MIPS: config: Remove left-over BACKLIGHT_LCD_SUPPORT
[linux/fpc-iii.git] / mm / memcontrol.c
blobe50a2db5b4ff0dcf4bf17b888ea12cfd1284dd56
1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
9 * Memory thresholds
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * Native page reclaim
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
37 #include <linux/mm.h>
38 #include <linux/sched/mm.h>
39 #include <linux/shmem_fs.h>
40 #include <linux/hugetlb.h>
41 #include <linux/pagemap.h>
42 #include <linux/vm_event_item.h>
43 #include <linux/smp.h>
44 #include <linux/page-flags.h>
45 #include <linux/backing-dev.h>
46 #include <linux/bit_spinlock.h>
47 #include <linux/rcupdate.h>
48 #include <linux/limits.h>
49 #include <linux/export.h>
50 #include <linux/mutex.h>
51 #include <linux/rbtree.h>
52 #include <linux/slab.h>
53 #include <linux/swap.h>
54 #include <linux/swapops.h>
55 #include <linux/spinlock.h>
56 #include <linux/eventfd.h>
57 #include <linux/poll.h>
58 #include <linux/sort.h>
59 #include <linux/fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/vmpressure.h>
62 #include <linux/mm_inline.h>
63 #include <linux/swap_cgroup.h>
64 #include <linux/cpu.h>
65 #include <linux/oom.h>
66 #include <linux/lockdep.h>
67 #include <linux/file.h>
68 #include <linux/tracehook.h>
69 #include "internal.h"
70 #include <net/sock.h>
71 #include <net/ip.h>
72 #include "slab.h"
74 #include <linux/uaccess.h>
76 #include <trace/events/vmscan.h>
78 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
79 EXPORT_SYMBOL(memory_cgrp_subsys);
81 struct mem_cgroup *root_mem_cgroup __read_mostly;
83 #define MEM_CGROUP_RECLAIM_RETRIES 5
85 /* Socket memory accounting disabled? */
86 static bool cgroup_memory_nosocket;
88 /* Kernel memory accounting disabled? */
89 static bool cgroup_memory_nokmem;
91 /* Whether the swap controller is active */
92 #ifdef CONFIG_MEMCG_SWAP
93 int do_swap_account __read_mostly;
94 #else
95 #define do_swap_account 0
96 #endif
98 /* Whether legacy memory+swap accounting is active */
99 static bool do_memsw_account(void)
101 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
104 static const char *const mem_cgroup_lru_names[] = {
105 "inactive_anon",
106 "active_anon",
107 "inactive_file",
108 "active_file",
109 "unevictable",
112 #define THRESHOLDS_EVENTS_TARGET 128
113 #define SOFTLIMIT_EVENTS_TARGET 1024
114 #define NUMAINFO_EVENTS_TARGET 1024
117 * Cgroups above their limits are maintained in a RB-Tree, independent of
118 * their hierarchy representation
121 struct mem_cgroup_tree_per_node {
122 struct rb_root rb_root;
123 struct rb_node *rb_rightmost;
124 spinlock_t lock;
127 struct mem_cgroup_tree {
128 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
131 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
133 /* for OOM */
134 struct mem_cgroup_eventfd_list {
135 struct list_head list;
136 struct eventfd_ctx *eventfd;
140 * cgroup_event represents events which userspace want to receive.
142 struct mem_cgroup_event {
144 * memcg which the event belongs to.
146 struct mem_cgroup *memcg;
148 * eventfd to signal userspace about the event.
150 struct eventfd_ctx *eventfd;
152 * Each of these stored in a list by the cgroup.
154 struct list_head list;
156 * register_event() callback will be used to add new userspace
157 * waiter for changes related to this event. Use eventfd_signal()
158 * on eventfd to send notification to userspace.
160 int (*register_event)(struct mem_cgroup *memcg,
161 struct eventfd_ctx *eventfd, const char *args);
163 * unregister_event() callback will be called when userspace closes
164 * the eventfd or on cgroup removing. This callback must be set,
165 * if you want provide notification functionality.
167 void (*unregister_event)(struct mem_cgroup *memcg,
168 struct eventfd_ctx *eventfd);
170 * All fields below needed to unregister event when
171 * userspace closes eventfd.
173 poll_table pt;
174 wait_queue_head_t *wqh;
175 wait_queue_entry_t wait;
176 struct work_struct remove;
179 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
180 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
182 /* Stuffs for move charges at task migration. */
184 * Types of charges to be moved.
186 #define MOVE_ANON 0x1U
187 #define MOVE_FILE 0x2U
188 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
190 /* "mc" and its members are protected by cgroup_mutex */
191 static struct move_charge_struct {
192 spinlock_t lock; /* for from, to */
193 struct mm_struct *mm;
194 struct mem_cgroup *from;
195 struct mem_cgroup *to;
196 unsigned long flags;
197 unsigned long precharge;
198 unsigned long moved_charge;
199 unsigned long moved_swap;
200 struct task_struct *moving_task; /* a task moving charges */
201 wait_queue_head_t waitq; /* a waitq for other context */
202 } mc = {
203 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
204 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
208 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
209 * limit reclaim to prevent infinite loops, if they ever occur.
211 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
212 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
214 enum charge_type {
215 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
216 MEM_CGROUP_CHARGE_TYPE_ANON,
217 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
218 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
219 NR_CHARGE_TYPE,
222 /* for encoding cft->private value on file */
223 enum res_type {
224 _MEM,
225 _MEMSWAP,
226 _OOM_TYPE,
227 _KMEM,
228 _TCP,
231 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
232 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
233 #define MEMFILE_ATTR(val) ((val) & 0xffff)
234 /* Used for OOM nofiier */
235 #define OOM_CONTROL (0)
238 * Iteration constructs for visiting all cgroups (under a tree). If
239 * loops are exited prematurely (break), mem_cgroup_iter_break() must
240 * be used for reference counting.
242 #define for_each_mem_cgroup_tree(iter, root) \
243 for (iter = mem_cgroup_iter(root, NULL, NULL); \
244 iter != NULL; \
245 iter = mem_cgroup_iter(root, iter, NULL))
247 #define for_each_mem_cgroup(iter) \
248 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
249 iter != NULL; \
250 iter = mem_cgroup_iter(NULL, iter, NULL))
252 static inline bool should_force_charge(void)
254 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
255 (current->flags & PF_EXITING);
258 /* Some nice accessors for the vmpressure. */
259 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
261 if (!memcg)
262 memcg = root_mem_cgroup;
263 return &memcg->vmpressure;
266 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
268 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
271 #ifdef CONFIG_MEMCG_KMEM
273 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
274 * The main reason for not using cgroup id for this:
275 * this works better in sparse environments, where we have a lot of memcgs,
276 * but only a few kmem-limited. Or also, if we have, for instance, 200
277 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
278 * 200 entry array for that.
280 * The current size of the caches array is stored in memcg_nr_cache_ids. It
281 * will double each time we have to increase it.
283 static DEFINE_IDA(memcg_cache_ida);
284 int memcg_nr_cache_ids;
286 /* Protects memcg_nr_cache_ids */
287 static DECLARE_RWSEM(memcg_cache_ids_sem);
289 void memcg_get_cache_ids(void)
291 down_read(&memcg_cache_ids_sem);
294 void memcg_put_cache_ids(void)
296 up_read(&memcg_cache_ids_sem);
300 * MIN_SIZE is different than 1, because we would like to avoid going through
301 * the alloc/free process all the time. In a small machine, 4 kmem-limited
302 * cgroups is a reasonable guess. In the future, it could be a parameter or
303 * tunable, but that is strictly not necessary.
305 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
306 * this constant directly from cgroup, but it is understandable that this is
307 * better kept as an internal representation in cgroup.c. In any case, the
308 * cgrp_id space is not getting any smaller, and we don't have to necessarily
309 * increase ours as well if it increases.
311 #define MEMCG_CACHES_MIN_SIZE 4
312 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
315 * A lot of the calls to the cache allocation functions are expected to be
316 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
317 * conditional to this static branch, we'll have to allow modules that does
318 * kmem_cache_alloc and the such to see this symbol as well
320 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
321 EXPORT_SYMBOL(memcg_kmem_enabled_key);
323 struct workqueue_struct *memcg_kmem_cache_wq;
325 static int memcg_shrinker_map_size;
326 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
328 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
330 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
333 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
334 int size, int old_size)
336 struct memcg_shrinker_map *new, *old;
337 int nid;
339 lockdep_assert_held(&memcg_shrinker_map_mutex);
341 for_each_node(nid) {
342 old = rcu_dereference_protected(
343 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
344 /* Not yet online memcg */
345 if (!old)
346 return 0;
348 new = kvmalloc(sizeof(*new) + size, GFP_KERNEL);
349 if (!new)
350 return -ENOMEM;
352 /* Set all old bits, clear all new bits */
353 memset(new->map, (int)0xff, old_size);
354 memset((void *)new->map + old_size, 0, size - old_size);
356 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
357 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
360 return 0;
363 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
365 struct mem_cgroup_per_node *pn;
366 struct memcg_shrinker_map *map;
367 int nid;
369 if (mem_cgroup_is_root(memcg))
370 return;
372 for_each_node(nid) {
373 pn = mem_cgroup_nodeinfo(memcg, nid);
374 map = rcu_dereference_protected(pn->shrinker_map, true);
375 if (map)
376 kvfree(map);
377 rcu_assign_pointer(pn->shrinker_map, NULL);
381 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
383 struct memcg_shrinker_map *map;
384 int nid, size, ret = 0;
386 if (mem_cgroup_is_root(memcg))
387 return 0;
389 mutex_lock(&memcg_shrinker_map_mutex);
390 size = memcg_shrinker_map_size;
391 for_each_node(nid) {
392 map = kvzalloc(sizeof(*map) + size, GFP_KERNEL);
393 if (!map) {
394 memcg_free_shrinker_maps(memcg);
395 ret = -ENOMEM;
396 break;
398 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
400 mutex_unlock(&memcg_shrinker_map_mutex);
402 return ret;
405 int memcg_expand_shrinker_maps(int new_id)
407 int size, old_size, ret = 0;
408 struct mem_cgroup *memcg;
410 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
411 old_size = memcg_shrinker_map_size;
412 if (size <= old_size)
413 return 0;
415 mutex_lock(&memcg_shrinker_map_mutex);
416 if (!root_mem_cgroup)
417 goto unlock;
419 for_each_mem_cgroup(memcg) {
420 if (mem_cgroup_is_root(memcg))
421 continue;
422 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
423 if (ret)
424 goto unlock;
426 unlock:
427 if (!ret)
428 memcg_shrinker_map_size = size;
429 mutex_unlock(&memcg_shrinker_map_mutex);
430 return ret;
433 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
435 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
436 struct memcg_shrinker_map *map;
438 rcu_read_lock();
439 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
440 /* Pairs with smp mb in shrink_slab() */
441 smp_mb__before_atomic();
442 set_bit(shrinker_id, map->map);
443 rcu_read_unlock();
447 #else /* CONFIG_MEMCG_KMEM */
448 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
450 return 0;
452 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg) { }
453 #endif /* CONFIG_MEMCG_KMEM */
456 * mem_cgroup_css_from_page - css of the memcg associated with a page
457 * @page: page of interest
459 * If memcg is bound to the default hierarchy, css of the memcg associated
460 * with @page is returned. The returned css remains associated with @page
461 * until it is released.
463 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
464 * is returned.
466 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
468 struct mem_cgroup *memcg;
470 memcg = page->mem_cgroup;
472 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
473 memcg = root_mem_cgroup;
475 return &memcg->css;
479 * page_cgroup_ino - return inode number of the memcg a page is charged to
480 * @page: the page
482 * Look up the closest online ancestor of the memory cgroup @page is charged to
483 * and return its inode number or 0 if @page is not charged to any cgroup. It
484 * is safe to call this function without holding a reference to @page.
486 * Note, this function is inherently racy, because there is nothing to prevent
487 * the cgroup inode from getting torn down and potentially reallocated a moment
488 * after page_cgroup_ino() returns, so it only should be used by callers that
489 * do not care (such as procfs interfaces).
491 ino_t page_cgroup_ino(struct page *page)
493 struct mem_cgroup *memcg;
494 unsigned long ino = 0;
496 rcu_read_lock();
497 memcg = READ_ONCE(page->mem_cgroup);
498 while (memcg && !(memcg->css.flags & CSS_ONLINE))
499 memcg = parent_mem_cgroup(memcg);
500 if (memcg)
501 ino = cgroup_ino(memcg->css.cgroup);
502 rcu_read_unlock();
503 return ino;
506 static struct mem_cgroup_per_node *
507 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
509 int nid = page_to_nid(page);
511 return memcg->nodeinfo[nid];
514 static struct mem_cgroup_tree_per_node *
515 soft_limit_tree_node(int nid)
517 return soft_limit_tree.rb_tree_per_node[nid];
520 static struct mem_cgroup_tree_per_node *
521 soft_limit_tree_from_page(struct page *page)
523 int nid = page_to_nid(page);
525 return soft_limit_tree.rb_tree_per_node[nid];
528 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
529 struct mem_cgroup_tree_per_node *mctz,
530 unsigned long new_usage_in_excess)
532 struct rb_node **p = &mctz->rb_root.rb_node;
533 struct rb_node *parent = NULL;
534 struct mem_cgroup_per_node *mz_node;
535 bool rightmost = true;
537 if (mz->on_tree)
538 return;
540 mz->usage_in_excess = new_usage_in_excess;
541 if (!mz->usage_in_excess)
542 return;
543 while (*p) {
544 parent = *p;
545 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
546 tree_node);
547 if (mz->usage_in_excess < mz_node->usage_in_excess) {
548 p = &(*p)->rb_left;
549 rightmost = false;
553 * We can't avoid mem cgroups that are over their soft
554 * limit by the same amount
556 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
557 p = &(*p)->rb_right;
560 if (rightmost)
561 mctz->rb_rightmost = &mz->tree_node;
563 rb_link_node(&mz->tree_node, parent, p);
564 rb_insert_color(&mz->tree_node, &mctz->rb_root);
565 mz->on_tree = true;
568 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
569 struct mem_cgroup_tree_per_node *mctz)
571 if (!mz->on_tree)
572 return;
574 if (&mz->tree_node == mctz->rb_rightmost)
575 mctz->rb_rightmost = rb_prev(&mz->tree_node);
577 rb_erase(&mz->tree_node, &mctz->rb_root);
578 mz->on_tree = false;
581 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
582 struct mem_cgroup_tree_per_node *mctz)
584 unsigned long flags;
586 spin_lock_irqsave(&mctz->lock, flags);
587 __mem_cgroup_remove_exceeded(mz, mctz);
588 spin_unlock_irqrestore(&mctz->lock, flags);
591 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
593 unsigned long nr_pages = page_counter_read(&memcg->memory);
594 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
595 unsigned long excess = 0;
597 if (nr_pages > soft_limit)
598 excess = nr_pages - soft_limit;
600 return excess;
603 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
605 unsigned long excess;
606 struct mem_cgroup_per_node *mz;
607 struct mem_cgroup_tree_per_node *mctz;
609 mctz = soft_limit_tree_from_page(page);
610 if (!mctz)
611 return;
613 * Necessary to update all ancestors when hierarchy is used.
614 * because their event counter is not touched.
616 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
617 mz = mem_cgroup_page_nodeinfo(memcg, page);
618 excess = soft_limit_excess(memcg);
620 * We have to update the tree if mz is on RB-tree or
621 * mem is over its softlimit.
623 if (excess || mz->on_tree) {
624 unsigned long flags;
626 spin_lock_irqsave(&mctz->lock, flags);
627 /* if on-tree, remove it */
628 if (mz->on_tree)
629 __mem_cgroup_remove_exceeded(mz, mctz);
631 * Insert again. mz->usage_in_excess will be updated.
632 * If excess is 0, no tree ops.
634 __mem_cgroup_insert_exceeded(mz, mctz, excess);
635 spin_unlock_irqrestore(&mctz->lock, flags);
640 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
642 struct mem_cgroup_tree_per_node *mctz;
643 struct mem_cgroup_per_node *mz;
644 int nid;
646 for_each_node(nid) {
647 mz = mem_cgroup_nodeinfo(memcg, nid);
648 mctz = soft_limit_tree_node(nid);
649 if (mctz)
650 mem_cgroup_remove_exceeded(mz, mctz);
654 static struct mem_cgroup_per_node *
655 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
657 struct mem_cgroup_per_node *mz;
659 retry:
660 mz = NULL;
661 if (!mctz->rb_rightmost)
662 goto done; /* Nothing to reclaim from */
664 mz = rb_entry(mctz->rb_rightmost,
665 struct mem_cgroup_per_node, tree_node);
667 * Remove the node now but someone else can add it back,
668 * we will to add it back at the end of reclaim to its correct
669 * position in the tree.
671 __mem_cgroup_remove_exceeded(mz, mctz);
672 if (!soft_limit_excess(mz->memcg) ||
673 !css_tryget_online(&mz->memcg->css))
674 goto retry;
675 done:
676 return mz;
679 static struct mem_cgroup_per_node *
680 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
682 struct mem_cgroup_per_node *mz;
684 spin_lock_irq(&mctz->lock);
685 mz = __mem_cgroup_largest_soft_limit_node(mctz);
686 spin_unlock_irq(&mctz->lock);
687 return mz;
691 * __mod_memcg_state - update cgroup memory statistics
692 * @memcg: the memory cgroup
693 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
694 * @val: delta to add to the counter, can be negative
696 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
698 long x;
700 if (mem_cgroup_disabled())
701 return;
703 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
704 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
705 struct mem_cgroup *mi;
707 atomic_long_add(x, &memcg->vmstats_local[idx]);
708 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
709 atomic_long_add(x, &mi->vmstats[idx]);
710 x = 0;
712 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
715 static struct mem_cgroup_per_node *
716 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
718 struct mem_cgroup *parent;
720 parent = parent_mem_cgroup(pn->memcg);
721 if (!parent)
722 return NULL;
723 return mem_cgroup_nodeinfo(parent, nid);
727 * __mod_lruvec_state - update lruvec memory statistics
728 * @lruvec: the lruvec
729 * @idx: the stat item
730 * @val: delta to add to the counter, can be negative
732 * The lruvec is the intersection of the NUMA node and a cgroup. This
733 * function updates the all three counters that are affected by a
734 * change of state at this level: per-node, per-cgroup, per-lruvec.
736 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
737 int val)
739 pg_data_t *pgdat = lruvec_pgdat(lruvec);
740 struct mem_cgroup_per_node *pn;
741 struct mem_cgroup *memcg;
742 long x;
744 /* Update node */
745 __mod_node_page_state(pgdat, idx, val);
747 if (mem_cgroup_disabled())
748 return;
750 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
751 memcg = pn->memcg;
753 /* Update memcg */
754 __mod_memcg_state(memcg, idx, val);
756 /* Update lruvec */
757 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
758 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
759 struct mem_cgroup_per_node *pi;
761 atomic_long_add(x, &pn->lruvec_stat_local[idx]);
762 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
763 atomic_long_add(x, &pi->lruvec_stat[idx]);
764 x = 0;
766 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
770 * __count_memcg_events - account VM events in a cgroup
771 * @memcg: the memory cgroup
772 * @idx: the event item
773 * @count: the number of events that occured
775 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
776 unsigned long count)
778 unsigned long x;
780 if (mem_cgroup_disabled())
781 return;
783 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
784 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
785 struct mem_cgroup *mi;
787 atomic_long_add(x, &memcg->vmevents_local[idx]);
788 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
789 atomic_long_add(x, &mi->vmevents[idx]);
790 x = 0;
792 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
795 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
797 return atomic_long_read(&memcg->vmevents[event]);
800 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
802 return atomic_long_read(&memcg->vmevents_local[event]);
805 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
806 struct page *page,
807 bool compound, int nr_pages)
810 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
811 * counted as CACHE even if it's on ANON LRU.
813 if (PageAnon(page))
814 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
815 else {
816 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
817 if (PageSwapBacked(page))
818 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
821 if (compound) {
822 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
823 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
826 /* pagein of a big page is an event. So, ignore page size */
827 if (nr_pages > 0)
828 __count_memcg_events(memcg, PGPGIN, 1);
829 else {
830 __count_memcg_events(memcg, PGPGOUT, 1);
831 nr_pages = -nr_pages; /* for event */
834 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
837 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
838 enum mem_cgroup_events_target target)
840 unsigned long val, next;
842 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
843 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
844 /* from time_after() in jiffies.h */
845 if ((long)(next - val) < 0) {
846 switch (target) {
847 case MEM_CGROUP_TARGET_THRESH:
848 next = val + THRESHOLDS_EVENTS_TARGET;
849 break;
850 case MEM_CGROUP_TARGET_SOFTLIMIT:
851 next = val + SOFTLIMIT_EVENTS_TARGET;
852 break;
853 case MEM_CGROUP_TARGET_NUMAINFO:
854 next = val + NUMAINFO_EVENTS_TARGET;
855 break;
856 default:
857 break;
859 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
860 return true;
862 return false;
866 * Check events in order.
869 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
871 /* threshold event is triggered in finer grain than soft limit */
872 if (unlikely(mem_cgroup_event_ratelimit(memcg,
873 MEM_CGROUP_TARGET_THRESH))) {
874 bool do_softlimit;
875 bool do_numainfo __maybe_unused;
877 do_softlimit = mem_cgroup_event_ratelimit(memcg,
878 MEM_CGROUP_TARGET_SOFTLIMIT);
879 #if MAX_NUMNODES > 1
880 do_numainfo = mem_cgroup_event_ratelimit(memcg,
881 MEM_CGROUP_TARGET_NUMAINFO);
882 #endif
883 mem_cgroup_threshold(memcg);
884 if (unlikely(do_softlimit))
885 mem_cgroup_update_tree(memcg, page);
886 #if MAX_NUMNODES > 1
887 if (unlikely(do_numainfo))
888 atomic_inc(&memcg->numainfo_events);
889 #endif
893 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
896 * mm_update_next_owner() may clear mm->owner to NULL
897 * if it races with swapoff, page migration, etc.
898 * So this can be called with p == NULL.
900 if (unlikely(!p))
901 return NULL;
903 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
905 EXPORT_SYMBOL(mem_cgroup_from_task);
908 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
909 * @mm: mm from which memcg should be extracted. It can be NULL.
911 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
912 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
913 * returned.
915 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
917 struct mem_cgroup *memcg;
919 if (mem_cgroup_disabled())
920 return NULL;
922 rcu_read_lock();
923 do {
925 * Page cache insertions can happen withou an
926 * actual mm context, e.g. during disk probing
927 * on boot, loopback IO, acct() writes etc.
929 if (unlikely(!mm))
930 memcg = root_mem_cgroup;
931 else {
932 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
933 if (unlikely(!memcg))
934 memcg = root_mem_cgroup;
936 } while (!css_tryget_online(&memcg->css));
937 rcu_read_unlock();
938 return memcg;
940 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
943 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
944 * @page: page from which memcg should be extracted.
946 * Obtain a reference on page->memcg and returns it if successful. Otherwise
947 * root_mem_cgroup is returned.
949 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
951 struct mem_cgroup *memcg = page->mem_cgroup;
953 if (mem_cgroup_disabled())
954 return NULL;
956 rcu_read_lock();
957 if (!memcg || !css_tryget_online(&memcg->css))
958 memcg = root_mem_cgroup;
959 rcu_read_unlock();
960 return memcg;
962 EXPORT_SYMBOL(get_mem_cgroup_from_page);
965 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
967 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
969 if (unlikely(current->active_memcg)) {
970 struct mem_cgroup *memcg = root_mem_cgroup;
972 rcu_read_lock();
973 if (css_tryget_online(&current->active_memcg->css))
974 memcg = current->active_memcg;
975 rcu_read_unlock();
976 return memcg;
978 return get_mem_cgroup_from_mm(current->mm);
982 * mem_cgroup_iter - iterate over memory cgroup hierarchy
983 * @root: hierarchy root
984 * @prev: previously returned memcg, NULL on first invocation
985 * @reclaim: cookie for shared reclaim walks, NULL for full walks
987 * Returns references to children of the hierarchy below @root, or
988 * @root itself, or %NULL after a full round-trip.
990 * Caller must pass the return value in @prev on subsequent
991 * invocations for reference counting, or use mem_cgroup_iter_break()
992 * to cancel a hierarchy walk before the round-trip is complete.
994 * Reclaimers can specify a node and a priority level in @reclaim to
995 * divide up the memcgs in the hierarchy among all concurrent
996 * reclaimers operating on the same node and priority.
998 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
999 struct mem_cgroup *prev,
1000 struct mem_cgroup_reclaim_cookie *reclaim)
1002 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1003 struct cgroup_subsys_state *css = NULL;
1004 struct mem_cgroup *memcg = NULL;
1005 struct mem_cgroup *pos = NULL;
1007 if (mem_cgroup_disabled())
1008 return NULL;
1010 if (!root)
1011 root = root_mem_cgroup;
1013 if (prev && !reclaim)
1014 pos = prev;
1016 if (!root->use_hierarchy && root != root_mem_cgroup) {
1017 if (prev)
1018 goto out;
1019 return root;
1022 rcu_read_lock();
1024 if (reclaim) {
1025 struct mem_cgroup_per_node *mz;
1027 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1028 iter = &mz->iter[reclaim->priority];
1030 if (prev && reclaim->generation != iter->generation)
1031 goto out_unlock;
1033 while (1) {
1034 pos = READ_ONCE(iter->position);
1035 if (!pos || css_tryget(&pos->css))
1036 break;
1038 * css reference reached zero, so iter->position will
1039 * be cleared by ->css_released. However, we should not
1040 * rely on this happening soon, because ->css_released
1041 * is called from a work queue, and by busy-waiting we
1042 * might block it. So we clear iter->position right
1043 * away.
1045 (void)cmpxchg(&iter->position, pos, NULL);
1049 if (pos)
1050 css = &pos->css;
1052 for (;;) {
1053 css = css_next_descendant_pre(css, &root->css);
1054 if (!css) {
1056 * Reclaimers share the hierarchy walk, and a
1057 * new one might jump in right at the end of
1058 * the hierarchy - make sure they see at least
1059 * one group and restart from the beginning.
1061 if (!prev)
1062 continue;
1063 break;
1067 * Verify the css and acquire a reference. The root
1068 * is provided by the caller, so we know it's alive
1069 * and kicking, and don't take an extra reference.
1071 memcg = mem_cgroup_from_css(css);
1073 if (css == &root->css)
1074 break;
1076 if (css_tryget(css))
1077 break;
1079 memcg = NULL;
1082 if (reclaim) {
1084 * The position could have already been updated by a competing
1085 * thread, so check that the value hasn't changed since we read
1086 * it to avoid reclaiming from the same cgroup twice.
1088 (void)cmpxchg(&iter->position, pos, memcg);
1090 if (pos)
1091 css_put(&pos->css);
1093 if (!memcg)
1094 iter->generation++;
1095 else if (!prev)
1096 reclaim->generation = iter->generation;
1099 out_unlock:
1100 rcu_read_unlock();
1101 out:
1102 if (prev && prev != root)
1103 css_put(&prev->css);
1105 return memcg;
1109 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1110 * @root: hierarchy root
1111 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1113 void mem_cgroup_iter_break(struct mem_cgroup *root,
1114 struct mem_cgroup *prev)
1116 if (!root)
1117 root = root_mem_cgroup;
1118 if (prev && prev != root)
1119 css_put(&prev->css);
1122 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1124 struct mem_cgroup *memcg = dead_memcg;
1125 struct mem_cgroup_reclaim_iter *iter;
1126 struct mem_cgroup_per_node *mz;
1127 int nid;
1128 int i;
1130 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1131 for_each_node(nid) {
1132 mz = mem_cgroup_nodeinfo(memcg, nid);
1133 for (i = 0; i <= DEF_PRIORITY; i++) {
1134 iter = &mz->iter[i];
1135 cmpxchg(&iter->position,
1136 dead_memcg, NULL);
1143 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1144 * @memcg: hierarchy root
1145 * @fn: function to call for each task
1146 * @arg: argument passed to @fn
1148 * This function iterates over tasks attached to @memcg or to any of its
1149 * descendants and calls @fn for each task. If @fn returns a non-zero
1150 * value, the function breaks the iteration loop and returns the value.
1151 * Otherwise, it will iterate over all tasks and return 0.
1153 * This function must not be called for the root memory cgroup.
1155 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1156 int (*fn)(struct task_struct *, void *), void *arg)
1158 struct mem_cgroup *iter;
1159 int ret = 0;
1161 BUG_ON(memcg == root_mem_cgroup);
1163 for_each_mem_cgroup_tree(iter, memcg) {
1164 struct css_task_iter it;
1165 struct task_struct *task;
1167 css_task_iter_start(&iter->css, 0, &it);
1168 while (!ret && (task = css_task_iter_next(&it)))
1169 ret = fn(task, arg);
1170 css_task_iter_end(&it);
1171 if (ret) {
1172 mem_cgroup_iter_break(memcg, iter);
1173 break;
1176 return ret;
1180 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1181 * @page: the page
1182 * @pgdat: pgdat of the page
1184 * This function is only safe when following the LRU page isolation
1185 * and putback protocol: the LRU lock must be held, and the page must
1186 * either be PageLRU() or the caller must have isolated/allocated it.
1188 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1190 struct mem_cgroup_per_node *mz;
1191 struct mem_cgroup *memcg;
1192 struct lruvec *lruvec;
1194 if (mem_cgroup_disabled()) {
1195 lruvec = &pgdat->lruvec;
1196 goto out;
1199 memcg = page->mem_cgroup;
1201 * Swapcache readahead pages are added to the LRU - and
1202 * possibly migrated - before they are charged.
1204 if (!memcg)
1205 memcg = root_mem_cgroup;
1207 mz = mem_cgroup_page_nodeinfo(memcg, page);
1208 lruvec = &mz->lruvec;
1209 out:
1211 * Since a node can be onlined after the mem_cgroup was created,
1212 * we have to be prepared to initialize lruvec->zone here;
1213 * and if offlined then reonlined, we need to reinitialize it.
1215 if (unlikely(lruvec->pgdat != pgdat))
1216 lruvec->pgdat = pgdat;
1217 return lruvec;
1221 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1222 * @lruvec: mem_cgroup per zone lru vector
1223 * @lru: index of lru list the page is sitting on
1224 * @zid: zone id of the accounted pages
1225 * @nr_pages: positive when adding or negative when removing
1227 * This function must be called under lru_lock, just before a page is added
1228 * to or just after a page is removed from an lru list (that ordering being
1229 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1231 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1232 int zid, int nr_pages)
1234 struct mem_cgroup_per_node *mz;
1235 unsigned long *lru_size;
1236 long size;
1238 if (mem_cgroup_disabled())
1239 return;
1241 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1242 lru_size = &mz->lru_zone_size[zid][lru];
1244 if (nr_pages < 0)
1245 *lru_size += nr_pages;
1247 size = *lru_size;
1248 if (WARN_ONCE(size < 0,
1249 "%s(%p, %d, %d): lru_size %ld\n",
1250 __func__, lruvec, lru, nr_pages, size)) {
1251 VM_BUG_ON(1);
1252 *lru_size = 0;
1255 if (nr_pages > 0)
1256 *lru_size += nr_pages;
1259 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1261 struct mem_cgroup *task_memcg;
1262 struct task_struct *p;
1263 bool ret;
1265 p = find_lock_task_mm(task);
1266 if (p) {
1267 task_memcg = get_mem_cgroup_from_mm(p->mm);
1268 task_unlock(p);
1269 } else {
1271 * All threads may have already detached their mm's, but the oom
1272 * killer still needs to detect if they have already been oom
1273 * killed to prevent needlessly killing additional tasks.
1275 rcu_read_lock();
1276 task_memcg = mem_cgroup_from_task(task);
1277 css_get(&task_memcg->css);
1278 rcu_read_unlock();
1280 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1281 css_put(&task_memcg->css);
1282 return ret;
1286 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1287 * @memcg: the memory cgroup
1289 * Returns the maximum amount of memory @mem can be charged with, in
1290 * pages.
1292 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1294 unsigned long margin = 0;
1295 unsigned long count;
1296 unsigned long limit;
1298 count = page_counter_read(&memcg->memory);
1299 limit = READ_ONCE(memcg->memory.max);
1300 if (count < limit)
1301 margin = limit - count;
1303 if (do_memsw_account()) {
1304 count = page_counter_read(&memcg->memsw);
1305 limit = READ_ONCE(memcg->memsw.max);
1306 if (count <= limit)
1307 margin = min(margin, limit - count);
1308 else
1309 margin = 0;
1312 return margin;
1316 * A routine for checking "mem" is under move_account() or not.
1318 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1319 * moving cgroups. This is for waiting at high-memory pressure
1320 * caused by "move".
1322 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1324 struct mem_cgroup *from;
1325 struct mem_cgroup *to;
1326 bool ret = false;
1328 * Unlike task_move routines, we access mc.to, mc.from not under
1329 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1331 spin_lock(&mc.lock);
1332 from = mc.from;
1333 to = mc.to;
1334 if (!from)
1335 goto unlock;
1337 ret = mem_cgroup_is_descendant(from, memcg) ||
1338 mem_cgroup_is_descendant(to, memcg);
1339 unlock:
1340 spin_unlock(&mc.lock);
1341 return ret;
1344 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1346 if (mc.moving_task && current != mc.moving_task) {
1347 if (mem_cgroup_under_move(memcg)) {
1348 DEFINE_WAIT(wait);
1349 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1350 /* moving charge context might have finished. */
1351 if (mc.moving_task)
1352 schedule();
1353 finish_wait(&mc.waitq, &wait);
1354 return true;
1357 return false;
1360 static const unsigned int memcg1_stats[] = {
1361 MEMCG_CACHE,
1362 MEMCG_RSS,
1363 MEMCG_RSS_HUGE,
1364 NR_SHMEM,
1365 NR_FILE_MAPPED,
1366 NR_FILE_DIRTY,
1367 NR_WRITEBACK,
1368 MEMCG_SWAP,
1371 static const char *const memcg1_stat_names[] = {
1372 "cache",
1373 "rss",
1374 "rss_huge",
1375 "shmem",
1376 "mapped_file",
1377 "dirty",
1378 "writeback",
1379 "swap",
1382 #define K(x) ((x) << (PAGE_SHIFT-10))
1384 * mem_cgroup_print_oom_context: Print OOM information relevant to
1385 * memory controller.
1386 * @memcg: The memory cgroup that went over limit
1387 * @p: Task that is going to be killed
1389 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1390 * enabled
1392 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1394 rcu_read_lock();
1396 if (memcg) {
1397 pr_cont(",oom_memcg=");
1398 pr_cont_cgroup_path(memcg->css.cgroup);
1399 } else
1400 pr_cont(",global_oom");
1401 if (p) {
1402 pr_cont(",task_memcg=");
1403 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1405 rcu_read_unlock();
1409 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1410 * memory controller.
1411 * @memcg: The memory cgroup that went over limit
1413 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1415 struct mem_cgroup *iter;
1416 unsigned int i;
1418 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1419 K((u64)page_counter_read(&memcg->memory)),
1420 K((u64)memcg->memory.max), memcg->memory.failcnt);
1421 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1422 K((u64)page_counter_read(&memcg->memsw)),
1423 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1424 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1425 K((u64)page_counter_read(&memcg->kmem)),
1426 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1428 for_each_mem_cgroup_tree(iter, memcg) {
1429 pr_info("Memory cgroup stats for ");
1430 pr_cont_cgroup_path(iter->css.cgroup);
1431 pr_cont(":");
1433 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
1434 if (memcg1_stats[i] == MEMCG_SWAP && !do_swap_account)
1435 continue;
1436 pr_cont(" %s:%luKB", memcg1_stat_names[i],
1437 K(memcg_page_state_local(iter,
1438 memcg1_stats[i])));
1441 for (i = 0; i < NR_LRU_LISTS; i++)
1442 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1443 K(memcg_page_state_local(iter,
1444 NR_LRU_BASE + i)));
1446 pr_cont("\n");
1451 * Return the memory (and swap, if configured) limit for a memcg.
1453 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1455 unsigned long max;
1457 max = memcg->memory.max;
1458 if (mem_cgroup_swappiness(memcg)) {
1459 unsigned long memsw_max;
1460 unsigned long swap_max;
1462 memsw_max = memcg->memsw.max;
1463 swap_max = memcg->swap.max;
1464 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1465 max = min(max + swap_max, memsw_max);
1467 return max;
1470 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1471 int order)
1473 struct oom_control oc = {
1474 .zonelist = NULL,
1475 .nodemask = NULL,
1476 .memcg = memcg,
1477 .gfp_mask = gfp_mask,
1478 .order = order,
1480 bool ret;
1482 if (mutex_lock_killable(&oom_lock))
1483 return true;
1485 * A few threads which were not waiting at mutex_lock_killable() can
1486 * fail to bail out. Therefore, check again after holding oom_lock.
1488 ret = should_force_charge() || out_of_memory(&oc);
1489 mutex_unlock(&oom_lock);
1490 return ret;
1493 #if MAX_NUMNODES > 1
1496 * test_mem_cgroup_node_reclaimable
1497 * @memcg: the target memcg
1498 * @nid: the node ID to be checked.
1499 * @noswap : specify true here if the user wants flle only information.
1501 * This function returns whether the specified memcg contains any
1502 * reclaimable pages on a node. Returns true if there are any reclaimable
1503 * pages in the node.
1505 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1506 int nid, bool noswap)
1508 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
1510 if (lruvec_page_state(lruvec, NR_INACTIVE_FILE) ||
1511 lruvec_page_state(lruvec, NR_ACTIVE_FILE))
1512 return true;
1513 if (noswap || !total_swap_pages)
1514 return false;
1515 if (lruvec_page_state(lruvec, NR_INACTIVE_ANON) ||
1516 lruvec_page_state(lruvec, NR_ACTIVE_ANON))
1517 return true;
1518 return false;
1523 * Always updating the nodemask is not very good - even if we have an empty
1524 * list or the wrong list here, we can start from some node and traverse all
1525 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1528 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1530 int nid;
1532 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1533 * pagein/pageout changes since the last update.
1535 if (!atomic_read(&memcg->numainfo_events))
1536 return;
1537 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1538 return;
1540 /* make a nodemask where this memcg uses memory from */
1541 memcg->scan_nodes = node_states[N_MEMORY];
1543 for_each_node_mask(nid, node_states[N_MEMORY]) {
1545 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1546 node_clear(nid, memcg->scan_nodes);
1549 atomic_set(&memcg->numainfo_events, 0);
1550 atomic_set(&memcg->numainfo_updating, 0);
1554 * Selecting a node where we start reclaim from. Because what we need is just
1555 * reducing usage counter, start from anywhere is O,K. Considering
1556 * memory reclaim from current node, there are pros. and cons.
1558 * Freeing memory from current node means freeing memory from a node which
1559 * we'll use or we've used. So, it may make LRU bad. And if several threads
1560 * hit limits, it will see a contention on a node. But freeing from remote
1561 * node means more costs for memory reclaim because of memory latency.
1563 * Now, we use round-robin. Better algorithm is welcomed.
1565 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1567 int node;
1569 mem_cgroup_may_update_nodemask(memcg);
1570 node = memcg->last_scanned_node;
1572 node = next_node_in(node, memcg->scan_nodes);
1574 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1575 * last time it really checked all the LRUs due to rate limiting.
1576 * Fallback to the current node in that case for simplicity.
1578 if (unlikely(node == MAX_NUMNODES))
1579 node = numa_node_id();
1581 memcg->last_scanned_node = node;
1582 return node;
1584 #else
1585 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1587 return 0;
1589 #endif
1591 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1592 pg_data_t *pgdat,
1593 gfp_t gfp_mask,
1594 unsigned long *total_scanned)
1596 struct mem_cgroup *victim = NULL;
1597 int total = 0;
1598 int loop = 0;
1599 unsigned long excess;
1600 unsigned long nr_scanned;
1601 struct mem_cgroup_reclaim_cookie reclaim = {
1602 .pgdat = pgdat,
1603 .priority = 0,
1606 excess = soft_limit_excess(root_memcg);
1608 while (1) {
1609 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1610 if (!victim) {
1611 loop++;
1612 if (loop >= 2) {
1614 * If we have not been able to reclaim
1615 * anything, it might because there are
1616 * no reclaimable pages under this hierarchy
1618 if (!total)
1619 break;
1621 * We want to do more targeted reclaim.
1622 * excess >> 2 is not to excessive so as to
1623 * reclaim too much, nor too less that we keep
1624 * coming back to reclaim from this cgroup
1626 if (total >= (excess >> 2) ||
1627 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1628 break;
1630 continue;
1632 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1633 pgdat, &nr_scanned);
1634 *total_scanned += nr_scanned;
1635 if (!soft_limit_excess(root_memcg))
1636 break;
1638 mem_cgroup_iter_break(root_memcg, victim);
1639 return total;
1642 #ifdef CONFIG_LOCKDEP
1643 static struct lockdep_map memcg_oom_lock_dep_map = {
1644 .name = "memcg_oom_lock",
1646 #endif
1648 static DEFINE_SPINLOCK(memcg_oom_lock);
1651 * Check OOM-Killer is already running under our hierarchy.
1652 * If someone is running, return false.
1654 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1656 struct mem_cgroup *iter, *failed = NULL;
1658 spin_lock(&memcg_oom_lock);
1660 for_each_mem_cgroup_tree(iter, memcg) {
1661 if (iter->oom_lock) {
1663 * this subtree of our hierarchy is already locked
1664 * so we cannot give a lock.
1666 failed = iter;
1667 mem_cgroup_iter_break(memcg, iter);
1668 break;
1669 } else
1670 iter->oom_lock = true;
1673 if (failed) {
1675 * OK, we failed to lock the whole subtree so we have
1676 * to clean up what we set up to the failing subtree
1678 for_each_mem_cgroup_tree(iter, memcg) {
1679 if (iter == failed) {
1680 mem_cgroup_iter_break(memcg, iter);
1681 break;
1683 iter->oom_lock = false;
1685 } else
1686 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1688 spin_unlock(&memcg_oom_lock);
1690 return !failed;
1693 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1695 struct mem_cgroup *iter;
1697 spin_lock(&memcg_oom_lock);
1698 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1699 for_each_mem_cgroup_tree(iter, memcg)
1700 iter->oom_lock = false;
1701 spin_unlock(&memcg_oom_lock);
1704 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1706 struct mem_cgroup *iter;
1708 spin_lock(&memcg_oom_lock);
1709 for_each_mem_cgroup_tree(iter, memcg)
1710 iter->under_oom++;
1711 spin_unlock(&memcg_oom_lock);
1714 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1716 struct mem_cgroup *iter;
1719 * When a new child is created while the hierarchy is under oom,
1720 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1722 spin_lock(&memcg_oom_lock);
1723 for_each_mem_cgroup_tree(iter, memcg)
1724 if (iter->under_oom > 0)
1725 iter->under_oom--;
1726 spin_unlock(&memcg_oom_lock);
1729 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1731 struct oom_wait_info {
1732 struct mem_cgroup *memcg;
1733 wait_queue_entry_t wait;
1736 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1737 unsigned mode, int sync, void *arg)
1739 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1740 struct mem_cgroup *oom_wait_memcg;
1741 struct oom_wait_info *oom_wait_info;
1743 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1744 oom_wait_memcg = oom_wait_info->memcg;
1746 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1747 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1748 return 0;
1749 return autoremove_wake_function(wait, mode, sync, arg);
1752 static void memcg_oom_recover(struct mem_cgroup *memcg)
1755 * For the following lockless ->under_oom test, the only required
1756 * guarantee is that it must see the state asserted by an OOM when
1757 * this function is called as a result of userland actions
1758 * triggered by the notification of the OOM. This is trivially
1759 * achieved by invoking mem_cgroup_mark_under_oom() before
1760 * triggering notification.
1762 if (memcg && memcg->under_oom)
1763 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1766 enum oom_status {
1767 OOM_SUCCESS,
1768 OOM_FAILED,
1769 OOM_ASYNC,
1770 OOM_SKIPPED
1773 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1775 enum oom_status ret;
1776 bool locked;
1778 if (order > PAGE_ALLOC_COSTLY_ORDER)
1779 return OOM_SKIPPED;
1781 memcg_memory_event(memcg, MEMCG_OOM);
1784 * We are in the middle of the charge context here, so we
1785 * don't want to block when potentially sitting on a callstack
1786 * that holds all kinds of filesystem and mm locks.
1788 * cgroup1 allows disabling the OOM killer and waiting for outside
1789 * handling until the charge can succeed; remember the context and put
1790 * the task to sleep at the end of the page fault when all locks are
1791 * released.
1793 * On the other hand, in-kernel OOM killer allows for an async victim
1794 * memory reclaim (oom_reaper) and that means that we are not solely
1795 * relying on the oom victim to make a forward progress and we can
1796 * invoke the oom killer here.
1798 * Please note that mem_cgroup_out_of_memory might fail to find a
1799 * victim and then we have to bail out from the charge path.
1801 if (memcg->oom_kill_disable) {
1802 if (!current->in_user_fault)
1803 return OOM_SKIPPED;
1804 css_get(&memcg->css);
1805 current->memcg_in_oom = memcg;
1806 current->memcg_oom_gfp_mask = mask;
1807 current->memcg_oom_order = order;
1809 return OOM_ASYNC;
1812 mem_cgroup_mark_under_oom(memcg);
1814 locked = mem_cgroup_oom_trylock(memcg);
1816 if (locked)
1817 mem_cgroup_oom_notify(memcg);
1819 mem_cgroup_unmark_under_oom(memcg);
1820 if (mem_cgroup_out_of_memory(memcg, mask, order))
1821 ret = OOM_SUCCESS;
1822 else
1823 ret = OOM_FAILED;
1825 if (locked)
1826 mem_cgroup_oom_unlock(memcg);
1828 return ret;
1832 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1833 * @handle: actually kill/wait or just clean up the OOM state
1835 * This has to be called at the end of a page fault if the memcg OOM
1836 * handler was enabled.
1838 * Memcg supports userspace OOM handling where failed allocations must
1839 * sleep on a waitqueue until the userspace task resolves the
1840 * situation. Sleeping directly in the charge context with all kinds
1841 * of locks held is not a good idea, instead we remember an OOM state
1842 * in the task and mem_cgroup_oom_synchronize() has to be called at
1843 * the end of the page fault to complete the OOM handling.
1845 * Returns %true if an ongoing memcg OOM situation was detected and
1846 * completed, %false otherwise.
1848 bool mem_cgroup_oom_synchronize(bool handle)
1850 struct mem_cgroup *memcg = current->memcg_in_oom;
1851 struct oom_wait_info owait;
1852 bool locked;
1854 /* OOM is global, do not handle */
1855 if (!memcg)
1856 return false;
1858 if (!handle)
1859 goto cleanup;
1861 owait.memcg = memcg;
1862 owait.wait.flags = 0;
1863 owait.wait.func = memcg_oom_wake_function;
1864 owait.wait.private = current;
1865 INIT_LIST_HEAD(&owait.wait.entry);
1867 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1868 mem_cgroup_mark_under_oom(memcg);
1870 locked = mem_cgroup_oom_trylock(memcg);
1872 if (locked)
1873 mem_cgroup_oom_notify(memcg);
1875 if (locked && !memcg->oom_kill_disable) {
1876 mem_cgroup_unmark_under_oom(memcg);
1877 finish_wait(&memcg_oom_waitq, &owait.wait);
1878 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1879 current->memcg_oom_order);
1880 } else {
1881 schedule();
1882 mem_cgroup_unmark_under_oom(memcg);
1883 finish_wait(&memcg_oom_waitq, &owait.wait);
1886 if (locked) {
1887 mem_cgroup_oom_unlock(memcg);
1889 * There is no guarantee that an OOM-lock contender
1890 * sees the wakeups triggered by the OOM kill
1891 * uncharges. Wake any sleepers explicitely.
1893 memcg_oom_recover(memcg);
1895 cleanup:
1896 current->memcg_in_oom = NULL;
1897 css_put(&memcg->css);
1898 return true;
1902 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1903 * @victim: task to be killed by the OOM killer
1904 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1906 * Returns a pointer to a memory cgroup, which has to be cleaned up
1907 * by killing all belonging OOM-killable tasks.
1909 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1911 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1912 struct mem_cgroup *oom_domain)
1914 struct mem_cgroup *oom_group = NULL;
1915 struct mem_cgroup *memcg;
1917 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1918 return NULL;
1920 if (!oom_domain)
1921 oom_domain = root_mem_cgroup;
1923 rcu_read_lock();
1925 memcg = mem_cgroup_from_task(victim);
1926 if (memcg == root_mem_cgroup)
1927 goto out;
1930 * Traverse the memory cgroup hierarchy from the victim task's
1931 * cgroup up to the OOMing cgroup (or root) to find the
1932 * highest-level memory cgroup with oom.group set.
1934 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1935 if (memcg->oom_group)
1936 oom_group = memcg;
1938 if (memcg == oom_domain)
1939 break;
1942 if (oom_group)
1943 css_get(&oom_group->css);
1944 out:
1945 rcu_read_unlock();
1947 return oom_group;
1950 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1952 pr_info("Tasks in ");
1953 pr_cont_cgroup_path(memcg->css.cgroup);
1954 pr_cont(" are going to be killed due to memory.oom.group set\n");
1958 * lock_page_memcg - lock a page->mem_cgroup binding
1959 * @page: the page
1961 * This function protects unlocked LRU pages from being moved to
1962 * another cgroup.
1964 * It ensures lifetime of the returned memcg. Caller is responsible
1965 * for the lifetime of the page; __unlock_page_memcg() is available
1966 * when @page might get freed inside the locked section.
1968 struct mem_cgroup *lock_page_memcg(struct page *page)
1970 struct mem_cgroup *memcg;
1971 unsigned long flags;
1974 * The RCU lock is held throughout the transaction. The fast
1975 * path can get away without acquiring the memcg->move_lock
1976 * because page moving starts with an RCU grace period.
1978 * The RCU lock also protects the memcg from being freed when
1979 * the page state that is going to change is the only thing
1980 * preventing the page itself from being freed. E.g. writeback
1981 * doesn't hold a page reference and relies on PG_writeback to
1982 * keep off truncation, migration and so forth.
1984 rcu_read_lock();
1986 if (mem_cgroup_disabled())
1987 return NULL;
1988 again:
1989 memcg = page->mem_cgroup;
1990 if (unlikely(!memcg))
1991 return NULL;
1993 if (atomic_read(&memcg->moving_account) <= 0)
1994 return memcg;
1996 spin_lock_irqsave(&memcg->move_lock, flags);
1997 if (memcg != page->mem_cgroup) {
1998 spin_unlock_irqrestore(&memcg->move_lock, flags);
1999 goto again;
2003 * When charge migration first begins, we can have locked and
2004 * unlocked page stat updates happening concurrently. Track
2005 * the task who has the lock for unlock_page_memcg().
2007 memcg->move_lock_task = current;
2008 memcg->move_lock_flags = flags;
2010 return memcg;
2012 EXPORT_SYMBOL(lock_page_memcg);
2015 * __unlock_page_memcg - unlock and unpin a memcg
2016 * @memcg: the memcg
2018 * Unlock and unpin a memcg returned by lock_page_memcg().
2020 void __unlock_page_memcg(struct mem_cgroup *memcg)
2022 if (memcg && memcg->move_lock_task == current) {
2023 unsigned long flags = memcg->move_lock_flags;
2025 memcg->move_lock_task = NULL;
2026 memcg->move_lock_flags = 0;
2028 spin_unlock_irqrestore(&memcg->move_lock, flags);
2031 rcu_read_unlock();
2035 * unlock_page_memcg - unlock a page->mem_cgroup binding
2036 * @page: the page
2038 void unlock_page_memcg(struct page *page)
2040 __unlock_page_memcg(page->mem_cgroup);
2042 EXPORT_SYMBOL(unlock_page_memcg);
2044 struct memcg_stock_pcp {
2045 struct mem_cgroup *cached; /* this never be root cgroup */
2046 unsigned int nr_pages;
2047 struct work_struct work;
2048 unsigned long flags;
2049 #define FLUSHING_CACHED_CHARGE 0
2051 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2052 static DEFINE_MUTEX(percpu_charge_mutex);
2055 * consume_stock: Try to consume stocked charge on this cpu.
2056 * @memcg: memcg to consume from.
2057 * @nr_pages: how many pages to charge.
2059 * The charges will only happen if @memcg matches the current cpu's memcg
2060 * stock, and at least @nr_pages are available in that stock. Failure to
2061 * service an allocation will refill the stock.
2063 * returns true if successful, false otherwise.
2065 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2067 struct memcg_stock_pcp *stock;
2068 unsigned long flags;
2069 bool ret = false;
2071 if (nr_pages > MEMCG_CHARGE_BATCH)
2072 return ret;
2074 local_irq_save(flags);
2076 stock = this_cpu_ptr(&memcg_stock);
2077 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2078 stock->nr_pages -= nr_pages;
2079 ret = true;
2082 local_irq_restore(flags);
2084 return ret;
2088 * Returns stocks cached in percpu and reset cached information.
2090 static void drain_stock(struct memcg_stock_pcp *stock)
2092 struct mem_cgroup *old = stock->cached;
2094 if (stock->nr_pages) {
2095 page_counter_uncharge(&old->memory, stock->nr_pages);
2096 if (do_memsw_account())
2097 page_counter_uncharge(&old->memsw, stock->nr_pages);
2098 css_put_many(&old->css, stock->nr_pages);
2099 stock->nr_pages = 0;
2101 stock->cached = NULL;
2104 static void drain_local_stock(struct work_struct *dummy)
2106 struct memcg_stock_pcp *stock;
2107 unsigned long flags;
2110 * The only protection from memory hotplug vs. drain_stock races is
2111 * that we always operate on local CPU stock here with IRQ disabled
2113 local_irq_save(flags);
2115 stock = this_cpu_ptr(&memcg_stock);
2116 drain_stock(stock);
2117 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2119 local_irq_restore(flags);
2123 * Cache charges(val) to local per_cpu area.
2124 * This will be consumed by consume_stock() function, later.
2126 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2128 struct memcg_stock_pcp *stock;
2129 unsigned long flags;
2131 local_irq_save(flags);
2133 stock = this_cpu_ptr(&memcg_stock);
2134 if (stock->cached != memcg) { /* reset if necessary */
2135 drain_stock(stock);
2136 stock->cached = memcg;
2138 stock->nr_pages += nr_pages;
2140 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2141 drain_stock(stock);
2143 local_irq_restore(flags);
2147 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2148 * of the hierarchy under it.
2150 static void drain_all_stock(struct mem_cgroup *root_memcg)
2152 int cpu, curcpu;
2154 /* If someone's already draining, avoid adding running more workers. */
2155 if (!mutex_trylock(&percpu_charge_mutex))
2156 return;
2158 * Notify other cpus that system-wide "drain" is running
2159 * We do not care about races with the cpu hotplug because cpu down
2160 * as well as workers from this path always operate on the local
2161 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2163 curcpu = get_cpu();
2164 for_each_online_cpu(cpu) {
2165 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2166 struct mem_cgroup *memcg;
2168 memcg = stock->cached;
2169 if (!memcg || !stock->nr_pages || !css_tryget(&memcg->css))
2170 continue;
2171 if (!mem_cgroup_is_descendant(memcg, root_memcg)) {
2172 css_put(&memcg->css);
2173 continue;
2175 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2176 if (cpu == curcpu)
2177 drain_local_stock(&stock->work);
2178 else
2179 schedule_work_on(cpu, &stock->work);
2181 css_put(&memcg->css);
2183 put_cpu();
2184 mutex_unlock(&percpu_charge_mutex);
2187 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2189 struct memcg_stock_pcp *stock;
2190 struct mem_cgroup *memcg, *mi;
2192 stock = &per_cpu(memcg_stock, cpu);
2193 drain_stock(stock);
2195 for_each_mem_cgroup(memcg) {
2196 int i;
2198 for (i = 0; i < MEMCG_NR_STAT; i++) {
2199 int nid;
2200 long x;
2202 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2203 if (x) {
2204 atomic_long_add(x, &memcg->vmstats_local[i]);
2205 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2206 atomic_long_add(x, &memcg->vmstats[i]);
2209 if (i >= NR_VM_NODE_STAT_ITEMS)
2210 continue;
2212 for_each_node(nid) {
2213 struct mem_cgroup_per_node *pn;
2215 pn = mem_cgroup_nodeinfo(memcg, nid);
2216 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2217 if (x) {
2218 atomic_long_add(x, &pn->lruvec_stat_local[i]);
2219 do {
2220 atomic_long_add(x, &pn->lruvec_stat[i]);
2221 } while ((pn = parent_nodeinfo(pn, nid)));
2226 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2227 long x;
2229 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2230 if (x) {
2231 atomic_long_add(x, &memcg->vmevents_local[i]);
2232 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2233 atomic_long_add(x, &memcg->vmevents[i]);
2238 return 0;
2241 static void reclaim_high(struct mem_cgroup *memcg,
2242 unsigned int nr_pages,
2243 gfp_t gfp_mask)
2245 do {
2246 if (page_counter_read(&memcg->memory) <= memcg->high)
2247 continue;
2248 memcg_memory_event(memcg, MEMCG_HIGH);
2249 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2250 } while ((memcg = parent_mem_cgroup(memcg)));
2253 static void high_work_func(struct work_struct *work)
2255 struct mem_cgroup *memcg;
2257 memcg = container_of(work, struct mem_cgroup, high_work);
2258 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2262 * Scheduled by try_charge() to be executed from the userland return path
2263 * and reclaims memory over the high limit.
2265 void mem_cgroup_handle_over_high(void)
2267 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2268 struct mem_cgroup *memcg;
2270 if (likely(!nr_pages))
2271 return;
2273 memcg = get_mem_cgroup_from_mm(current->mm);
2274 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2275 css_put(&memcg->css);
2276 current->memcg_nr_pages_over_high = 0;
2279 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2280 unsigned int nr_pages)
2282 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2283 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2284 struct mem_cgroup *mem_over_limit;
2285 struct page_counter *counter;
2286 unsigned long nr_reclaimed;
2287 bool may_swap = true;
2288 bool drained = false;
2289 bool oomed = false;
2290 enum oom_status oom_status;
2292 if (mem_cgroup_is_root(memcg))
2293 return 0;
2294 retry:
2295 if (consume_stock(memcg, nr_pages))
2296 return 0;
2298 if (!do_memsw_account() ||
2299 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2300 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2301 goto done_restock;
2302 if (do_memsw_account())
2303 page_counter_uncharge(&memcg->memsw, batch);
2304 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2305 } else {
2306 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2307 may_swap = false;
2310 if (batch > nr_pages) {
2311 batch = nr_pages;
2312 goto retry;
2316 * Unlike in global OOM situations, memcg is not in a physical
2317 * memory shortage. Allow dying and OOM-killed tasks to
2318 * bypass the last charges so that they can exit quickly and
2319 * free their memory.
2321 if (unlikely(should_force_charge()))
2322 goto force;
2325 * Prevent unbounded recursion when reclaim operations need to
2326 * allocate memory. This might exceed the limits temporarily,
2327 * but we prefer facilitating memory reclaim and getting back
2328 * under the limit over triggering OOM kills in these cases.
2330 if (unlikely(current->flags & PF_MEMALLOC))
2331 goto force;
2333 if (unlikely(task_in_memcg_oom(current)))
2334 goto nomem;
2336 if (!gfpflags_allow_blocking(gfp_mask))
2337 goto nomem;
2339 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2341 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2342 gfp_mask, may_swap);
2344 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2345 goto retry;
2347 if (!drained) {
2348 drain_all_stock(mem_over_limit);
2349 drained = true;
2350 goto retry;
2353 if (gfp_mask & __GFP_NORETRY)
2354 goto nomem;
2356 * Even though the limit is exceeded at this point, reclaim
2357 * may have been able to free some pages. Retry the charge
2358 * before killing the task.
2360 * Only for regular pages, though: huge pages are rather
2361 * unlikely to succeed so close to the limit, and we fall back
2362 * to regular pages anyway in case of failure.
2364 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2365 goto retry;
2367 * At task move, charge accounts can be doubly counted. So, it's
2368 * better to wait until the end of task_move if something is going on.
2370 if (mem_cgroup_wait_acct_move(mem_over_limit))
2371 goto retry;
2373 if (nr_retries--)
2374 goto retry;
2376 if (gfp_mask & __GFP_RETRY_MAYFAIL && oomed)
2377 goto nomem;
2379 if (gfp_mask & __GFP_NOFAIL)
2380 goto force;
2382 if (fatal_signal_pending(current))
2383 goto force;
2386 * keep retrying as long as the memcg oom killer is able to make
2387 * a forward progress or bypass the charge if the oom killer
2388 * couldn't make any progress.
2390 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2391 get_order(nr_pages * PAGE_SIZE));
2392 switch (oom_status) {
2393 case OOM_SUCCESS:
2394 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2395 oomed = true;
2396 goto retry;
2397 case OOM_FAILED:
2398 goto force;
2399 default:
2400 goto nomem;
2402 nomem:
2403 if (!(gfp_mask & __GFP_NOFAIL))
2404 return -ENOMEM;
2405 force:
2407 * The allocation either can't fail or will lead to more memory
2408 * being freed very soon. Allow memory usage go over the limit
2409 * temporarily by force charging it.
2411 page_counter_charge(&memcg->memory, nr_pages);
2412 if (do_memsw_account())
2413 page_counter_charge(&memcg->memsw, nr_pages);
2414 css_get_many(&memcg->css, nr_pages);
2416 return 0;
2418 done_restock:
2419 css_get_many(&memcg->css, batch);
2420 if (batch > nr_pages)
2421 refill_stock(memcg, batch - nr_pages);
2424 * If the hierarchy is above the normal consumption range, schedule
2425 * reclaim on returning to userland. We can perform reclaim here
2426 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2427 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2428 * not recorded as it most likely matches current's and won't
2429 * change in the meantime. As high limit is checked again before
2430 * reclaim, the cost of mismatch is negligible.
2432 do {
2433 if (page_counter_read(&memcg->memory) > memcg->high) {
2434 /* Don't bother a random interrupted task */
2435 if (in_interrupt()) {
2436 schedule_work(&memcg->high_work);
2437 break;
2439 current->memcg_nr_pages_over_high += batch;
2440 set_notify_resume(current);
2441 break;
2443 } while ((memcg = parent_mem_cgroup(memcg)));
2445 return 0;
2448 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2450 if (mem_cgroup_is_root(memcg))
2451 return;
2453 page_counter_uncharge(&memcg->memory, nr_pages);
2454 if (do_memsw_account())
2455 page_counter_uncharge(&memcg->memsw, nr_pages);
2457 css_put_many(&memcg->css, nr_pages);
2460 static void lock_page_lru(struct page *page, int *isolated)
2462 pg_data_t *pgdat = page_pgdat(page);
2464 spin_lock_irq(&pgdat->lru_lock);
2465 if (PageLRU(page)) {
2466 struct lruvec *lruvec;
2468 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2469 ClearPageLRU(page);
2470 del_page_from_lru_list(page, lruvec, page_lru(page));
2471 *isolated = 1;
2472 } else
2473 *isolated = 0;
2476 static void unlock_page_lru(struct page *page, int isolated)
2478 pg_data_t *pgdat = page_pgdat(page);
2480 if (isolated) {
2481 struct lruvec *lruvec;
2483 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2484 VM_BUG_ON_PAGE(PageLRU(page), page);
2485 SetPageLRU(page);
2486 add_page_to_lru_list(page, lruvec, page_lru(page));
2488 spin_unlock_irq(&pgdat->lru_lock);
2491 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2492 bool lrucare)
2494 int isolated;
2496 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2499 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2500 * may already be on some other mem_cgroup's LRU. Take care of it.
2502 if (lrucare)
2503 lock_page_lru(page, &isolated);
2506 * Nobody should be changing or seriously looking at
2507 * page->mem_cgroup at this point:
2509 * - the page is uncharged
2511 * - the page is off-LRU
2513 * - an anonymous fault has exclusive page access, except for
2514 * a locked page table
2516 * - a page cache insertion, a swapin fault, or a migration
2517 * have the page locked
2519 page->mem_cgroup = memcg;
2521 if (lrucare)
2522 unlock_page_lru(page, isolated);
2525 #ifdef CONFIG_MEMCG_KMEM
2526 static int memcg_alloc_cache_id(void)
2528 int id, size;
2529 int err;
2531 id = ida_simple_get(&memcg_cache_ida,
2532 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2533 if (id < 0)
2534 return id;
2536 if (id < memcg_nr_cache_ids)
2537 return id;
2540 * There's no space for the new id in memcg_caches arrays,
2541 * so we have to grow them.
2543 down_write(&memcg_cache_ids_sem);
2545 size = 2 * (id + 1);
2546 if (size < MEMCG_CACHES_MIN_SIZE)
2547 size = MEMCG_CACHES_MIN_SIZE;
2548 else if (size > MEMCG_CACHES_MAX_SIZE)
2549 size = MEMCG_CACHES_MAX_SIZE;
2551 err = memcg_update_all_caches(size);
2552 if (!err)
2553 err = memcg_update_all_list_lrus(size);
2554 if (!err)
2555 memcg_nr_cache_ids = size;
2557 up_write(&memcg_cache_ids_sem);
2559 if (err) {
2560 ida_simple_remove(&memcg_cache_ida, id);
2561 return err;
2563 return id;
2566 static void memcg_free_cache_id(int id)
2568 ida_simple_remove(&memcg_cache_ida, id);
2571 struct memcg_kmem_cache_create_work {
2572 struct mem_cgroup *memcg;
2573 struct kmem_cache *cachep;
2574 struct work_struct work;
2577 static void memcg_kmem_cache_create_func(struct work_struct *w)
2579 struct memcg_kmem_cache_create_work *cw =
2580 container_of(w, struct memcg_kmem_cache_create_work, work);
2581 struct mem_cgroup *memcg = cw->memcg;
2582 struct kmem_cache *cachep = cw->cachep;
2584 memcg_create_kmem_cache(memcg, cachep);
2586 css_put(&memcg->css);
2587 kfree(cw);
2591 * Enqueue the creation of a per-memcg kmem_cache.
2593 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2594 struct kmem_cache *cachep)
2596 struct memcg_kmem_cache_create_work *cw;
2598 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2599 if (!cw)
2600 return;
2602 css_get(&memcg->css);
2604 cw->memcg = memcg;
2605 cw->cachep = cachep;
2606 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2608 queue_work(memcg_kmem_cache_wq, &cw->work);
2611 static inline bool memcg_kmem_bypass(void)
2613 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2614 return true;
2615 return false;
2619 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2620 * @cachep: the original global kmem cache
2622 * Return the kmem_cache we're supposed to use for a slab allocation.
2623 * We try to use the current memcg's version of the cache.
2625 * If the cache does not exist yet, if we are the first user of it, we
2626 * create it asynchronously in a workqueue and let the current allocation
2627 * go through with the original cache.
2629 * This function takes a reference to the cache it returns to assure it
2630 * won't get destroyed while we are working with it. Once the caller is
2631 * done with it, memcg_kmem_put_cache() must be called to release the
2632 * reference.
2634 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2636 struct mem_cgroup *memcg;
2637 struct kmem_cache *memcg_cachep;
2638 int kmemcg_id;
2640 VM_BUG_ON(!is_root_cache(cachep));
2642 if (memcg_kmem_bypass())
2643 return cachep;
2645 memcg = get_mem_cgroup_from_current();
2646 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2647 if (kmemcg_id < 0)
2648 goto out;
2650 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2651 if (likely(memcg_cachep))
2652 return memcg_cachep;
2655 * If we are in a safe context (can wait, and not in interrupt
2656 * context), we could be be predictable and return right away.
2657 * This would guarantee that the allocation being performed
2658 * already belongs in the new cache.
2660 * However, there are some clashes that can arrive from locking.
2661 * For instance, because we acquire the slab_mutex while doing
2662 * memcg_create_kmem_cache, this means no further allocation
2663 * could happen with the slab_mutex held. So it's better to
2664 * defer everything.
2666 memcg_schedule_kmem_cache_create(memcg, cachep);
2667 out:
2668 css_put(&memcg->css);
2669 return cachep;
2673 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2674 * @cachep: the cache returned by memcg_kmem_get_cache
2676 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2678 if (!is_root_cache(cachep))
2679 css_put(&cachep->memcg_params.memcg->css);
2683 * __memcg_kmem_charge_memcg: charge a kmem page
2684 * @page: page to charge
2685 * @gfp: reclaim mode
2686 * @order: allocation order
2687 * @memcg: memory cgroup to charge
2689 * Returns 0 on success, an error code on failure.
2691 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2692 struct mem_cgroup *memcg)
2694 unsigned int nr_pages = 1 << order;
2695 struct page_counter *counter;
2696 int ret;
2698 ret = try_charge(memcg, gfp, nr_pages);
2699 if (ret)
2700 return ret;
2702 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2703 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2704 cancel_charge(memcg, nr_pages);
2705 return -ENOMEM;
2708 page->mem_cgroup = memcg;
2710 return 0;
2714 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2715 * @page: page to charge
2716 * @gfp: reclaim mode
2717 * @order: allocation order
2719 * Returns 0 on success, an error code on failure.
2721 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2723 struct mem_cgroup *memcg;
2724 int ret = 0;
2726 if (memcg_kmem_bypass())
2727 return 0;
2729 memcg = get_mem_cgroup_from_current();
2730 if (!mem_cgroup_is_root(memcg)) {
2731 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2732 if (!ret)
2733 __SetPageKmemcg(page);
2735 css_put(&memcg->css);
2736 return ret;
2739 * __memcg_kmem_uncharge: uncharge a kmem page
2740 * @page: page to uncharge
2741 * @order: allocation order
2743 void __memcg_kmem_uncharge(struct page *page, int order)
2745 struct mem_cgroup *memcg = page->mem_cgroup;
2746 unsigned int nr_pages = 1 << order;
2748 if (!memcg)
2749 return;
2751 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2753 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2754 page_counter_uncharge(&memcg->kmem, nr_pages);
2756 page_counter_uncharge(&memcg->memory, nr_pages);
2757 if (do_memsw_account())
2758 page_counter_uncharge(&memcg->memsw, nr_pages);
2760 page->mem_cgroup = NULL;
2762 /* slab pages do not have PageKmemcg flag set */
2763 if (PageKmemcg(page))
2764 __ClearPageKmemcg(page);
2766 css_put_many(&memcg->css, nr_pages);
2768 #endif /* CONFIG_MEMCG_KMEM */
2770 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2773 * Because tail pages are not marked as "used", set it. We're under
2774 * pgdat->lru_lock and migration entries setup in all page mappings.
2776 void mem_cgroup_split_huge_fixup(struct page *head)
2778 int i;
2780 if (mem_cgroup_disabled())
2781 return;
2783 for (i = 1; i < HPAGE_PMD_NR; i++)
2784 head[i].mem_cgroup = head->mem_cgroup;
2786 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2788 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2790 #ifdef CONFIG_MEMCG_SWAP
2792 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2793 * @entry: swap entry to be moved
2794 * @from: mem_cgroup which the entry is moved from
2795 * @to: mem_cgroup which the entry is moved to
2797 * It succeeds only when the swap_cgroup's record for this entry is the same
2798 * as the mem_cgroup's id of @from.
2800 * Returns 0 on success, -EINVAL on failure.
2802 * The caller must have charged to @to, IOW, called page_counter_charge() about
2803 * both res and memsw, and called css_get().
2805 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2806 struct mem_cgroup *from, struct mem_cgroup *to)
2808 unsigned short old_id, new_id;
2810 old_id = mem_cgroup_id(from);
2811 new_id = mem_cgroup_id(to);
2813 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2814 mod_memcg_state(from, MEMCG_SWAP, -1);
2815 mod_memcg_state(to, MEMCG_SWAP, 1);
2816 return 0;
2818 return -EINVAL;
2820 #else
2821 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2822 struct mem_cgroup *from, struct mem_cgroup *to)
2824 return -EINVAL;
2826 #endif
2828 static DEFINE_MUTEX(memcg_max_mutex);
2830 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
2831 unsigned long max, bool memsw)
2833 bool enlarge = false;
2834 bool drained = false;
2835 int ret;
2836 bool limits_invariant;
2837 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
2839 do {
2840 if (signal_pending(current)) {
2841 ret = -EINTR;
2842 break;
2845 mutex_lock(&memcg_max_mutex);
2847 * Make sure that the new limit (memsw or memory limit) doesn't
2848 * break our basic invariant rule memory.max <= memsw.max.
2850 limits_invariant = memsw ? max >= memcg->memory.max :
2851 max <= memcg->memsw.max;
2852 if (!limits_invariant) {
2853 mutex_unlock(&memcg_max_mutex);
2854 ret = -EINVAL;
2855 break;
2857 if (max > counter->max)
2858 enlarge = true;
2859 ret = page_counter_set_max(counter, max);
2860 mutex_unlock(&memcg_max_mutex);
2862 if (!ret)
2863 break;
2865 if (!drained) {
2866 drain_all_stock(memcg);
2867 drained = true;
2868 continue;
2871 if (!try_to_free_mem_cgroup_pages(memcg, 1,
2872 GFP_KERNEL, !memsw)) {
2873 ret = -EBUSY;
2874 break;
2876 } while (true);
2878 if (!ret && enlarge)
2879 memcg_oom_recover(memcg);
2881 return ret;
2884 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2885 gfp_t gfp_mask,
2886 unsigned long *total_scanned)
2888 unsigned long nr_reclaimed = 0;
2889 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2890 unsigned long reclaimed;
2891 int loop = 0;
2892 struct mem_cgroup_tree_per_node *mctz;
2893 unsigned long excess;
2894 unsigned long nr_scanned;
2896 if (order > 0)
2897 return 0;
2899 mctz = soft_limit_tree_node(pgdat->node_id);
2902 * Do not even bother to check the largest node if the root
2903 * is empty. Do it lockless to prevent lock bouncing. Races
2904 * are acceptable as soft limit is best effort anyway.
2906 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
2907 return 0;
2910 * This loop can run a while, specially if mem_cgroup's continuously
2911 * keep exceeding their soft limit and putting the system under
2912 * pressure
2914 do {
2915 if (next_mz)
2916 mz = next_mz;
2917 else
2918 mz = mem_cgroup_largest_soft_limit_node(mctz);
2919 if (!mz)
2920 break;
2922 nr_scanned = 0;
2923 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2924 gfp_mask, &nr_scanned);
2925 nr_reclaimed += reclaimed;
2926 *total_scanned += nr_scanned;
2927 spin_lock_irq(&mctz->lock);
2928 __mem_cgroup_remove_exceeded(mz, mctz);
2931 * If we failed to reclaim anything from this memory cgroup
2932 * it is time to move on to the next cgroup
2934 next_mz = NULL;
2935 if (!reclaimed)
2936 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2938 excess = soft_limit_excess(mz->memcg);
2940 * One school of thought says that we should not add
2941 * back the node to the tree if reclaim returns 0.
2942 * But our reclaim could return 0, simply because due
2943 * to priority we are exposing a smaller subset of
2944 * memory to reclaim from. Consider this as a longer
2945 * term TODO.
2947 /* If excess == 0, no tree ops */
2948 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2949 spin_unlock_irq(&mctz->lock);
2950 css_put(&mz->memcg->css);
2951 loop++;
2953 * Could not reclaim anything and there are no more
2954 * mem cgroups to try or we seem to be looping without
2955 * reclaiming anything.
2957 if (!nr_reclaimed &&
2958 (next_mz == NULL ||
2959 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2960 break;
2961 } while (!nr_reclaimed);
2962 if (next_mz)
2963 css_put(&next_mz->memcg->css);
2964 return nr_reclaimed;
2968 * Test whether @memcg has children, dead or alive. Note that this
2969 * function doesn't care whether @memcg has use_hierarchy enabled and
2970 * returns %true if there are child csses according to the cgroup
2971 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2973 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2975 bool ret;
2977 rcu_read_lock();
2978 ret = css_next_child(NULL, &memcg->css);
2979 rcu_read_unlock();
2980 return ret;
2984 * Reclaims as many pages from the given memcg as possible.
2986 * Caller is responsible for holding css reference for memcg.
2988 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2990 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2992 /* we call try-to-free pages for make this cgroup empty */
2993 lru_add_drain_all();
2995 drain_all_stock(memcg);
2997 /* try to free all pages in this cgroup */
2998 while (nr_retries && page_counter_read(&memcg->memory)) {
2999 int progress;
3001 if (signal_pending(current))
3002 return -EINTR;
3004 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3005 GFP_KERNEL, true);
3006 if (!progress) {
3007 nr_retries--;
3008 /* maybe some writeback is necessary */
3009 congestion_wait(BLK_RW_ASYNC, HZ/10);
3014 return 0;
3017 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3018 char *buf, size_t nbytes,
3019 loff_t off)
3021 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3023 if (mem_cgroup_is_root(memcg))
3024 return -EINVAL;
3025 return mem_cgroup_force_empty(memcg) ?: nbytes;
3028 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3029 struct cftype *cft)
3031 return mem_cgroup_from_css(css)->use_hierarchy;
3034 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3035 struct cftype *cft, u64 val)
3037 int retval = 0;
3038 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3039 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3041 if (memcg->use_hierarchy == val)
3042 return 0;
3045 * If parent's use_hierarchy is set, we can't make any modifications
3046 * in the child subtrees. If it is unset, then the change can
3047 * occur, provided the current cgroup has no children.
3049 * For the root cgroup, parent_mem is NULL, we allow value to be
3050 * set if there are no children.
3052 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3053 (val == 1 || val == 0)) {
3054 if (!memcg_has_children(memcg))
3055 memcg->use_hierarchy = val;
3056 else
3057 retval = -EBUSY;
3058 } else
3059 retval = -EINVAL;
3061 return retval;
3064 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3066 unsigned long val;
3068 if (mem_cgroup_is_root(memcg)) {
3069 val = memcg_page_state(memcg, MEMCG_CACHE) +
3070 memcg_page_state(memcg, MEMCG_RSS);
3071 if (swap)
3072 val += memcg_page_state(memcg, MEMCG_SWAP);
3073 } else {
3074 if (!swap)
3075 val = page_counter_read(&memcg->memory);
3076 else
3077 val = page_counter_read(&memcg->memsw);
3079 return val;
3082 enum {
3083 RES_USAGE,
3084 RES_LIMIT,
3085 RES_MAX_USAGE,
3086 RES_FAILCNT,
3087 RES_SOFT_LIMIT,
3090 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3091 struct cftype *cft)
3093 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3094 struct page_counter *counter;
3096 switch (MEMFILE_TYPE(cft->private)) {
3097 case _MEM:
3098 counter = &memcg->memory;
3099 break;
3100 case _MEMSWAP:
3101 counter = &memcg->memsw;
3102 break;
3103 case _KMEM:
3104 counter = &memcg->kmem;
3105 break;
3106 case _TCP:
3107 counter = &memcg->tcpmem;
3108 break;
3109 default:
3110 BUG();
3113 switch (MEMFILE_ATTR(cft->private)) {
3114 case RES_USAGE:
3115 if (counter == &memcg->memory)
3116 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3117 if (counter == &memcg->memsw)
3118 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3119 return (u64)page_counter_read(counter) * PAGE_SIZE;
3120 case RES_LIMIT:
3121 return (u64)counter->max * PAGE_SIZE;
3122 case RES_MAX_USAGE:
3123 return (u64)counter->watermark * PAGE_SIZE;
3124 case RES_FAILCNT:
3125 return counter->failcnt;
3126 case RES_SOFT_LIMIT:
3127 return (u64)memcg->soft_limit * PAGE_SIZE;
3128 default:
3129 BUG();
3133 #ifdef CONFIG_MEMCG_KMEM
3134 static int memcg_online_kmem(struct mem_cgroup *memcg)
3136 int memcg_id;
3138 if (cgroup_memory_nokmem)
3139 return 0;
3141 BUG_ON(memcg->kmemcg_id >= 0);
3142 BUG_ON(memcg->kmem_state);
3144 memcg_id = memcg_alloc_cache_id();
3145 if (memcg_id < 0)
3146 return memcg_id;
3148 static_branch_inc(&memcg_kmem_enabled_key);
3150 * A memory cgroup is considered kmem-online as soon as it gets
3151 * kmemcg_id. Setting the id after enabling static branching will
3152 * guarantee no one starts accounting before all call sites are
3153 * patched.
3155 memcg->kmemcg_id = memcg_id;
3156 memcg->kmem_state = KMEM_ONLINE;
3157 INIT_LIST_HEAD(&memcg->kmem_caches);
3159 return 0;
3162 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3164 struct cgroup_subsys_state *css;
3165 struct mem_cgroup *parent, *child;
3166 int kmemcg_id;
3168 if (memcg->kmem_state != KMEM_ONLINE)
3169 return;
3171 * Clear the online state before clearing memcg_caches array
3172 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3173 * guarantees that no cache will be created for this cgroup
3174 * after we are done (see memcg_create_kmem_cache()).
3176 memcg->kmem_state = KMEM_ALLOCATED;
3178 memcg_deactivate_kmem_caches(memcg);
3180 kmemcg_id = memcg->kmemcg_id;
3181 BUG_ON(kmemcg_id < 0);
3183 parent = parent_mem_cgroup(memcg);
3184 if (!parent)
3185 parent = root_mem_cgroup;
3188 * Change kmemcg_id of this cgroup and all its descendants to the
3189 * parent's id, and then move all entries from this cgroup's list_lrus
3190 * to ones of the parent. After we have finished, all list_lrus
3191 * corresponding to this cgroup are guaranteed to remain empty. The
3192 * ordering is imposed by list_lru_node->lock taken by
3193 * memcg_drain_all_list_lrus().
3195 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3196 css_for_each_descendant_pre(css, &memcg->css) {
3197 child = mem_cgroup_from_css(css);
3198 BUG_ON(child->kmemcg_id != kmemcg_id);
3199 child->kmemcg_id = parent->kmemcg_id;
3200 if (!memcg->use_hierarchy)
3201 break;
3203 rcu_read_unlock();
3205 memcg_drain_all_list_lrus(kmemcg_id, parent);
3207 memcg_free_cache_id(kmemcg_id);
3210 static void memcg_free_kmem(struct mem_cgroup *memcg)
3212 /* css_alloc() failed, offlining didn't happen */
3213 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3214 memcg_offline_kmem(memcg);
3216 if (memcg->kmem_state == KMEM_ALLOCATED) {
3217 memcg_destroy_kmem_caches(memcg);
3218 static_branch_dec(&memcg_kmem_enabled_key);
3219 WARN_ON(page_counter_read(&memcg->kmem));
3222 #else
3223 static int memcg_online_kmem(struct mem_cgroup *memcg)
3225 return 0;
3227 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3230 static void memcg_free_kmem(struct mem_cgroup *memcg)
3233 #endif /* CONFIG_MEMCG_KMEM */
3235 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3236 unsigned long max)
3238 int ret;
3240 mutex_lock(&memcg_max_mutex);
3241 ret = page_counter_set_max(&memcg->kmem, max);
3242 mutex_unlock(&memcg_max_mutex);
3243 return ret;
3246 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3248 int ret;
3250 mutex_lock(&memcg_max_mutex);
3252 ret = page_counter_set_max(&memcg->tcpmem, max);
3253 if (ret)
3254 goto out;
3256 if (!memcg->tcpmem_active) {
3258 * The active flag needs to be written after the static_key
3259 * update. This is what guarantees that the socket activation
3260 * function is the last one to run. See mem_cgroup_sk_alloc()
3261 * for details, and note that we don't mark any socket as
3262 * belonging to this memcg until that flag is up.
3264 * We need to do this, because static_keys will span multiple
3265 * sites, but we can't control their order. If we mark a socket
3266 * as accounted, but the accounting functions are not patched in
3267 * yet, we'll lose accounting.
3269 * We never race with the readers in mem_cgroup_sk_alloc(),
3270 * because when this value change, the code to process it is not
3271 * patched in yet.
3273 static_branch_inc(&memcg_sockets_enabled_key);
3274 memcg->tcpmem_active = true;
3276 out:
3277 mutex_unlock(&memcg_max_mutex);
3278 return ret;
3282 * The user of this function is...
3283 * RES_LIMIT.
3285 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3286 char *buf, size_t nbytes, loff_t off)
3288 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3289 unsigned long nr_pages;
3290 int ret;
3292 buf = strstrip(buf);
3293 ret = page_counter_memparse(buf, "-1", &nr_pages);
3294 if (ret)
3295 return ret;
3297 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3298 case RES_LIMIT:
3299 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3300 ret = -EINVAL;
3301 break;
3303 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3304 case _MEM:
3305 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3306 break;
3307 case _MEMSWAP:
3308 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3309 break;
3310 case _KMEM:
3311 ret = memcg_update_kmem_max(memcg, nr_pages);
3312 break;
3313 case _TCP:
3314 ret = memcg_update_tcp_max(memcg, nr_pages);
3315 break;
3317 break;
3318 case RES_SOFT_LIMIT:
3319 memcg->soft_limit = nr_pages;
3320 ret = 0;
3321 break;
3323 return ret ?: nbytes;
3326 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3327 size_t nbytes, loff_t off)
3329 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3330 struct page_counter *counter;
3332 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3333 case _MEM:
3334 counter = &memcg->memory;
3335 break;
3336 case _MEMSWAP:
3337 counter = &memcg->memsw;
3338 break;
3339 case _KMEM:
3340 counter = &memcg->kmem;
3341 break;
3342 case _TCP:
3343 counter = &memcg->tcpmem;
3344 break;
3345 default:
3346 BUG();
3349 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3350 case RES_MAX_USAGE:
3351 page_counter_reset_watermark(counter);
3352 break;
3353 case RES_FAILCNT:
3354 counter->failcnt = 0;
3355 break;
3356 default:
3357 BUG();
3360 return nbytes;
3363 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3364 struct cftype *cft)
3366 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3369 #ifdef CONFIG_MMU
3370 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3371 struct cftype *cft, u64 val)
3373 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3375 if (val & ~MOVE_MASK)
3376 return -EINVAL;
3379 * No kind of locking is needed in here, because ->can_attach() will
3380 * check this value once in the beginning of the process, and then carry
3381 * on with stale data. This means that changes to this value will only
3382 * affect task migrations starting after the change.
3384 memcg->move_charge_at_immigrate = val;
3385 return 0;
3387 #else
3388 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3389 struct cftype *cft, u64 val)
3391 return -ENOSYS;
3393 #endif
3395 #ifdef CONFIG_NUMA
3397 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3398 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3399 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3401 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3402 int nid, unsigned int lru_mask)
3404 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
3405 unsigned long nr = 0;
3406 enum lru_list lru;
3408 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3410 for_each_lru(lru) {
3411 if (!(BIT(lru) & lru_mask))
3412 continue;
3413 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3415 return nr;
3418 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3419 unsigned int lru_mask)
3421 unsigned long nr = 0;
3422 enum lru_list lru;
3424 for_each_lru(lru) {
3425 if (!(BIT(lru) & lru_mask))
3426 continue;
3427 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3429 return nr;
3432 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3434 struct numa_stat {
3435 const char *name;
3436 unsigned int lru_mask;
3439 static const struct numa_stat stats[] = {
3440 { "total", LRU_ALL },
3441 { "file", LRU_ALL_FILE },
3442 { "anon", LRU_ALL_ANON },
3443 { "unevictable", BIT(LRU_UNEVICTABLE) },
3445 const struct numa_stat *stat;
3446 int nid;
3447 unsigned long nr;
3448 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3450 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3451 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3452 seq_printf(m, "%s=%lu", stat->name, nr);
3453 for_each_node_state(nid, N_MEMORY) {
3454 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3455 stat->lru_mask);
3456 seq_printf(m, " N%d=%lu", nid, nr);
3458 seq_putc(m, '\n');
3461 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3462 struct mem_cgroup *iter;
3464 nr = 0;
3465 for_each_mem_cgroup_tree(iter, memcg)
3466 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3467 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3468 for_each_node_state(nid, N_MEMORY) {
3469 nr = 0;
3470 for_each_mem_cgroup_tree(iter, memcg)
3471 nr += mem_cgroup_node_nr_lru_pages(
3472 iter, nid, stat->lru_mask);
3473 seq_printf(m, " N%d=%lu", nid, nr);
3475 seq_putc(m, '\n');
3478 return 0;
3480 #endif /* CONFIG_NUMA */
3482 /* Universal VM events cgroup1 shows, original sort order */
3483 static const unsigned int memcg1_events[] = {
3484 PGPGIN,
3485 PGPGOUT,
3486 PGFAULT,
3487 PGMAJFAULT,
3490 static const char *const memcg1_event_names[] = {
3491 "pgpgin",
3492 "pgpgout",
3493 "pgfault",
3494 "pgmajfault",
3497 static int memcg_stat_show(struct seq_file *m, void *v)
3499 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3500 unsigned long memory, memsw;
3501 struct mem_cgroup *mi;
3502 unsigned int i;
3504 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3505 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3507 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3508 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3509 continue;
3510 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3511 memcg_page_state_local(memcg, memcg1_stats[i]) *
3512 PAGE_SIZE);
3515 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3516 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3517 memcg_events_local(memcg, memcg1_events[i]));
3519 for (i = 0; i < NR_LRU_LISTS; i++)
3520 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3521 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3522 PAGE_SIZE);
3524 /* Hierarchical information */
3525 memory = memsw = PAGE_COUNTER_MAX;
3526 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3527 memory = min(memory, mi->memory.max);
3528 memsw = min(memsw, mi->memsw.max);
3530 seq_printf(m, "hierarchical_memory_limit %llu\n",
3531 (u64)memory * PAGE_SIZE);
3532 if (do_memsw_account())
3533 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3534 (u64)memsw * PAGE_SIZE);
3536 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3537 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3538 continue;
3539 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3540 (u64)memcg_page_state(memcg, i) * PAGE_SIZE);
3543 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3544 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3545 (u64)memcg_events(memcg, i));
3547 for (i = 0; i < NR_LRU_LISTS; i++)
3548 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3549 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
3550 PAGE_SIZE);
3552 #ifdef CONFIG_DEBUG_VM
3554 pg_data_t *pgdat;
3555 struct mem_cgroup_per_node *mz;
3556 struct zone_reclaim_stat *rstat;
3557 unsigned long recent_rotated[2] = {0, 0};
3558 unsigned long recent_scanned[2] = {0, 0};
3560 for_each_online_pgdat(pgdat) {
3561 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3562 rstat = &mz->lruvec.reclaim_stat;
3564 recent_rotated[0] += rstat->recent_rotated[0];
3565 recent_rotated[1] += rstat->recent_rotated[1];
3566 recent_scanned[0] += rstat->recent_scanned[0];
3567 recent_scanned[1] += rstat->recent_scanned[1];
3569 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3570 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3571 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3572 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3574 #endif
3576 return 0;
3579 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3580 struct cftype *cft)
3582 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3584 return mem_cgroup_swappiness(memcg);
3587 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3588 struct cftype *cft, u64 val)
3590 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3592 if (val > 100)
3593 return -EINVAL;
3595 if (css->parent)
3596 memcg->swappiness = val;
3597 else
3598 vm_swappiness = val;
3600 return 0;
3603 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3605 struct mem_cgroup_threshold_ary *t;
3606 unsigned long usage;
3607 int i;
3609 rcu_read_lock();
3610 if (!swap)
3611 t = rcu_dereference(memcg->thresholds.primary);
3612 else
3613 t = rcu_dereference(memcg->memsw_thresholds.primary);
3615 if (!t)
3616 goto unlock;
3618 usage = mem_cgroup_usage(memcg, swap);
3621 * current_threshold points to threshold just below or equal to usage.
3622 * If it's not true, a threshold was crossed after last
3623 * call of __mem_cgroup_threshold().
3625 i = t->current_threshold;
3628 * Iterate backward over array of thresholds starting from
3629 * current_threshold and check if a threshold is crossed.
3630 * If none of thresholds below usage is crossed, we read
3631 * only one element of the array here.
3633 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3634 eventfd_signal(t->entries[i].eventfd, 1);
3636 /* i = current_threshold + 1 */
3637 i++;
3640 * Iterate forward over array of thresholds starting from
3641 * current_threshold+1 and check if a threshold is crossed.
3642 * If none of thresholds above usage is crossed, we read
3643 * only one element of the array here.
3645 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3646 eventfd_signal(t->entries[i].eventfd, 1);
3648 /* Update current_threshold */
3649 t->current_threshold = i - 1;
3650 unlock:
3651 rcu_read_unlock();
3654 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3656 while (memcg) {
3657 __mem_cgroup_threshold(memcg, false);
3658 if (do_memsw_account())
3659 __mem_cgroup_threshold(memcg, true);
3661 memcg = parent_mem_cgroup(memcg);
3665 static int compare_thresholds(const void *a, const void *b)
3667 const struct mem_cgroup_threshold *_a = a;
3668 const struct mem_cgroup_threshold *_b = b;
3670 if (_a->threshold > _b->threshold)
3671 return 1;
3673 if (_a->threshold < _b->threshold)
3674 return -1;
3676 return 0;
3679 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3681 struct mem_cgroup_eventfd_list *ev;
3683 spin_lock(&memcg_oom_lock);
3685 list_for_each_entry(ev, &memcg->oom_notify, list)
3686 eventfd_signal(ev->eventfd, 1);
3688 spin_unlock(&memcg_oom_lock);
3689 return 0;
3692 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3694 struct mem_cgroup *iter;
3696 for_each_mem_cgroup_tree(iter, memcg)
3697 mem_cgroup_oom_notify_cb(iter);
3700 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3701 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3703 struct mem_cgroup_thresholds *thresholds;
3704 struct mem_cgroup_threshold_ary *new;
3705 unsigned long threshold;
3706 unsigned long usage;
3707 int i, size, ret;
3709 ret = page_counter_memparse(args, "-1", &threshold);
3710 if (ret)
3711 return ret;
3713 mutex_lock(&memcg->thresholds_lock);
3715 if (type == _MEM) {
3716 thresholds = &memcg->thresholds;
3717 usage = mem_cgroup_usage(memcg, false);
3718 } else if (type == _MEMSWAP) {
3719 thresholds = &memcg->memsw_thresholds;
3720 usage = mem_cgroup_usage(memcg, true);
3721 } else
3722 BUG();
3724 /* Check if a threshold crossed before adding a new one */
3725 if (thresholds->primary)
3726 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3728 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3730 /* Allocate memory for new array of thresholds */
3731 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
3732 if (!new) {
3733 ret = -ENOMEM;
3734 goto unlock;
3736 new->size = size;
3738 /* Copy thresholds (if any) to new array */
3739 if (thresholds->primary) {
3740 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3741 sizeof(struct mem_cgroup_threshold));
3744 /* Add new threshold */
3745 new->entries[size - 1].eventfd = eventfd;
3746 new->entries[size - 1].threshold = threshold;
3748 /* Sort thresholds. Registering of new threshold isn't time-critical */
3749 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3750 compare_thresholds, NULL);
3752 /* Find current threshold */
3753 new->current_threshold = -1;
3754 for (i = 0; i < size; i++) {
3755 if (new->entries[i].threshold <= usage) {
3757 * new->current_threshold will not be used until
3758 * rcu_assign_pointer(), so it's safe to increment
3759 * it here.
3761 ++new->current_threshold;
3762 } else
3763 break;
3766 /* Free old spare buffer and save old primary buffer as spare */
3767 kfree(thresholds->spare);
3768 thresholds->spare = thresholds->primary;
3770 rcu_assign_pointer(thresholds->primary, new);
3772 /* To be sure that nobody uses thresholds */
3773 synchronize_rcu();
3775 unlock:
3776 mutex_unlock(&memcg->thresholds_lock);
3778 return ret;
3781 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3782 struct eventfd_ctx *eventfd, const char *args)
3784 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3787 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3788 struct eventfd_ctx *eventfd, const char *args)
3790 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3793 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3794 struct eventfd_ctx *eventfd, enum res_type type)
3796 struct mem_cgroup_thresholds *thresholds;
3797 struct mem_cgroup_threshold_ary *new;
3798 unsigned long usage;
3799 int i, j, size;
3801 mutex_lock(&memcg->thresholds_lock);
3803 if (type == _MEM) {
3804 thresholds = &memcg->thresholds;
3805 usage = mem_cgroup_usage(memcg, false);
3806 } else if (type == _MEMSWAP) {
3807 thresholds = &memcg->memsw_thresholds;
3808 usage = mem_cgroup_usage(memcg, true);
3809 } else
3810 BUG();
3812 if (!thresholds->primary)
3813 goto unlock;
3815 /* Check if a threshold crossed before removing */
3816 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3818 /* Calculate new number of threshold */
3819 size = 0;
3820 for (i = 0; i < thresholds->primary->size; i++) {
3821 if (thresholds->primary->entries[i].eventfd != eventfd)
3822 size++;
3825 new = thresholds->spare;
3827 /* Set thresholds array to NULL if we don't have thresholds */
3828 if (!size) {
3829 kfree(new);
3830 new = NULL;
3831 goto swap_buffers;
3834 new->size = size;
3836 /* Copy thresholds and find current threshold */
3837 new->current_threshold = -1;
3838 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3839 if (thresholds->primary->entries[i].eventfd == eventfd)
3840 continue;
3842 new->entries[j] = thresholds->primary->entries[i];
3843 if (new->entries[j].threshold <= usage) {
3845 * new->current_threshold will not be used
3846 * until rcu_assign_pointer(), so it's safe to increment
3847 * it here.
3849 ++new->current_threshold;
3851 j++;
3854 swap_buffers:
3855 /* Swap primary and spare array */
3856 thresholds->spare = thresholds->primary;
3858 rcu_assign_pointer(thresholds->primary, new);
3860 /* To be sure that nobody uses thresholds */
3861 synchronize_rcu();
3863 /* If all events are unregistered, free the spare array */
3864 if (!new) {
3865 kfree(thresholds->spare);
3866 thresholds->spare = NULL;
3868 unlock:
3869 mutex_unlock(&memcg->thresholds_lock);
3872 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3873 struct eventfd_ctx *eventfd)
3875 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3878 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3879 struct eventfd_ctx *eventfd)
3881 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3884 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3885 struct eventfd_ctx *eventfd, const char *args)
3887 struct mem_cgroup_eventfd_list *event;
3889 event = kmalloc(sizeof(*event), GFP_KERNEL);
3890 if (!event)
3891 return -ENOMEM;
3893 spin_lock(&memcg_oom_lock);
3895 event->eventfd = eventfd;
3896 list_add(&event->list, &memcg->oom_notify);
3898 /* already in OOM ? */
3899 if (memcg->under_oom)
3900 eventfd_signal(eventfd, 1);
3901 spin_unlock(&memcg_oom_lock);
3903 return 0;
3906 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3907 struct eventfd_ctx *eventfd)
3909 struct mem_cgroup_eventfd_list *ev, *tmp;
3911 spin_lock(&memcg_oom_lock);
3913 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3914 if (ev->eventfd == eventfd) {
3915 list_del(&ev->list);
3916 kfree(ev);
3920 spin_unlock(&memcg_oom_lock);
3923 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3925 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
3927 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3928 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3929 seq_printf(sf, "oom_kill %lu\n",
3930 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
3931 return 0;
3934 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3935 struct cftype *cft, u64 val)
3937 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3939 /* cannot set to root cgroup and only 0 and 1 are allowed */
3940 if (!css->parent || !((val == 0) || (val == 1)))
3941 return -EINVAL;
3943 memcg->oom_kill_disable = val;
3944 if (!val)
3945 memcg_oom_recover(memcg);
3947 return 0;
3950 #ifdef CONFIG_CGROUP_WRITEBACK
3952 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3954 return wb_domain_init(&memcg->cgwb_domain, gfp);
3957 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3959 wb_domain_exit(&memcg->cgwb_domain);
3962 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3964 wb_domain_size_changed(&memcg->cgwb_domain);
3967 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3969 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3971 if (!memcg->css.parent)
3972 return NULL;
3974 return &memcg->cgwb_domain;
3978 * idx can be of type enum memcg_stat_item or node_stat_item.
3979 * Keep in sync with memcg_exact_page().
3981 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
3983 long x = atomic_long_read(&memcg->vmstats[idx]);
3984 int cpu;
3986 for_each_online_cpu(cpu)
3987 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
3988 if (x < 0)
3989 x = 0;
3990 return x;
3994 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3995 * @wb: bdi_writeback in question
3996 * @pfilepages: out parameter for number of file pages
3997 * @pheadroom: out parameter for number of allocatable pages according to memcg
3998 * @pdirty: out parameter for number of dirty pages
3999 * @pwriteback: out parameter for number of pages under writeback
4001 * Determine the numbers of file, headroom, dirty, and writeback pages in
4002 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4003 * is a bit more involved.
4005 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4006 * headroom is calculated as the lowest headroom of itself and the
4007 * ancestors. Note that this doesn't consider the actual amount of
4008 * available memory in the system. The caller should further cap
4009 * *@pheadroom accordingly.
4011 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4012 unsigned long *pheadroom, unsigned long *pdirty,
4013 unsigned long *pwriteback)
4015 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4016 struct mem_cgroup *parent;
4018 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4020 /* this should eventually include NR_UNSTABLE_NFS */
4021 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4022 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4023 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4024 *pheadroom = PAGE_COUNTER_MAX;
4026 while ((parent = parent_mem_cgroup(memcg))) {
4027 unsigned long ceiling = min(memcg->memory.max, memcg->high);
4028 unsigned long used = page_counter_read(&memcg->memory);
4030 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4031 memcg = parent;
4035 #else /* CONFIG_CGROUP_WRITEBACK */
4037 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4039 return 0;
4042 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4046 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4050 #endif /* CONFIG_CGROUP_WRITEBACK */
4053 * DO NOT USE IN NEW FILES.
4055 * "cgroup.event_control" implementation.
4057 * This is way over-engineered. It tries to support fully configurable
4058 * events for each user. Such level of flexibility is completely
4059 * unnecessary especially in the light of the planned unified hierarchy.
4061 * Please deprecate this and replace with something simpler if at all
4062 * possible.
4066 * Unregister event and free resources.
4068 * Gets called from workqueue.
4070 static void memcg_event_remove(struct work_struct *work)
4072 struct mem_cgroup_event *event =
4073 container_of(work, struct mem_cgroup_event, remove);
4074 struct mem_cgroup *memcg = event->memcg;
4076 remove_wait_queue(event->wqh, &event->wait);
4078 event->unregister_event(memcg, event->eventfd);
4080 /* Notify userspace the event is going away. */
4081 eventfd_signal(event->eventfd, 1);
4083 eventfd_ctx_put(event->eventfd);
4084 kfree(event);
4085 css_put(&memcg->css);
4089 * Gets called on EPOLLHUP on eventfd when user closes it.
4091 * Called with wqh->lock held and interrupts disabled.
4093 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4094 int sync, void *key)
4096 struct mem_cgroup_event *event =
4097 container_of(wait, struct mem_cgroup_event, wait);
4098 struct mem_cgroup *memcg = event->memcg;
4099 __poll_t flags = key_to_poll(key);
4101 if (flags & EPOLLHUP) {
4103 * If the event has been detached at cgroup removal, we
4104 * can simply return knowing the other side will cleanup
4105 * for us.
4107 * We can't race against event freeing since the other
4108 * side will require wqh->lock via remove_wait_queue(),
4109 * which we hold.
4111 spin_lock(&memcg->event_list_lock);
4112 if (!list_empty(&event->list)) {
4113 list_del_init(&event->list);
4115 * We are in atomic context, but cgroup_event_remove()
4116 * may sleep, so we have to call it in workqueue.
4118 schedule_work(&event->remove);
4120 spin_unlock(&memcg->event_list_lock);
4123 return 0;
4126 static void memcg_event_ptable_queue_proc(struct file *file,
4127 wait_queue_head_t *wqh, poll_table *pt)
4129 struct mem_cgroup_event *event =
4130 container_of(pt, struct mem_cgroup_event, pt);
4132 event->wqh = wqh;
4133 add_wait_queue(wqh, &event->wait);
4137 * DO NOT USE IN NEW FILES.
4139 * Parse input and register new cgroup event handler.
4141 * Input must be in format '<event_fd> <control_fd> <args>'.
4142 * Interpretation of args is defined by control file implementation.
4144 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4145 char *buf, size_t nbytes, loff_t off)
4147 struct cgroup_subsys_state *css = of_css(of);
4148 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4149 struct mem_cgroup_event *event;
4150 struct cgroup_subsys_state *cfile_css;
4151 unsigned int efd, cfd;
4152 struct fd efile;
4153 struct fd cfile;
4154 const char *name;
4155 char *endp;
4156 int ret;
4158 buf = strstrip(buf);
4160 efd = simple_strtoul(buf, &endp, 10);
4161 if (*endp != ' ')
4162 return -EINVAL;
4163 buf = endp + 1;
4165 cfd = simple_strtoul(buf, &endp, 10);
4166 if ((*endp != ' ') && (*endp != '\0'))
4167 return -EINVAL;
4168 buf = endp + 1;
4170 event = kzalloc(sizeof(*event), GFP_KERNEL);
4171 if (!event)
4172 return -ENOMEM;
4174 event->memcg = memcg;
4175 INIT_LIST_HEAD(&event->list);
4176 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4177 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4178 INIT_WORK(&event->remove, memcg_event_remove);
4180 efile = fdget(efd);
4181 if (!efile.file) {
4182 ret = -EBADF;
4183 goto out_kfree;
4186 event->eventfd = eventfd_ctx_fileget(efile.file);
4187 if (IS_ERR(event->eventfd)) {
4188 ret = PTR_ERR(event->eventfd);
4189 goto out_put_efile;
4192 cfile = fdget(cfd);
4193 if (!cfile.file) {
4194 ret = -EBADF;
4195 goto out_put_eventfd;
4198 /* the process need read permission on control file */
4199 /* AV: shouldn't we check that it's been opened for read instead? */
4200 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4201 if (ret < 0)
4202 goto out_put_cfile;
4205 * Determine the event callbacks and set them in @event. This used
4206 * to be done via struct cftype but cgroup core no longer knows
4207 * about these events. The following is crude but the whole thing
4208 * is for compatibility anyway.
4210 * DO NOT ADD NEW FILES.
4212 name = cfile.file->f_path.dentry->d_name.name;
4214 if (!strcmp(name, "memory.usage_in_bytes")) {
4215 event->register_event = mem_cgroup_usage_register_event;
4216 event->unregister_event = mem_cgroup_usage_unregister_event;
4217 } else if (!strcmp(name, "memory.oom_control")) {
4218 event->register_event = mem_cgroup_oom_register_event;
4219 event->unregister_event = mem_cgroup_oom_unregister_event;
4220 } else if (!strcmp(name, "memory.pressure_level")) {
4221 event->register_event = vmpressure_register_event;
4222 event->unregister_event = vmpressure_unregister_event;
4223 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4224 event->register_event = memsw_cgroup_usage_register_event;
4225 event->unregister_event = memsw_cgroup_usage_unregister_event;
4226 } else {
4227 ret = -EINVAL;
4228 goto out_put_cfile;
4232 * Verify @cfile should belong to @css. Also, remaining events are
4233 * automatically removed on cgroup destruction but the removal is
4234 * asynchronous, so take an extra ref on @css.
4236 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4237 &memory_cgrp_subsys);
4238 ret = -EINVAL;
4239 if (IS_ERR(cfile_css))
4240 goto out_put_cfile;
4241 if (cfile_css != css) {
4242 css_put(cfile_css);
4243 goto out_put_cfile;
4246 ret = event->register_event(memcg, event->eventfd, buf);
4247 if (ret)
4248 goto out_put_css;
4250 vfs_poll(efile.file, &event->pt);
4252 spin_lock(&memcg->event_list_lock);
4253 list_add(&event->list, &memcg->event_list);
4254 spin_unlock(&memcg->event_list_lock);
4256 fdput(cfile);
4257 fdput(efile);
4259 return nbytes;
4261 out_put_css:
4262 css_put(css);
4263 out_put_cfile:
4264 fdput(cfile);
4265 out_put_eventfd:
4266 eventfd_ctx_put(event->eventfd);
4267 out_put_efile:
4268 fdput(efile);
4269 out_kfree:
4270 kfree(event);
4272 return ret;
4275 static struct cftype mem_cgroup_legacy_files[] = {
4277 .name = "usage_in_bytes",
4278 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4279 .read_u64 = mem_cgroup_read_u64,
4282 .name = "max_usage_in_bytes",
4283 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4284 .write = mem_cgroup_reset,
4285 .read_u64 = mem_cgroup_read_u64,
4288 .name = "limit_in_bytes",
4289 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4290 .write = mem_cgroup_write,
4291 .read_u64 = mem_cgroup_read_u64,
4294 .name = "soft_limit_in_bytes",
4295 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4296 .write = mem_cgroup_write,
4297 .read_u64 = mem_cgroup_read_u64,
4300 .name = "failcnt",
4301 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4302 .write = mem_cgroup_reset,
4303 .read_u64 = mem_cgroup_read_u64,
4306 .name = "stat",
4307 .seq_show = memcg_stat_show,
4310 .name = "force_empty",
4311 .write = mem_cgroup_force_empty_write,
4314 .name = "use_hierarchy",
4315 .write_u64 = mem_cgroup_hierarchy_write,
4316 .read_u64 = mem_cgroup_hierarchy_read,
4319 .name = "cgroup.event_control", /* XXX: for compat */
4320 .write = memcg_write_event_control,
4321 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4324 .name = "swappiness",
4325 .read_u64 = mem_cgroup_swappiness_read,
4326 .write_u64 = mem_cgroup_swappiness_write,
4329 .name = "move_charge_at_immigrate",
4330 .read_u64 = mem_cgroup_move_charge_read,
4331 .write_u64 = mem_cgroup_move_charge_write,
4334 .name = "oom_control",
4335 .seq_show = mem_cgroup_oom_control_read,
4336 .write_u64 = mem_cgroup_oom_control_write,
4337 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4340 .name = "pressure_level",
4342 #ifdef CONFIG_NUMA
4344 .name = "numa_stat",
4345 .seq_show = memcg_numa_stat_show,
4347 #endif
4349 .name = "kmem.limit_in_bytes",
4350 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4351 .write = mem_cgroup_write,
4352 .read_u64 = mem_cgroup_read_u64,
4355 .name = "kmem.usage_in_bytes",
4356 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4357 .read_u64 = mem_cgroup_read_u64,
4360 .name = "kmem.failcnt",
4361 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4362 .write = mem_cgroup_reset,
4363 .read_u64 = mem_cgroup_read_u64,
4366 .name = "kmem.max_usage_in_bytes",
4367 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4368 .write = mem_cgroup_reset,
4369 .read_u64 = mem_cgroup_read_u64,
4371 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4373 .name = "kmem.slabinfo",
4374 .seq_start = memcg_slab_start,
4375 .seq_next = memcg_slab_next,
4376 .seq_stop = memcg_slab_stop,
4377 .seq_show = memcg_slab_show,
4379 #endif
4381 .name = "kmem.tcp.limit_in_bytes",
4382 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4383 .write = mem_cgroup_write,
4384 .read_u64 = mem_cgroup_read_u64,
4387 .name = "kmem.tcp.usage_in_bytes",
4388 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4389 .read_u64 = mem_cgroup_read_u64,
4392 .name = "kmem.tcp.failcnt",
4393 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4394 .write = mem_cgroup_reset,
4395 .read_u64 = mem_cgroup_read_u64,
4398 .name = "kmem.tcp.max_usage_in_bytes",
4399 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4400 .write = mem_cgroup_reset,
4401 .read_u64 = mem_cgroup_read_u64,
4403 { }, /* terminate */
4407 * Private memory cgroup IDR
4409 * Swap-out records and page cache shadow entries need to store memcg
4410 * references in constrained space, so we maintain an ID space that is
4411 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4412 * memory-controlled cgroups to 64k.
4414 * However, there usually are many references to the oflline CSS after
4415 * the cgroup has been destroyed, such as page cache or reclaimable
4416 * slab objects, that don't need to hang on to the ID. We want to keep
4417 * those dead CSS from occupying IDs, or we might quickly exhaust the
4418 * relatively small ID space and prevent the creation of new cgroups
4419 * even when there are much fewer than 64k cgroups - possibly none.
4421 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4422 * be freed and recycled when it's no longer needed, which is usually
4423 * when the CSS is offlined.
4425 * The only exception to that are records of swapped out tmpfs/shmem
4426 * pages that need to be attributed to live ancestors on swapin. But
4427 * those references are manageable from userspace.
4430 static DEFINE_IDR(mem_cgroup_idr);
4432 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4434 if (memcg->id.id > 0) {
4435 idr_remove(&mem_cgroup_idr, memcg->id.id);
4436 memcg->id.id = 0;
4440 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4442 refcount_add(n, &memcg->id.ref);
4445 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4447 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4448 mem_cgroup_id_remove(memcg);
4450 /* Memcg ID pins CSS */
4451 css_put(&memcg->css);
4455 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4457 mem_cgroup_id_get_many(memcg, 1);
4460 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4462 mem_cgroup_id_put_many(memcg, 1);
4466 * mem_cgroup_from_id - look up a memcg from a memcg id
4467 * @id: the memcg id to look up
4469 * Caller must hold rcu_read_lock().
4471 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4473 WARN_ON_ONCE(!rcu_read_lock_held());
4474 return idr_find(&mem_cgroup_idr, id);
4477 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4479 struct mem_cgroup_per_node *pn;
4480 int tmp = node;
4482 * This routine is called against possible nodes.
4483 * But it's BUG to call kmalloc() against offline node.
4485 * TODO: this routine can waste much memory for nodes which will
4486 * never be onlined. It's better to use memory hotplug callback
4487 * function.
4489 if (!node_state(node, N_NORMAL_MEMORY))
4490 tmp = -1;
4491 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4492 if (!pn)
4493 return 1;
4495 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4496 if (!pn->lruvec_stat_cpu) {
4497 kfree(pn);
4498 return 1;
4501 lruvec_init(&pn->lruvec);
4502 pn->usage_in_excess = 0;
4503 pn->on_tree = false;
4504 pn->memcg = memcg;
4506 memcg->nodeinfo[node] = pn;
4507 return 0;
4510 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4512 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4514 if (!pn)
4515 return;
4517 free_percpu(pn->lruvec_stat_cpu);
4518 kfree(pn);
4521 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4523 int node;
4525 for_each_node(node)
4526 free_mem_cgroup_per_node_info(memcg, node);
4527 free_percpu(memcg->vmstats_percpu);
4528 kfree(memcg);
4531 static void mem_cgroup_free(struct mem_cgroup *memcg)
4533 memcg_wb_domain_exit(memcg);
4534 __mem_cgroup_free(memcg);
4537 static struct mem_cgroup *mem_cgroup_alloc(void)
4539 struct mem_cgroup *memcg;
4540 unsigned int size;
4541 int node;
4543 size = sizeof(struct mem_cgroup);
4544 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4546 memcg = kzalloc(size, GFP_KERNEL);
4547 if (!memcg)
4548 return NULL;
4550 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4551 1, MEM_CGROUP_ID_MAX,
4552 GFP_KERNEL);
4553 if (memcg->id.id < 0)
4554 goto fail;
4556 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
4557 if (!memcg->vmstats_percpu)
4558 goto fail;
4560 for_each_node(node)
4561 if (alloc_mem_cgroup_per_node_info(memcg, node))
4562 goto fail;
4564 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4565 goto fail;
4567 INIT_WORK(&memcg->high_work, high_work_func);
4568 memcg->last_scanned_node = MAX_NUMNODES;
4569 INIT_LIST_HEAD(&memcg->oom_notify);
4570 mutex_init(&memcg->thresholds_lock);
4571 spin_lock_init(&memcg->move_lock);
4572 vmpressure_init(&memcg->vmpressure);
4573 INIT_LIST_HEAD(&memcg->event_list);
4574 spin_lock_init(&memcg->event_list_lock);
4575 memcg->socket_pressure = jiffies;
4576 #ifdef CONFIG_MEMCG_KMEM
4577 memcg->kmemcg_id = -1;
4578 #endif
4579 #ifdef CONFIG_CGROUP_WRITEBACK
4580 INIT_LIST_HEAD(&memcg->cgwb_list);
4581 #endif
4582 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4583 return memcg;
4584 fail:
4585 mem_cgroup_id_remove(memcg);
4586 __mem_cgroup_free(memcg);
4587 return NULL;
4590 static struct cgroup_subsys_state * __ref
4591 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4593 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4594 struct mem_cgroup *memcg;
4595 long error = -ENOMEM;
4597 memcg = mem_cgroup_alloc();
4598 if (!memcg)
4599 return ERR_PTR(error);
4601 memcg->high = PAGE_COUNTER_MAX;
4602 memcg->soft_limit = PAGE_COUNTER_MAX;
4603 if (parent) {
4604 memcg->swappiness = mem_cgroup_swappiness(parent);
4605 memcg->oom_kill_disable = parent->oom_kill_disable;
4607 if (parent && parent->use_hierarchy) {
4608 memcg->use_hierarchy = true;
4609 page_counter_init(&memcg->memory, &parent->memory);
4610 page_counter_init(&memcg->swap, &parent->swap);
4611 page_counter_init(&memcg->memsw, &parent->memsw);
4612 page_counter_init(&memcg->kmem, &parent->kmem);
4613 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4614 } else {
4615 page_counter_init(&memcg->memory, NULL);
4616 page_counter_init(&memcg->swap, NULL);
4617 page_counter_init(&memcg->memsw, NULL);
4618 page_counter_init(&memcg->kmem, NULL);
4619 page_counter_init(&memcg->tcpmem, NULL);
4621 * Deeper hierachy with use_hierarchy == false doesn't make
4622 * much sense so let cgroup subsystem know about this
4623 * unfortunate state in our controller.
4625 if (parent != root_mem_cgroup)
4626 memory_cgrp_subsys.broken_hierarchy = true;
4629 /* The following stuff does not apply to the root */
4630 if (!parent) {
4631 root_mem_cgroup = memcg;
4632 return &memcg->css;
4635 error = memcg_online_kmem(memcg);
4636 if (error)
4637 goto fail;
4639 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4640 static_branch_inc(&memcg_sockets_enabled_key);
4642 return &memcg->css;
4643 fail:
4644 mem_cgroup_id_remove(memcg);
4645 mem_cgroup_free(memcg);
4646 return ERR_PTR(-ENOMEM);
4649 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4651 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4654 * A memcg must be visible for memcg_expand_shrinker_maps()
4655 * by the time the maps are allocated. So, we allocate maps
4656 * here, when for_each_mem_cgroup() can't skip it.
4658 if (memcg_alloc_shrinker_maps(memcg)) {
4659 mem_cgroup_id_remove(memcg);
4660 return -ENOMEM;
4663 /* Online state pins memcg ID, memcg ID pins CSS */
4664 refcount_set(&memcg->id.ref, 1);
4665 css_get(css);
4666 return 0;
4669 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4671 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4672 struct mem_cgroup_event *event, *tmp;
4675 * Unregister events and notify userspace.
4676 * Notify userspace about cgroup removing only after rmdir of cgroup
4677 * directory to avoid race between userspace and kernelspace.
4679 spin_lock(&memcg->event_list_lock);
4680 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4681 list_del_init(&event->list);
4682 schedule_work(&event->remove);
4684 spin_unlock(&memcg->event_list_lock);
4686 page_counter_set_min(&memcg->memory, 0);
4687 page_counter_set_low(&memcg->memory, 0);
4689 memcg_offline_kmem(memcg);
4690 wb_memcg_offline(memcg);
4692 drain_all_stock(memcg);
4694 mem_cgroup_id_put(memcg);
4697 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4699 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4701 invalidate_reclaim_iterators(memcg);
4704 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4706 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4708 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4709 static_branch_dec(&memcg_sockets_enabled_key);
4711 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4712 static_branch_dec(&memcg_sockets_enabled_key);
4714 vmpressure_cleanup(&memcg->vmpressure);
4715 cancel_work_sync(&memcg->high_work);
4716 mem_cgroup_remove_from_trees(memcg);
4717 memcg_free_shrinker_maps(memcg);
4718 memcg_free_kmem(memcg);
4719 mem_cgroup_free(memcg);
4723 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4724 * @css: the target css
4726 * Reset the states of the mem_cgroup associated with @css. This is
4727 * invoked when the userland requests disabling on the default hierarchy
4728 * but the memcg is pinned through dependency. The memcg should stop
4729 * applying policies and should revert to the vanilla state as it may be
4730 * made visible again.
4732 * The current implementation only resets the essential configurations.
4733 * This needs to be expanded to cover all the visible parts.
4735 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4737 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4739 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
4740 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
4741 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
4742 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
4743 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
4744 page_counter_set_min(&memcg->memory, 0);
4745 page_counter_set_low(&memcg->memory, 0);
4746 memcg->high = PAGE_COUNTER_MAX;
4747 memcg->soft_limit = PAGE_COUNTER_MAX;
4748 memcg_wb_domain_size_changed(memcg);
4751 #ifdef CONFIG_MMU
4752 /* Handlers for move charge at task migration. */
4753 static int mem_cgroup_do_precharge(unsigned long count)
4755 int ret;
4757 /* Try a single bulk charge without reclaim first, kswapd may wake */
4758 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4759 if (!ret) {
4760 mc.precharge += count;
4761 return ret;
4764 /* Try charges one by one with reclaim, but do not retry */
4765 while (count--) {
4766 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4767 if (ret)
4768 return ret;
4769 mc.precharge++;
4770 cond_resched();
4772 return 0;
4775 union mc_target {
4776 struct page *page;
4777 swp_entry_t ent;
4780 enum mc_target_type {
4781 MC_TARGET_NONE = 0,
4782 MC_TARGET_PAGE,
4783 MC_TARGET_SWAP,
4784 MC_TARGET_DEVICE,
4787 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4788 unsigned long addr, pte_t ptent)
4790 struct page *page = _vm_normal_page(vma, addr, ptent, true);
4792 if (!page || !page_mapped(page))
4793 return NULL;
4794 if (PageAnon(page)) {
4795 if (!(mc.flags & MOVE_ANON))
4796 return NULL;
4797 } else {
4798 if (!(mc.flags & MOVE_FILE))
4799 return NULL;
4801 if (!get_page_unless_zero(page))
4802 return NULL;
4804 return page;
4807 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
4808 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4809 pte_t ptent, swp_entry_t *entry)
4811 struct page *page = NULL;
4812 swp_entry_t ent = pte_to_swp_entry(ptent);
4814 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4815 return NULL;
4818 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
4819 * a device and because they are not accessible by CPU they are store
4820 * as special swap entry in the CPU page table.
4822 if (is_device_private_entry(ent)) {
4823 page = device_private_entry_to_page(ent);
4825 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
4826 * a refcount of 1 when free (unlike normal page)
4828 if (!page_ref_add_unless(page, 1, 1))
4829 return NULL;
4830 return page;
4834 * Because lookup_swap_cache() updates some statistics counter,
4835 * we call find_get_page() with swapper_space directly.
4837 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4838 if (do_memsw_account())
4839 entry->val = ent.val;
4841 return page;
4843 #else
4844 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4845 pte_t ptent, swp_entry_t *entry)
4847 return NULL;
4849 #endif
4851 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4852 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4854 struct page *page = NULL;
4855 struct address_space *mapping;
4856 pgoff_t pgoff;
4858 if (!vma->vm_file) /* anonymous vma */
4859 return NULL;
4860 if (!(mc.flags & MOVE_FILE))
4861 return NULL;
4863 mapping = vma->vm_file->f_mapping;
4864 pgoff = linear_page_index(vma, addr);
4866 /* page is moved even if it's not RSS of this task(page-faulted). */
4867 #ifdef CONFIG_SWAP
4868 /* shmem/tmpfs may report page out on swap: account for that too. */
4869 if (shmem_mapping(mapping)) {
4870 page = find_get_entry(mapping, pgoff);
4871 if (xa_is_value(page)) {
4872 swp_entry_t swp = radix_to_swp_entry(page);
4873 if (do_memsw_account())
4874 *entry = swp;
4875 page = find_get_page(swap_address_space(swp),
4876 swp_offset(swp));
4878 } else
4879 page = find_get_page(mapping, pgoff);
4880 #else
4881 page = find_get_page(mapping, pgoff);
4882 #endif
4883 return page;
4887 * mem_cgroup_move_account - move account of the page
4888 * @page: the page
4889 * @compound: charge the page as compound or small page
4890 * @from: mem_cgroup which the page is moved from.
4891 * @to: mem_cgroup which the page is moved to. @from != @to.
4893 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4895 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4896 * from old cgroup.
4898 static int mem_cgroup_move_account(struct page *page,
4899 bool compound,
4900 struct mem_cgroup *from,
4901 struct mem_cgroup *to)
4903 unsigned long flags;
4904 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4905 int ret;
4906 bool anon;
4908 VM_BUG_ON(from == to);
4909 VM_BUG_ON_PAGE(PageLRU(page), page);
4910 VM_BUG_ON(compound && !PageTransHuge(page));
4913 * Prevent mem_cgroup_migrate() from looking at
4914 * page->mem_cgroup of its source page while we change it.
4916 ret = -EBUSY;
4917 if (!trylock_page(page))
4918 goto out;
4920 ret = -EINVAL;
4921 if (page->mem_cgroup != from)
4922 goto out_unlock;
4924 anon = PageAnon(page);
4926 spin_lock_irqsave(&from->move_lock, flags);
4928 if (!anon && page_mapped(page)) {
4929 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
4930 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
4934 * move_lock grabbed above and caller set from->moving_account, so
4935 * mod_memcg_page_state will serialize updates to PageDirty.
4936 * So mapping should be stable for dirty pages.
4938 if (!anon && PageDirty(page)) {
4939 struct address_space *mapping = page_mapping(page);
4941 if (mapping_cap_account_dirty(mapping)) {
4942 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
4943 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
4947 if (PageWriteback(page)) {
4948 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
4949 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
4953 * It is safe to change page->mem_cgroup here because the page
4954 * is referenced, charged, and isolated - we can't race with
4955 * uncharging, charging, migration, or LRU putback.
4958 /* caller should have done css_get */
4959 page->mem_cgroup = to;
4960 spin_unlock_irqrestore(&from->move_lock, flags);
4962 ret = 0;
4964 local_irq_disable();
4965 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4966 memcg_check_events(to, page);
4967 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4968 memcg_check_events(from, page);
4969 local_irq_enable();
4970 out_unlock:
4971 unlock_page(page);
4972 out:
4973 return ret;
4977 * get_mctgt_type - get target type of moving charge
4978 * @vma: the vma the pte to be checked belongs
4979 * @addr: the address corresponding to the pte to be checked
4980 * @ptent: the pte to be checked
4981 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4983 * Returns
4984 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4985 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4986 * move charge. if @target is not NULL, the page is stored in target->page
4987 * with extra refcnt got(Callers should handle it).
4988 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4989 * target for charge migration. if @target is not NULL, the entry is stored
4990 * in target->ent.
4991 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PUBLIC
4992 * or MEMORY_DEVICE_PRIVATE (so ZONE_DEVICE page and thus not on the lru).
4993 * For now we such page is charge like a regular page would be as for all
4994 * intent and purposes it is just special memory taking the place of a
4995 * regular page.
4997 * See Documentations/vm/hmm.txt and include/linux/hmm.h
4999 * Called with pte lock held.
5002 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5003 unsigned long addr, pte_t ptent, union mc_target *target)
5005 struct page *page = NULL;
5006 enum mc_target_type ret = MC_TARGET_NONE;
5007 swp_entry_t ent = { .val = 0 };
5009 if (pte_present(ptent))
5010 page = mc_handle_present_pte(vma, addr, ptent);
5011 else if (is_swap_pte(ptent))
5012 page = mc_handle_swap_pte(vma, ptent, &ent);
5013 else if (pte_none(ptent))
5014 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5016 if (!page && !ent.val)
5017 return ret;
5018 if (page) {
5020 * Do only loose check w/o serialization.
5021 * mem_cgroup_move_account() checks the page is valid or
5022 * not under LRU exclusion.
5024 if (page->mem_cgroup == mc.from) {
5025 ret = MC_TARGET_PAGE;
5026 if (is_device_private_page(page) ||
5027 is_device_public_page(page))
5028 ret = MC_TARGET_DEVICE;
5029 if (target)
5030 target->page = page;
5032 if (!ret || !target)
5033 put_page(page);
5036 * There is a swap entry and a page doesn't exist or isn't charged.
5037 * But we cannot move a tail-page in a THP.
5039 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5040 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5041 ret = MC_TARGET_SWAP;
5042 if (target)
5043 target->ent = ent;
5045 return ret;
5048 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5050 * We don't consider PMD mapped swapping or file mapped pages because THP does
5051 * not support them for now.
5052 * Caller should make sure that pmd_trans_huge(pmd) is true.
5054 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5055 unsigned long addr, pmd_t pmd, union mc_target *target)
5057 struct page *page = NULL;
5058 enum mc_target_type ret = MC_TARGET_NONE;
5060 if (unlikely(is_swap_pmd(pmd))) {
5061 VM_BUG_ON(thp_migration_supported() &&
5062 !is_pmd_migration_entry(pmd));
5063 return ret;
5065 page = pmd_page(pmd);
5066 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5067 if (!(mc.flags & MOVE_ANON))
5068 return ret;
5069 if (page->mem_cgroup == mc.from) {
5070 ret = MC_TARGET_PAGE;
5071 if (target) {
5072 get_page(page);
5073 target->page = page;
5076 return ret;
5078 #else
5079 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5080 unsigned long addr, pmd_t pmd, union mc_target *target)
5082 return MC_TARGET_NONE;
5084 #endif
5086 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5087 unsigned long addr, unsigned long end,
5088 struct mm_walk *walk)
5090 struct vm_area_struct *vma = walk->vma;
5091 pte_t *pte;
5092 spinlock_t *ptl;
5094 ptl = pmd_trans_huge_lock(pmd, vma);
5095 if (ptl) {
5097 * Note their can not be MC_TARGET_DEVICE for now as we do not
5098 * support transparent huge page with MEMORY_DEVICE_PUBLIC or
5099 * MEMORY_DEVICE_PRIVATE but this might change.
5101 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5102 mc.precharge += HPAGE_PMD_NR;
5103 spin_unlock(ptl);
5104 return 0;
5107 if (pmd_trans_unstable(pmd))
5108 return 0;
5109 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5110 for (; addr != end; pte++, addr += PAGE_SIZE)
5111 if (get_mctgt_type(vma, addr, *pte, NULL))
5112 mc.precharge++; /* increment precharge temporarily */
5113 pte_unmap_unlock(pte - 1, ptl);
5114 cond_resched();
5116 return 0;
5119 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5121 unsigned long precharge;
5123 struct mm_walk mem_cgroup_count_precharge_walk = {
5124 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5125 .mm = mm,
5127 down_read(&mm->mmap_sem);
5128 walk_page_range(0, mm->highest_vm_end,
5129 &mem_cgroup_count_precharge_walk);
5130 up_read(&mm->mmap_sem);
5132 precharge = mc.precharge;
5133 mc.precharge = 0;
5135 return precharge;
5138 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5140 unsigned long precharge = mem_cgroup_count_precharge(mm);
5142 VM_BUG_ON(mc.moving_task);
5143 mc.moving_task = current;
5144 return mem_cgroup_do_precharge(precharge);
5147 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5148 static void __mem_cgroup_clear_mc(void)
5150 struct mem_cgroup *from = mc.from;
5151 struct mem_cgroup *to = mc.to;
5153 /* we must uncharge all the leftover precharges from mc.to */
5154 if (mc.precharge) {
5155 cancel_charge(mc.to, mc.precharge);
5156 mc.precharge = 0;
5159 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5160 * we must uncharge here.
5162 if (mc.moved_charge) {
5163 cancel_charge(mc.from, mc.moved_charge);
5164 mc.moved_charge = 0;
5166 /* we must fixup refcnts and charges */
5167 if (mc.moved_swap) {
5168 /* uncharge swap account from the old cgroup */
5169 if (!mem_cgroup_is_root(mc.from))
5170 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5172 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5175 * we charged both to->memory and to->memsw, so we
5176 * should uncharge to->memory.
5178 if (!mem_cgroup_is_root(mc.to))
5179 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5181 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5182 css_put_many(&mc.to->css, mc.moved_swap);
5184 mc.moved_swap = 0;
5186 memcg_oom_recover(from);
5187 memcg_oom_recover(to);
5188 wake_up_all(&mc.waitq);
5191 static void mem_cgroup_clear_mc(void)
5193 struct mm_struct *mm = mc.mm;
5196 * we must clear moving_task before waking up waiters at the end of
5197 * task migration.
5199 mc.moving_task = NULL;
5200 __mem_cgroup_clear_mc();
5201 spin_lock(&mc.lock);
5202 mc.from = NULL;
5203 mc.to = NULL;
5204 mc.mm = NULL;
5205 spin_unlock(&mc.lock);
5207 mmput(mm);
5210 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5212 struct cgroup_subsys_state *css;
5213 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5214 struct mem_cgroup *from;
5215 struct task_struct *leader, *p;
5216 struct mm_struct *mm;
5217 unsigned long move_flags;
5218 int ret = 0;
5220 /* charge immigration isn't supported on the default hierarchy */
5221 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5222 return 0;
5225 * Multi-process migrations only happen on the default hierarchy
5226 * where charge immigration is not used. Perform charge
5227 * immigration if @tset contains a leader and whine if there are
5228 * multiple.
5230 p = NULL;
5231 cgroup_taskset_for_each_leader(leader, css, tset) {
5232 WARN_ON_ONCE(p);
5233 p = leader;
5234 memcg = mem_cgroup_from_css(css);
5236 if (!p)
5237 return 0;
5240 * We are now commited to this value whatever it is. Changes in this
5241 * tunable will only affect upcoming migrations, not the current one.
5242 * So we need to save it, and keep it going.
5244 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5245 if (!move_flags)
5246 return 0;
5248 from = mem_cgroup_from_task(p);
5250 VM_BUG_ON(from == memcg);
5252 mm = get_task_mm(p);
5253 if (!mm)
5254 return 0;
5255 /* We move charges only when we move a owner of the mm */
5256 if (mm->owner == p) {
5257 VM_BUG_ON(mc.from);
5258 VM_BUG_ON(mc.to);
5259 VM_BUG_ON(mc.precharge);
5260 VM_BUG_ON(mc.moved_charge);
5261 VM_BUG_ON(mc.moved_swap);
5263 spin_lock(&mc.lock);
5264 mc.mm = mm;
5265 mc.from = from;
5266 mc.to = memcg;
5267 mc.flags = move_flags;
5268 spin_unlock(&mc.lock);
5269 /* We set mc.moving_task later */
5271 ret = mem_cgroup_precharge_mc(mm);
5272 if (ret)
5273 mem_cgroup_clear_mc();
5274 } else {
5275 mmput(mm);
5277 return ret;
5280 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5282 if (mc.to)
5283 mem_cgroup_clear_mc();
5286 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5287 unsigned long addr, unsigned long end,
5288 struct mm_walk *walk)
5290 int ret = 0;
5291 struct vm_area_struct *vma = walk->vma;
5292 pte_t *pte;
5293 spinlock_t *ptl;
5294 enum mc_target_type target_type;
5295 union mc_target target;
5296 struct page *page;
5298 ptl = pmd_trans_huge_lock(pmd, vma);
5299 if (ptl) {
5300 if (mc.precharge < HPAGE_PMD_NR) {
5301 spin_unlock(ptl);
5302 return 0;
5304 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5305 if (target_type == MC_TARGET_PAGE) {
5306 page = target.page;
5307 if (!isolate_lru_page(page)) {
5308 if (!mem_cgroup_move_account(page, true,
5309 mc.from, mc.to)) {
5310 mc.precharge -= HPAGE_PMD_NR;
5311 mc.moved_charge += HPAGE_PMD_NR;
5313 putback_lru_page(page);
5315 put_page(page);
5316 } else if (target_type == MC_TARGET_DEVICE) {
5317 page = target.page;
5318 if (!mem_cgroup_move_account(page, true,
5319 mc.from, mc.to)) {
5320 mc.precharge -= HPAGE_PMD_NR;
5321 mc.moved_charge += HPAGE_PMD_NR;
5323 put_page(page);
5325 spin_unlock(ptl);
5326 return 0;
5329 if (pmd_trans_unstable(pmd))
5330 return 0;
5331 retry:
5332 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5333 for (; addr != end; addr += PAGE_SIZE) {
5334 pte_t ptent = *(pte++);
5335 bool device = false;
5336 swp_entry_t ent;
5338 if (!mc.precharge)
5339 break;
5341 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5342 case MC_TARGET_DEVICE:
5343 device = true;
5344 /* fall through */
5345 case MC_TARGET_PAGE:
5346 page = target.page;
5348 * We can have a part of the split pmd here. Moving it
5349 * can be done but it would be too convoluted so simply
5350 * ignore such a partial THP and keep it in original
5351 * memcg. There should be somebody mapping the head.
5353 if (PageTransCompound(page))
5354 goto put;
5355 if (!device && isolate_lru_page(page))
5356 goto put;
5357 if (!mem_cgroup_move_account(page, false,
5358 mc.from, mc.to)) {
5359 mc.precharge--;
5360 /* we uncharge from mc.from later. */
5361 mc.moved_charge++;
5363 if (!device)
5364 putback_lru_page(page);
5365 put: /* get_mctgt_type() gets the page */
5366 put_page(page);
5367 break;
5368 case MC_TARGET_SWAP:
5369 ent = target.ent;
5370 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5371 mc.precharge--;
5372 /* we fixup refcnts and charges later. */
5373 mc.moved_swap++;
5375 break;
5376 default:
5377 break;
5380 pte_unmap_unlock(pte - 1, ptl);
5381 cond_resched();
5383 if (addr != end) {
5385 * We have consumed all precharges we got in can_attach().
5386 * We try charge one by one, but don't do any additional
5387 * charges to mc.to if we have failed in charge once in attach()
5388 * phase.
5390 ret = mem_cgroup_do_precharge(1);
5391 if (!ret)
5392 goto retry;
5395 return ret;
5398 static void mem_cgroup_move_charge(void)
5400 struct mm_walk mem_cgroup_move_charge_walk = {
5401 .pmd_entry = mem_cgroup_move_charge_pte_range,
5402 .mm = mc.mm,
5405 lru_add_drain_all();
5407 * Signal lock_page_memcg() to take the memcg's move_lock
5408 * while we're moving its pages to another memcg. Then wait
5409 * for already started RCU-only updates to finish.
5411 atomic_inc(&mc.from->moving_account);
5412 synchronize_rcu();
5413 retry:
5414 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5416 * Someone who are holding the mmap_sem might be waiting in
5417 * waitq. So we cancel all extra charges, wake up all waiters,
5418 * and retry. Because we cancel precharges, we might not be able
5419 * to move enough charges, but moving charge is a best-effort
5420 * feature anyway, so it wouldn't be a big problem.
5422 __mem_cgroup_clear_mc();
5423 cond_resched();
5424 goto retry;
5427 * When we have consumed all precharges and failed in doing
5428 * additional charge, the page walk just aborts.
5430 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5432 up_read(&mc.mm->mmap_sem);
5433 atomic_dec(&mc.from->moving_account);
5436 static void mem_cgroup_move_task(void)
5438 if (mc.to) {
5439 mem_cgroup_move_charge();
5440 mem_cgroup_clear_mc();
5443 #else /* !CONFIG_MMU */
5444 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5446 return 0;
5448 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5451 static void mem_cgroup_move_task(void)
5454 #endif
5457 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5458 * to verify whether we're attached to the default hierarchy on each mount
5459 * attempt.
5461 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5464 * use_hierarchy is forced on the default hierarchy. cgroup core
5465 * guarantees that @root doesn't have any children, so turning it
5466 * on for the root memcg is enough.
5468 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5469 root_mem_cgroup->use_hierarchy = true;
5470 else
5471 root_mem_cgroup->use_hierarchy = false;
5474 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
5476 if (value == PAGE_COUNTER_MAX)
5477 seq_puts(m, "max\n");
5478 else
5479 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
5481 return 0;
5484 static u64 memory_current_read(struct cgroup_subsys_state *css,
5485 struct cftype *cft)
5487 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5489 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5492 static int memory_min_show(struct seq_file *m, void *v)
5494 return seq_puts_memcg_tunable(m,
5495 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
5498 static ssize_t memory_min_write(struct kernfs_open_file *of,
5499 char *buf, size_t nbytes, loff_t off)
5501 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5502 unsigned long min;
5503 int err;
5505 buf = strstrip(buf);
5506 err = page_counter_memparse(buf, "max", &min);
5507 if (err)
5508 return err;
5510 page_counter_set_min(&memcg->memory, min);
5512 return nbytes;
5515 static int memory_low_show(struct seq_file *m, void *v)
5517 return seq_puts_memcg_tunable(m,
5518 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
5521 static ssize_t memory_low_write(struct kernfs_open_file *of,
5522 char *buf, size_t nbytes, loff_t off)
5524 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5525 unsigned long low;
5526 int err;
5528 buf = strstrip(buf);
5529 err = page_counter_memparse(buf, "max", &low);
5530 if (err)
5531 return err;
5533 page_counter_set_low(&memcg->memory, low);
5535 return nbytes;
5538 static int memory_high_show(struct seq_file *m, void *v)
5540 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
5543 static ssize_t memory_high_write(struct kernfs_open_file *of,
5544 char *buf, size_t nbytes, loff_t off)
5546 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5547 unsigned long nr_pages;
5548 unsigned long high;
5549 int err;
5551 buf = strstrip(buf);
5552 err = page_counter_memparse(buf, "max", &high);
5553 if (err)
5554 return err;
5556 memcg->high = high;
5558 nr_pages = page_counter_read(&memcg->memory);
5559 if (nr_pages > high)
5560 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5561 GFP_KERNEL, true);
5563 memcg_wb_domain_size_changed(memcg);
5564 return nbytes;
5567 static int memory_max_show(struct seq_file *m, void *v)
5569 return seq_puts_memcg_tunable(m,
5570 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
5573 static ssize_t memory_max_write(struct kernfs_open_file *of,
5574 char *buf, size_t nbytes, loff_t off)
5576 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5577 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5578 bool drained = false;
5579 unsigned long max;
5580 int err;
5582 buf = strstrip(buf);
5583 err = page_counter_memparse(buf, "max", &max);
5584 if (err)
5585 return err;
5587 xchg(&memcg->memory.max, max);
5589 for (;;) {
5590 unsigned long nr_pages = page_counter_read(&memcg->memory);
5592 if (nr_pages <= max)
5593 break;
5595 if (signal_pending(current)) {
5596 err = -EINTR;
5597 break;
5600 if (!drained) {
5601 drain_all_stock(memcg);
5602 drained = true;
5603 continue;
5606 if (nr_reclaims) {
5607 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5608 GFP_KERNEL, true))
5609 nr_reclaims--;
5610 continue;
5613 memcg_memory_event(memcg, MEMCG_OOM);
5614 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5615 break;
5618 memcg_wb_domain_size_changed(memcg);
5619 return nbytes;
5622 static int memory_events_show(struct seq_file *m, void *v)
5624 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5626 seq_printf(m, "low %lu\n",
5627 atomic_long_read(&memcg->memory_events[MEMCG_LOW]));
5628 seq_printf(m, "high %lu\n",
5629 atomic_long_read(&memcg->memory_events[MEMCG_HIGH]));
5630 seq_printf(m, "max %lu\n",
5631 atomic_long_read(&memcg->memory_events[MEMCG_MAX]));
5632 seq_printf(m, "oom %lu\n",
5633 atomic_long_read(&memcg->memory_events[MEMCG_OOM]));
5634 seq_printf(m, "oom_kill %lu\n",
5635 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
5637 return 0;
5640 static int memory_stat_show(struct seq_file *m, void *v)
5642 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5643 int i;
5646 * Provide statistics on the state of the memory subsystem as
5647 * well as cumulative event counters that show past behavior.
5649 * This list is ordered following a combination of these gradients:
5650 * 1) generic big picture -> specifics and details
5651 * 2) reflecting userspace activity -> reflecting kernel heuristics
5653 * Current memory state:
5656 seq_printf(m, "anon %llu\n",
5657 (u64)memcg_page_state(memcg, MEMCG_RSS) * PAGE_SIZE);
5658 seq_printf(m, "file %llu\n",
5659 (u64)memcg_page_state(memcg, MEMCG_CACHE) * PAGE_SIZE);
5660 seq_printf(m, "kernel_stack %llu\n",
5661 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) * 1024);
5662 seq_printf(m, "slab %llu\n",
5663 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) +
5664 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE)) *
5665 PAGE_SIZE);
5666 seq_printf(m, "sock %llu\n",
5667 (u64)memcg_page_state(memcg, MEMCG_SOCK) * PAGE_SIZE);
5669 seq_printf(m, "shmem %llu\n",
5670 (u64)memcg_page_state(memcg, NR_SHMEM) * PAGE_SIZE);
5671 seq_printf(m, "file_mapped %llu\n",
5672 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) * PAGE_SIZE);
5673 seq_printf(m, "file_dirty %llu\n",
5674 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) * PAGE_SIZE);
5675 seq_printf(m, "file_writeback %llu\n",
5676 (u64)memcg_page_state(memcg, NR_WRITEBACK) * PAGE_SIZE);
5679 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
5680 * with the NR_ANON_THP vm counter, but right now it's a pain in the
5681 * arse because it requires migrating the work out of rmap to a place
5682 * where the page->mem_cgroup is set up and stable.
5684 seq_printf(m, "anon_thp %llu\n",
5685 (u64)memcg_page_state(memcg, MEMCG_RSS_HUGE) * PAGE_SIZE);
5687 for (i = 0; i < NR_LRU_LISTS; i++)
5688 seq_printf(m, "%s %llu\n", mem_cgroup_lru_names[i],
5689 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
5690 PAGE_SIZE);
5692 seq_printf(m, "slab_reclaimable %llu\n",
5693 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) *
5694 PAGE_SIZE);
5695 seq_printf(m, "slab_unreclaimable %llu\n",
5696 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE) *
5697 PAGE_SIZE);
5699 /* Accumulated memory events */
5701 seq_printf(m, "pgfault %lu\n", memcg_events(memcg, PGFAULT));
5702 seq_printf(m, "pgmajfault %lu\n", memcg_events(memcg, PGMAJFAULT));
5704 seq_printf(m, "workingset_refault %lu\n",
5705 memcg_page_state(memcg, WORKINGSET_REFAULT));
5706 seq_printf(m, "workingset_activate %lu\n",
5707 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
5708 seq_printf(m, "workingset_nodereclaim %lu\n",
5709 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
5711 seq_printf(m, "pgrefill %lu\n", memcg_events(memcg, PGREFILL));
5712 seq_printf(m, "pgscan %lu\n", memcg_events(memcg, PGSCAN_KSWAPD) +
5713 memcg_events(memcg, PGSCAN_DIRECT));
5714 seq_printf(m, "pgsteal %lu\n", memcg_events(memcg, PGSTEAL_KSWAPD) +
5715 memcg_events(memcg, PGSTEAL_DIRECT));
5716 seq_printf(m, "pgactivate %lu\n", memcg_events(memcg, PGACTIVATE));
5717 seq_printf(m, "pgdeactivate %lu\n", memcg_events(memcg, PGDEACTIVATE));
5718 seq_printf(m, "pglazyfree %lu\n", memcg_events(memcg, PGLAZYFREE));
5719 seq_printf(m, "pglazyfreed %lu\n", memcg_events(memcg, PGLAZYFREED));
5721 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5722 seq_printf(m, "thp_fault_alloc %lu\n",
5723 memcg_events(memcg, THP_FAULT_ALLOC));
5724 seq_printf(m, "thp_collapse_alloc %lu\n",
5725 memcg_events(memcg, THP_COLLAPSE_ALLOC));
5726 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
5728 return 0;
5731 static int memory_oom_group_show(struct seq_file *m, void *v)
5733 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5735 seq_printf(m, "%d\n", memcg->oom_group);
5737 return 0;
5740 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
5741 char *buf, size_t nbytes, loff_t off)
5743 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5744 int ret, oom_group;
5746 buf = strstrip(buf);
5747 if (!buf)
5748 return -EINVAL;
5750 ret = kstrtoint(buf, 0, &oom_group);
5751 if (ret)
5752 return ret;
5754 if (oom_group != 0 && oom_group != 1)
5755 return -EINVAL;
5757 memcg->oom_group = oom_group;
5759 return nbytes;
5762 static struct cftype memory_files[] = {
5764 .name = "current",
5765 .flags = CFTYPE_NOT_ON_ROOT,
5766 .read_u64 = memory_current_read,
5769 .name = "min",
5770 .flags = CFTYPE_NOT_ON_ROOT,
5771 .seq_show = memory_min_show,
5772 .write = memory_min_write,
5775 .name = "low",
5776 .flags = CFTYPE_NOT_ON_ROOT,
5777 .seq_show = memory_low_show,
5778 .write = memory_low_write,
5781 .name = "high",
5782 .flags = CFTYPE_NOT_ON_ROOT,
5783 .seq_show = memory_high_show,
5784 .write = memory_high_write,
5787 .name = "max",
5788 .flags = CFTYPE_NOT_ON_ROOT,
5789 .seq_show = memory_max_show,
5790 .write = memory_max_write,
5793 .name = "events",
5794 .flags = CFTYPE_NOT_ON_ROOT,
5795 .file_offset = offsetof(struct mem_cgroup, events_file),
5796 .seq_show = memory_events_show,
5799 .name = "stat",
5800 .flags = CFTYPE_NOT_ON_ROOT,
5801 .seq_show = memory_stat_show,
5804 .name = "oom.group",
5805 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
5806 .seq_show = memory_oom_group_show,
5807 .write = memory_oom_group_write,
5809 { } /* terminate */
5812 struct cgroup_subsys memory_cgrp_subsys = {
5813 .css_alloc = mem_cgroup_css_alloc,
5814 .css_online = mem_cgroup_css_online,
5815 .css_offline = mem_cgroup_css_offline,
5816 .css_released = mem_cgroup_css_released,
5817 .css_free = mem_cgroup_css_free,
5818 .css_reset = mem_cgroup_css_reset,
5819 .can_attach = mem_cgroup_can_attach,
5820 .cancel_attach = mem_cgroup_cancel_attach,
5821 .post_attach = mem_cgroup_move_task,
5822 .bind = mem_cgroup_bind,
5823 .dfl_cftypes = memory_files,
5824 .legacy_cftypes = mem_cgroup_legacy_files,
5825 .early_init = 0,
5829 * mem_cgroup_protected - check if memory consumption is in the normal range
5830 * @root: the top ancestor of the sub-tree being checked
5831 * @memcg: the memory cgroup to check
5833 * WARNING: This function is not stateless! It can only be used as part
5834 * of a top-down tree iteration, not for isolated queries.
5836 * Returns one of the following:
5837 * MEMCG_PROT_NONE: cgroup memory is not protected
5838 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
5839 * an unprotected supply of reclaimable memory from other cgroups.
5840 * MEMCG_PROT_MIN: cgroup memory is protected
5842 * @root is exclusive; it is never protected when looked at directly
5844 * To provide a proper hierarchical behavior, effective memory.min/low values
5845 * are used. Below is the description of how effective memory.low is calculated.
5846 * Effective memory.min values is calculated in the same way.
5848 * Effective memory.low is always equal or less than the original memory.low.
5849 * If there is no memory.low overcommittment (which is always true for
5850 * top-level memory cgroups), these two values are equal.
5851 * Otherwise, it's a part of parent's effective memory.low,
5852 * calculated as a cgroup's memory.low usage divided by sum of sibling's
5853 * memory.low usages, where memory.low usage is the size of actually
5854 * protected memory.
5856 * low_usage
5857 * elow = min( memory.low, parent->elow * ------------------ ),
5858 * siblings_low_usage
5860 * | memory.current, if memory.current < memory.low
5861 * low_usage = |
5862 * | 0, otherwise.
5865 * Such definition of the effective memory.low provides the expected
5866 * hierarchical behavior: parent's memory.low value is limiting
5867 * children, unprotected memory is reclaimed first and cgroups,
5868 * which are not using their guarantee do not affect actual memory
5869 * distribution.
5871 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
5873 * A A/memory.low = 2G, A/memory.current = 6G
5874 * //\\
5875 * BC DE B/memory.low = 3G B/memory.current = 2G
5876 * C/memory.low = 1G C/memory.current = 2G
5877 * D/memory.low = 0 D/memory.current = 2G
5878 * E/memory.low = 10G E/memory.current = 0
5880 * and the memory pressure is applied, the following memory distribution
5881 * is expected (approximately):
5883 * A/memory.current = 2G
5885 * B/memory.current = 1.3G
5886 * C/memory.current = 0.6G
5887 * D/memory.current = 0
5888 * E/memory.current = 0
5890 * These calculations require constant tracking of the actual low usages
5891 * (see propagate_protected_usage()), as well as recursive calculation of
5892 * effective memory.low values. But as we do call mem_cgroup_protected()
5893 * path for each memory cgroup top-down from the reclaim,
5894 * it's possible to optimize this part, and save calculated elow
5895 * for next usage. This part is intentionally racy, but it's ok,
5896 * as memory.low is a best-effort mechanism.
5898 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
5899 struct mem_cgroup *memcg)
5901 struct mem_cgroup *parent;
5902 unsigned long emin, parent_emin;
5903 unsigned long elow, parent_elow;
5904 unsigned long usage;
5906 if (mem_cgroup_disabled())
5907 return MEMCG_PROT_NONE;
5909 if (!root)
5910 root = root_mem_cgroup;
5911 if (memcg == root)
5912 return MEMCG_PROT_NONE;
5914 usage = page_counter_read(&memcg->memory);
5915 if (!usage)
5916 return MEMCG_PROT_NONE;
5918 emin = memcg->memory.min;
5919 elow = memcg->memory.low;
5921 parent = parent_mem_cgroup(memcg);
5922 /* No parent means a non-hierarchical mode on v1 memcg */
5923 if (!parent)
5924 return MEMCG_PROT_NONE;
5926 if (parent == root)
5927 goto exit;
5929 parent_emin = READ_ONCE(parent->memory.emin);
5930 emin = min(emin, parent_emin);
5931 if (emin && parent_emin) {
5932 unsigned long min_usage, siblings_min_usage;
5934 min_usage = min(usage, memcg->memory.min);
5935 siblings_min_usage = atomic_long_read(
5936 &parent->memory.children_min_usage);
5938 if (min_usage && siblings_min_usage)
5939 emin = min(emin, parent_emin * min_usage /
5940 siblings_min_usage);
5943 parent_elow = READ_ONCE(parent->memory.elow);
5944 elow = min(elow, parent_elow);
5945 if (elow && parent_elow) {
5946 unsigned long low_usage, siblings_low_usage;
5948 low_usage = min(usage, memcg->memory.low);
5949 siblings_low_usage = atomic_long_read(
5950 &parent->memory.children_low_usage);
5952 if (low_usage && siblings_low_usage)
5953 elow = min(elow, parent_elow * low_usage /
5954 siblings_low_usage);
5957 exit:
5958 memcg->memory.emin = emin;
5959 memcg->memory.elow = elow;
5961 if (usage <= emin)
5962 return MEMCG_PROT_MIN;
5963 else if (usage <= elow)
5964 return MEMCG_PROT_LOW;
5965 else
5966 return MEMCG_PROT_NONE;
5970 * mem_cgroup_try_charge - try charging a page
5971 * @page: page to charge
5972 * @mm: mm context of the victim
5973 * @gfp_mask: reclaim mode
5974 * @memcgp: charged memcg return
5975 * @compound: charge the page as compound or small page
5977 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5978 * pages according to @gfp_mask if necessary.
5980 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5981 * Otherwise, an error code is returned.
5983 * After page->mapping has been set up, the caller must finalize the
5984 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5985 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5987 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5988 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5989 bool compound)
5991 struct mem_cgroup *memcg = NULL;
5992 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5993 int ret = 0;
5995 if (mem_cgroup_disabled())
5996 goto out;
5998 if (PageSwapCache(page)) {
6000 * Every swap fault against a single page tries to charge the
6001 * page, bail as early as possible. shmem_unuse() encounters
6002 * already charged pages, too. The USED bit is protected by
6003 * the page lock, which serializes swap cache removal, which
6004 * in turn serializes uncharging.
6006 VM_BUG_ON_PAGE(!PageLocked(page), page);
6007 if (compound_head(page)->mem_cgroup)
6008 goto out;
6010 if (do_swap_account) {
6011 swp_entry_t ent = { .val = page_private(page), };
6012 unsigned short id = lookup_swap_cgroup_id(ent);
6014 rcu_read_lock();
6015 memcg = mem_cgroup_from_id(id);
6016 if (memcg && !css_tryget_online(&memcg->css))
6017 memcg = NULL;
6018 rcu_read_unlock();
6022 if (!memcg)
6023 memcg = get_mem_cgroup_from_mm(mm);
6025 ret = try_charge(memcg, gfp_mask, nr_pages);
6027 css_put(&memcg->css);
6028 out:
6029 *memcgp = memcg;
6030 return ret;
6033 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6034 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6035 bool compound)
6037 struct mem_cgroup *memcg;
6038 int ret;
6040 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6041 memcg = *memcgp;
6042 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6043 return ret;
6047 * mem_cgroup_commit_charge - commit a page charge
6048 * @page: page to charge
6049 * @memcg: memcg to charge the page to
6050 * @lrucare: page might be on LRU already
6051 * @compound: charge the page as compound or small page
6053 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6054 * after page->mapping has been set up. This must happen atomically
6055 * as part of the page instantiation, i.e. under the page table lock
6056 * for anonymous pages, under the page lock for page and swap cache.
6058 * In addition, the page must not be on the LRU during the commit, to
6059 * prevent racing with task migration. If it might be, use @lrucare.
6061 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6063 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6064 bool lrucare, bool compound)
6066 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6068 VM_BUG_ON_PAGE(!page->mapping, page);
6069 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6071 if (mem_cgroup_disabled())
6072 return;
6074 * Swap faults will attempt to charge the same page multiple
6075 * times. But reuse_swap_page() might have removed the page
6076 * from swapcache already, so we can't check PageSwapCache().
6078 if (!memcg)
6079 return;
6081 commit_charge(page, memcg, lrucare);
6083 local_irq_disable();
6084 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6085 memcg_check_events(memcg, page);
6086 local_irq_enable();
6088 if (do_memsw_account() && PageSwapCache(page)) {
6089 swp_entry_t entry = { .val = page_private(page) };
6091 * The swap entry might not get freed for a long time,
6092 * let's not wait for it. The page already received a
6093 * memory+swap charge, drop the swap entry duplicate.
6095 mem_cgroup_uncharge_swap(entry, nr_pages);
6100 * mem_cgroup_cancel_charge - cancel a page charge
6101 * @page: page to charge
6102 * @memcg: memcg to charge the page to
6103 * @compound: charge the page as compound or small page
6105 * Cancel a charge transaction started by mem_cgroup_try_charge().
6107 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6108 bool compound)
6110 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6112 if (mem_cgroup_disabled())
6113 return;
6115 * Swap faults will attempt to charge the same page multiple
6116 * times. But reuse_swap_page() might have removed the page
6117 * from swapcache already, so we can't check PageSwapCache().
6119 if (!memcg)
6120 return;
6122 cancel_charge(memcg, nr_pages);
6125 struct uncharge_gather {
6126 struct mem_cgroup *memcg;
6127 unsigned long pgpgout;
6128 unsigned long nr_anon;
6129 unsigned long nr_file;
6130 unsigned long nr_kmem;
6131 unsigned long nr_huge;
6132 unsigned long nr_shmem;
6133 struct page *dummy_page;
6136 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6138 memset(ug, 0, sizeof(*ug));
6141 static void uncharge_batch(const struct uncharge_gather *ug)
6143 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6144 unsigned long flags;
6146 if (!mem_cgroup_is_root(ug->memcg)) {
6147 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6148 if (do_memsw_account())
6149 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6150 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6151 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6152 memcg_oom_recover(ug->memcg);
6155 local_irq_save(flags);
6156 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6157 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6158 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6159 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6160 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6161 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, nr_pages);
6162 memcg_check_events(ug->memcg, ug->dummy_page);
6163 local_irq_restore(flags);
6165 if (!mem_cgroup_is_root(ug->memcg))
6166 css_put_many(&ug->memcg->css, nr_pages);
6169 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6171 VM_BUG_ON_PAGE(PageLRU(page), page);
6172 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6173 !PageHWPoison(page) , page);
6175 if (!page->mem_cgroup)
6176 return;
6179 * Nobody should be changing or seriously looking at
6180 * page->mem_cgroup at this point, we have fully
6181 * exclusive access to the page.
6184 if (ug->memcg != page->mem_cgroup) {
6185 if (ug->memcg) {
6186 uncharge_batch(ug);
6187 uncharge_gather_clear(ug);
6189 ug->memcg = page->mem_cgroup;
6192 if (!PageKmemcg(page)) {
6193 unsigned int nr_pages = 1;
6195 if (PageTransHuge(page)) {
6196 nr_pages <<= compound_order(page);
6197 ug->nr_huge += nr_pages;
6199 if (PageAnon(page))
6200 ug->nr_anon += nr_pages;
6201 else {
6202 ug->nr_file += nr_pages;
6203 if (PageSwapBacked(page))
6204 ug->nr_shmem += nr_pages;
6206 ug->pgpgout++;
6207 } else {
6208 ug->nr_kmem += 1 << compound_order(page);
6209 __ClearPageKmemcg(page);
6212 ug->dummy_page = page;
6213 page->mem_cgroup = NULL;
6216 static void uncharge_list(struct list_head *page_list)
6218 struct uncharge_gather ug;
6219 struct list_head *next;
6221 uncharge_gather_clear(&ug);
6224 * Note that the list can be a single page->lru; hence the
6225 * do-while loop instead of a simple list_for_each_entry().
6227 next = page_list->next;
6228 do {
6229 struct page *page;
6231 page = list_entry(next, struct page, lru);
6232 next = page->lru.next;
6234 uncharge_page(page, &ug);
6235 } while (next != page_list);
6237 if (ug.memcg)
6238 uncharge_batch(&ug);
6242 * mem_cgroup_uncharge - uncharge a page
6243 * @page: page to uncharge
6245 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6246 * mem_cgroup_commit_charge().
6248 void mem_cgroup_uncharge(struct page *page)
6250 struct uncharge_gather ug;
6252 if (mem_cgroup_disabled())
6253 return;
6255 /* Don't touch page->lru of any random page, pre-check: */
6256 if (!page->mem_cgroup)
6257 return;
6259 uncharge_gather_clear(&ug);
6260 uncharge_page(page, &ug);
6261 uncharge_batch(&ug);
6265 * mem_cgroup_uncharge_list - uncharge a list of page
6266 * @page_list: list of pages to uncharge
6268 * Uncharge a list of pages previously charged with
6269 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6271 void mem_cgroup_uncharge_list(struct list_head *page_list)
6273 if (mem_cgroup_disabled())
6274 return;
6276 if (!list_empty(page_list))
6277 uncharge_list(page_list);
6281 * mem_cgroup_migrate - charge a page's replacement
6282 * @oldpage: currently circulating page
6283 * @newpage: replacement page
6285 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6286 * be uncharged upon free.
6288 * Both pages must be locked, @newpage->mapping must be set up.
6290 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6292 struct mem_cgroup *memcg;
6293 unsigned int nr_pages;
6294 bool compound;
6295 unsigned long flags;
6297 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6298 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6299 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6300 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6301 newpage);
6303 if (mem_cgroup_disabled())
6304 return;
6306 /* Page cache replacement: new page already charged? */
6307 if (newpage->mem_cgroup)
6308 return;
6310 /* Swapcache readahead pages can get replaced before being charged */
6311 memcg = oldpage->mem_cgroup;
6312 if (!memcg)
6313 return;
6315 /* Force-charge the new page. The old one will be freed soon */
6316 compound = PageTransHuge(newpage);
6317 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6319 page_counter_charge(&memcg->memory, nr_pages);
6320 if (do_memsw_account())
6321 page_counter_charge(&memcg->memsw, nr_pages);
6322 css_get_many(&memcg->css, nr_pages);
6324 commit_charge(newpage, memcg, false);
6326 local_irq_save(flags);
6327 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6328 memcg_check_events(memcg, newpage);
6329 local_irq_restore(flags);
6332 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6333 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6335 void mem_cgroup_sk_alloc(struct sock *sk)
6337 struct mem_cgroup *memcg;
6339 if (!mem_cgroup_sockets_enabled)
6340 return;
6343 * Socket cloning can throw us here with sk_memcg already
6344 * filled. It won't however, necessarily happen from
6345 * process context. So the test for root memcg given
6346 * the current task's memcg won't help us in this case.
6348 * Respecting the original socket's memcg is a better
6349 * decision in this case.
6351 if (sk->sk_memcg) {
6352 css_get(&sk->sk_memcg->css);
6353 return;
6356 rcu_read_lock();
6357 memcg = mem_cgroup_from_task(current);
6358 if (memcg == root_mem_cgroup)
6359 goto out;
6360 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6361 goto out;
6362 if (css_tryget_online(&memcg->css))
6363 sk->sk_memcg = memcg;
6364 out:
6365 rcu_read_unlock();
6368 void mem_cgroup_sk_free(struct sock *sk)
6370 if (sk->sk_memcg)
6371 css_put(&sk->sk_memcg->css);
6375 * mem_cgroup_charge_skmem - charge socket memory
6376 * @memcg: memcg to charge
6377 * @nr_pages: number of pages to charge
6379 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6380 * @memcg's configured limit, %false if the charge had to be forced.
6382 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6384 gfp_t gfp_mask = GFP_KERNEL;
6386 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6387 struct page_counter *fail;
6389 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6390 memcg->tcpmem_pressure = 0;
6391 return true;
6393 page_counter_charge(&memcg->tcpmem, nr_pages);
6394 memcg->tcpmem_pressure = 1;
6395 return false;
6398 /* Don't block in the packet receive path */
6399 if (in_softirq())
6400 gfp_mask = GFP_NOWAIT;
6402 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6404 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6405 return true;
6407 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6408 return false;
6412 * mem_cgroup_uncharge_skmem - uncharge socket memory
6413 * @memcg: memcg to uncharge
6414 * @nr_pages: number of pages to uncharge
6416 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6418 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6419 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6420 return;
6423 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6425 refill_stock(memcg, nr_pages);
6428 static int __init cgroup_memory(char *s)
6430 char *token;
6432 while ((token = strsep(&s, ",")) != NULL) {
6433 if (!*token)
6434 continue;
6435 if (!strcmp(token, "nosocket"))
6436 cgroup_memory_nosocket = true;
6437 if (!strcmp(token, "nokmem"))
6438 cgroup_memory_nokmem = true;
6440 return 0;
6442 __setup("cgroup.memory=", cgroup_memory);
6445 * subsys_initcall() for memory controller.
6447 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6448 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6449 * basically everything that doesn't depend on a specific mem_cgroup structure
6450 * should be initialized from here.
6452 static int __init mem_cgroup_init(void)
6454 int cpu, node;
6456 #ifdef CONFIG_MEMCG_KMEM
6458 * Kmem cache creation is mostly done with the slab_mutex held,
6459 * so use a workqueue with limited concurrency to avoid stalling
6460 * all worker threads in case lots of cgroups are created and
6461 * destroyed simultaneously.
6463 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6464 BUG_ON(!memcg_kmem_cache_wq);
6465 #endif
6467 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6468 memcg_hotplug_cpu_dead);
6470 for_each_possible_cpu(cpu)
6471 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6472 drain_local_stock);
6474 for_each_node(node) {
6475 struct mem_cgroup_tree_per_node *rtpn;
6477 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6478 node_online(node) ? node : NUMA_NO_NODE);
6480 rtpn->rb_root = RB_ROOT;
6481 rtpn->rb_rightmost = NULL;
6482 spin_lock_init(&rtpn->lock);
6483 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6486 return 0;
6488 subsys_initcall(mem_cgroup_init);
6490 #ifdef CONFIG_MEMCG_SWAP
6491 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6493 while (!refcount_inc_not_zero(&memcg->id.ref)) {
6495 * The root cgroup cannot be destroyed, so it's refcount must
6496 * always be >= 1.
6498 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6499 VM_BUG_ON(1);
6500 break;
6502 memcg = parent_mem_cgroup(memcg);
6503 if (!memcg)
6504 memcg = root_mem_cgroup;
6506 return memcg;
6510 * mem_cgroup_swapout - transfer a memsw charge to swap
6511 * @page: page whose memsw charge to transfer
6512 * @entry: swap entry to move the charge to
6514 * Transfer the memsw charge of @page to @entry.
6516 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6518 struct mem_cgroup *memcg, *swap_memcg;
6519 unsigned int nr_entries;
6520 unsigned short oldid;
6522 VM_BUG_ON_PAGE(PageLRU(page), page);
6523 VM_BUG_ON_PAGE(page_count(page), page);
6525 if (!do_memsw_account())
6526 return;
6528 memcg = page->mem_cgroup;
6530 /* Readahead page, never charged */
6531 if (!memcg)
6532 return;
6535 * In case the memcg owning these pages has been offlined and doesn't
6536 * have an ID allocated to it anymore, charge the closest online
6537 * ancestor for the swap instead and transfer the memory+swap charge.
6539 swap_memcg = mem_cgroup_id_get_online(memcg);
6540 nr_entries = hpage_nr_pages(page);
6541 /* Get references for the tail pages, too */
6542 if (nr_entries > 1)
6543 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6544 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6545 nr_entries);
6546 VM_BUG_ON_PAGE(oldid, page);
6547 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6549 page->mem_cgroup = NULL;
6551 if (!mem_cgroup_is_root(memcg))
6552 page_counter_uncharge(&memcg->memory, nr_entries);
6554 if (memcg != swap_memcg) {
6555 if (!mem_cgroup_is_root(swap_memcg))
6556 page_counter_charge(&swap_memcg->memsw, nr_entries);
6557 page_counter_uncharge(&memcg->memsw, nr_entries);
6561 * Interrupts should be disabled here because the caller holds the
6562 * i_pages lock which is taken with interrupts-off. It is
6563 * important here to have the interrupts disabled because it is the
6564 * only synchronisation we have for updating the per-CPU variables.
6566 VM_BUG_ON(!irqs_disabled());
6567 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6568 -nr_entries);
6569 memcg_check_events(memcg, page);
6571 if (!mem_cgroup_is_root(memcg))
6572 css_put_many(&memcg->css, nr_entries);
6576 * mem_cgroup_try_charge_swap - try charging swap space for a page
6577 * @page: page being added to swap
6578 * @entry: swap entry to charge
6580 * Try to charge @page's memcg for the swap space at @entry.
6582 * Returns 0 on success, -ENOMEM on failure.
6584 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6586 unsigned int nr_pages = hpage_nr_pages(page);
6587 struct page_counter *counter;
6588 struct mem_cgroup *memcg;
6589 unsigned short oldid;
6591 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6592 return 0;
6594 memcg = page->mem_cgroup;
6596 /* Readahead page, never charged */
6597 if (!memcg)
6598 return 0;
6600 if (!entry.val) {
6601 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6602 return 0;
6605 memcg = mem_cgroup_id_get_online(memcg);
6607 if (!mem_cgroup_is_root(memcg) &&
6608 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6609 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
6610 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6611 mem_cgroup_id_put(memcg);
6612 return -ENOMEM;
6615 /* Get references for the tail pages, too */
6616 if (nr_pages > 1)
6617 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6618 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6619 VM_BUG_ON_PAGE(oldid, page);
6620 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
6622 return 0;
6626 * mem_cgroup_uncharge_swap - uncharge swap space
6627 * @entry: swap entry to uncharge
6628 * @nr_pages: the amount of swap space to uncharge
6630 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6632 struct mem_cgroup *memcg;
6633 unsigned short id;
6635 if (!do_swap_account)
6636 return;
6638 id = swap_cgroup_record(entry, 0, nr_pages);
6639 rcu_read_lock();
6640 memcg = mem_cgroup_from_id(id);
6641 if (memcg) {
6642 if (!mem_cgroup_is_root(memcg)) {
6643 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6644 page_counter_uncharge(&memcg->swap, nr_pages);
6645 else
6646 page_counter_uncharge(&memcg->memsw, nr_pages);
6648 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
6649 mem_cgroup_id_put_many(memcg, nr_pages);
6651 rcu_read_unlock();
6654 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6656 long nr_swap_pages = get_nr_swap_pages();
6658 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6659 return nr_swap_pages;
6660 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6661 nr_swap_pages = min_t(long, nr_swap_pages,
6662 READ_ONCE(memcg->swap.max) -
6663 page_counter_read(&memcg->swap));
6664 return nr_swap_pages;
6667 bool mem_cgroup_swap_full(struct page *page)
6669 struct mem_cgroup *memcg;
6671 VM_BUG_ON_PAGE(!PageLocked(page), page);
6673 if (vm_swap_full())
6674 return true;
6675 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6676 return false;
6678 memcg = page->mem_cgroup;
6679 if (!memcg)
6680 return false;
6682 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6683 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
6684 return true;
6686 return false;
6689 /* for remember boot option*/
6690 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6691 static int really_do_swap_account __initdata = 1;
6692 #else
6693 static int really_do_swap_account __initdata;
6694 #endif
6696 static int __init enable_swap_account(char *s)
6698 if (!strcmp(s, "1"))
6699 really_do_swap_account = 1;
6700 else if (!strcmp(s, "0"))
6701 really_do_swap_account = 0;
6702 return 1;
6704 __setup("swapaccount=", enable_swap_account);
6706 static u64 swap_current_read(struct cgroup_subsys_state *css,
6707 struct cftype *cft)
6709 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6711 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6714 static int swap_max_show(struct seq_file *m, void *v)
6716 return seq_puts_memcg_tunable(m,
6717 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
6720 static ssize_t swap_max_write(struct kernfs_open_file *of,
6721 char *buf, size_t nbytes, loff_t off)
6723 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6724 unsigned long max;
6725 int err;
6727 buf = strstrip(buf);
6728 err = page_counter_memparse(buf, "max", &max);
6729 if (err)
6730 return err;
6732 xchg(&memcg->swap.max, max);
6734 return nbytes;
6737 static int swap_events_show(struct seq_file *m, void *v)
6739 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6741 seq_printf(m, "max %lu\n",
6742 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
6743 seq_printf(m, "fail %lu\n",
6744 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
6746 return 0;
6749 static struct cftype swap_files[] = {
6751 .name = "swap.current",
6752 .flags = CFTYPE_NOT_ON_ROOT,
6753 .read_u64 = swap_current_read,
6756 .name = "swap.max",
6757 .flags = CFTYPE_NOT_ON_ROOT,
6758 .seq_show = swap_max_show,
6759 .write = swap_max_write,
6762 .name = "swap.events",
6763 .flags = CFTYPE_NOT_ON_ROOT,
6764 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
6765 .seq_show = swap_events_show,
6767 { } /* terminate */
6770 static struct cftype memsw_cgroup_files[] = {
6772 .name = "memsw.usage_in_bytes",
6773 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6774 .read_u64 = mem_cgroup_read_u64,
6777 .name = "memsw.max_usage_in_bytes",
6778 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6779 .write = mem_cgroup_reset,
6780 .read_u64 = mem_cgroup_read_u64,
6783 .name = "memsw.limit_in_bytes",
6784 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6785 .write = mem_cgroup_write,
6786 .read_u64 = mem_cgroup_read_u64,
6789 .name = "memsw.failcnt",
6790 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6791 .write = mem_cgroup_reset,
6792 .read_u64 = mem_cgroup_read_u64,
6794 { }, /* terminate */
6797 static int __init mem_cgroup_swap_init(void)
6799 if (!mem_cgroup_disabled() && really_do_swap_account) {
6800 do_swap_account = 1;
6801 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6802 swap_files));
6803 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6804 memsw_cgroup_files));
6806 return 0;
6808 subsys_initcall(mem_cgroup_swap_init);
6810 #endif /* CONFIG_MEMCG_SWAP */