548 return wb_domain_init(&memcg->cgwb_domain, gfp);
4551 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4553 wb_domain_exit(&memcg->cgwb_domain);
4556 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4558 wb_domain_size_changed(&memcg->cgwb_domain);
4561 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4563 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4565 if (!memcg->css.parent)
4568 return &memcg->cgwb_domain;
4572 * idx can be of type enum memcg_stat_item or node_stat_item.
4573 * Keep in sync with memcg_exact_page().
4575 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4577 long x = atomic_long_read(&memcg->vmstats[idx]);
4580 for_each_online_cpu(cpu)
4581 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4588 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4589 * @wb: bdi_writeback in question
4590 * @pfilepages: out parameter for number of file pages
4591 * @pheadroom: out parameter for number of allocatable pages according to memcg
4592 * @pdirty: out parameter for number of dirty pages
4593 * @pwriteback: out parameter for number of pages under writeback
4595 * Determine the numbers of file, headroom, dirty, and writeback pages in
4596 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4597 * is a bit more involved.
4599 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4600 * headroom is calculated as the lowest headroom of itself and the
4601 * ancestors. Note that this doesn't consider the actual amount of
4602 * available memory in the system. The caller should further cap
4603 * *@pheadroom accordingly.
4605 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4606 unsigned long *pheadroom, unsigned long *pdirty,
4607 unsigned long *pwriteback)
4609 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4610 struct mem_cgroup *parent;
4612 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4614 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4615 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4616 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4617 *pheadroom = PAGE_COUNTER_MAX;
4619 while ((parent = parent_mem_cgroup(memcg))) {
4620 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4621 READ_ONCE(memcg->memory.high));
4622 unsigned long used = page_counter_read(&memcg->memory);
4624 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4630 * Foreign dirty flushing
4632 * There's an inherent mismatch between memcg and writeback. The former
4633 * trackes ownership per-page while the latter per-inode. This was a
4634 * deliberate design decision because honoring per-page ownership in the
4635 * writeback path is complicated, may lead to higher CPU and IO overheads
4636 * and deemed unnecessary given that write-sharing an inode across
4637 * different cgroups isn't a common use-case.
4639 * Combined with inode majority-writer ownership switching, this works well
4640 * enough in most cases but there are some pathological cases. For
4641 * example, let's say there are two cgroups A and B which keep writing to
4642 * different but confined parts of the same inode. B owns the inode and
4643 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4644 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4645 * triggering background writeback. A will be slowed down without a way to
4646 * make writeback of the dirty pages happen.
4648 * Conditions like the above can lead to a cgroup getting repatedly and
4649 * severely throttled after making some progress after each
4650 * dirty_expire_interval while the underyling IO device is almost
4653 * Solving this problem completely requires matching the ownership tracking
4654 * granularities between memcg and writeback in either direction. However,
4655 * the more egregious behaviors can be avoided by simply remembering the
4656 * most recent foreign dirtying events and initiating remote flushes on
4657 * them when local writeback isn't enough to keep the memory clean enough.
4659 * The following two functions implement such mechanism. When a foreign
4660 * page - a page whose memcg and writeback ownerships don't match - is
4661 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4662 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4663 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4664 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4665 * foreign bdi_writebacks which haven't expired. Both the numbers of
4666 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4667 * limited to MEMCG_CGWB_FRN_CNT.
4669 * The mechanism only remembers IDs and doesn't hold any object references.
4670 * As being wrong occasionally doesn't matter, updates and accesses to the
4671 * records are lockless and racy.
4673 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4674 struct bdi_writeback *wb)
4676 struct mem_cgroup *memcg = page->mem_cgroup;
4677 struct memcg_cgwb_frn *frn;
4678 u64 now = get_jiffies_64();
4679 u64 oldest_at = now;
4683 trace_track_foreign_dirty(page, wb);
4686 * Pick the slot to use. If there is already a slot for @wb, keep
4687 * using it. If not replace the oldest one which isn't being
4690 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4691 frn = &memcg->cgwb_frn[i];
4692 if (frn->bdi_id == wb->bdi->id &&
4693 frn->memcg_id == wb->memcg_css->id)
4695 if (time_before64(frn->at, oldest_at) &&
4696 atomic_read(&frn->done.cnt) == 1) {
4698 oldest_at = frn->at;
4702 if (i < MEMCG_CGWB_FRN_CNT) {
4704 * Re-using an existing one. Update timestamp lazily to
4705 * avoid making the cacheline hot. We want them to be
4706 * reasonably up-to-date and significantly shorter than
4707 * dirty_expire_interval as that's what expires the record.
4708 * Use the shorter of 1s and dirty_expire_interval / 8.
4710 unsigned long update_intv =
4711 min_t(unsigned long, HZ,
4712 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4714 if (time_before64(frn->at, now - update_intv))
4716 } else if (oldest >= 0) {
4717 /* replace the oldest free one */
4718 frn = &memcg->cgwb_frn[oldest];
4719 frn->bdi_id = wb->bdi->id;
4720 frn->memcg_id = wb->memcg_css->id;
4725 /* issue foreign writeback flushes for recorded foreign dirtying events */
4726 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4728 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4729 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4730 u64 now = jiffies_64;
4733 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4734 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4737 * If the record is older than dirty_expire_interval,
4738 * writeback on it has already started. No need to kick it
4739 * off again. Also, don't start a new one if there's
4740 * already one in flight.
4742 if (time_after64(frn->at, now - intv) &&
4743 atomic_read(&frn->done.cnt) == 1) {
4745 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4746 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4747 WB_REASON_FOREIGN_FLUSH,
4753 #else /* CONFIG_CGROUP_WRITEBACK */
4755 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4760 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4764 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4768 #endif /* CONFIG_CGROUP_WRITEBACK */
4771 * DO NOT USE IN NEW FILES.
4773 * "cgroup.event_control" implementation.
4775 * This is way over-engineered. It tries to support fully configurable
4776 * events for each user. Such level of flexibility is completely
4777 * unnecessary especially in the light of the planned unified hierarchy.
4779 * Please deprecate this and replace with something simpler if at all
4784 * Unregister event and free resources.
4786 * Gets called from workqueue.
4788 static void memcg_event_remove(struct work_struct *work)
4790 struct mem_cgroup_event *event =
4791 container_of(work, struct mem_cgroup_event, remove);
4792 struct mem_cgroup *memcg = event->memcg;
4794 remove_wait_queue(event->wqh, &event->wait);
4796 event->unregister_event(memcg, event->eventfd);
4798 /* Notify userspace the event is going away. */
4799 eventfd_signal(event->eventfd, 1);
4801 eventfd_ctx_put(event->eventfd);
4803 css_put(&memcg->css);
4807 * Gets called on EPOLLHUP on eventfd when user closes it.
4809 * Called with wqh->lock held and interrupts disabled.
4811 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4812 int sync, void *key)
4814 struct mem_cgroup_event *event =
4815 container_of(wait, struct mem_cgroup_event, wait);
4816 struct mem_cgroup *memcg = event->memcg;
4817 __poll_t flags = key_to_poll(key);
4819 if (flags & EPOLLHUP) {
4821 * If the event has been detached at cgroup removal, we
4822 * can simply return knowing the other side will cleanup
4825 * We can't race against event freeing since the other
4826 * side will require wqh->lock via remove_wait_queue(),
4829 spin_lock(&memcg->event_list_lock);
4830 if (!list_empty(&event->list)) {
4831 list_del_init(&event->list);
4833 * We are in atomic context, but cgroup_event_remove()
4834 * may sleep, so we have to call it in workqueue.
4836 schedule_work(&event->remove);
4838 spin_unlock(&memcg->event_list_lock);
4844 static void memcg_event_ptable_queue_proc(struct file *file,
4845 wait_queue_head_t *wqh, poll_table *pt)
4847 struct mem_cgroup_event *event =
4848 container_of(pt, struct mem_cgroup_event, pt);
4851 add_wait_queue(wqh, &event->wait);
4855 * DO NOT USE IN NEW FILES.
4857 * Parse input and register new cgroup event handler.
4859 * Input must be in format '<event_fd> <control_fd> <args>'.
4860 * Interpretation of args is defined by control file implementation.
4862 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4863 char *buf, size_t nbytes, loff_t off)
4865 struct cgroup_subsys_state *css = of_css(of);
4866 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4867 struct mem_cgroup_event *event;
4868 struct cgroup_subsys_state *cfile_css;
4869 unsigned int efd, cfd;
4876 buf = strstrip(buf);
4878 efd = simple_strtoul(buf, &endp, 10);
4883 cfd = simple_strtoul(buf, &endp, 10);
4884 if ((*endp != ' ') && (*endp != '\0'))
4888 event = kzalloc(sizeof(*event), GFP_KERNEL);
4892 event->memcg = memcg;
4893 INIT_LIST_HEAD(&event->list);
4894 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4895 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4896 INIT_WORK(&event->remove, memcg_event_remove);
4904 event->eventfd = eventfd_ctx_fileget(efile.file);
4905 if (IS_ERR(event->eventfd)) {
4906 ret = PTR_ERR(event->eventfd);
4913 goto out_put_eventfd;
4916 /* the process need read permission on control file */
4917 /* AV: shouldn't we check that it's been opened for read instead? */
4918 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4923 * Determine the event callbacks and set them in @event. This used
4924 * to be done via struct cftype but cgroup core no longer knows
4925 * about these events. The following is crude but the whole thing
4926 * is for compatibility anyway.
4928 * DO NOT ADD NEW FILES.
4930 name = cfile.file->f_path.dentry->d_name.name;
4932 if (!strcmp(name, "memory.usage_in_bytes")) {
4933 event->register_event = mem_cgroup_usage_register_event;
4934 event->unregister_event = mem_cgroup_usage_unregister_event;
4935 } else if (!strcmp(name, "memory.oom_control")) {
4936 event->register_event = mem_cgroup_oom_register_event;
4937 event->unregister_event = mem_cgroup_oom_unregister_event;
4938 } else if (!strcmp(name, "memory.pressure_level")) {
4939 event->register_event = vmpressure_register_event;
4940 event->unregister_event = vmpressure_unregister_event;
4941 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4942 event->register_event = memsw_cgroup_usage_register_event;
4943 event->unregister_event = memsw_cgroup_usage_unregister_event;
4950 * Verify @cfile should belong to @css. Also, remaining events are
4951 * automatically removed on cgroup destruction but the removal is
4952 * asynchronous, so take an extra ref on @css.
4954 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4955 &memory_cgrp_subsys);
4957 if (IS_ERR(cfile_css))
4959 if (cfile_css != css) {
4964 ret = event->register_event(memcg, event->eventfd, buf);
4968 vfs_poll(efile.file, &event->pt);
4970 spin_lock(&memcg->event_list_lock);
4971 list_add(&event->list, &memcg->event_list);
4972 spin_unlock(&memcg->event_list_lock);
4984 eventfd_ctx_put(event->eventfd);
4993 static struct cftype mem_cgroup_legacy_files[] = {
4995 .name = "usage_in_bytes",
4996 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4997 .read_u64 = mem_cgroup_read_u64,
5000 .name = "max_usage_in_bytes",
5001 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5002 .write = mem_cgroup_reset,
5003 .read_u64 = mem_cgroup_read_u64,
5006 .name = "limit_in_bytes",
5007 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5008 .write = mem_cgroup_write,
5009 .read_u64 = mem_cgroup_read_u64,
5012 .name = "soft_limit_in_bytes",
5013 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5014 .write = mem_cgroup_write,
5015 .read_u64 = mem_cgroup_read_u64,
5019 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5020 .write = mem_cgroup_reset,
5021 .read_u64 = mem_cgroup_read_u64,
5025 .seq_show = memcg_stat_show,
5028 .name = "force_empty",
5029 .write = mem_cgroup_force_empty_write,
5032 .name = "use_hierarchy",
5033 .write_u64 = mem_cgroup_hierarchy_write,
5034 .read_u64 = mem_cgroup_hierarchy_read,
5037 .name = "cgroup.event_control", /* XXX: for compat */
5038 .write = memcg_write_event_control,
5039 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5042 .name = "swappiness",
5043 .read_u64 = mem_cgroup_swappiness_read,
5044 .write_u64 = mem_cgroup_swappiness_write,
5047 .name = "move_charge_at_immigrate",
5048 .read_u64 = mem_cgroup_move_charge_read,
5049 .write_u64 = mem_cgroup_move_charge_write,
5052 .name = "oom_control",
5053 .seq_show = mem_cgroup_oom_control_read,
5054 .write_u64 = mem_cgroup_oom_control_write,
5055 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5058 .name = "pressure_level",
5062 .name = "numa_stat",
5063 .seq_show = memcg_numa_stat_show,
5067 .name = "kmem.limit_in_bytes",
5068 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5069 .write = mem_cgroup_write,
5070 .read_u64 = mem_cgroup_read_u64,
5073 .name = "kmem.usage_in_bytes",
5074 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5075 .read_u64 = mem_cgroup_read_u64,
5078 .name = "kmem.failcnt",
5079 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5080 .write = mem_cgroup_reset,
5081 .read_u64 = mem_cgroup_read_u64,
5084 .name = "kmem.max_usage_in_bytes",
5085 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5086 .write = mem_cgroup_reset,
5087 .read_u64 = mem_cgroup_read_u64,
5089 #if defined(CONFIG_MEMCG_KMEM) && \
5090 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5092 .name = "kmem.slabinfo",
5093 .seq_show = memcg_slab_show,
5097 .name = "kmem.tcp.limit_in_bytes",
5098 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5099 .write = mem_cgroup_write,
5100 .read_u64 = mem_cgroup_read_u64,
5103 .name = "kmem.tcp.usage_in_bytes",
5104 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5105 .read_u64 = mem_cgroup_read_u64,
5108 .name = "kmem.tcp.failcnt",
5109 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5110 .write = mem_cgroup_reset,
5111 .read_u64 = mem_cgroup_read_u64,
5114 .name = "kmem.tcp.max_usage_in_bytes",
5115 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5116 .write = mem_cgroup_reset,
5117 .read_u64 = mem_cgroup_read_u64,
5119 { }, /* terminate */
5123 * Private memory cgroup IDR
5125 * Swap-out records and page cache shadow entries need to store memcg
5126 * references in constrained space, so we maintain an ID space that is
5127 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5128 * memory-controlled cgroups to 64k.
5130 * However, there usually are many references to the offline CSS after
5131 * the cgroup has been destroyed, such as page cache or reclaimable
5132 * slab objects, that don't need to hang on to the ID. We want to keep
5133 * those dead CSS from occupying IDs, or we might quickly exhaust the
5134 * relatively small ID space and prevent the creation of new cgroups
5135 * even when there are much fewer than 64k cgroups - possibly none.
5137 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5138 * be freed and recycled when it's no longer needed, which is usually
5139 * when the CSS is offlined.
5141 * The only exception to that are records of swapped out tmpfs/shmem
5142 * pages that need to be attributed to live ancestors on swapin. But
5143 * those references are manageable from userspace.
5146 static DEFINE_IDR(mem_cgroup_idr);
5148 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5150 if (memcg->id.id > 0) {
5151 idr_remove(&mem_cgroup_idr, memcg->id.id);
5156 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5159 refcount_add(n, &memcg->id.ref);
5162 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5164 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5165 mem_cgroup_id_remove(memcg);
5167 /* Memcg ID pins CSS */
5168 css_put(&memcg->css);
5172 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5174 mem_cgroup_id_put_many(memcg, 1);
5178 * mem_cgroup_from_id - look up a memcg from a memcg id
5179 * @id: the memcg id to look up
5181 * Caller must hold rcu_read_lock().
5183 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5185 WARN_ON_ONCE(!rcu_read_lock_held());
5186 return idr_find(&mem_cgroup_idr, id);
5189 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5191 struct mem_cgroup_per_node *pn;
5194 * This routine is called against possible nodes.
5195 * But it's BUG to call kmalloc() against offline node.
5197 * TODO: this routine can waste much memory for nodes which will
5198 * never be onlined. It's better to use memory hotplug callback
5201 if (!node_state(node, N_NORMAL_MEMORY))
5203 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5207 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5208 GFP_KERNEL_ACCOUNT);
5209 if (!pn->lruvec_stat_local) {
5214 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct lruvec_stat,
5215 GFP_KERNEL_ACCOUNT);
5216 if (!pn->lruvec_stat_cpu) {
5217 free_percpu(pn->lruvec_stat_local);
5222 lruvec_init(&pn->lruvec);
5223 pn->usage_in_excess = 0;
5224 pn->on_tree = false;
5227 memcg->nodeinfo[node] = pn;
5231 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5233 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5238 free_percpu(pn->lruvec_stat_cpu);
5239 free_percpu(pn->lruvec_stat_local);
5243 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5248 free_mem_cgroup_per_node_info(memcg, node);
5249 free_percpu(memcg->vmstats_percpu);
5250 free_percpu(memcg->vmstats_local);
5254 static void mem_cgroup_free(struct mem_cgroup *memcg)
5256 memcg_wb_domain_exit(memcg);
5258 * Flush percpu vmstats and vmevents to guarantee the value correctness
5259 * on parent's and all ancestor levels.
5261 memcg_flush_percpu_vmstats(memcg);
5262 memcg_flush_percpu_vmevents(memcg);
5263 __mem_cgroup_free(memcg);
5266 static struct mem_cgroup *mem_cgroup_alloc(void)
5268 struct mem_cgroup *memcg;
5271 int __maybe_unused i;
5272 long error = -ENOMEM;
5274 size = sizeof(struct mem_cgroup);
5275 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5277 memcg = kzalloc(size, GFP_KERNEL);
5279 return ERR_PTR(error);
5281 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5282 1, MEM_CGROUP_ID_MAX,
5284 if (memcg->id.id < 0) {
5285 error = memcg->id.id;
5289 memcg->vmstats_local = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5290 GFP_KERNEL_ACCOUNT);
5291 if (!memcg->vmstats_local)
5294 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5295 GFP_KERNEL_ACCOUNT);
5296 if (!memcg->vmstats_percpu)
5300 if (alloc_mem_cgroup_per_node_info(memcg, node))
5303 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5306 INIT_WORK(&memcg->high_work, high_work_func);
5307 INIT_LIST_HEAD(&memcg->oom_notify);
5308 mutex_init(&memcg->thresholds_lock);
5309 spin_lock_init(&memcg->move_lock);
5310 vmpressure_init(&memcg->vmpressure);
5311 INIT_LIST_HEAD(&memcg->event_list);
5312 spin_lock_init(&memcg->event_list_lock);
5313 memcg->socket_pressure = jiffies;
5314 #ifdef CONFIG_MEMCG_KMEM
5315 memcg->kmemcg_id = -1;
5316 INIT_LIST_HEAD(&memcg->objcg_list);
5318 #ifdef CONFIG_CGROUP_WRITEBACK
5319 INIT_LIST_HEAD(&memcg->cgwb_list);
5320 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5321 memcg->cgwb_frn[i].done =
5322 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5324 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5325 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5326 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5327 memcg->deferred_split_queue.split_queue_len = 0;
5329 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5332 mem_cgroup_id_remove(memcg);
5333 __mem_cgroup_free(memcg);
5334 return ERR_PTR(error);
5337 static struct cgroup_subsys_state * __ref
5338 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5340 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5341 struct mem_cgroup *memcg, *old_memcg;
5342 long error = -ENOMEM;
5344 old_memcg = set_active_memcg(parent);
5345 memcg = mem_cgroup_alloc();
5346 set_active_memcg(old_memcg);
5348 return ERR_CAST(memcg);
5350 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5351 memcg->soft_limit = PAGE_COUNTER_MAX;
5352 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5354 memcg->swappiness = mem_cgroup_swappiness(parent);
5355 memcg->oom_kill_disable = parent->oom_kill_disable;
5358 page_counter_init(&memcg->memory, NULL);
5359 page_counter_init(&memcg->swap, NULL);
5360 page_counter_init(&memcg->kmem, NULL);
5361 page_counter_init(&memcg->tcpmem, NULL);
5362 } else if (parent->use_hierarchy) {
5363 memcg->use_hierarchy = true;
5364 page_counter_init(&memcg->memory, &parent->memory);
5365 page_counter_init(&memcg->swap, &parent->swap);
5366 page_counter_init(&memcg->kmem, &parent->kmem);
5367 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5369 page_counter_init(&memcg->memory, &root_mem_cgroup->memory);
5370 page_counter_init(&memcg->swap, &root_mem_cgroup->swap);
5371 page_counter_init(&memcg->kmem, &root_mem_cgroup->kmem);
5372 page_counter_init(&memcg->tcpmem, &root_mem_cgroup->tcpmem);
5374 * Deeper hierachy with use_hierarchy == false doesn't make
5375 * much sense so let cgroup subsystem know about this
5376 * unfortunate state in our controller.
5378 if (parent != root_mem_cgroup)
5379 memory_cgrp_subsys.broken_hierarchy = true;
5382 /* The following stuff does not apply to the root */
5384 root_mem_cgroup = memcg;
5388 error = memcg_online_kmem(memcg);
5392 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5393 static_branch_inc(&memcg_sockets_enabled_key);
5397 mem_cgroup_id_remove(memcg);
5398 mem_cgroup_free(memcg);
5399 return ERR_PTR(error);
5402 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5404 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5407 * A memcg must be visible for memcg_expand_shrinker_maps()
5408 * by the time the maps are allocated. So, we allocate maps
5409 * here, when for_each_mem_cgroup() can't skip it.
5411 if (memcg_alloc_shrinker_maps(memcg)) {
5412 mem_cgroup_id_remove(memcg);
5416 /* Online state pins memcg ID, memcg ID pins CSS */
5417 refcount_set(&memcg->id.ref, 1);
5422 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5424 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5425 struct mem_cgroup_event *event, *tmp;
5428 * Unregister events and notify userspace.
5429 * Notify userspace about cgroup removing only after rmdir of cgroup
5430 * directory to avoid race between userspace and kernelspace.
5432 spin_lock(&memcg->event_list_lock);
5433 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5434 list_del_init(&event->list);
5435 schedule_work(&event->remove);
5437 spin_unlock(&memcg->event_list_lock);
5439 page_counter_set_min(&memcg->memory, 0);
5440 page_counter_set_low(&memcg->memory, 0);
5442 memcg_offline_kmem(memcg);
5443 wb_memcg_offline(memcg);
5445 drain_all_stock(memcg);
5447 mem_cgroup_id_put(memcg);
5450 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5452 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5454 invalidate_reclaim_iterators(memcg);
5457 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5459 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5460 int __maybe_unused i;
5462 #ifdef CONFIG_CGROUP_WRITEBACK
5463 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5464 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5466 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5467 static_branch_dec(&memcg_sockets_enabled_key);
5469 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5470 static_branch_dec(&memcg_sockets_enabled_key);
5472 vmpressure_cleanup(&memcg->vmpressure);
5473 cancel_work_sync(&memcg->high_work);
5474 mem_cgroup_remove_from_trees(memcg);
5475 memcg_free_shrinker_maps(memcg);
5476 memcg_free_kmem(memcg);
5477 mem_cgroup_free(memcg);
5481 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5482 * @css: the target css
5484 * Reset the states of the mem_cgroup associated with @css. This is
5485 * invoked when the userland requests disabling on the default hierarchy
5486 * but the memcg is pinned through dependency. The memcg should stop
5487 * applying policies and should revert to the vanilla state as it may be
5488 * made visible again.
5490 * The current implementation only resets the essential configurations.
5491 * This needs to be expanded to cover all the visible parts.
5493 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5495 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5497 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5498 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5499 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5500 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5501 page_counter_set_min(&memcg->memory, 0);
5502 page_counter_set_low(&memcg->memory, 0);
5503 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5504 memcg->soft_limit = PAGE_COUNTER_MAX;
5505 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5506 memcg_wb_domain_size_changed(memcg);
5510 /* Handlers for move charge at task migration. */
5511 static int mem_cgroup_do_precharge(unsigned long count)
5515 /* Try a single bulk charge without reclaim first, kswapd may wake */
5516 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5518 mc.precharge += count;
5522 /* Try charges one by one with reclaim, but do not retry */
5524 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5538 enum mc_target_type {
5545 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5546 unsigned long addr, pte_t ptent)
5548 struct page *page = vm_normal_page(vma, addr, ptent);
5550 if (!page || !page_mapped(page))
5552 if (PageAnon(page)) {
5553 if (!(mc.flags & MOVE_ANON))
5556 if (!(mc.flags & MOVE_FILE))
5559 if (!get_page_unless_zero(page))
5565 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5566 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5567 pte_t ptent, swp_entry_t *entry)
5569 struct page *page = NULL;
5570 swp_entry_t ent = pte_to_swp_entry(ptent);
5572 if (!(mc.flags & MOVE_ANON))
5576 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5577 * a device and because they are not accessible by CPU they are store
5578 * as special swap entry in the CPU page table.
5580 if (is_device_private_entry(ent)) {
5581 page = device_private_entry_to_page(ent);
5583 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5584 * a refcount of 1 when free (unlike normal page)
5586 if (!page_ref_add_unless(page, 1, 1))
5591 if (non_swap_entry(ent))
5595 * Because lookup_swap_cache() updates some statistics counter,
5596 * we call find_get_page() with swapper_space directly.
5598 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5599 entry->val = ent.val;
5604 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5605 pte_t ptent, swp_entry_t *entry)
5611 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5612 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5614 if (!vma->vm_file) /* anonymous vma */
5616 if (!(mc.flags & MOVE_FILE))
5619 /* page is moved even if it's not RSS of this task(page-faulted). */
5620 /* shmem/tmpfs may report page out on swap: account for that too. */
5621 return find_get_incore_page(vma->vm_file->f_mapping,
5622 linear_page_index(vma, addr));
5626 * mem_cgroup_move_account - move account of the page
5628 * @compound: charge the page as compound or small page
5629 * @from: mem_cgroup which the page is moved from.
5630 * @to: mem_cgroup which the page is moved to. @from != @to.
5632 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5634 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5637 static int mem_cgroup_move_account(struct page *page,
5639 struct mem_cgroup *from,
5640 struct mem_cgroup *to)
5642 struct lruvec *from_vec, *to_vec;
5643 struct pglist_data *pgdat;
5644 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5647 VM_BUG_ON(from == to);
5648 VM_BUG_ON_PAGE(PageLRU(page), page);
5649 VM_BUG_ON(compound && !PageTransHuge(page));
5652 * Prevent mem_cgroup_migrate() from looking at
5653 * page->mem_cgroup of its source page while we change it.
5656 if (!trylock_page(page))
5660 if (page->mem_cgroup != from)
5663 pgdat = page_pgdat(page);
5664 from_vec = mem_cgroup_lruvec(from, pgdat);
5665 to_vec = mem_cgroup_lruvec(to, pgdat);
5667 lock_page_memcg(page);
5669 if (PageAnon(page)) {
5670 if (page_mapped(page)) {
5671 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5672 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5673 if (PageTransHuge(page)) {
5674 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5676 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5682 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5683 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5685 if (PageSwapBacked(page)) {
5686 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5687 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5690 if (page_mapped(page)) {
5691 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5692 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5695 if (PageDirty(page)) {
5696 struct address_space *mapping = page_mapping(page);
5698 if (mapping_can_writeback(mapping)) {
5699 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5701 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5707 if (PageWriteback(page)) {
5708 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5709 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5713 * All state has been migrated, let's switch to the new memcg.
5715 * It is safe to change page->mem_cgroup here because the page
5716 * is referenced, charged, isolated, and locked: we can't race
5717 * with (un)charging, migration, LRU putback, or anything else
5718 * that would rely on a stable page->mem_cgroup.
5720 * Note that lock_page_memcg is a memcg lock, not a page lock,
5721 * to save space. As soon as we switch page->mem_cgroup to a
5722 * new memcg that isn't locked, the above state can change
5723 * concurrently again. Make sure we're truly done with it.
5728 css_put(&from->css);
5730 page->mem_cgroup = to;
5732 __unlock_page_memcg(from);
5736 local_irq_disable();
5737 mem_cgroup_charge_statistics(to, page, nr_pages);
5738 memcg_check_events(to, page);
5739 mem_cgroup_charge_statistics(from, page, -nr_pages);
5740 memcg_check_events(from, page);
5749 * get_mctgt_type - get target type of moving charge
5750 * @vma: the vma the pte to be checked belongs
5751 * @addr: the address corresponding to the pte to be checked
5752 * @ptent: the pte to be checked
5753 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5756 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5757 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5758 * move charge. if @target is not NULL, the page is stored in target->page
5759 * with extra refcnt got(Callers should handle it).
5760 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5761 * target for charge migration. if @target is not NULL, the entry is stored
5763 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5764 * (so ZONE_DEVICE page and thus not on the lru).
5765 * For now we such page is charge like a regular page would be as for all
5766 * intent and purposes it is just special memory taking the place of a
5769 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5771 * Called with pte lock held.
5774 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5775 unsigned long addr, pte_t ptent, union mc_target *target)
5777 struct page *page = NULL;
5778 enum mc_target_type ret = MC_TARGET_NONE;
5779 swp_entry_t ent = { .val = 0 };
5781 if (pte_present(ptent))
5782 page = mc_handle_present_pte(vma, addr, ptent);
5783 else if (is_swap_pte(ptent))
5784 page = mc_handle_swap_pte(vma, ptent, &ent);
5785 else if (pte_none(ptent))
5786 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5788 if (!page && !ent.val)
5792 * Do only loose check w/o serialization.
5793 * mem_cgroup_move_account() checks the page is valid or
5794 * not under LRU exclusion.
5796 if (page->mem_cgroup == mc.from) {
5797 ret = MC_TARGET_PAGE;
5798 if (is_device_private_page(page))
5799 ret = MC_TARGET_DEVICE;
5801 target->page = page;
5803 if (!ret || !target)
5807 * There is a swap entry and a page doesn't exist or isn't charged.
5808 * But we cannot move a tail-page in a THP.
5810 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5811 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5812 ret = MC_TARGET_SWAP;
5819 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5821 * We don't consider PMD mapped swapping or file mapped pages because THP does
5822 * not support them for now.
5823 * Caller should make sure that pmd_trans_huge(pmd) is true.
5825 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5826 unsigned long addr, pmd_t pmd, union mc_target *target)
5828 struct page *page = NULL;
5829 enum mc_target_type ret = MC_TARGET_NONE;
5831 if (unlikely(is_swap_pmd(pmd))) {
5832 VM_BUG_ON(thp_migration_supported() &&
5833 !is_pmd_migration_entry(pmd));
5836 page = pmd_page(pmd);
5837 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5838 if (!(mc.flags & MOVE_ANON))
5840 if (page->mem_cgroup == mc.from) {
5841 ret = MC_TARGET_PAGE;
5844 target->page = page;
5850 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5851 unsigned long addr, pmd_t pmd, union mc_target *target)
5853 return MC_TARGET_NONE;
5857 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5858 unsigned long addr, unsigned long end,
5859 struct mm_walk *walk)
5861 struct vm_area_struct *vma = walk->vma;
5865 ptl = pmd_trans_huge_lock(pmd, vma);
5868 * Note their can not be MC_TARGET_DEVICE for now as we do not
5869 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5870 * this might change.
5872 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5873 mc.precharge += HPAGE_PMD_NR;
5878 if (pmd_trans_unstable(pmd))
5880 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5881 for (; addr != end; pte++, addr += PAGE_SIZE)
5882 if (get_mctgt_type(vma, addr, *pte, NULL))
5883 mc.precharge++; /* increment precharge temporarily */
5884 pte_unmap_unlock(pte - 1, ptl);
5890 static const struct mm_walk_ops precharge_walk_ops = {
5891 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5894 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5896 unsigned long precharge;
5899 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5900 mmap_read_unlock(mm);
5902 precharge = mc.precharge;
5908 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5910 unsigned long precharge = mem_cgroup_count_precharge(mm);
5912 VM_BUG_ON(mc.moving_task);
5913 mc.moving_task = current;
5914 return mem_cgroup_do_precharge(precharge);
5917 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5918 static void __mem_cgroup_clear_mc(void)
5920 struct mem_cgroup *from = mc.from;
5921 struct mem_cgroup *to = mc.to;
5923 /* we must uncharge all the leftover precharges from mc.to */
5925 cancel_charge(mc.to, mc.precharge);
5929 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5930 * we must uncharge here.
5932 if (mc.moved_charge) {
5933 cancel_charge(mc.from, mc.moved_charge);
5934 mc.moved_charge = 0;
5936 /* we must fixup refcnts and charges */
5937 if (mc.moved_swap) {
5938 /* uncharge swap account from the old cgroup */
5939 if (!mem_cgroup_is_root(mc.from))
5940 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5942 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5945 * we charged both to->memory and to->memsw, so we
5946 * should uncharge to->memory.
5948 if (!mem_cgroup_is_root(mc.to))
5949 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5953 memcg_oom_recover(from);
5954 memcg_oom_recover(to);
5955 wake_up_all(&mc.waitq);
5958 static void mem_cgroup_clear_mc(void)
5960 struct mm_struct *mm = mc.mm;
5963 * we must clear moving_task before waking up waiters at the end of
5966 mc.moving_task = NULL;
5967 __mem_cgroup_clear_mc();
5968 spin_lock(&mc.lock);
5972 spin_unlock(&mc.lock);
5977 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5979 struct cgroup_subsys_state *css;
5980 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5981 struct mem_cgroup *from;
5982 struct task_struct *leader, *p;
5983 struct mm_struct *mm;
5984 unsigned long move_flags;
5987 /* charge immigration isn't supported on the default hierarchy */
5988 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5992 * Multi-process migrations only happen on the default hierarchy
5993 * where charge immigration is not used. Perform charge
5994 * immigration if @tset contains a leader and whine if there are
5998 cgroup_taskset_for_each_leader(leader, css, tset) {
6001 memcg = mem_cgroup_from_css(css);
6007 * We are now commited to this value whatever it is. Changes in this
6008 * tunable will only affect upcoming migrations, not the current one.
6009 * So we need to save it, and keep it going.
6011 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
6015 from = mem_cgroup_from_task(p);
6017 VM_BUG_ON(from == memcg);
6019 mm = get_task_mm(p);
6022 /* We move charges only when we move a owner of the mm */
6023 if (mm->owner == p) {
6026 VM_BUG_ON(mc.precharge);
6027 VM_BUG_ON(mc.moved_charge);
6028 VM_BUG_ON(mc.moved_swap);
6030 spin_lock(&mc.lock);
6034 mc.flags = move_flags;
6035 spin_unlock(&mc.lock);
6036 /* We set mc.moving_task later */
6038 ret = mem_cgroup_precharge_mc(mm);
6040 mem_cgroup_clear_mc();
6047 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6050 mem_cgroup_clear_mc();
6053 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6054 unsigned long addr, unsigned long end,
6055 struct mm_walk *walk)
6058 struct vm_area_struct *vma = walk->vma;
6061 enum mc_target_type target_type;
6062 union mc_target target;
6065 ptl = pmd_trans_huge_lock(pmd, vma);
6067 if (mc.precharge < HPAGE_PMD_NR) {
6071 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6072 if (target_type == MC_TARGET_PAGE) {
6074 if (!isolate_lru_page(page)) {
6075 if (!mem_cgroup_move_account(page, true,
6077 mc.precharge -= HPAGE_PMD_NR;
6078 mc.moved_charge += HPAGE_PMD_NR;
6080 putback_lru_page(page);
6083 } else if (target_type == MC_TARGET_DEVICE) {
6085 if (!mem_cgroup_move_account(page, true,
6087 mc.precharge -= HPAGE_PMD_NR;
6088 mc.moved_charge += HPAGE_PMD_NR;
6096 if (pmd_trans_unstable(pmd))
6099 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6100 for (; addr != end; addr += PAGE_SIZE) {
6101 pte_t ptent = *(pte++);
6102 bool device = false;
6108 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6109 case MC_TARGET_DEVICE:
6112 case MC_TARGET_PAGE:
6115 * We can have a part of the split pmd here. Moving it
6116 * can be done but it would be too convoluted so simply
6117 * ignore such a partial THP and keep it in original
6118 * memcg. There should be somebody mapping the head.
6120 if (PageTransCompound(page))
6122 if (!device && isolate_lru_page(page))
6124 if (!mem_cgroup_move_account(page, false,
6127 /* we uncharge from mc.from later. */
6131 putback_lru_page(page);
6132 put: /* get_mctgt_type() gets the page */
6135 case MC_TARGET_SWAP:
6137 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6139 mem_cgroup_id_get_many(mc.to, 1);
6140 /* we fixup other refcnts and charges later. */
6148 pte_unmap_unlock(pte - 1, ptl);
6153 * We have consumed all precharges we got in can_attach().
6154 * We try charge one by one, but don't do any additional
6155 * charges to mc.to if we have failed in charge once in attach()
6158 ret = mem_cgroup_do_precharge(1);
6166 static const struct mm_walk_ops charge_walk_ops = {
6167 .pmd_entry = mem_cgroup_move_charge_pte_range,
6170 static void mem_cgroup_move_charge(void)
6172 lru_add_drain_all();
6174 * Signal lock_page_memcg() to take the memcg's move_lock
6175 * while we're moving its pages to another memcg. Then wait
6176 * for already started RCU-only updates to finish.
6178 atomic_inc(&mc.from->moving_account);
6181 if (unlikely(!mmap_read_trylock(mc.mm))) {
6183 * Someone who are holding the mmap_lock might be waiting in
6184 * waitq. So we cancel all extra charges, wake up all waiters,
6185 * and retry. Because we cancel precharges, we might not be able
6186 * to move enough charges, but moving charge is a best-effort
6187 * feature anyway, so it wouldn't be a big problem.
6189 __mem_cgroup_clear_mc();
6194 * When we have consumed all precharges and failed in doing
6195 * additional charge, the page walk just aborts.
6197 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6200 mmap_read_unlock(mc.mm);
6201 atomic_dec(&mc.from->moving_account);
6204 static void mem_cgroup_move_task(void)
6207 mem_cgroup_move_charge();
6208 mem_cgroup_clear_mc();
6211 #else /* !CONFIG_MMU */
6212 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6216 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6219 static void mem_cgroup_move_task(void)
6225 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6226 * to verify whether we're attached to the default hierarchy on each mount
6229 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6232 * use_hierarchy is forced on the default hierarchy. cgroup core
6233 * guarantees that @root doesn't have any children, so turning it
6234 * on for the root memcg is enough.
6236 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6237 root_mem_cgroup->use_hierarchy = true;
6239 root_mem_cgroup->use_hierarchy = false;
6242 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6244 if (value == PAGE_COUNTER_MAX)
6245 seq_puts(m, "max\n");
6247 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6252 static u64 memory_current_read(struct cgroup_subsys_state *css,
6255 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6257 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6260 static int memory_min_show(struct seq_file *m, void *v)
6262 return seq_puts_memcg_tunable(m,
6263 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6266 static ssize_t memory_min_write(struct kernfs_open_file *of,
6267 char *buf, size_t nbytes, loff_t off)
6269 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6273 buf = strstrip(buf);
6274 err = page_counter_memparse(buf, "max", &min);
6278 page_counter_set_min(&memcg->memory, min);
6283 static int memory_low_show(struct seq_file *m, void *v)
6285 return seq_puts_memcg_tunable(m,
6286 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6289 static ssize_t memory_low_write(struct kernfs_open_file *of,
6290 char *buf, size_t nbytes, loff_t off)
6292 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6296 buf = strstrip(buf);
6297 err = page_counter_memparse(buf, "max", &low);
6301 page_counter_set_low(&memcg->memory, low);
6306 static int memory_high_show(struct seq_file *m, void *v)
6308 return seq_puts_memcg_tunable(m,
6309 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6312 static ssize_t memory_high_write(struct kernfs_open_file *of,
6313 char *buf, size_t nbytes, loff_t off)
6315 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6316 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6317 bool drained = false;
6321 buf = strstrip(buf);
6322 err = page_counter_memparse(buf, "max", &high);
6327 unsigned long nr_pages = page_counter_read(&memcg->memory);
6328 unsigned long reclaimed;
6330 if (nr_pages <= high)
6333 if (signal_pending(current))
6337 drain_all_stock(memcg);
6342 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6345 if (!reclaimed && !nr_retries--)
6349 page_counter_set_high(&memcg->memory, high);
6351 memcg_wb_domain_size_changed(memcg);
6356 static int memory_max_show(struct seq_file *m, void *v)
6358 return seq_puts_memcg_tunable(m,
6359 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6362 static ssize_t memory_max_write(struct kernfs_open_file *of,
6363 char *buf, size_t nbytes, loff_t off)
6365 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6366 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6367 bool drained = false;
6371 buf = strstrip(buf);
6372 err = page_counter_memparse(buf, "max", &max);
6376 xchg(&memcg->memory.max, max);
6379 unsigned long nr_pages = page_counter_read(&memcg->memory);
6381 if (nr_pages <= max)
6384 if (signal_pending(current))
6388 drain_all_stock(memcg);
6394 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6400 memcg_memory_event(memcg, MEMCG_OOM);
6401 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6405 memcg_wb_domain_size_changed(memcg);
6409 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6411 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6412 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6413 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6414 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6415 seq_printf(m, "oom_kill %lu\n",
6416 atomic_long_read(&events[MEMCG_OOM_KILL]));
6419 static int memory_events_show(struct seq_file *m, void *v)
6421 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6423 __memory_events_show(m, memcg->memory_events);
6427 static int memory_events_local_show(struct seq_file *m, void *v)
6429 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6431 __memory_events_show(m, memcg->memory_events_local);
6435 static int memory_stat_show(struct seq_file *m, void *v)
6437 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6440 buf = memory_stat_format(memcg);
6449 static int memory_numa_stat_show(struct seq_file *m, void *v)
6452 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6454 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6457 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6460 seq_printf(m, "%s", memory_stats[i].name);
6461 for_each_node_state(nid, N_MEMORY) {
6463 struct lruvec *lruvec;
6465 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6466 size = lruvec_page_state(lruvec, memory_stats[i].idx);
6467 size *= memory_stats[i].ratio;
6468 seq_printf(m, " N%d=%llu", nid, size);
6477 static int memory_oom_group_show(struct seq_file *m, void *v)
6479 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6481 seq_printf(m, "%d\n", memcg->oom_group);
6486 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6487 char *buf, size_t nbytes, loff_t off)
6489 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6492 buf = strstrip(buf);
6496 ret = kstrtoint(buf, 0, &oom_group);
6500 if (oom_group != 0 && oom_group != 1)
6503 memcg->oom_group = oom_group;
6508 static struct cftype memory_files[] = {
6511 .flags = CFTYPE_NOT_ON_ROOT,
6512 .read_u64 = memory_current_read,
6516 .flags = CFTYPE_NOT_ON_ROOT,
6517 .seq_show = memory_min_show,
6518 .write = memory_min_write,
6522 .flags = CFTYPE_NOT_ON_ROOT,
6523 .seq_show = memory_low_show,
6524 .write = memory_low_write,
6528 .flags = CFTYPE_NOT_ON_ROOT,
6529 .seq_show = memory_high_show,
6530 .write = memory_high_write,
6534 .flags = CFTYPE_NOT_ON_ROOT,
6535 .seq_show = memory_max_show,
6536 .write = memory_max_write,
6540 .flags = CFTYPE_NOT_ON_ROOT,
6541 .file_offset = offsetof(struct mem_cgroup, events_file),
6542 .seq_show = memory_events_show,
6545 .name = "events.local",
6546 .flags = CFTYPE_NOT_ON_ROOT,
6547 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6548 .seq_show = memory_events_local_show,
6552 .seq_show = memory_stat_show,
6556 .name = "numa_stat",
6557 .seq_show = memory_numa_stat_show,
6561 .name = "oom.group",
6562 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6563 .seq_show = memory_oom_group_show,
6564 .write = memory_oom_group_write,
6569 struct cgroup_subsys memory_cgrp_subsys = {
6570 .css_alloc = mem_cgroup_css_alloc,
6571 .css_online = mem_cgroup_css_online,
6572 .css_offline = mem_cgroup_css_offline,
6573 .css_released = mem_cgroup_css_released,
6574 .css_free = mem_cgroup_css_free,
6575 .css_reset = mem_cgroup_css_reset,
6576 .can_attach = mem_cgroup_can_attach,
6577 .cancel_attach = mem_cgroup_cancel_attach,
6578 .post_attach = mem_cgroup_move_task,
6579 .bind = mem_cgroup_bind,
6580 .dfl_cftypes = memory_files,
6581 .legacy_cftypes = mem_cgroup_legacy_files,
6586 * This function calculates an individual cgroup's effective
6587 * protection which is derived from its own memory.min/low, its
6588 * parent's and siblings' settings, as well as the actual memory
6589 * distribution in the tree.
6591 * The following rules apply to the effective protection values:
6593 * 1. At the first level of reclaim, effective protection is equal to
6594 * the declared protection in memory.min and memory.low.
6596 * 2. To enable safe delegation of the protection configuration, at
6597 * subsequent levels the effective protection is capped to the
6598 * parent's effective protection.
6600 * 3. To make complex and dynamic subtrees easier to configure, the
6601 * user is allowed to overcommit the declared protection at a given
6602 * level. If that is the case, the parent's effective protection is
6603 * distributed to the children in proportion to how much protection
6604 * they have declared and how much of it they are utilizing.
6606 * This makes distribution proportional, but also work-conserving:
6607 * if one cgroup claims much more protection than it uses memory,
6608 * the unused remainder is available to its siblings.
6610 * 4. Conversely, when the declared protection is undercommitted at a
6611 * given level, the distribution of the larger parental protection
6612 * budget is NOT proportional. A cgroup's protection from a sibling
6613 * is capped to its own memory.min/low setting.
6615 * 5. However, to allow protecting recursive subtrees from each other
6616 * without having to declare each individual cgroup's fixed share
6617 * of the ancestor's claim to protection, any unutilized -
6618 * "floating" - protection from up the tree is distributed in
6619 * proportion to each cgroup's *usage*. This makes the protection
6620 * neutral wrt sibling cgroups and lets them compete freely over
6621 * the shared parental protection budget, but it protects the
6622 * subtree as a whole from neighboring subtrees.
6624 * Note that 4. and 5. are not in conflict: 4. is about protecting
6625 * against immediate siblings whereas 5. is about protecting against
6626 * neighboring subtrees.
6628 static unsigned long effective_protection(unsigned long usage,
6629 unsigned long parent_usage,
6630 unsigned long setting,
6631 unsigned long parent_effective,
6632 unsigned long siblings_protected)
6634 unsigned long protected;
6637 protected = min(usage, setting);
6639 * If all cgroups at this level combined claim and use more
6640 * protection then what the parent affords them, distribute
6641 * shares in proportion to utilization.
6643 * We are using actual utilization rather than the statically
6644 * claimed protection in order to be work-conserving: claimed
6645 * but unused protection is available to siblings that would
6646 * otherwise get a smaller chunk than what they claimed.
6648 if (siblings_protected > parent_effective)
6649 return protected * parent_effective / siblings_protected;
6652 * Ok, utilized protection of all children is within what the
6653 * parent affords them, so we know whatever this child claims
6654 * and utilizes is effectively protected.
6656 * If there is unprotected usage beyond this value, reclaim
6657 * will apply pressure in proportion to that amount.
6659 * If there is unutilized protection, the cgroup will be fully
6660 * shielded from reclaim, but we do return a smaller value for
6661 * protection than what the group could enjoy in theory. This
6662 * is okay. With the overcommit distribution above, effective
6663 * protection is always dependent on how memory is actually
6664 * consumed among the siblings anyway.
6669 * If the children aren't claiming (all of) the protection
6670 * afforded to them by the parent, distribute the remainder in
6671 * proportion to the (unprotected) memory of each cgroup. That
6672 * way, cgroups that aren't explicitly prioritized wrt each
6673 * other compete freely over the allowance, but they are
6674 * collectively protected from neighboring trees.
6676 * We're using unprotected memory for the weight so that if
6677 * some cgroups DO claim explicit protection, we don't protect
6678 * the same bytes twice.
6680 * Check both usage and parent_usage against the respective
6681 * protected values. One should imply the other, but they
6682 * aren't read atomically - make sure the division is sane.
6684 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6686 if (parent_effective > siblings_protected &&
6687 parent_usage > siblings_protected &&
6688 usage > protected) {
6689 unsigned long unclaimed;
6691 unclaimed = parent_effective - siblings_protected;
6692 unclaimed *= usage - protected;
6693 unclaimed /= parent_usage - siblings_protected;
6702 * mem_cgroup_protected - check if memory consumption is in the normal range
6703 * @root: the top ancestor of the sub-tree being checked
6704 * @memcg: the memory cgroup to check
6706 * WARNING: This function is not stateless! It can only be used as part
6707 * of a top-down tree iteration, not for isolated queries.
6709 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6710 struct mem_cgroup *memcg)
6712 unsigned long usage, parent_usage;
6713 struct mem_cgroup *parent;
6715 if (mem_cgroup_disabled())
6719 root = root_mem_cgroup;
6722 * Effective values of the reclaim targets are ignored so they
6723 * can be stale. Have a look at mem_cgroup_protection for more
6725 * TODO: calculation should be more robust so that we do not need
6726 * that special casing.
6731 usage = page_counter_read(&memcg->memory);
6735 parent = parent_mem_cgroup(memcg);
6736 /* No parent means a non-hierarchical mode on v1 memcg */
6740 if (parent == root) {
6741 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6742 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6746 parent_usage = page_counter_read(&parent->memory);
6748 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6749 READ_ONCE(memcg->memory.min),
6750 READ_ONCE(parent->memory.emin),
6751 atomic_long_read(&parent->memory.children_min_usage)));
6753 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6754 READ_ONCE(memcg->memory.low),
6755 READ_ONCE(parent->memory.elow),
6756 atomic_long_read(&parent->memory.children_low_usage)));
6760 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6761 * @page: page to charge
6762 * @mm: mm context of the victim
6763 * @gfp_mask: reclaim mode
6765 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6766 * pages according to @gfp_mask if necessary.
6768 * Returns 0 on success. Otherwise, an error code is returned.
6770 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6772 unsigned int nr_pages = thp_nr_pages(page);
6773 struct mem_cgroup *memcg = NULL;
6776 if (mem_cgroup_disabled())
6779 if (PageSwapCache(page)) {
6780 swp_entry_t ent = { .val = page_private(page), };
6784 * Every swap fault against a single page tries to charge the
6785 * page, bail as early as possible. shmem_unuse() encounters
6786 * already charged pages, too. page->mem_cgroup is protected
6787 * by the page lock, which serializes swap cache removal, which
6788 * in turn serializes uncharging.
6790 VM_BUG_ON_PAGE(!PageLocked(page), page);
6791 if (compound_head(page)->mem_cgroup)
6794 id = lookup_swap_cgroup_id(ent);
6796 memcg = mem_cgroup_from_id(id);
6797 if (memcg && !css_tryget_online(&memcg->css))
6803 memcg = get_mem_cgroup_from_mm(mm);
6805 ret = try_charge(memcg, gfp_mask, nr_pages);
6809 css_get(&memcg->css);
6810 commit_charge(page, memcg);
6812 local_irq_disable();
6813 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6814 memcg_check_events(memcg, page);
6817 if (PageSwapCache(page)) {
6818 swp_entry_t entry = { .val = page_private(page) };
6820 * The swap entry might not get freed for a long time,
6821 * let's not wait for it. The page already received a
6822 * memory+swap charge, drop the swap entry duplicate.
6824 mem_cgroup_uncharge_swap(entry, nr_pages);
6828 css_put(&memcg->css);
6833 struct uncharge_gather {
6834 struct mem_cgroup *memcg;
6835 unsigned long nr_pages;
6836 unsigned long pgpgout;
6837 unsigned long nr_kmem;
6838 struct page *dummy_page;
6841 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6843 memset(ug, 0, sizeof(*ug));
6846 static void uncharge_batch(const struct uncharge_gather *ug)
6848 unsigned long flags;
6850 if (!mem_cgroup_is_root(ug->memcg)) {
6851 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6852 if (do_memsw_account())
6853 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6854 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6855 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6856 memcg_oom_recover(ug->memcg);
6859 local_irq_save(flags);
6860 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6861 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6862 memcg_check_events(ug->memcg, ug->dummy_page);
6863 local_irq_restore(flags);
6865 /* drop reference from uncharge_page */
6866 css_put(&ug->memcg->css);
6869 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6871 unsigned long nr_pages;
6873 VM_BUG_ON_PAGE(PageLRU(page), page);
6875 if (!page->mem_cgroup)
6879 * Nobody should be changing or seriously looking at
6880 * page->mem_cgroup at this point, we have fully
6881 * exclusive access to the page.
6884 if (ug->memcg != page->mem_cgroup) {
6887 uncharge_gather_clear(ug);
6889 ug->memcg = page->mem_cgroup;
6891 /* pairs with css_put in uncharge_batch */
6892 css_get(&ug->memcg->css);
6895 nr_pages = compound_nr(page);
6896 ug->nr_pages += nr_pages;
6898 if (!PageKmemcg(page)) {
6901 ug->nr_kmem += nr_pages;
6902 __ClearPageKmemcg(page);
6905 ug->dummy_page = page;
6906 page->mem_cgroup = NULL;
6907 css_put(&ug->memcg->css);
6910 static void uncharge_list(struct list_head *page_list)
6912 struct uncharge_gather ug;
6913 struct list_head *next;
6915 uncharge_gather_clear(&ug);
6918 * Note that the list can be a single page->lru; hence the
6919 * do-while loop instead of a simple list_for_each_entry().
6921 next = page_list->next;
6925 page = list_entry(next, struct page, lru);
6926 next = page->lru.next;
6928 uncharge_page(page, &ug);
6929 } while (next != page_list);
6932 uncharge_batch(&ug);
6936 * mem_cgroup_uncharge - uncharge a page
6937 * @page: page to uncharge
6939 * Uncharge a page previously charged with mem_cgroup_charge().
6941 void mem_cgroup_uncharge(struct page *page)
6943 struct uncharge_gather ug;
6945 if (mem_cgroup_disabled())
6948 /* Don't touch page->lru of any random page, pre-check: */
6949 if (!page->mem_cgroup)
6952 uncharge_gather_clear(&ug);
6953 uncharge_page(page, &ug);
6954 uncharge_batch(&ug);
6958 * mem_cgroup_uncharge_list - uncharge a list of page
6959 * @page_list: list of pages to uncharge
6961 * Uncharge a list of pages previously charged with
6962 * mem_cgroup_charge().
6964 void mem_cgroup_uncharge_list(struct list_head *page_list)
6966 if (mem_cgroup_disabled())
6969 if (!list_empty(page_list))
6970 uncharge_list(page_list);
6974 * mem_cgroup_migrate - charge a page's replacement
6975 * @oldpage: currently circulating page
6976 * @newpage: replacement page
6978 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6979 * be uncharged upon free.
6981 * Both pages must be locked, @newpage->mapping must be set up.
6983 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6985 struct mem_cgroup *memcg;
6986 unsigned int nr_pages;
6987 unsigned long flags;
6989 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6990 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6991 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6992 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6995 if (mem_cgroup_disabled())
6998 /* Page cache replacement: new page already charged? */
6999 if (newpage->mem_cgroup)
7002 /* Swapcache readahead pages can get replaced before being charged */
7003 memcg = oldpage->mem_cgroup;
7007 /* Force-charge the new page. The old one will be freed soon */
7008 nr_pages = thp_nr_pages(newpage);
7010 page_counter_charge(&memcg->memory, nr_pages);
7011 if (do_memsw_account())
7012 page_counter_charge(&memcg->memsw, nr_pages);
7014 css_get(&memcg->css);
7015 commit_charge(newpage, memcg);
7017 local_irq_save(flags);
7018 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
7019 memcg_check_events(memcg, newpage);
7020 local_irq_restore(flags);
7023 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7024 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7026 void mem_cgroup_sk_alloc(struct sock *sk)
7028 struct mem_cgroup *memcg;
7030 if (!mem_cgroup_sockets_enabled)
7033 /* Do not associate the sock with unrelated interrupted task's memcg. */
7038 memcg = mem_cgroup_from_task(current);
7039 if (memcg == root_mem_cgroup)
7041 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7043 if (css_tryget(&memcg->css))
7044 sk->sk_memcg = memcg;
7049 void mem_cgroup_sk_free(struct sock *sk)
7052 css_put(&sk->sk_memcg->css);
7056 * mem_cgroup_charge_skmem - charge socket memory
7057 * @memcg: memcg to charge
7058 * @nr_pages: number of pages to charge
7060 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7061 * @memcg's configured limit, %false if the charge had to be forced.
7063 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7065 gfp_t gfp_mask = GFP_KERNEL;
7067 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7068 struct page_counter *fail;
7070 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7071 memcg->tcpmem_pressure = 0;
7074 page_counter_charge(&memcg->tcpmem, nr_pages);
7075 memcg->tcpmem_pressure = 1;
7079 /* Don't block in the packet receive path */
7081 gfp_mask = GFP_NOWAIT;
7083 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7085 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7088 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7093 * mem_cgroup_uncharge_skmem - uncharge socket memory
7094 * @memcg: memcg to uncharge
7095 * @nr_pages: number of pages to uncharge
7097 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7099 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7100 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7104 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7106 refill_stock(memcg, nr_pages);
7109 static int __init cgroup_memory(char *s)
7113 while ((token = strsep(&s, ",")) != NULL) {
7116 if (!strcmp(token, "nosocket"))
7117 cgroup_memory_nosocket = true;
7118 if (!strcmp(token, "nokmem"))
7119 cgroup_memory_nokmem = true;
7123 __setup("cgroup.memory=", cgroup_memory);
7126 * subsys_initcall() for memory controller.
7128 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7129 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7130 * basically everything that doesn't depend on a specific mem_cgroup structure
7131 * should be initialized from here.
7133 static int __init mem_cgroup_init(void)
7137 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7138 memcg_hotplug_cpu_dead);
7140 for_each_possible_cpu(cpu)
7141 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7144 for_each_node(node) {
7145 struct mem_cgroup_tree_per_node *rtpn;
7147 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7148 node_online(node) ? node : NUMA_NO_NODE);
7150 rtpn->rb_root = RB_ROOT;
7151 rtpn->rb_rightmost = NULL;
7152 spin_lock_init(&rtpn->lock);
7153 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7158 subsys_initcall(mem_cgroup_init);
7160 #ifdef CONFIG_MEMCG_SWAP
7161 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7163 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7165 * The root cgroup cannot be destroyed, so it's refcount must
7168 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7172 memcg = parent_mem_cgroup(memcg);
7174 memcg = root_mem_cgroup;
7180 * mem_cgroup_swapout - transfer a memsw charge to swap
7181 * @page: page whose memsw charge to transfer
7182 * @entry: swap entry to move the charge to
7184 * Transfer the memsw charge of @page to @entry.
7186 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7188 struct mem_cgroup *memcg, *swap_memcg;
7189 unsigned int nr_entries;
7190 unsigned short oldid;
7192 VM_BUG_ON_PAGE(PageLRU(page), page);
7193 VM_BUG_ON_PAGE(page_count(page), page);
7195 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7198 memcg = page->mem_cgroup;
7200 /* Readahead page, never charged */
7205 * In case the memcg owning these pages has been offlined and doesn't
7206 * have an ID allocated to it anymore, charge the closest online
7207 * ancestor for the swap instead and transfer the memory+swap charge.
7209 swap_memcg = mem_cgroup_id_get_online(memcg);
7210 nr_entries = thp_nr_pages(page);
7211 /* Get references for the tail pages, too */
7213 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7214 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7216 VM_BUG_ON_PAGE(oldid, page);
7217 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7219 page->mem_cgroup = NULL;
7221 if (!mem_cgroup_is_root(memcg))
7222 page_counter_uncharge(&memcg->memory, nr_entries);
7224 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7225 if (!mem_cgroup_is_root(swap_memcg))
7226 page_counter_charge(&swap_memcg->memsw, nr_entries);
7227 page_counter_uncharge(&memcg->memsw, nr_entries);
7231 * Interrupts should be disabled here because the caller holds the
7232 * i_pages lock which is taken with interrupts-off. It is
7233 * important here to have the interrupts disabled because it is the
7234 * only synchronisation we have for updating the per-CPU variables.
7236 VM_BUG_ON(!irqs_disabled());
7237 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7238 memcg_check_events(memcg, page);
7240 css_put(&memcg->css);
7244 * mem_cgroup_try_charge_swap - try charging swap space for a page
7245 * @page: page being added to swap
7246 * @entry: swap entry to charge
7248 * Try to charge @page's memcg for the swap space at @entry.
7250 * Returns 0 on success, -ENOMEM on failure.
7252 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7254 unsigned int nr_pages = thp_nr_pages(page);
7255 struct page_counter *counter;
7256 struct mem_cgroup *memcg;
7257 unsigned short oldid;
7259 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7262 memcg = page->mem_cgroup;
7264 /* Readahead page, never charged */
7269 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7273 memcg = mem_cgroup_id_get_online(memcg);
7275 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7276 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7277 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7278 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7279 mem_cgroup_id_put(memcg);
7283 /* Get references for the tail pages, too */
7285 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7286 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7287 VM_BUG_ON_PAGE(oldid, page);
7288 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7294 * mem_cgroup_uncharge_swap - uncharge swap space
7295 * @entry: swap entry to uncharge
7296 * @nr_pages: the amount of swap space to uncharge
7298 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7300 struct mem_cgroup *memcg;
7303 id = swap_cgroup_record(entry, 0, nr_pages);
7305 memcg = mem_cgroup_from_id(id);
7307 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7308 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7309 page_counter_uncharge(&memcg->swap, nr_pages);
7311 page_counter_uncharge(&memcg->memsw, nr_pages);
7313 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7314 mem_cgroup_id_put_many(memcg, nr_pages);
7319 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7321 long nr_swap_pages = get_nr_swap_pages();
7323 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7324 return nr_swap_pages;
7325 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7326 nr_swap_pages = min_t(long, nr_swap_pages,
7327 READ_ONCE(memcg->swap.max) -
7328 page_counter_read(&memcg->swap));
7329 return nr_swap_pages;
7332 bool mem_cgroup_swap_full(struct page *page)
7334 struct mem_cgroup *memcg;
7336 VM_BUG_ON_PAGE(!PageLocked(page), page);
7340 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7343 memcg = page->mem_cgroup;
7347 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7348 unsigned long usage = page_counter_read(&memcg->swap);
7350 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7351 usage * 2 >= READ_ONCE(memcg->swap.max))
7358 static int __init setup_swap_account(char *s)
7360 if (!strcmp(s, "1"))
7361 cgroup_memory_noswap = 0;
7362 else if (!strcmp(s, "0"))
7363 cgroup_memory_noswap = 1;
7366 __setup("swapaccount=", setup_swap_account);
7368 static u64 swap_current_read(struct cgroup_subsys_state *css,
7371 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7373 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7376 static int swap_high_show(struct seq_file *m, void *v)
7378 return seq_puts_memcg_tunable(m,
7379 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7382 static ssize_t swap_high_write(struct kernfs_open_file *of,
7383 char *buf, size_t nbytes, loff_t off)
7385 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7389 buf = strstrip(buf);
7390 err = page_counter_memparse(buf, "max", &high);
7394 page_counter_set_high(&memcg->swap, high);
7399 static int swap_max_show(struct seq_file *m, void *v)
7401 return seq_puts_memcg_tunable(m,
7402 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7405 static ssize_t swap_max_write(struct kernfs_open_file *of,
7406 char *buf, size_t nbytes, loff_t off)
7408 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7412 buf = strstrip(buf);
7413 err = page_counter_memparse(buf, "max", &max);
7417 xchg(&memcg->swap.max, max);
7422 static int swap_events_show(struct seq_file *m, void *v)
7424 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7426 seq_printf(m, "high %lu\n",
7427 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7428 seq_printf(m, "max %lu\n",
7429 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7430 seq_printf(m, "fail %lu\n",
7431 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7436 static struct cftype swap_files[] = {
7438 .name = "swap.current",
7439 .flags = CFTYPE_NOT_ON_ROOT,
7440 .read_u64 = swap_current_read,
7443 .name = "swap.high",
7444 .flags = CFTYPE_NOT_ON_ROOT,
7445 .seq_show = swap_high_show,
7446 .write = swap_high_write,
7450 .flags = CFTYPE_NOT_ON_ROOT,
7451 .seq_show = swap_max_show,
7452 .write = swap_max_write,
7455 .name = "swap.events",
7456 .flags = CFTYPE_NOT_ON_ROOT,
7457 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7458 .seq_show = swap_events_show,
7463 static struct cftype memsw_files[] = {
7465 .name = "memsw.usage_in_bytes",
7466 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7467 .read_u64 = mem_cgroup_read_u64,
7470 .name = "memsw.max_usage_in_bytes",
7471 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7472 .write = mem_cgroup_reset,
7473 .read_u64 = mem_cgroup_read_u64,
7476 .name = "memsw.limit_in_bytes",
7477 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7478 .write = mem_cgroup_write,
7479 .read_u64 = mem_cgroup_read_u64,
7482 .name = "memsw.failcnt",
7483 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7484 .write = mem_cgroup_reset,
7485 .read_u64 = mem_cgroup_read_u64,
7487 { }, /* terminate */
7491 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7492 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7493 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7494 * boot parameter. This may result in premature OOPS inside
7495 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7497 static int __init mem_cgroup_swap_init(void)
7499 /* No memory control -> no swap control */
7500 if (mem_cgroup_disabled())
7501 cgroup_memory_noswap = true;
7503 if (cgroup_memory_noswap)
7506 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7507 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7511 core_initcall(mem_cgroup_swap_init);
7513 #endif /* CONFIG_MEMCG_SWAP */