1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
25 #include <linux/page_counter.h>
26 #include <linux/memcontrol.h>
27 #include <linux/cgroup.h>
28 #include <linux/pagewalk.h>
29 #include <linux/sched/mm.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/hugetlb.h>
32 #include <linux/pagemap.h>
33 #include <linux/vm_event_item.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
59 #include <linux/tracehook.h>
60 #include <linux/psi.h>
61 #include <linux/seq_buf.h>
67 #include <linux/uaccess.h>
69 #include <trace/events/vmscan.h>
71 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
72 EXPORT_SYMBOL(memory_cgrp_subsys);
74 struct mem_cgroup *root_mem_cgroup __read_mostly;
76 #define MEM_CGROUP_RECLAIM_RETRIES 5
78 /* Socket memory accounting disabled? */
79 static bool cgroup_memory_nosocket;
81 /* Kernel memory accounting disabled? */
82 static bool cgroup_memory_nokmem;
84 /* Whether the swap controller is active */
85 #ifdef CONFIG_MEMCG_SWAP
86 int do_swap_account __read_mostly;
88 #define do_swap_account 0
91 #ifdef CONFIG_CGROUP_WRITEBACK
92 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
101 #define THRESHOLDS_EVENTS_TARGET 128
102 #define SOFTLIMIT_EVENTS_TARGET 1024
105 * Cgroups above their limits are maintained in a RB-Tree, independent of
106 * their hierarchy representation
109 struct mem_cgroup_tree_per_node {
110 struct rb_root rb_root;
111 struct rb_node *rb_rightmost;
115 struct mem_cgroup_tree {
116 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
119 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
122 struct mem_cgroup_eventfd_list {
123 struct list_head list;
124 struct eventfd_ctx *eventfd;
128 * cgroup_event represents events which userspace want to receive.
130 struct mem_cgroup_event {
132 * memcg which the event belongs to.
134 struct mem_cgroup *memcg;
136 * eventfd to signal userspace about the event.
138 struct eventfd_ctx *eventfd;
140 * Each of these stored in a list by the cgroup.
142 struct list_head list;
144 * register_event() callback will be used to add new userspace
145 * waiter for changes related to this event. Use eventfd_signal()
146 * on eventfd to send notification to userspace.
148 int (*register_event)(struct mem_cgroup *memcg,
149 struct eventfd_ctx *eventfd, const char *args);
151 * unregister_event() callback will be called when userspace closes
152 * the eventfd or on cgroup removing. This callback must be set,
153 * if you want provide notification functionality.
155 void (*unregister_event)(struct mem_cgroup *memcg,
156 struct eventfd_ctx *eventfd);
158 * All fields below needed to unregister event when
159 * userspace closes eventfd.
162 wait_queue_head_t *wqh;
163 wait_queue_entry_t wait;
164 struct work_struct remove;
167 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
168 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
170 /* Stuffs for move charges at task migration. */
172 * Types of charges to be moved.
174 #define MOVE_ANON 0x1U
175 #define MOVE_FILE 0x2U
176 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
178 /* "mc" and its members are protected by cgroup_mutex */
179 static struct move_charge_struct {
180 spinlock_t lock; /* for from, to */
181 struct mm_struct *mm;
182 struct mem_cgroup *from;
183 struct mem_cgroup *to;
185 unsigned long precharge;
186 unsigned long moved_charge;
187 unsigned long moved_swap;
188 struct task_struct *moving_task; /* a task moving charges */
189 wait_queue_head_t waitq; /* a waitq for other context */
191 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
192 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
196 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
197 * limit reclaim to prevent infinite loops, if they ever occur.
199 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
200 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
203 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
204 MEM_CGROUP_CHARGE_TYPE_ANON,
205 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
206 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
210 /* for encoding cft->private value on file */
219 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
220 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
221 #define MEMFILE_ATTR(val) ((val) & 0xffff)
222 /* Used for OOM nofiier */
223 #define OOM_CONTROL (0)
226 * Iteration constructs for visiting all cgroups (under a tree). If
227 * loops are exited prematurely (break), mem_cgroup_iter_break() must
228 * be used for reference counting.
230 #define for_each_mem_cgroup_tree(iter, root) \
231 for (iter = mem_cgroup_iter(root, NULL, NULL); \
233 iter = mem_cgroup_iter(root, iter, NULL))
235 #define for_each_mem_cgroup(iter) \
236 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
238 iter = mem_cgroup_iter(NULL, iter, NULL))
240 static inline bool should_force_charge(void)
242 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
243 (current->flags & PF_EXITING);
246 /* Some nice accessors for the vmpressure. */
247 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
250 memcg = root_mem_cgroup;
251 return &memcg->vmpressure;
254 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
256 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
259 #ifdef CONFIG_MEMCG_KMEM
261 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
262 * The main reason for not using cgroup id for this:
263 * this works better in sparse environments, where we have a lot of memcgs,
264 * but only a few kmem-limited. Or also, if we have, for instance, 200
265 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
266 * 200 entry array for that.
268 * The current size of the caches array is stored in memcg_nr_cache_ids. It
269 * will double each time we have to increase it.
271 static DEFINE_IDA(memcg_cache_ida);
272 int memcg_nr_cache_ids;
274 /* Protects memcg_nr_cache_ids */
275 static DECLARE_RWSEM(memcg_cache_ids_sem);
277 void memcg_get_cache_ids(void)
279 down_read(&memcg_cache_ids_sem);
282 void memcg_put_cache_ids(void)
284 up_read(&memcg_cache_ids_sem);
288 * MIN_SIZE is different than 1, because we would like to avoid going through
289 * the alloc/free process all the time. In a small machine, 4 kmem-limited
290 * cgroups is a reasonable guess. In the future, it could be a parameter or
291 * tunable, but that is strictly not necessary.
293 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
294 * this constant directly from cgroup, but it is understandable that this is
295 * better kept as an internal representation in cgroup.c. In any case, the
296 * cgrp_id space is not getting any smaller, and we don't have to necessarily
297 * increase ours as well if it increases.
299 #define MEMCG_CACHES_MIN_SIZE 4
300 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
303 * A lot of the calls to the cache allocation functions are expected to be
304 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
305 * conditional to this static branch, we'll have to allow modules that does
306 * kmem_cache_alloc and the such to see this symbol as well
308 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
309 EXPORT_SYMBOL(memcg_kmem_enabled_key);
311 struct workqueue_struct *memcg_kmem_cache_wq;
314 static int memcg_shrinker_map_size;
315 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
317 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
319 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
322 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
323 int size, int old_size)
325 struct memcg_shrinker_map *new, *old;
328 lockdep_assert_held(&memcg_shrinker_map_mutex);
331 old = rcu_dereference_protected(
332 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
333 /* Not yet online memcg */
337 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
341 /* Set all old bits, clear all new bits */
342 memset(new->map, (int)0xff, old_size);
343 memset((void *)new->map + old_size, 0, size - old_size);
345 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
346 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
352 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
354 struct mem_cgroup_per_node *pn;
355 struct memcg_shrinker_map *map;
358 if (mem_cgroup_is_root(memcg))
362 pn = mem_cgroup_nodeinfo(memcg, nid);
363 map = rcu_dereference_protected(pn->shrinker_map, true);
366 rcu_assign_pointer(pn->shrinker_map, NULL);
370 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
372 struct memcg_shrinker_map *map;
373 int nid, size, ret = 0;
375 if (mem_cgroup_is_root(memcg))
378 mutex_lock(&memcg_shrinker_map_mutex);
379 size = memcg_shrinker_map_size;
381 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
383 memcg_free_shrinker_maps(memcg);
387 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
389 mutex_unlock(&memcg_shrinker_map_mutex);
394 int memcg_expand_shrinker_maps(int new_id)
396 int size, old_size, ret = 0;
397 struct mem_cgroup *memcg;
399 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
400 old_size = memcg_shrinker_map_size;
401 if (size <= old_size)
404 mutex_lock(&memcg_shrinker_map_mutex);
405 if (!root_mem_cgroup)
408 for_each_mem_cgroup(memcg) {
409 if (mem_cgroup_is_root(memcg))
411 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
413 mem_cgroup_iter_break(NULL, memcg);
419 memcg_shrinker_map_size = size;
420 mutex_unlock(&memcg_shrinker_map_mutex);
424 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
426 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
427 struct memcg_shrinker_map *map;
430 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
431 /* Pairs with smp mb in shrink_slab() */
432 smp_mb__before_atomic();
433 set_bit(shrinker_id, map->map);
439 * mem_cgroup_css_from_page - css of the memcg associated with a page
440 * @page: page of interest
442 * If memcg is bound to the default hierarchy, css of the memcg associated
443 * with @page is returned. The returned css remains associated with @page
444 * until it is released.
446 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
449 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
451 struct mem_cgroup *memcg;
453 memcg = page->mem_cgroup;
455 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
456 memcg = root_mem_cgroup;
462 * page_cgroup_ino - return inode number of the memcg a page is charged to
465 * Look up the closest online ancestor of the memory cgroup @page is charged to
466 * and return its inode number or 0 if @page is not charged to any cgroup. It
467 * is safe to call this function without holding a reference to @page.
469 * Note, this function is inherently racy, because there is nothing to prevent
470 * the cgroup inode from getting torn down and potentially reallocated a moment
471 * after page_cgroup_ino() returns, so it only should be used by callers that
472 * do not care (such as procfs interfaces).
474 ino_t page_cgroup_ino(struct page *page)
476 struct mem_cgroup *memcg;
477 unsigned long ino = 0;
480 if (PageSlab(page) && !PageTail(page))
481 memcg = memcg_from_slab_page(page);
483 memcg = READ_ONCE(page->mem_cgroup);
484 while (memcg && !(memcg->css.flags & CSS_ONLINE))
485 memcg = parent_mem_cgroup(memcg);
487 ino = cgroup_ino(memcg->css.cgroup);
492 static struct mem_cgroup_per_node *
493 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
495 int nid = page_to_nid(page);
497 return memcg->nodeinfo[nid];
500 static struct mem_cgroup_tree_per_node *
501 soft_limit_tree_node(int nid)
503 return soft_limit_tree.rb_tree_per_node[nid];
506 static struct mem_cgroup_tree_per_node *
507 soft_limit_tree_from_page(struct page *page)
509 int nid = page_to_nid(page);
511 return soft_limit_tree.rb_tree_per_node[nid];
514 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
515 struct mem_cgroup_tree_per_node *mctz,
516 unsigned long new_usage_in_excess)
518 struct rb_node **p = &mctz->rb_root.rb_node;
519 struct rb_node *parent = NULL;
520 struct mem_cgroup_per_node *mz_node;
521 bool rightmost = true;
526 mz->usage_in_excess = new_usage_in_excess;
527 if (!mz->usage_in_excess)
531 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
533 if (mz->usage_in_excess < mz_node->usage_in_excess) {
539 * We can't avoid mem cgroups that are over their soft
540 * limit by the same amount
542 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
547 mctz->rb_rightmost = &mz->tree_node;
549 rb_link_node(&mz->tree_node, parent, p);
550 rb_insert_color(&mz->tree_node, &mctz->rb_root);
554 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
555 struct mem_cgroup_tree_per_node *mctz)
560 if (&mz->tree_node == mctz->rb_rightmost)
561 mctz->rb_rightmost = rb_prev(&mz->tree_node);
563 rb_erase(&mz->tree_node, &mctz->rb_root);
567 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
568 struct mem_cgroup_tree_per_node *mctz)
572 spin_lock_irqsave(&mctz->lock, flags);
573 __mem_cgroup_remove_exceeded(mz, mctz);
574 spin_unlock_irqrestore(&mctz->lock, flags);
577 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
579 unsigned long nr_pages = page_counter_read(&memcg->memory);
580 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
581 unsigned long excess = 0;
583 if (nr_pages > soft_limit)
584 excess = nr_pages - soft_limit;
589 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
591 unsigned long excess;
592 struct mem_cgroup_per_node *mz;
593 struct mem_cgroup_tree_per_node *mctz;
595 mctz = soft_limit_tree_from_page(page);
599 * Necessary to update all ancestors when hierarchy is used.
600 * because their event counter is not touched.
602 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
603 mz = mem_cgroup_page_nodeinfo(memcg, page);
604 excess = soft_limit_excess(memcg);
606 * We have to update the tree if mz is on RB-tree or
607 * mem is over its softlimit.
609 if (excess || mz->on_tree) {
612 spin_lock_irqsave(&mctz->lock, flags);
613 /* if on-tree, remove it */
615 __mem_cgroup_remove_exceeded(mz, mctz);
617 * Insert again. mz->usage_in_excess will be updated.
618 * If excess is 0, no tree ops.
620 __mem_cgroup_insert_exceeded(mz, mctz, excess);
621 spin_unlock_irqrestore(&mctz->lock, flags);
626 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
628 struct mem_cgroup_tree_per_node *mctz;
629 struct mem_cgroup_per_node *mz;
633 mz = mem_cgroup_nodeinfo(memcg, nid);
634 mctz = soft_limit_tree_node(nid);
636 mem_cgroup_remove_exceeded(mz, mctz);
640 static struct mem_cgroup_per_node *
641 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
643 struct mem_cgroup_per_node *mz;
647 if (!mctz->rb_rightmost)
648 goto done; /* Nothing to reclaim from */
650 mz = rb_entry(mctz->rb_rightmost,
651 struct mem_cgroup_per_node, tree_node);
653 * Remove the node now but someone else can add it back,
654 * we will to add it back at the end of reclaim to its correct
655 * position in the tree.
657 __mem_cgroup_remove_exceeded(mz, mctz);
658 if (!soft_limit_excess(mz->memcg) ||
659 !css_tryget(&mz->memcg->css))
665 static struct mem_cgroup_per_node *
666 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
668 struct mem_cgroup_per_node *mz;
670 spin_lock_irq(&mctz->lock);
671 mz = __mem_cgroup_largest_soft_limit_node(mctz);
672 spin_unlock_irq(&mctz->lock);
677 * __mod_memcg_state - update cgroup memory statistics
678 * @memcg: the memory cgroup
679 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
680 * @val: delta to add to the counter, can be negative
682 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
686 if (mem_cgroup_disabled())
689 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
690 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
691 struct mem_cgroup *mi;
694 * Batch local counters to keep them in sync with
695 * the hierarchical ones.
697 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
698 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
699 atomic_long_add(x, &mi->vmstats[idx]);
702 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
705 static struct mem_cgroup_per_node *
706 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
708 struct mem_cgroup *parent;
710 parent = parent_mem_cgroup(pn->memcg);
713 return mem_cgroup_nodeinfo(parent, nid);
717 * __mod_lruvec_state - update lruvec memory statistics
718 * @lruvec: the lruvec
719 * @idx: the stat item
720 * @val: delta to add to the counter, can be negative
722 * The lruvec is the intersection of the NUMA node and a cgroup. This
723 * function updates the all three counters that are affected by a
724 * change of state at this level: per-node, per-cgroup, per-lruvec.
726 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
729 pg_data_t *pgdat = lruvec_pgdat(lruvec);
730 struct mem_cgroup_per_node *pn;
731 struct mem_cgroup *memcg;
735 __mod_node_page_state(pgdat, idx, val);
737 if (mem_cgroup_disabled())
740 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
744 __mod_memcg_state(memcg, idx, val);
747 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
749 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
750 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
751 struct mem_cgroup_per_node *pi;
753 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
754 atomic_long_add(x, &pi->lruvec_stat[idx]);
757 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
760 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
762 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
763 struct mem_cgroup *memcg;
764 struct lruvec *lruvec;
767 memcg = mem_cgroup_from_obj(p);
769 /* Untracked pages have no memcg, no lruvec. Update only the node */
770 if (!memcg || memcg == root_mem_cgroup) {
771 __mod_node_page_state(pgdat, idx, val);
773 lruvec = mem_cgroup_lruvec(memcg, pgdat);
774 __mod_lruvec_state(lruvec, idx, val);
779 void mod_memcg_obj_state(void *p, int idx, int val)
781 struct mem_cgroup *memcg;
784 memcg = mem_cgroup_from_obj(p);
786 mod_memcg_state(memcg, idx, val);
791 * __count_memcg_events - account VM events in a cgroup
792 * @memcg: the memory cgroup
793 * @idx: the event item
794 * @count: the number of events that occured
796 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
801 if (mem_cgroup_disabled())
804 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
805 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
806 struct mem_cgroup *mi;
809 * Batch local counters to keep them in sync with
810 * the hierarchical ones.
812 __this_cpu_add(memcg->vmstats_local->events[idx], x);
813 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
814 atomic_long_add(x, &mi->vmevents[idx]);
817 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
820 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
822 return atomic_long_read(&memcg->vmevents[event]);
825 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
830 for_each_possible_cpu(cpu)
831 x += per_cpu(memcg->vmstats_local->events[event], cpu);
835 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
839 if (abs(nr_pages) > 1) {
840 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
841 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
844 /* pagein of a big page is an event. So, ignore page size */
846 __count_memcg_events(memcg, PGPGIN, 1);
848 __count_memcg_events(memcg, PGPGOUT, 1);
849 nr_pages = -nr_pages; /* for event */
852 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
855 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
856 enum mem_cgroup_events_target target)
858 unsigned long val, next;
860 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
861 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
862 /* from time_after() in jiffies.h */
863 if ((long)(next - val) < 0) {
865 case MEM_CGROUP_TARGET_THRESH:
866 next = val + THRESHOLDS_EVENTS_TARGET;
868 case MEM_CGROUP_TARGET_SOFTLIMIT:
869 next = val + SOFTLIMIT_EVENTS_TARGET;
874 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
881 * Check events in order.
884 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
886 /* threshold event is triggered in finer grain than soft limit */
887 if (unlikely(mem_cgroup_event_ratelimit(memcg,
888 MEM_CGROUP_TARGET_THRESH))) {
891 do_softlimit = mem_cgroup_event_ratelimit(memcg,
892 MEM_CGROUP_TARGET_SOFTLIMIT);
893 mem_cgroup_threshold(memcg);
894 if (unlikely(do_softlimit))
895 mem_cgroup_update_tree(memcg, page);
899 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
902 * mm_update_next_owner() may clear mm->owner to NULL
903 * if it races with swapoff, page migration, etc.
904 * So this can be called with p == NULL.
909 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
911 EXPORT_SYMBOL(mem_cgroup_from_task);
914 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
915 * @mm: mm from which memcg should be extracted. It can be NULL.
917 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
918 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
921 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
923 struct mem_cgroup *memcg;
925 if (mem_cgroup_disabled())
931 * Page cache insertions can happen withou an
932 * actual mm context, e.g. during disk probing
933 * on boot, loopback IO, acct() writes etc.
936 memcg = root_mem_cgroup;
938 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
939 if (unlikely(!memcg))
940 memcg = root_mem_cgroup;
942 } while (!css_tryget(&memcg->css));
946 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
949 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
950 * @page: page from which memcg should be extracted.
952 * Obtain a reference on page->memcg and returns it if successful. Otherwise
953 * root_mem_cgroup is returned.
955 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
957 struct mem_cgroup *memcg = page->mem_cgroup;
959 if (mem_cgroup_disabled())
963 /* Page should not get uncharged and freed memcg under us. */
964 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
965 memcg = root_mem_cgroup;
969 EXPORT_SYMBOL(get_mem_cgroup_from_page);
972 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
974 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
976 if (unlikely(current->active_memcg)) {
977 struct mem_cgroup *memcg;
980 /* current->active_memcg must hold a ref. */
981 if (WARN_ON_ONCE(!css_tryget(¤t->active_memcg->css)))
982 memcg = root_mem_cgroup;
984 memcg = current->active_memcg;
988 return get_mem_cgroup_from_mm(current->mm);
992 * mem_cgroup_iter - iterate over memory cgroup hierarchy
993 * @root: hierarchy root
994 * @prev: previously returned memcg, NULL on first invocation
995 * @reclaim: cookie for shared reclaim walks, NULL for full walks
997 * Returns references to children of the hierarchy below @root, or
998 * @root itself, or %NULL after a full round-trip.
1000 * Caller must pass the return value in @prev on subsequent
1001 * invocations for reference counting, or use mem_cgroup_iter_break()
1002 * to cancel a hierarchy walk before the round-trip is complete.
1004 * Reclaimers can specify a node and a priority level in @reclaim to
1005 * divide up the memcgs in the hierarchy among all concurrent
1006 * reclaimers operating on the same node and priority.
1008 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1009 struct mem_cgroup *prev,
1010 struct mem_cgroup_reclaim_cookie *reclaim)
1012 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1013 struct cgroup_subsys_state *css = NULL;
1014 struct mem_cgroup *memcg = NULL;
1015 struct mem_cgroup *pos = NULL;
1017 if (mem_cgroup_disabled())
1021 root = root_mem_cgroup;
1023 if (prev && !reclaim)
1026 if (!root->use_hierarchy && root != root_mem_cgroup) {
1035 struct mem_cgroup_per_node *mz;
1037 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1040 if (prev && reclaim->generation != iter->generation)
1044 pos = READ_ONCE(iter->position);
1045 if (!pos || css_tryget(&pos->css))
1048 * css reference reached zero, so iter->position will
1049 * be cleared by ->css_released. However, we should not
1050 * rely on this happening soon, because ->css_released
1051 * is called from a work queue, and by busy-waiting we
1052 * might block it. So we clear iter->position right
1055 (void)cmpxchg(&iter->position, pos, NULL);
1063 css = css_next_descendant_pre(css, &root->css);
1066 * Reclaimers share the hierarchy walk, and a
1067 * new one might jump in right at the end of
1068 * the hierarchy - make sure they see at least
1069 * one group and restart from the beginning.
1077 * Verify the css and acquire a reference. The root
1078 * is provided by the caller, so we know it's alive
1079 * and kicking, and don't take an extra reference.
1081 memcg = mem_cgroup_from_css(css);
1083 if (css == &root->css)
1086 if (css_tryget(css))
1094 * The position could have already been updated by a competing
1095 * thread, so check that the value hasn't changed since we read
1096 * it to avoid reclaiming from the same cgroup twice.
1098 (void)cmpxchg(&iter->position, pos, memcg);
1106 reclaim->generation = iter->generation;
1112 if (prev && prev != root)
1113 css_put(&prev->css);
1119 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1120 * @root: hierarchy root
1121 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1123 void mem_cgroup_iter_break(struct mem_cgroup *root,
1124 struct mem_cgroup *prev)
1127 root = root_mem_cgroup;
1128 if (prev && prev != root)
1129 css_put(&prev->css);
1132 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1133 struct mem_cgroup *dead_memcg)
1135 struct mem_cgroup_reclaim_iter *iter;
1136 struct mem_cgroup_per_node *mz;
1139 for_each_node(nid) {
1140 mz = mem_cgroup_nodeinfo(from, nid);
1142 cmpxchg(&iter->position, dead_memcg, NULL);
1146 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1148 struct mem_cgroup *memcg = dead_memcg;
1149 struct mem_cgroup *last;
1152 __invalidate_reclaim_iterators(memcg, dead_memcg);
1154 } while ((memcg = parent_mem_cgroup(memcg)));
1157 * When cgruop1 non-hierarchy mode is used,
1158 * parent_mem_cgroup() does not walk all the way up to the
1159 * cgroup root (root_mem_cgroup). So we have to handle
1160 * dead_memcg from cgroup root separately.
1162 if (last != root_mem_cgroup)
1163 __invalidate_reclaim_iterators(root_mem_cgroup,
1168 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1169 * @memcg: hierarchy root
1170 * @fn: function to call for each task
1171 * @arg: argument passed to @fn
1173 * This function iterates over tasks attached to @memcg or to any of its
1174 * descendants and calls @fn for each task. If @fn returns a non-zero
1175 * value, the function breaks the iteration loop and returns the value.
1176 * Otherwise, it will iterate over all tasks and return 0.
1178 * This function must not be called for the root memory cgroup.
1180 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1181 int (*fn)(struct task_struct *, void *), void *arg)
1183 struct mem_cgroup *iter;
1186 BUG_ON(memcg == root_mem_cgroup);
1188 for_each_mem_cgroup_tree(iter, memcg) {
1189 struct css_task_iter it;
1190 struct task_struct *task;
1192 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1193 while (!ret && (task = css_task_iter_next(&it)))
1194 ret = fn(task, arg);
1195 css_task_iter_end(&it);
1197 mem_cgroup_iter_break(memcg, iter);
1205 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1207 * @pgdat: pgdat of the page
1209 * This function is only safe when following the LRU page isolation
1210 * and putback protocol: the LRU lock must be held, and the page must
1211 * either be PageLRU() or the caller must have isolated/allocated it.
1213 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1215 struct mem_cgroup_per_node *mz;
1216 struct mem_cgroup *memcg;
1217 struct lruvec *lruvec;
1219 if (mem_cgroup_disabled()) {
1220 lruvec = &pgdat->__lruvec;
1224 memcg = page->mem_cgroup;
1226 * Swapcache readahead pages are added to the LRU - and
1227 * possibly migrated - before they are charged.
1230 memcg = root_mem_cgroup;
1232 mz = mem_cgroup_page_nodeinfo(memcg, page);
1233 lruvec = &mz->lruvec;
1236 * Since a node can be onlined after the mem_cgroup was created,
1237 * we have to be prepared to initialize lruvec->zone here;
1238 * and if offlined then reonlined, we need to reinitialize it.
1240 if (unlikely(lruvec->pgdat != pgdat))
1241 lruvec->pgdat = pgdat;
1246 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1247 * @lruvec: mem_cgroup per zone lru vector
1248 * @lru: index of lru list the page is sitting on
1249 * @zid: zone id of the accounted pages
1250 * @nr_pages: positive when adding or negative when removing
1252 * This function must be called under lru_lock, just before a page is added
1253 * to or just after a page is removed from an lru list (that ordering being
1254 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1256 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1257 int zid, int nr_pages)
1259 struct mem_cgroup_per_node *mz;
1260 unsigned long *lru_size;
1263 if (mem_cgroup_disabled())
1266 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1267 lru_size = &mz->lru_zone_size[zid][lru];
1270 *lru_size += nr_pages;
1273 if (WARN_ONCE(size < 0,
1274 "%s(%p, %d, %d): lru_size %ld\n",
1275 __func__, lruvec, lru, nr_pages, size)) {
1281 *lru_size += nr_pages;
1285 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1286 * @memcg: the memory cgroup
1288 * Returns the maximum amount of memory @mem can be charged with, in
1291 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1293 unsigned long margin = 0;
1294 unsigned long count;
1295 unsigned long limit;
1297 count = page_counter_read(&memcg->memory);
1298 limit = READ_ONCE(memcg->memory.max);
1300 margin = limit - count;
1302 if (do_memsw_account()) {
1303 count = page_counter_read(&memcg->memsw);
1304 limit = READ_ONCE(memcg->memsw.max);
1306 margin = min(margin, limit - count);
1315 * A routine for checking "mem" is under move_account() or not.
1317 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1318 * moving cgroups. This is for waiting at high-memory pressure
1321 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1323 struct mem_cgroup *from;
1324 struct mem_cgroup *to;
1327 * Unlike task_move routines, we access mc.to, mc.from not under
1328 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1330 spin_lock(&mc.lock);
1336 ret = mem_cgroup_is_descendant(from, memcg) ||
1337 mem_cgroup_is_descendant(to, memcg);
1339 spin_unlock(&mc.lock);
1343 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1345 if (mc.moving_task && current != mc.moving_task) {
1346 if (mem_cgroup_under_move(memcg)) {
1348 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1349 /* moving charge context might have finished. */
1352 finish_wait(&mc.waitq, &wait);
1359 static char *memory_stat_format(struct mem_cgroup *memcg)
1364 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1369 * Provide statistics on the state of the memory subsystem as
1370 * well as cumulative event counters that show past behavior.
1372 * This list is ordered following a combination of these gradients:
1373 * 1) generic big picture -> specifics and details
1374 * 2) reflecting userspace activity -> reflecting kernel heuristics
1376 * Current memory state:
1379 seq_buf_printf(&s, "anon %llu\n",
1380 (u64)memcg_page_state(memcg, NR_ANON_MAPPED) *
1382 seq_buf_printf(&s, "file %llu\n",
1383 (u64)memcg_page_state(memcg, NR_FILE_PAGES) *
1385 seq_buf_printf(&s, "kernel_stack %llu\n",
1386 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) *
1388 seq_buf_printf(&s, "slab %llu\n",
1389 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) +
1390 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE)) *
1392 seq_buf_printf(&s, "sock %llu\n",
1393 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1396 seq_buf_printf(&s, "shmem %llu\n",
1397 (u64)memcg_page_state(memcg, NR_SHMEM) *
1399 seq_buf_printf(&s, "file_mapped %llu\n",
1400 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1402 seq_buf_printf(&s, "file_dirty %llu\n",
1403 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1405 seq_buf_printf(&s, "file_writeback %llu\n",
1406 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1410 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1411 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1412 * arse because it requires migrating the work out of rmap to a place
1413 * where the page->mem_cgroup is set up and stable.
1415 seq_buf_printf(&s, "anon_thp %llu\n",
1416 (u64)memcg_page_state(memcg, MEMCG_RSS_HUGE) *
1419 for (i = 0; i < NR_LRU_LISTS; i++)
1420 seq_buf_printf(&s, "%s %llu\n", lru_list_name(i),
1421 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1424 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1425 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) *
1427 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1428 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE) *
1431 /* Accumulated memory events */
1433 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1434 memcg_events(memcg, PGFAULT));
1435 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1436 memcg_events(memcg, PGMAJFAULT));
1438 seq_buf_printf(&s, "workingset_refault %lu\n",
1439 memcg_page_state(memcg, WORKINGSET_REFAULT));
1440 seq_buf_printf(&s, "workingset_activate %lu\n",
1441 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1442 seq_buf_printf(&s, "workingset_restore %lu\n",
1443 memcg_page_state(memcg, WORKINGSET_RESTORE));
1444 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1445 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1447 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1448 memcg_events(memcg, PGREFILL));
1449 seq_buf_printf(&s, "pgscan %lu\n",
1450 memcg_events(memcg, PGSCAN_KSWAPD) +
1451 memcg_events(memcg, PGSCAN_DIRECT));
1452 seq_buf_printf(&s, "pgsteal %lu\n",
1453 memcg_events(memcg, PGSTEAL_KSWAPD) +
1454 memcg_events(memcg, PGSTEAL_DIRECT));
1455 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1456 memcg_events(memcg, PGACTIVATE));
1457 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1458 memcg_events(memcg, PGDEACTIVATE));
1459 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1460 memcg_events(memcg, PGLAZYFREE));
1461 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1462 memcg_events(memcg, PGLAZYFREED));
1464 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1465 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1466 memcg_events(memcg, THP_FAULT_ALLOC));
1467 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1468 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1469 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1471 /* The above should easily fit into one page */
1472 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1477 #define K(x) ((x) << (PAGE_SHIFT-10))
1479 * mem_cgroup_print_oom_context: Print OOM information relevant to
1480 * memory controller.
1481 * @memcg: The memory cgroup that went over limit
1482 * @p: Task that is going to be killed
1484 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1487 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1492 pr_cont(",oom_memcg=");
1493 pr_cont_cgroup_path(memcg->css.cgroup);
1495 pr_cont(",global_oom");
1497 pr_cont(",task_memcg=");
1498 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1504 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1505 * memory controller.
1506 * @memcg: The memory cgroup that went over limit
1508 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1512 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1513 K((u64)page_counter_read(&memcg->memory)),
1514 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1515 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1516 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1517 K((u64)page_counter_read(&memcg->swap)),
1518 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1520 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1521 K((u64)page_counter_read(&memcg->memsw)),
1522 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1523 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1524 K((u64)page_counter_read(&memcg->kmem)),
1525 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1528 pr_info("Memory cgroup stats for ");
1529 pr_cont_cgroup_path(memcg->css.cgroup);
1531 buf = memory_stat_format(memcg);
1539 * Return the memory (and swap, if configured) limit for a memcg.
1541 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1545 max = READ_ONCE(memcg->memory.max);
1546 if (mem_cgroup_swappiness(memcg)) {
1547 unsigned long memsw_max;
1548 unsigned long swap_max;
1550 memsw_max = memcg->memsw.max;
1551 swap_max = READ_ONCE(memcg->swap.max);
1552 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1553 max = min(max + swap_max, memsw_max);
1558 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1560 return page_counter_read(&memcg->memory);
1563 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1566 struct oom_control oc = {
1570 .gfp_mask = gfp_mask,
1575 if (mutex_lock_killable(&oom_lock))
1578 * A few threads which were not waiting at mutex_lock_killable() can
1579 * fail to bail out. Therefore, check again after holding oom_lock.
1581 ret = should_force_charge() || out_of_memory(&oc);
1582 mutex_unlock(&oom_lock);
1586 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1589 unsigned long *total_scanned)
1591 struct mem_cgroup *victim = NULL;
1594 unsigned long excess;
1595 unsigned long nr_scanned;
1596 struct mem_cgroup_reclaim_cookie reclaim = {
1600 excess = soft_limit_excess(root_memcg);
1603 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1608 * If we have not been able to reclaim
1609 * anything, it might because there are
1610 * no reclaimable pages under this hierarchy
1615 * We want to do more targeted reclaim.
1616 * excess >> 2 is not to excessive so as to
1617 * reclaim too much, nor too less that we keep
1618 * coming back to reclaim from this cgroup
1620 if (total >= (excess >> 2) ||
1621 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1626 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1627 pgdat, &nr_scanned);
1628 *total_scanned += nr_scanned;
1629 if (!soft_limit_excess(root_memcg))
1632 mem_cgroup_iter_break(root_memcg, victim);
1636 #ifdef CONFIG_LOCKDEP
1637 static struct lockdep_map memcg_oom_lock_dep_map = {
1638 .name = "memcg_oom_lock",
1642 static DEFINE_SPINLOCK(memcg_oom_lock);
1645 * Check OOM-Killer is already running under our hierarchy.
1646 * If someone is running, return false.
1648 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1650 struct mem_cgroup *iter, *failed = NULL;
1652 spin_lock(&memcg_oom_lock);
1654 for_each_mem_cgroup_tree(iter, memcg) {
1655 if (iter->oom_lock) {
1657 * this subtree of our hierarchy is already locked
1658 * so we cannot give a lock.
1661 mem_cgroup_iter_break(memcg, iter);
1664 iter->oom_lock = true;
1669 * OK, we failed to lock the whole subtree so we have
1670 * to clean up what we set up to the failing subtree
1672 for_each_mem_cgroup_tree(iter, memcg) {
1673 if (iter == failed) {
1674 mem_cgroup_iter_break(memcg, iter);
1677 iter->oom_lock = false;
1680 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1682 spin_unlock(&memcg_oom_lock);
1687 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1689 struct mem_cgroup *iter;
1691 spin_lock(&memcg_oom_lock);
1692 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1693 for_each_mem_cgroup_tree(iter, memcg)
1694 iter->oom_lock = false;
1695 spin_unlock(&memcg_oom_lock);
1698 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1700 struct mem_cgroup *iter;
1702 spin_lock(&memcg_oom_lock);
1703 for_each_mem_cgroup_tree(iter, memcg)
1705 spin_unlock(&memcg_oom_lock);
1708 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1710 struct mem_cgroup *iter;
1713 * When a new child is created while the hierarchy is under oom,
1714 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1716 spin_lock(&memcg_oom_lock);
1717 for_each_mem_cgroup_tree(iter, memcg)
1718 if (iter->under_oom > 0)
1720 spin_unlock(&memcg_oom_lock);
1723 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1725 struct oom_wait_info {
1726 struct mem_cgroup *memcg;
1727 wait_queue_entry_t wait;
1730 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1731 unsigned mode, int sync, void *arg)
1733 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1734 struct mem_cgroup *oom_wait_memcg;
1735 struct oom_wait_info *oom_wait_info;
1737 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1738 oom_wait_memcg = oom_wait_info->memcg;
1740 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1741 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1743 return autoremove_wake_function(wait, mode, sync, arg);
1746 static void memcg_oom_recover(struct mem_cgroup *memcg)
1749 * For the following lockless ->under_oom test, the only required
1750 * guarantee is that it must see the state asserted by an OOM when
1751 * this function is called as a result of userland actions
1752 * triggered by the notification of the OOM. This is trivially
1753 * achieved by invoking mem_cgroup_mark_under_oom() before
1754 * triggering notification.
1756 if (memcg && memcg->under_oom)
1757 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1767 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1769 enum oom_status ret;
1772 if (order > PAGE_ALLOC_COSTLY_ORDER)
1775 memcg_memory_event(memcg, MEMCG_OOM);
1778 * We are in the middle of the charge context here, so we
1779 * don't want to block when potentially sitting on a callstack
1780 * that holds all kinds of filesystem and mm locks.
1782 * cgroup1 allows disabling the OOM killer and waiting for outside
1783 * handling until the charge can succeed; remember the context and put
1784 * the task to sleep at the end of the page fault when all locks are
1787 * On the other hand, in-kernel OOM killer allows for an async victim
1788 * memory reclaim (oom_reaper) and that means that we are not solely
1789 * relying on the oom victim to make a forward progress and we can
1790 * invoke the oom killer here.
1792 * Please note that mem_cgroup_out_of_memory might fail to find a
1793 * victim and then we have to bail out from the charge path.
1795 if (memcg->oom_kill_disable) {
1796 if (!current->in_user_fault)
1798 css_get(&memcg->css);
1799 current->memcg_in_oom = memcg;
1800 current->memcg_oom_gfp_mask = mask;
1801 current->memcg_oom_order = order;
1806 mem_cgroup_mark_under_oom(memcg);
1808 locked = mem_cgroup_oom_trylock(memcg);
1811 mem_cgroup_oom_notify(memcg);
1813 mem_cgroup_unmark_under_oom(memcg);
1814 if (mem_cgroup_out_of_memory(memcg, mask, order))
1820 mem_cgroup_oom_unlock(memcg);
1826 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1827 * @handle: actually kill/wait or just clean up the OOM state
1829 * This has to be called at the end of a page fault if the memcg OOM
1830 * handler was enabled.
1832 * Memcg supports userspace OOM handling where failed allocations must
1833 * sleep on a waitqueue until the userspace task resolves the
1834 * situation. Sleeping directly in the charge context with all kinds
1835 * of locks held is not a good idea, instead we remember an OOM state
1836 * in the task and mem_cgroup_oom_synchronize() has to be called at
1837 * the end of the page fault to complete the OOM handling.
1839 * Returns %true if an ongoing memcg OOM situation was detected and
1840 * completed, %false otherwise.
1842 bool mem_cgroup_oom_synchronize(bool handle)
1844 struct mem_cgroup *memcg = current->memcg_in_oom;
1845 struct oom_wait_info owait;
1848 /* OOM is global, do not handle */
1855 owait.memcg = memcg;
1856 owait.wait.flags = 0;
1857 owait.wait.func = memcg_oom_wake_function;
1858 owait.wait.private = current;
1859 INIT_LIST_HEAD(&owait.wait.entry);
1861 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1862 mem_cgroup_mark_under_oom(memcg);
1864 locked = mem_cgroup_oom_trylock(memcg);
1867 mem_cgroup_oom_notify(memcg);
1869 if (locked && !memcg->oom_kill_disable) {
1870 mem_cgroup_unmark_under_oom(memcg);
1871 finish_wait(&memcg_oom_waitq, &owait.wait);
1872 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1873 current->memcg_oom_order);
1876 mem_cgroup_unmark_under_oom(memcg);
1877 finish_wait(&memcg_oom_waitq, &owait.wait);
1881 mem_cgroup_oom_unlock(memcg);
1883 * There is no guarantee that an OOM-lock contender
1884 * sees the wakeups triggered by the OOM kill
1885 * uncharges. Wake any sleepers explicitely.
1887 memcg_oom_recover(memcg);
1890 current->memcg_in_oom = NULL;
1891 css_put(&memcg->css);
1896 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1897 * @victim: task to be killed by the OOM killer
1898 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1900 * Returns a pointer to a memory cgroup, which has to be cleaned up
1901 * by killing all belonging OOM-killable tasks.
1903 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1905 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1906 struct mem_cgroup *oom_domain)
1908 struct mem_cgroup *oom_group = NULL;
1909 struct mem_cgroup *memcg;
1911 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1915 oom_domain = root_mem_cgroup;
1919 memcg = mem_cgroup_from_task(victim);
1920 if (memcg == root_mem_cgroup)
1924 * If the victim task has been asynchronously moved to a different
1925 * memory cgroup, we might end up killing tasks outside oom_domain.
1926 * In this case it's better to ignore memory.group.oom.
1928 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
1932 * Traverse the memory cgroup hierarchy from the victim task's
1933 * cgroup up to the OOMing cgroup (or root) to find the
1934 * highest-level memory cgroup with oom.group set.
1936 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1937 if (memcg->oom_group)
1940 if (memcg == oom_domain)
1945 css_get(&oom_group->css);
1952 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1954 pr_info("Tasks in ");
1955 pr_cont_cgroup_path(memcg->css.cgroup);
1956 pr_cont(" are going to be killed due to memory.oom.group set\n");
1960 * lock_page_memcg - lock a page->mem_cgroup binding
1963 * This function protects unlocked LRU pages from being moved to
1966 * It ensures lifetime of the returned memcg. Caller is responsible
1967 * for the lifetime of the page; __unlock_page_memcg() is available
1968 * when @page might get freed inside the locked section.
1970 struct mem_cgroup *lock_page_memcg(struct page *page)
1972 struct page *head = compound_head(page); /* rmap on tail pages */
1973 struct mem_cgroup *memcg;
1974 unsigned long flags;
1977 * The RCU lock is held throughout the transaction. The fast
1978 * path can get away without acquiring the memcg->move_lock
1979 * because page moving starts with an RCU grace period.
1981 * The RCU lock also protects the memcg from being freed when
1982 * the page state that is going to change is the only thing
1983 * preventing the page itself from being freed. E.g. writeback
1984 * doesn't hold a page reference and relies on PG_writeback to
1985 * keep off truncation, migration and so forth.
1989 if (mem_cgroup_disabled())
1992 memcg = head->mem_cgroup;
1993 if (unlikely(!memcg))
1996 if (atomic_read(&memcg->moving_account) <= 0)
1999 spin_lock_irqsave(&memcg->move_lock, flags);
2000 if (memcg != head->mem_cgroup) {
2001 spin_unlock_irqrestore(&memcg->move_lock, flags);
2006 * When charge migration first begins, we can have locked and
2007 * unlocked page stat updates happening concurrently. Track
2008 * the task who has the lock for unlock_page_memcg().
2010 memcg->move_lock_task = current;
2011 memcg->move_lock_flags = flags;
2015 EXPORT_SYMBOL(lock_page_memcg);
2018 * __unlock_page_memcg - unlock and unpin a memcg
2021 * Unlock and unpin a memcg returned by lock_page_memcg().
2023 void __unlock_page_memcg(struct mem_cgroup *memcg)
2025 if (memcg && memcg->move_lock_task == current) {
2026 unsigned long flags = memcg->move_lock_flags;
2028 memcg->move_lock_task = NULL;
2029 memcg->move_lock_flags = 0;
2031 spin_unlock_irqrestore(&memcg->move_lock, flags);
2038 * unlock_page_memcg - unlock a page->mem_cgroup binding
2041 void unlock_page_memcg(struct page *page)
2043 struct page *head = compound_head(page);
2045 __unlock_page_memcg(head->mem_cgroup);
2047 EXPORT_SYMBOL(unlock_page_memcg);
2049 struct memcg_stock_pcp {
2050 struct mem_cgroup *cached; /* this never be root cgroup */
2051 unsigned int nr_pages;
2052 struct work_struct work;
2053 unsigned long flags;
2054 #define FLUSHING_CACHED_CHARGE 0
2056 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2057 static DEFINE_MUTEX(percpu_charge_mutex);
2060 * consume_stock: Try to consume stocked charge on this cpu.
2061 * @memcg: memcg to consume from.
2062 * @nr_pages: how many pages to charge.
2064 * The charges will only happen if @memcg matches the current cpu's memcg
2065 * stock, and at least @nr_pages are available in that stock. Failure to
2066 * service an allocation will refill the stock.
2068 * returns true if successful, false otherwise.
2070 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2072 struct memcg_stock_pcp *stock;
2073 unsigned long flags;
2076 if (nr_pages > MEMCG_CHARGE_BATCH)
2079 local_irq_save(flags);
2081 stock = this_cpu_ptr(&memcg_stock);
2082 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2083 stock->nr_pages -= nr_pages;
2087 local_irq_restore(flags);
2093 * Returns stocks cached in percpu and reset cached information.
2095 static void drain_stock(struct memcg_stock_pcp *stock)
2097 struct mem_cgroup *old = stock->cached;
2099 if (stock->nr_pages) {
2100 page_counter_uncharge(&old->memory, stock->nr_pages);
2101 if (do_memsw_account())
2102 page_counter_uncharge(&old->memsw, stock->nr_pages);
2103 css_put_many(&old->css, stock->nr_pages);
2104 stock->nr_pages = 0;
2106 stock->cached = NULL;
2109 static void drain_local_stock(struct work_struct *dummy)
2111 struct memcg_stock_pcp *stock;
2112 unsigned long flags;
2115 * The only protection from memory hotplug vs. drain_stock races is
2116 * that we always operate on local CPU stock here with IRQ disabled
2118 local_irq_save(flags);
2120 stock = this_cpu_ptr(&memcg_stock);
2122 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2124 local_irq_restore(flags);
2128 * Cache charges(val) to local per_cpu area.
2129 * This will be consumed by consume_stock() function, later.
2131 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2133 struct memcg_stock_pcp *stock;
2134 unsigned long flags;
2136 local_irq_save(flags);
2138 stock = this_cpu_ptr(&memcg_stock);
2139 if (stock->cached != memcg) { /* reset if necessary */
2141 stock->cached = memcg;
2143 stock->nr_pages += nr_pages;
2145 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2148 local_irq_restore(flags);
2152 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2153 * of the hierarchy under it.
2155 static void drain_all_stock(struct mem_cgroup *root_memcg)
2159 /* If someone's already draining, avoid adding running more workers. */
2160 if (!mutex_trylock(&percpu_charge_mutex))
2163 * Notify other cpus that system-wide "drain" is running
2164 * We do not care about races with the cpu hotplug because cpu down
2165 * as well as workers from this path always operate on the local
2166 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2169 for_each_online_cpu(cpu) {
2170 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2171 struct mem_cgroup *memcg;
2175 memcg = stock->cached;
2176 if (memcg && stock->nr_pages &&
2177 mem_cgroup_is_descendant(memcg, root_memcg))
2182 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2184 drain_local_stock(&stock->work);
2186 schedule_work_on(cpu, &stock->work);
2190 mutex_unlock(&percpu_charge_mutex);
2193 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2195 struct memcg_stock_pcp *stock;
2196 struct mem_cgroup *memcg, *mi;
2198 stock = &per_cpu(memcg_stock, cpu);
2201 for_each_mem_cgroup(memcg) {
2204 for (i = 0; i < MEMCG_NR_STAT; i++) {
2208 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2210 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2211 atomic_long_add(x, &memcg->vmstats[i]);
2213 if (i >= NR_VM_NODE_STAT_ITEMS)
2216 for_each_node(nid) {
2217 struct mem_cgroup_per_node *pn;
2219 pn = mem_cgroup_nodeinfo(memcg, nid);
2220 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2223 atomic_long_add(x, &pn->lruvec_stat[i]);
2224 } while ((pn = parent_nodeinfo(pn, nid)));
2228 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2231 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2233 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2234 atomic_long_add(x, &memcg->vmevents[i]);
2241 static void reclaim_high(struct mem_cgroup *memcg,
2242 unsigned int nr_pages,
2246 if (page_counter_read(&memcg->memory) <=
2247 READ_ONCE(memcg->memory.high))
2249 memcg_memory_event(memcg, MEMCG_HIGH);
2250 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2251 } while ((memcg = parent_mem_cgroup(memcg)) &&
2252 !mem_cgroup_is_root(memcg));
2255 static void high_work_func(struct work_struct *work)
2257 struct mem_cgroup *memcg;
2259 memcg = container_of(work, struct mem_cgroup, high_work);
2260 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2264 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2265 * enough to still cause a significant slowdown in most cases, while still
2266 * allowing diagnostics and tracing to proceed without becoming stuck.
2268 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2271 * When calculating the delay, we use these either side of the exponentiation to
2272 * maintain precision and scale to a reasonable number of jiffies (see the table
2275 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2276 * overage ratio to a delay.
2277 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2278 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2279 * to produce a reasonable delay curve.
2281 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2282 * reasonable delay curve compared to precision-adjusted overage, not
2283 * penalising heavily at first, but still making sure that growth beyond the
2284 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2285 * example, with a high of 100 megabytes:
2287 * +-------+------------------------+
2288 * | usage | time to allocate in ms |
2289 * +-------+------------------------+
2311 * +-------+------------------------+
2313 #define MEMCG_DELAY_PRECISION_SHIFT 20
2314 #define MEMCG_DELAY_SCALING_SHIFT 14
2316 static u64 calculate_overage(unsigned long usage, unsigned long high)
2324 * Prevent division by 0 in overage calculation by acting as if
2325 * it was a threshold of 1 page
2327 high = max(high, 1UL);
2329 overage = usage - high;
2330 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2331 return div64_u64(overage, high);
2334 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2336 u64 overage, max_overage = 0;
2339 overage = calculate_overage(page_counter_read(&memcg->memory),
2340 READ_ONCE(memcg->memory.high));
2341 max_overage = max(overage, max_overage);
2342 } while ((memcg = parent_mem_cgroup(memcg)) &&
2343 !mem_cgroup_is_root(memcg));
2348 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2350 u64 overage, max_overage = 0;
2353 overage = calculate_overage(page_counter_read(&memcg->swap),
2354 READ_ONCE(memcg->swap.high));
2356 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2357 max_overage = max(overage, max_overage);
2358 } while ((memcg = parent_mem_cgroup(memcg)) &&
2359 !mem_cgroup_is_root(memcg));
2365 * Get the number of jiffies that we should penalise a mischievous cgroup which
2366 * is exceeding its memory.high by checking both it and its ancestors.
2368 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2369 unsigned int nr_pages,
2372 unsigned long penalty_jiffies;
2378 * We use overage compared to memory.high to calculate the number of
2379 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2380 * fairly lenient on small overages, and increasingly harsh when the
2381 * memcg in question makes it clear that it has no intention of stopping
2382 * its crazy behaviour, so we exponentially increase the delay based on
2385 penalty_jiffies = max_overage * max_overage * HZ;
2386 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2387 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2390 * Factor in the task's own contribution to the overage, such that four
2391 * N-sized allocations are throttled approximately the same as one
2392 * 4N-sized allocation.
2394 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2395 * larger the current charge patch is than that.
2397 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2401 * Scheduled by try_charge() to be executed from the userland return path
2402 * and reclaims memory over the high limit.
2404 void mem_cgroup_handle_over_high(void)
2406 unsigned long penalty_jiffies;
2407 unsigned long pflags;
2408 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2409 struct mem_cgroup *memcg;
2411 if (likely(!nr_pages))
2414 memcg = get_mem_cgroup_from_mm(current->mm);
2415 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2416 current->memcg_nr_pages_over_high = 0;
2419 * memory.high is breached and reclaim is unable to keep up. Throttle
2420 * allocators proactively to slow down excessive growth.
2422 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2423 mem_find_max_overage(memcg));
2425 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2426 swap_find_max_overage(memcg));
2429 * Clamp the max delay per usermode return so as to still keep the
2430 * application moving forwards and also permit diagnostics, albeit
2433 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2436 * Don't sleep if the amount of jiffies this memcg owes us is so low
2437 * that it's not even worth doing, in an attempt to be nice to those who
2438 * go only a small amount over their memory.high value and maybe haven't
2439 * been aggressively reclaimed enough yet.
2441 if (penalty_jiffies <= HZ / 100)
2445 * If we exit early, we're guaranteed to die (since
2446 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2447 * need to account for any ill-begotten jiffies to pay them off later.
2449 psi_memstall_enter(&pflags);
2450 schedule_timeout_killable(penalty_jiffies);
2451 psi_memstall_leave(&pflags);
2454 css_put(&memcg->css);
2457 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2458 unsigned int nr_pages)
2460 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2461 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2462 struct mem_cgroup *mem_over_limit;
2463 struct page_counter *counter;
2464 unsigned long nr_reclaimed;
2465 bool may_swap = true;
2466 bool drained = false;
2467 enum oom_status oom_status;
2469 if (mem_cgroup_is_root(memcg))
2472 if (consume_stock(memcg, nr_pages))
2475 if (!do_memsw_account() ||
2476 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2477 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2479 if (do_memsw_account())
2480 page_counter_uncharge(&memcg->memsw, batch);
2481 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2483 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2487 if (batch > nr_pages) {
2493 * Memcg doesn't have a dedicated reserve for atomic
2494 * allocations. But like the global atomic pool, we need to
2495 * put the burden of reclaim on regular allocation requests
2496 * and let these go through as privileged allocations.
2498 if (gfp_mask & __GFP_ATOMIC)
2502 * Unlike in global OOM situations, memcg is not in a physical
2503 * memory shortage. Allow dying and OOM-killed tasks to
2504 * bypass the last charges so that they can exit quickly and
2505 * free their memory.
2507 if (unlikely(should_force_charge()))
2511 * Prevent unbounded recursion when reclaim operations need to
2512 * allocate memory. This might exceed the limits temporarily,
2513 * but we prefer facilitating memory reclaim and getting back
2514 * under the limit over triggering OOM kills in these cases.
2516 if (unlikely(current->flags & PF_MEMALLOC))
2519 if (unlikely(task_in_memcg_oom(current)))
2522 if (!gfpflags_allow_blocking(gfp_mask))
2525 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2527 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2528 gfp_mask, may_swap);
2530 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2534 drain_all_stock(mem_over_limit);
2539 if (gfp_mask & __GFP_NORETRY)
2542 * Even though the limit is exceeded at this point, reclaim
2543 * may have been able to free some pages. Retry the charge
2544 * before killing the task.
2546 * Only for regular pages, though: huge pages are rather
2547 * unlikely to succeed so close to the limit, and we fall back
2548 * to regular pages anyway in case of failure.
2550 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2553 * At task move, charge accounts can be doubly counted. So, it's
2554 * better to wait until the end of task_move if something is going on.
2556 if (mem_cgroup_wait_acct_move(mem_over_limit))
2562 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2565 if (gfp_mask & __GFP_NOFAIL)
2568 if (fatal_signal_pending(current))
2572 * keep retrying as long as the memcg oom killer is able to make
2573 * a forward progress or bypass the charge if the oom killer
2574 * couldn't make any progress.
2576 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2577 get_order(nr_pages * PAGE_SIZE));
2578 switch (oom_status) {
2580 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2588 if (!(gfp_mask & __GFP_NOFAIL))
2592 * The allocation either can't fail or will lead to more memory
2593 * being freed very soon. Allow memory usage go over the limit
2594 * temporarily by force charging it.
2596 page_counter_charge(&memcg->memory, nr_pages);
2597 if (do_memsw_account())
2598 page_counter_charge(&memcg->memsw, nr_pages);
2599 css_get_many(&memcg->css, nr_pages);
2604 css_get_many(&memcg->css, batch);
2605 if (batch > nr_pages)
2606 refill_stock(memcg, batch - nr_pages);
2609 * If the hierarchy is above the normal consumption range, schedule
2610 * reclaim on returning to userland. We can perform reclaim here
2611 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2612 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2613 * not recorded as it most likely matches current's and won't
2614 * change in the meantime. As high limit is checked again before
2615 * reclaim, the cost of mismatch is negligible.
2618 bool mem_high, swap_high;
2620 mem_high = page_counter_read(&memcg->memory) >
2621 READ_ONCE(memcg->memory.high);
2622 swap_high = page_counter_read(&memcg->swap) >
2623 READ_ONCE(memcg->swap.high);
2625 /* Don't bother a random interrupted task */
2626 if (in_interrupt()) {
2628 schedule_work(&memcg->high_work);
2634 if (mem_high || swap_high) {
2636 * The allocating tasks in this cgroup will need to do
2637 * reclaim or be throttled to prevent further growth
2638 * of the memory or swap footprints.
2640 * Target some best-effort fairness between the tasks,
2641 * and distribute reclaim work and delay penalties
2642 * based on how much each task is actually allocating.
2644 current->memcg_nr_pages_over_high += batch;
2645 set_notify_resume(current);
2648 } while ((memcg = parent_mem_cgroup(memcg)));
2653 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2655 if (mem_cgroup_is_root(memcg))
2658 page_counter_uncharge(&memcg->memory, nr_pages);
2659 if (do_memsw_account())
2660 page_counter_uncharge(&memcg->memsw, nr_pages);
2662 css_put_many(&memcg->css, nr_pages);
2665 static void lock_page_lru(struct page *page, int *isolated)
2667 pg_data_t *pgdat = page_pgdat(page);
2669 spin_lock_irq(&pgdat->lru_lock);
2670 if (PageLRU(page)) {
2671 struct lruvec *lruvec;
2673 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2675 del_page_from_lru_list(page, lruvec, page_lru(page));
2681 static void unlock_page_lru(struct page *page, int isolated)
2683 pg_data_t *pgdat = page_pgdat(page);
2686 struct lruvec *lruvec;
2688 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2689 VM_BUG_ON_PAGE(PageLRU(page), page);
2691 add_page_to_lru_list(page, lruvec, page_lru(page));
2693 spin_unlock_irq(&pgdat->lru_lock);
2696 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2701 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2704 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2705 * may already be on some other mem_cgroup's LRU. Take care of it.
2708 lock_page_lru(page, &isolated);
2711 * Nobody should be changing or seriously looking at
2712 * page->mem_cgroup at this point:
2714 * - the page is uncharged
2716 * - the page is off-LRU
2718 * - an anonymous fault has exclusive page access, except for
2719 * a locked page table
2721 * - a page cache insertion, a swapin fault, or a migration
2722 * have the page locked
2724 page->mem_cgroup = memcg;
2727 unlock_page_lru(page, isolated);
2730 #ifdef CONFIG_MEMCG_KMEM
2732 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2734 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2735 * cgroup_mutex, etc.
2737 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2741 if (mem_cgroup_disabled())
2744 page = virt_to_head_page(p);
2747 * Slab pages don't have page->mem_cgroup set because corresponding
2748 * kmem caches can be reparented during the lifetime. That's why
2749 * memcg_from_slab_page() should be used instead.
2752 return memcg_from_slab_page(page);
2754 /* All other pages use page->mem_cgroup */
2755 return page->mem_cgroup;
2758 static int memcg_alloc_cache_id(void)
2763 id = ida_simple_get(&memcg_cache_ida,
2764 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2768 if (id < memcg_nr_cache_ids)
2772 * There's no space for the new id in memcg_caches arrays,
2773 * so we have to grow them.
2775 down_write(&memcg_cache_ids_sem);
2777 size = 2 * (id + 1);
2778 if (size < MEMCG_CACHES_MIN_SIZE)
2779 size = MEMCG_CACHES_MIN_SIZE;
2780 else if (size > MEMCG_CACHES_MAX_SIZE)
2781 size = MEMCG_CACHES_MAX_SIZE;
2783 err = memcg_update_all_caches(size);
2785 err = memcg_update_all_list_lrus(size);
2787 memcg_nr_cache_ids = size;
2789 up_write(&memcg_cache_ids_sem);
2792 ida_simple_remove(&memcg_cache_ida, id);
2798 static void memcg_free_cache_id(int id)
2800 ida_simple_remove(&memcg_cache_ida, id);
2803 struct memcg_kmem_cache_create_work {
2804 struct mem_cgroup *memcg;
2805 struct kmem_cache *cachep;
2806 struct work_struct work;
2809 static void memcg_kmem_cache_create_func(struct work_struct *w)
2811 struct memcg_kmem_cache_create_work *cw =
2812 container_of(w, struct memcg_kmem_cache_create_work, work);
2813 struct mem_cgroup *memcg = cw->memcg;
2814 struct kmem_cache *cachep = cw->cachep;
2816 memcg_create_kmem_cache(memcg, cachep);
2818 css_put(&memcg->css);
2823 * Enqueue the creation of a per-memcg kmem_cache.
2825 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2826 struct kmem_cache *cachep)
2828 struct memcg_kmem_cache_create_work *cw;
2830 if (!css_tryget_online(&memcg->css))
2833 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2838 cw->cachep = cachep;
2839 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2841 queue_work(memcg_kmem_cache_wq, &cw->work);
2844 static inline bool memcg_kmem_bypass(void)
2849 /* Allow remote memcg charging in kthread contexts. */
2850 if ((!current->mm || (current->flags & PF_KTHREAD)) &&
2851 !current->active_memcg)
2857 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2858 * @cachep: the original global kmem cache
2860 * Return the kmem_cache we're supposed to use for a slab allocation.
2861 * We try to use the current memcg's version of the cache.
2863 * If the cache does not exist yet, if we are the first user of it, we
2864 * create it asynchronously in a workqueue and let the current allocation
2865 * go through with the original cache.
2867 * This function takes a reference to the cache it returns to assure it
2868 * won't get destroyed while we are working with it. Once the caller is
2869 * done with it, memcg_kmem_put_cache() must be called to release the
2872 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2874 struct mem_cgroup *memcg;
2875 struct kmem_cache *memcg_cachep;
2876 struct memcg_cache_array *arr;
2879 VM_BUG_ON(!is_root_cache(cachep));
2881 if (memcg_kmem_bypass())
2886 if (unlikely(current->active_memcg))
2887 memcg = current->active_memcg;
2889 memcg = mem_cgroup_from_task(current);
2891 if (!memcg || memcg == root_mem_cgroup)
2894 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2898 arr = rcu_dereference(cachep->memcg_params.memcg_caches);
2901 * Make sure we will access the up-to-date value. The code updating
2902 * memcg_caches issues a write barrier to match the data dependency
2903 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2905 memcg_cachep = READ_ONCE(arr->entries[kmemcg_id]);
2908 * If we are in a safe context (can wait, and not in interrupt
2909 * context), we could be be predictable and return right away.
2910 * This would guarantee that the allocation being performed
2911 * already belongs in the new cache.
2913 * However, there are some clashes that can arrive from locking.
2914 * For instance, because we acquire the slab_mutex while doing
2915 * memcg_create_kmem_cache, this means no further allocation
2916 * could happen with the slab_mutex held. So it's better to
2919 * If the memcg is dying or memcg_cache is about to be released,
2920 * don't bother creating new kmem_caches. Because memcg_cachep
2921 * is ZEROed as the fist step of kmem offlining, we don't need
2922 * percpu_ref_tryget_live() here. css_tryget_online() check in
2923 * memcg_schedule_kmem_cache_create() will prevent us from
2924 * creation of a new kmem_cache.
2926 if (unlikely(!memcg_cachep))
2927 memcg_schedule_kmem_cache_create(memcg, cachep);
2928 else if (percpu_ref_tryget(&memcg_cachep->memcg_params.refcnt))
2929 cachep = memcg_cachep;
2936 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2937 * @cachep: the cache returned by memcg_kmem_get_cache
2939 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2941 if (!is_root_cache(cachep))
2942 percpu_ref_put(&cachep->memcg_params.refcnt);
2946 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
2947 * @memcg: memory cgroup to charge
2948 * @gfp: reclaim mode
2949 * @nr_pages: number of pages to charge
2951 * Returns 0 on success, an error code on failure.
2953 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
2954 unsigned int nr_pages)
2956 struct page_counter *counter;
2959 ret = try_charge(memcg, gfp, nr_pages);
2963 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2964 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2967 * Enforce __GFP_NOFAIL allocation because callers are not
2968 * prepared to see failures and likely do not have any failure
2971 if (gfp & __GFP_NOFAIL) {
2972 page_counter_charge(&memcg->kmem, nr_pages);
2975 cancel_charge(memcg, nr_pages);
2982 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
2983 * @memcg: memcg to uncharge
2984 * @nr_pages: number of pages to uncharge
2986 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
2988 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2989 page_counter_uncharge(&memcg->kmem, nr_pages);
2991 page_counter_uncharge(&memcg->memory, nr_pages);
2992 if (do_memsw_account())
2993 page_counter_uncharge(&memcg->memsw, nr_pages);
2997 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
2998 * @page: page to charge
2999 * @gfp: reclaim mode
3000 * @order: allocation order
3002 * Returns 0 on success, an error code on failure.
3004 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3006 struct mem_cgroup *memcg;
3009 if (memcg_kmem_bypass())
3012 memcg = get_mem_cgroup_from_current();
3013 if (!mem_cgroup_is_root(memcg)) {
3014 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3016 page->mem_cgroup = memcg;
3017 __SetPageKmemcg(page);
3020 css_put(&memcg->css);
3025 * __memcg_kmem_uncharge_page: uncharge a kmem page
3026 * @page: page to uncharge
3027 * @order: allocation order
3029 void __memcg_kmem_uncharge_page(struct page *page, int order)
3031 struct mem_cgroup *memcg = page->mem_cgroup;
3032 unsigned int nr_pages = 1 << order;
3037 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3038 __memcg_kmem_uncharge(memcg, nr_pages);
3039 page->mem_cgroup = NULL;
3041 /* slab pages do not have PageKmemcg flag set */
3042 if (PageKmemcg(page))
3043 __ClearPageKmemcg(page);
3045 css_put_many(&memcg->css, nr_pages);
3047 #endif /* CONFIG_MEMCG_KMEM */
3049 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3052 * Because tail pages are not marked as "used", set it. We're under
3053 * pgdat->lru_lock and migration entries setup in all page mappings.
3055 void mem_cgroup_split_huge_fixup(struct page *head)
3059 if (mem_cgroup_disabled())
3062 for (i = 1; i < HPAGE_PMD_NR; i++)
3063 head[i].mem_cgroup = head->mem_cgroup;
3065 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
3067 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3069 #ifdef CONFIG_MEMCG_SWAP
3071 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3072 * @entry: swap entry to be moved
3073 * @from: mem_cgroup which the entry is moved from
3074 * @to: mem_cgroup which the entry is moved to
3076 * It succeeds only when the swap_cgroup's record for this entry is the same
3077 * as the mem_cgroup's id of @from.
3079 * Returns 0 on success, -EINVAL on failure.
3081 * The caller must have charged to @to, IOW, called page_counter_charge() about
3082 * both res and memsw, and called css_get().
3084 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3085 struct mem_cgroup *from, struct mem_cgroup *to)
3087 unsigned short old_id, new_id;
3089 old_id = mem_cgroup_id(from);
3090 new_id = mem_cgroup_id(to);
3092 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3093 mod_memcg_state(from, MEMCG_SWAP, -1);
3094 mod_memcg_state(to, MEMCG_SWAP, 1);
3100 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3101 struct mem_cgroup *from, struct mem_cgroup *to)
3107 static DEFINE_MUTEX(memcg_max_mutex);
3109 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3110 unsigned long max, bool memsw)
3112 bool enlarge = false;
3113 bool drained = false;
3115 bool limits_invariant;
3116 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3119 if (signal_pending(current)) {
3124 mutex_lock(&memcg_max_mutex);
3126 * Make sure that the new limit (memsw or memory limit) doesn't
3127 * break our basic invariant rule memory.max <= memsw.max.
3129 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3130 max <= memcg->memsw.max;
3131 if (!limits_invariant) {
3132 mutex_unlock(&memcg_max_mutex);
3136 if (max > counter->max)
3138 ret = page_counter_set_max(counter, max);
3139 mutex_unlock(&memcg_max_mutex);
3145 drain_all_stock(memcg);
3150 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3151 GFP_KERNEL, !memsw)) {
3157 if (!ret && enlarge)
3158 memcg_oom_recover(memcg);
3163 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3165 unsigned long *total_scanned)
3167 unsigned long nr_reclaimed = 0;
3168 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3169 unsigned long reclaimed;
3171 struct mem_cgroup_tree_per_node *mctz;
3172 unsigned long excess;
3173 unsigned long nr_scanned;
3178 mctz = soft_limit_tree_node(pgdat->node_id);
3181 * Do not even bother to check the largest node if the root
3182 * is empty. Do it lockless to prevent lock bouncing. Races
3183 * are acceptable as soft limit is best effort anyway.
3185 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3189 * This loop can run a while, specially if mem_cgroup's continuously
3190 * keep exceeding their soft limit and putting the system under
3197 mz = mem_cgroup_largest_soft_limit_node(mctz);
3202 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3203 gfp_mask, &nr_scanned);
3204 nr_reclaimed += reclaimed;
3205 *total_scanned += nr_scanned;
3206 spin_lock_irq(&mctz->lock);
3207 __mem_cgroup_remove_exceeded(mz, mctz);
3210 * If we failed to reclaim anything from this memory cgroup
3211 * it is time to move on to the next cgroup
3215 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3217 excess = soft_limit_excess(mz->memcg);
3219 * One school of thought says that we should not add
3220 * back the node to the tree if reclaim returns 0.
3221 * But our reclaim could return 0, simply because due
3222 * to priority we are exposing a smaller subset of
3223 * memory to reclaim from. Consider this as a longer
3226 /* If excess == 0, no tree ops */
3227 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3228 spin_unlock_irq(&mctz->lock);
3229 css_put(&mz->memcg->css);
3232 * Could not reclaim anything and there are no more
3233 * mem cgroups to try or we seem to be looping without
3234 * reclaiming anything.
3236 if (!nr_reclaimed &&
3238 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3240 } while (!nr_reclaimed);
3242 css_put(&next_mz->memcg->css);
3243 return nr_reclaimed;
3247 * Test whether @memcg has children, dead or alive. Note that this
3248 * function doesn't care whether @memcg has use_hierarchy enabled and
3249 * returns %true if there are child csses according to the cgroup
3250 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3252 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3257 ret = css_next_child(NULL, &memcg->css);
3263 * Reclaims as many pages from the given memcg as possible.
3265 * Caller is responsible for holding css reference for memcg.
3267 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3269 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3271 /* we call try-to-free pages for make this cgroup empty */
3272 lru_add_drain_all();
3274 drain_all_stock(memcg);
3276 /* try to free all pages in this cgroup */
3277 while (nr_retries && page_counter_read(&memcg->memory)) {
3280 if (signal_pending(current))
3283 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3287 /* maybe some writeback is necessary */
3288 congestion_wait(BLK_RW_ASYNC, HZ/10);
3296 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3297 char *buf, size_t nbytes,
3300 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3302 if (mem_cgroup_is_root(memcg))
3304 return mem_cgroup_force_empty(memcg) ?: nbytes;
3307 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3310 return mem_cgroup_from_css(css)->use_hierarchy;
3313 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3314 struct cftype *cft, u64 val)
3317 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3318 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3320 if (memcg->use_hierarchy == val)
3324 * If parent's use_hierarchy is set, we can't make any modifications
3325 * in the child subtrees. If it is unset, then the change can
3326 * occur, provided the current cgroup has no children.
3328 * For the root cgroup, parent_mem is NULL, we allow value to be
3329 * set if there are no children.
3331 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3332 (val == 1 || val == 0)) {
3333 if (!memcg_has_children(memcg))
3334 memcg->use_hierarchy = val;
3343 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3347 if (mem_cgroup_is_root(memcg)) {
3348 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3349 memcg_page_state(memcg, NR_ANON_MAPPED);
3351 val += memcg_page_state(memcg, MEMCG_SWAP);
3354 val = page_counter_read(&memcg->memory);
3356 val = page_counter_read(&memcg->memsw);
3369 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3372 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3373 struct page_counter *counter;
3375 switch (MEMFILE_TYPE(cft->private)) {
3377 counter = &memcg->memory;
3380 counter = &memcg->memsw;
3383 counter = &memcg->kmem;
3386 counter = &memcg->tcpmem;
3392 switch (MEMFILE_ATTR(cft->private)) {
3394 if (counter == &memcg->memory)
3395 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3396 if (counter == &memcg->memsw)
3397 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3398 return (u64)page_counter_read(counter) * PAGE_SIZE;
3400 return (u64)counter->max * PAGE_SIZE;
3402 return (u64)counter->watermark * PAGE_SIZE;
3404 return counter->failcnt;
3405 case RES_SOFT_LIMIT:
3406 return (u64)memcg->soft_limit * PAGE_SIZE;
3412 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3414 unsigned long stat[MEMCG_NR_STAT] = {0};
3415 struct mem_cgroup *mi;
3418 for_each_online_cpu(cpu)
3419 for (i = 0; i < MEMCG_NR_STAT; i++)
3420 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3422 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3423 for (i = 0; i < MEMCG_NR_STAT; i++)
3424 atomic_long_add(stat[i], &mi->vmstats[i]);
3426 for_each_node(node) {
3427 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3428 struct mem_cgroup_per_node *pi;
3430 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3433 for_each_online_cpu(cpu)
3434 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3436 pn->lruvec_stat_cpu->count[i], cpu);
3438 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3439 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3440 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3444 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3446 unsigned long events[NR_VM_EVENT_ITEMS];
3447 struct mem_cgroup *mi;
3450 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3453 for_each_online_cpu(cpu)
3454 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3455 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3458 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3459 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3460 atomic_long_add(events[i], &mi->vmevents[i]);
3463 #ifdef CONFIG_MEMCG_KMEM
3464 static int memcg_online_kmem(struct mem_cgroup *memcg)
3468 if (cgroup_memory_nokmem)
3471 BUG_ON(memcg->kmemcg_id >= 0);
3472 BUG_ON(memcg->kmem_state);
3474 memcg_id = memcg_alloc_cache_id();
3478 static_branch_inc(&memcg_kmem_enabled_key);
3480 * A memory cgroup is considered kmem-online as soon as it gets
3481 * kmemcg_id. Setting the id after enabling static branching will
3482 * guarantee no one starts accounting before all call sites are
3485 memcg->kmemcg_id = memcg_id;
3486 memcg->kmem_state = KMEM_ONLINE;
3487 INIT_LIST_HEAD(&memcg->kmem_caches);
3492 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3494 struct cgroup_subsys_state *css;
3495 struct mem_cgroup *parent, *child;
3498 if (memcg->kmem_state != KMEM_ONLINE)
3501 * Clear the online state before clearing memcg_caches array
3502 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3503 * guarantees that no cache will be created for this cgroup
3504 * after we are done (see memcg_create_kmem_cache()).
3506 memcg->kmem_state = KMEM_ALLOCATED;
3508 parent = parent_mem_cgroup(memcg);
3510 parent = root_mem_cgroup;
3513 * Deactivate and reparent kmem_caches.
3515 memcg_deactivate_kmem_caches(memcg, parent);
3517 kmemcg_id = memcg->kmemcg_id;
3518 BUG_ON(kmemcg_id < 0);
3521 * Change kmemcg_id of this cgroup and all its descendants to the
3522 * parent's id, and then move all entries from this cgroup's list_lrus
3523 * to ones of the parent. After we have finished, all list_lrus
3524 * corresponding to this cgroup are guaranteed to remain empty. The
3525 * ordering is imposed by list_lru_node->lock taken by
3526 * memcg_drain_all_list_lrus().
3528 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3529 css_for_each_descendant_pre(css, &memcg->css) {
3530 child = mem_cgroup_from_css(css);
3531 BUG_ON(child->kmemcg_id != kmemcg_id);
3532 child->kmemcg_id = parent->kmemcg_id;
3533 if (!memcg->use_hierarchy)
3538 memcg_drain_all_list_lrus(kmemcg_id, parent);
3540 memcg_free_cache_id(kmemcg_id);
3543 static void memcg_free_kmem(struct mem_cgroup *memcg)
3545 /* css_alloc() failed, offlining didn't happen */
3546 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3547 memcg_offline_kmem(memcg);
3549 if (memcg->kmem_state == KMEM_ALLOCATED) {
3550 WARN_ON(!list_empty(&memcg->kmem_caches));
3551 static_branch_dec(&memcg_kmem_enabled_key);
3555 static int memcg_online_kmem(struct mem_cgroup *memcg)
3559 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3562 static void memcg_free_kmem(struct mem_cgroup *memcg)
3565 #endif /* CONFIG_MEMCG_KMEM */
3567 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3572 mutex_lock(&memcg_max_mutex);
3573 ret = page_counter_set_max(&memcg->kmem, max);
3574 mutex_unlock(&memcg_max_mutex);
3578 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3582 mutex_lock(&memcg_max_mutex);
3584 ret = page_counter_set_max(&memcg->tcpmem, max);
3588 if (!memcg->tcpmem_active) {
3590 * The active flag needs to be written after the static_key
3591 * update. This is what guarantees that the socket activation
3592 * function is the last one to run. See mem_cgroup_sk_alloc()
3593 * for details, and note that we don't mark any socket as
3594 * belonging to this memcg until that flag is up.
3596 * We need to do this, because static_keys will span multiple
3597 * sites, but we can't control their order. If we mark a socket
3598 * as accounted, but the accounting functions are not patched in
3599 * yet, we'll lose accounting.
3601 * We never race with the readers in mem_cgroup_sk_alloc(),
3602 * because when this value change, the code to process it is not
3605 static_branch_inc(&memcg_sockets_enabled_key);
3606 memcg->tcpmem_active = true;
3609 mutex_unlock(&memcg_max_mutex);
3614 * The user of this function is...
3617 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3618 char *buf, size_t nbytes, loff_t off)
3620 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3621 unsigned long nr_pages;
3624 buf = strstrip(buf);
3625 ret = page_counter_memparse(buf, "-1", &nr_pages);
3629 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3631 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3635 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3637 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3640 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3643 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3644 "Please report your usecase to linux-mm@kvack.org if you "
3645 "depend on this functionality.\n");
3646 ret = memcg_update_kmem_max(memcg, nr_pages);
3649 ret = memcg_update_tcp_max(memcg, nr_pages);
3653 case RES_SOFT_LIMIT:
3654 memcg->soft_limit = nr_pages;
3658 return ret ?: nbytes;
3661 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3662 size_t nbytes, loff_t off)
3664 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3665 struct page_counter *counter;
3667 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3669 counter = &memcg->memory;
3672 counter = &memcg->memsw;
3675 counter = &memcg->kmem;
3678 counter = &memcg->tcpmem;
3684 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3686 page_counter_reset_watermark(counter);
3689 counter->failcnt = 0;
3698 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3701 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3705 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3706 struct cftype *cft, u64 val)
3708 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3710 if (val & ~MOVE_MASK)
3714 * No kind of locking is needed in here, because ->can_attach() will
3715 * check this value once in the beginning of the process, and then carry
3716 * on with stale data. This means that changes to this value will only
3717 * affect task migrations starting after the change.
3719 memcg->move_charge_at_immigrate = val;
3723 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3724 struct cftype *cft, u64 val)
3732 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3733 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3734 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3736 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3737 int nid, unsigned int lru_mask, bool tree)
3739 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3740 unsigned long nr = 0;
3743 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3746 if (!(BIT(lru) & lru_mask))
3749 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3751 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3756 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3757 unsigned int lru_mask,
3760 unsigned long nr = 0;
3764 if (!(BIT(lru) & lru_mask))
3767 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3769 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3774 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3778 unsigned int lru_mask;
3781 static const struct numa_stat stats[] = {
3782 { "total", LRU_ALL },
3783 { "file", LRU_ALL_FILE },
3784 { "anon", LRU_ALL_ANON },
3785 { "unevictable", BIT(LRU_UNEVICTABLE) },
3787 const struct numa_stat *stat;
3789 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3791 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3792 seq_printf(m, "%s=%lu", stat->name,
3793 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3795 for_each_node_state(nid, N_MEMORY)
3796 seq_printf(m, " N%d=%lu", nid,
3797 mem_cgroup_node_nr_lru_pages(memcg, nid,
3798 stat->lru_mask, false));
3802 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3804 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3805 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3807 for_each_node_state(nid, N_MEMORY)
3808 seq_printf(m, " N%d=%lu", nid,
3809 mem_cgroup_node_nr_lru_pages(memcg, nid,
3810 stat->lru_mask, true));
3816 #endif /* CONFIG_NUMA */
3818 static const unsigned int memcg1_stats[] = {
3829 static const char *const memcg1_stat_names[] = {
3840 /* Universal VM events cgroup1 shows, original sort order */
3841 static const unsigned int memcg1_events[] = {
3848 static int memcg_stat_show(struct seq_file *m, void *v)
3850 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3851 unsigned long memory, memsw;
3852 struct mem_cgroup *mi;
3855 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3857 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3858 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3860 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3861 memcg_page_state_local(memcg, memcg1_stats[i]) *
3865 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3866 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
3867 memcg_events_local(memcg, memcg1_events[i]));
3869 for (i = 0; i < NR_LRU_LISTS; i++)
3870 seq_printf(m, "%s %lu\n", lru_list_name(i),
3871 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3874 /* Hierarchical information */
3875 memory = memsw = PAGE_COUNTER_MAX;
3876 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3877 memory = min(memory, READ_ONCE(mi->memory.max));
3878 memsw = min(memsw, READ_ONCE(mi->memsw.max));
3880 seq_printf(m, "hierarchical_memory_limit %llu\n",
3881 (u64)memory * PAGE_SIZE);
3882 if (do_memsw_account())
3883 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3884 (u64)memsw * PAGE_SIZE);
3886 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3887 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3889 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3890 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
3894 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3895 seq_printf(m, "total_%s %llu\n",
3896 vm_event_name(memcg1_events[i]),
3897 (u64)memcg_events(memcg, memcg1_events[i]));
3899 for (i = 0; i < NR_LRU_LISTS; i++)
3900 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
3901 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
3904 #ifdef CONFIG_DEBUG_VM
3907 struct mem_cgroup_per_node *mz;
3908 struct zone_reclaim_stat *rstat;
3909 unsigned long recent_rotated[2] = {0, 0};
3910 unsigned long recent_scanned[2] = {0, 0};
3912 for_each_online_pgdat(pgdat) {
3913 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3914 rstat = &mz->lruvec.reclaim_stat;
3916 recent_rotated[0] += rstat->recent_rotated[0];
3917 recent_rotated[1] += rstat->recent_rotated[1];
3918 recent_scanned[0] += rstat->recent_scanned[0];
3919 recent_scanned[1] += rstat->recent_scanned[1];
3921 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3922 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3923 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3924 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3931 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3934 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3936 return mem_cgroup_swappiness(memcg);
3939 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3940 struct cftype *cft, u64 val)
3942 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3948 memcg->swappiness = val;
3950 vm_swappiness = val;
3955 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3957 struct mem_cgroup_threshold_ary *t;
3958 unsigned long usage;
3963 t = rcu_dereference(memcg->thresholds.primary);
3965 t = rcu_dereference(memcg->memsw_thresholds.primary);
3970 usage = mem_cgroup_usage(memcg, swap);
3973 * current_threshold points to threshold just below or equal to usage.
3974 * If it's not true, a threshold was crossed after last
3975 * call of __mem_cgroup_threshold().
3977 i = t->current_threshold;
3980 * Iterate backward over array of thresholds starting from
3981 * current_threshold and check if a threshold is crossed.
3982 * If none of thresholds below usage is crossed, we read
3983 * only one element of the array here.
3985 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3986 eventfd_signal(t->entries[i].eventfd, 1);
3988 /* i = current_threshold + 1 */
3992 * Iterate forward over array of thresholds starting from
3993 * current_threshold+1 and check if a threshold is crossed.
3994 * If none of thresholds above usage is crossed, we read
3995 * only one element of the array here.
3997 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3998 eventfd_signal(t->entries[i].eventfd, 1);
4000 /* Update current_threshold */
4001 t->current_threshold = i - 1;
4006 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4009 __mem_cgroup_threshold(memcg, false);
4010 if (do_memsw_account())
4011 __mem_cgroup_threshold(memcg, true);
4013 memcg = parent_mem_cgroup(memcg);
4017 static int compare_thresholds(const void *a, const void *b)
4019 const struct mem_cgroup_threshold *_a = a;
4020 const struct mem_cgroup_threshold *_b = b;
4022 if (_a->threshold > _b->threshold)
4025 if (_a->threshold < _b->threshold)
4031 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4033 struct mem_cgroup_eventfd_list *ev;
4035 spin_lock(&memcg_oom_lock);
4037 list_for_each_entry(ev, &memcg->oom_notify, list)
4038 eventfd_signal(ev->eventfd, 1);
4040 spin_unlock(&memcg_oom_lock);
4044 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4046 struct mem_cgroup *iter;
4048 for_each_mem_cgroup_tree(iter, memcg)
4049 mem_cgroup_oom_notify_cb(iter);
4052 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4053 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4055 struct mem_cgroup_thresholds *thresholds;
4056 struct mem_cgroup_threshold_ary *new;
4057 unsigned long threshold;
4058 unsigned long usage;
4061 ret = page_counter_memparse(args, "-1", &threshold);
4065 mutex_lock(&memcg->thresholds_lock);
4068 thresholds = &memcg->thresholds;
4069 usage = mem_cgroup_usage(memcg, false);
4070 } else if (type == _MEMSWAP) {
4071 thresholds = &memcg->memsw_thresholds;
4072 usage = mem_cgroup_usage(memcg, true);
4076 /* Check if a threshold crossed before adding a new one */
4077 if (thresholds->primary)
4078 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4080 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4082 /* Allocate memory for new array of thresholds */
4083 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4090 /* Copy thresholds (if any) to new array */
4091 if (thresholds->primary) {
4092 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4093 sizeof(struct mem_cgroup_threshold));
4096 /* Add new threshold */
4097 new->entries[size - 1].eventfd = eventfd;
4098 new->entries[size - 1].threshold = threshold;
4100 /* Sort thresholds. Registering of new threshold isn't time-critical */
4101 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4102 compare_thresholds, NULL);
4104 /* Find current threshold */
4105 new->current_threshold = -1;
4106 for (i = 0; i < size; i++) {
4107 if (new->entries[i].threshold <= usage) {
4109 * new->current_threshold will not be used until
4110 * rcu_assign_pointer(), so it's safe to increment
4113 ++new->current_threshold;
4118 /* Free old spare buffer and save old primary buffer as spare */
4119 kfree(thresholds->spare);
4120 thresholds->spare = thresholds->primary;
4122 rcu_assign_pointer(thresholds->primary, new);
4124 /* To be sure that nobody uses thresholds */
4128 mutex_unlock(&memcg->thresholds_lock);
4133 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4134 struct eventfd_ctx *eventfd, const char *args)
4136 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4139 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4140 struct eventfd_ctx *eventfd, const char *args)
4142 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4145 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4146 struct eventfd_ctx *eventfd, enum res_type type)
4148 struct mem_cgroup_thresholds *thresholds;
4149 struct mem_cgroup_threshold_ary *new;
4150 unsigned long usage;
4151 int i, j, size, entries;
4153 mutex_lock(&memcg->thresholds_lock);
4156 thresholds = &memcg->thresholds;
4157 usage = mem_cgroup_usage(memcg, false);
4158 } else if (type == _MEMSWAP) {
4159 thresholds = &memcg->memsw_thresholds;
4160 usage = mem_cgroup_usage(memcg, true);
4164 if (!thresholds->primary)
4167 /* Check if a threshold crossed before removing */
4168 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4170 /* Calculate new number of threshold */
4172 for (i = 0; i < thresholds->primary->size; i++) {
4173 if (thresholds->primary->entries[i].eventfd != eventfd)
4179 new = thresholds->spare;
4181 /* If no items related to eventfd have been cleared, nothing to do */
4185 /* Set thresholds array to NULL if we don't have thresholds */
4194 /* Copy thresholds and find current threshold */
4195 new->current_threshold = -1;
4196 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4197 if (thresholds->primary->entries[i].eventfd == eventfd)
4200 new->entries[j] = thresholds->primary->entries[i];
4201 if (new->entries[j].threshold <= usage) {
4203 * new->current_threshold will not be used
4204 * until rcu_assign_pointer(), so it's safe to increment
4207 ++new->current_threshold;
4213 /* Swap primary and spare array */
4214 thresholds->spare = thresholds->primary;
4216 rcu_assign_pointer(thresholds->primary, new);
4218 /* To be sure that nobody uses thresholds */
4221 /* If all events are unregistered, free the spare array */
4223 kfree(thresholds->spare);
4224 thresholds->spare = NULL;
4227 mutex_unlock(&memcg->thresholds_lock);
4230 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4231 struct eventfd_ctx *eventfd)
4233 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4236 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4237 struct eventfd_ctx *eventfd)
4239 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4242 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4243 struct eventfd_ctx *eventfd, const char *args)
4245 struct mem_cgroup_eventfd_list *event;
4247 event = kmalloc(sizeof(*event), GFP_KERNEL);
4251 spin_lock(&memcg_oom_lock);
4253 event->eventfd = eventfd;
4254 list_add(&event->list, &memcg->oom_notify);
4256 /* already in OOM ? */
4257 if (memcg->under_oom)
4258 eventfd_signal(eventfd, 1);
4259 spin_unlock(&memcg_oom_lock);
4264 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4265 struct eventfd_ctx *eventfd)
4267 struct mem_cgroup_eventfd_list *ev, *tmp;
4269 spin_lock(&memcg_oom_lock);
4271 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4272 if (ev->eventfd == eventfd) {
4273 list_del(&ev->list);
4278 spin_unlock(&memcg_oom_lock);
4281 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4283 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4285 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4286 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4287 seq_printf(sf, "oom_kill %lu\n",
4288 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4292 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4293 struct cftype *cft, u64 val)
4295 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4297 /* cannot set to root cgroup and only 0 and 1 are allowed */
4298 if (!css->parent || !((val == 0) || (val == 1)))
4301 memcg->oom_kill_disable = val;
4303 memcg_oom_recover(memcg);
4308 #ifdef CONFIG_CGROUP_WRITEBACK
4310 #include <trace/events/writeback.h>
4312 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4314 return wb_domain_init(&memcg->cgwb_domain, gfp);
4317 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4319 wb_domain_exit(&memcg->cgwb_domain);
4322 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4324 wb_domain_size_changed(&memcg->cgwb_domain);
4327 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4329 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4331 if (!memcg->css.parent)
4334 return &memcg->cgwb_domain;
4338 * idx can be of type enum memcg_stat_item or node_stat_item.
4339 * Keep in sync with memcg_exact_page().
4341 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4343 long x = atomic_long_read(&memcg->vmstats[idx]);
4346 for_each_online_cpu(cpu)
4347 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4354 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4355 * @wb: bdi_writeback in question
4356 * @pfilepages: out parameter for number of file pages
4357 * @pheadroom: out parameter for number of allocatable pages according to memcg
4358 * @pdirty: out parameter for number of dirty pages
4359 * @pwriteback: out parameter for number of pages under writeback
4361 * Determine the numbers of file, headroom, dirty, and writeback pages in
4362 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4363 * is a bit more involved.
4365 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4366 * headroom is calculated as the lowest headroom of itself and the
4367 * ancestors. Note that this doesn't consider the actual amount of
4368 * available memory in the system. The caller should further cap
4369 * *@pheadroom accordingly.
4371 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4372 unsigned long *pheadroom, unsigned long *pdirty,
4373 unsigned long *pwriteback)
4375 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4376 struct mem_cgroup *parent;
4378 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4380 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4381 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4382 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4383 *pheadroom = PAGE_COUNTER_MAX;
4385 while ((parent = parent_mem_cgroup(memcg))) {
4386 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4387 READ_ONCE(memcg->memory.high));
4388 unsigned long used = page_counter_read(&memcg->memory);
4390 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4396 * Foreign dirty flushing
4398 * There's an inherent mismatch between memcg and writeback. The former
4399 * trackes ownership per-page while the latter per-inode. This was a
4400 * deliberate design decision because honoring per-page ownership in the
4401 * writeback path is complicated, may lead to higher CPU and IO overheads
4402 * and deemed unnecessary given that write-sharing an inode across
4403 * different cgroups isn't a common use-case.
4405 * Combined with inode majority-writer ownership switching, this works well
4406 * enough in most cases but there are some pathological cases. For
4407 * example, let's say there are two cgroups A and B which keep writing to
4408 * different but confined parts of the same inode. B owns the inode and
4409 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4410 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4411 * triggering background writeback. A will be slowed down without a way to
4412 * make writeback of the dirty pages happen.
4414 * Conditions like the above can lead to a cgroup getting repatedly and
4415 * severely throttled after making some progress after each
4416 * dirty_expire_interval while the underyling IO device is almost
4419 * Solving this problem completely requires matching the ownership tracking
4420 * granularities between memcg and writeback in either direction. However,
4421 * the more egregious behaviors can be avoided by simply remembering the
4422 * most recent foreign dirtying events and initiating remote flushes on
4423 * them when local writeback isn't enough to keep the memory clean enough.
4425 * The following two functions implement such mechanism. When a foreign
4426 * page - a page whose memcg and writeback ownerships don't match - is
4427 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4428 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4429 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4430 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4431 * foreign bdi_writebacks which haven't expired. Both the numbers of
4432 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4433 * limited to MEMCG_CGWB_FRN_CNT.
4435 * The mechanism only remembers IDs and doesn't hold any object references.
4436 * As being wrong occasionally doesn't matter, updates and accesses to the
4437 * records are lockless and racy.
4439 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4440 struct bdi_writeback *wb)
4442 struct mem_cgroup *memcg = page->mem_cgroup;
4443 struct memcg_cgwb_frn *frn;
4444 u64 now = get_jiffies_64();
4445 u64 oldest_at = now;
4449 trace_track_foreign_dirty(page, wb);
4452 * Pick the slot to use. If there is already a slot for @wb, keep
4453 * using it. If not replace the oldest one which isn't being
4456 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4457 frn = &memcg->cgwb_frn[i];
4458 if (frn->bdi_id == wb->bdi->id &&
4459 frn->memcg_id == wb->memcg_css->id)
4461 if (time_before64(frn->at, oldest_at) &&
4462 atomic_read(&frn->done.cnt) == 1) {
4464 oldest_at = frn->at;
4468 if (i < MEMCG_CGWB_FRN_CNT) {
4470 * Re-using an existing one. Update timestamp lazily to
4471 * avoid making the cacheline hot. We want them to be
4472 * reasonably up-to-date and significantly shorter than
4473 * dirty_expire_interval as that's what expires the record.
4474 * Use the shorter of 1s and dirty_expire_interval / 8.
4476 unsigned long update_intv =
4477 min_t(unsigned long, HZ,
4478 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4480 if (time_before64(frn->at, now - update_intv))
4482 } else if (oldest >= 0) {
4483 /* replace the oldest free one */
4484 frn = &memcg->cgwb_frn[oldest];
4485 frn->bdi_id = wb->bdi->id;
4486 frn->memcg_id = wb->memcg_css->id;
4491 /* issue foreign writeback flushes for recorded foreign dirtying events */
4492 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4494 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4495 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4496 u64 now = jiffies_64;
4499 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4500 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4503 * If the record is older than dirty_expire_interval,
4504 * writeback on it has already started. No need to kick it
4505 * off again. Also, don't start a new one if there's
4506 * already one in flight.
4508 if (time_after64(frn->at, now - intv) &&
4509 atomic_read(&frn->done.cnt) == 1) {
4511 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4512 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4513 WB_REASON_FOREIGN_FLUSH,
4519 #else /* CONFIG_CGROUP_WRITEBACK */
4521 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4526 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4530 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4534 #endif /* CONFIG_CGROUP_WRITEBACK */
4537 * DO NOT USE IN NEW FILES.
4539 * "cgroup.event_control" implementation.
4541 * This is way over-engineered. It tries to support fully configurable
4542 * events for each user. Such level of flexibility is completely
4543 * unnecessary especially in the light of the planned unified hierarchy.
4545 * Please deprecate this and replace with something simpler if at all
4550 * Unregister event and free resources.
4552 * Gets called from workqueue.
4554 static void memcg_event_remove(struct work_struct *work)
4556 struct mem_cgroup_event *event =
4557 container_of(work, struct mem_cgroup_event, remove);
4558 struct mem_cgroup *memcg = event->memcg;
4560 remove_wait_queue(event->wqh, &event->wait);
4562 event->unregister_event(memcg, event->eventfd);
4564 /* Notify userspace the event is going away. */
4565 eventfd_signal(event->eventfd, 1);
4567 eventfd_ctx_put(event->eventfd);
4569 css_put(&memcg->css);
4573 * Gets called on EPOLLHUP on eventfd when user closes it.
4575 * Called with wqh->lock held and interrupts disabled.
4577 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4578 int sync, void *key)
4580 struct mem_cgroup_event *event =
4581 container_of(wait, struct mem_cgroup_event, wait);
4582 struct mem_cgroup *memcg = event->memcg;
4583 __poll_t flags = key_to_poll(key);
4585 if (flags & EPOLLHUP) {
4587 * If the event has been detached at cgroup removal, we
4588 * can simply return knowing the other side will cleanup
4591 * We can't race against event freeing since the other
4592 * side will require wqh->lock via remove_wait_queue(),
4595 spin_lock(&memcg->event_list_lock);
4596 if (!list_empty(&event->list)) {
4597 list_del_init(&event->list);
4599 * We are in atomic context, but cgroup_event_remove()
4600 * may sleep, so we have to call it in workqueue.
4602 schedule_work(&event->remove);
4604 spin_unlock(&memcg->event_list_lock);
4610 static void memcg_event_ptable_queue_proc(struct file *file,
4611 wait_queue_head_t *wqh, poll_table *pt)
4613 struct mem_cgroup_event *event =
4614 container_of(pt, struct mem_cgroup_event, pt);
4617 add_wait_queue(wqh, &event->wait);
4621 * DO NOT USE IN NEW FILES.
4623 * Parse input and register new cgroup event handler.
4625 * Input must be in format '<event_fd> <control_fd> <args>'.
4626 * Interpretation of args is defined by control file implementation.
4628 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4629 char *buf, size_t nbytes, loff_t off)
4631 struct cgroup_subsys_state *css = of_css(of);
4632 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4633 struct mem_cgroup_event *event;
4634 struct cgroup_subsys_state *cfile_css;
4635 unsigned int efd, cfd;
4642 buf = strstrip(buf);
4644 efd = simple_strtoul(buf, &endp, 10);
4649 cfd = simple_strtoul(buf, &endp, 10);
4650 if ((*endp != ' ') && (*endp != '\0'))
4654 event = kzalloc(sizeof(*event), GFP_KERNEL);
4658 event->memcg = memcg;
4659 INIT_LIST_HEAD(&event->list);
4660 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4661 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4662 INIT_WORK(&event->remove, memcg_event_remove);
4670 event->eventfd = eventfd_ctx_fileget(efile.file);
4671 if (IS_ERR(event->eventfd)) {
4672 ret = PTR_ERR(event->eventfd);
4679 goto out_put_eventfd;
4682 /* the process need read permission on control file */
4683 /* AV: shouldn't we check that it's been opened for read instead? */
4684 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4689 * Determine the event callbacks and set them in @event. This used
4690 * to be done via struct cftype but cgroup core no longer knows
4691 * about these events. The following is crude but the whole thing
4692 * is for compatibility anyway.
4694 * DO NOT ADD NEW FILES.
4696 name = cfile.file->f_path.dentry->d_name.name;
4698 if (!strcmp(name, "memory.usage_in_bytes")) {
4699 event->register_event = mem_cgroup_usage_register_event;
4700 event->unregister_event = mem_cgroup_usage_unregister_event;
4701 } else if (!strcmp(name, "memory.oom_control")) {
4702 event->register_event = mem_cgroup_oom_register_event;
4703 event->unregister_event = mem_cgroup_oom_unregister_event;
4704 } else if (!strcmp(name, "memory.pressure_level")) {
4705 event->register_event = vmpressure_register_event;
4706 event->unregister_event = vmpressure_unregister_event;
4707 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4708 event->register_event = memsw_cgroup_usage_register_event;
4709 event->unregister_event = memsw_cgroup_usage_unregister_event;
4716 * Verify @cfile should belong to @css. Also, remaining events are
4717 * automatically removed on cgroup destruction but the removal is
4718 * asynchronous, so take an extra ref on @css.
4720 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4721 &memory_cgrp_subsys);
4723 if (IS_ERR(cfile_css))
4725 if (cfile_css != css) {
4730 ret = event->register_event(memcg, event->eventfd, buf);
4734 vfs_poll(efile.file, &event->pt);
4736 spin_lock(&memcg->event_list_lock);
4737 list_add(&event->list, &memcg->event_list);
4738 spin_unlock(&memcg->event_list_lock);
4750 eventfd_ctx_put(event->eventfd);
4759 static struct cftype mem_cgroup_legacy_files[] = {
4761 .name = "usage_in_bytes",
4762 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4763 .read_u64 = mem_cgroup_read_u64,
4766 .name = "max_usage_in_bytes",
4767 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4768 .write = mem_cgroup_reset,
4769 .read_u64 = mem_cgroup_read_u64,
4772 .name = "limit_in_bytes",
4773 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4774 .write = mem_cgroup_write,
4775 .read_u64 = mem_cgroup_read_u64,
4778 .name = "soft_limit_in_bytes",
4779 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4780 .write = mem_cgroup_write,
4781 .read_u64 = mem_cgroup_read_u64,
4785 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4786 .write = mem_cgroup_reset,
4787 .read_u64 = mem_cgroup_read_u64,
4791 .seq_show = memcg_stat_show,
4794 .name = "force_empty",
4795 .write = mem_cgroup_force_empty_write,
4798 .name = "use_hierarchy",
4799 .write_u64 = mem_cgroup_hierarchy_write,
4800 .read_u64 = mem_cgroup_hierarchy_read,
4803 .name = "cgroup.event_control", /* XXX: for compat */
4804 .write = memcg_write_event_control,
4805 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4808 .name = "swappiness",
4809 .read_u64 = mem_cgroup_swappiness_read,
4810 .write_u64 = mem_cgroup_swappiness_write,
4813 .name = "move_charge_at_immigrate",
4814 .read_u64 = mem_cgroup_move_charge_read,
4815 .write_u64 = mem_cgroup_move_charge_write,
4818 .name = "oom_control",
4819 .seq_show = mem_cgroup_oom_control_read,
4820 .write_u64 = mem_cgroup_oom_control_write,
4821 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4824 .name = "pressure_level",
4828 .name = "numa_stat",
4829 .seq_show = memcg_numa_stat_show,
4833 .name = "kmem.limit_in_bytes",
4834 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4835 .write = mem_cgroup_write,
4836 .read_u64 = mem_cgroup_read_u64,
4839 .name = "kmem.usage_in_bytes",
4840 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4841 .read_u64 = mem_cgroup_read_u64,
4844 .name = "kmem.failcnt",
4845 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4846 .write = mem_cgroup_reset,
4847 .read_u64 = mem_cgroup_read_u64,
4850 .name = "kmem.max_usage_in_bytes",
4851 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4852 .write = mem_cgroup_reset,
4853 .read_u64 = mem_cgroup_read_u64,
4855 #if defined(CONFIG_MEMCG_KMEM) && \
4856 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4858 .name = "kmem.slabinfo",
4859 .seq_start = memcg_slab_start,
4860 .seq_next = memcg_slab_next,
4861 .seq_stop = memcg_slab_stop,
4862 .seq_show = memcg_slab_show,
4866 .name = "kmem.tcp.limit_in_bytes",
4867 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4868 .write = mem_cgroup_write,
4869 .read_u64 = mem_cgroup_read_u64,
4872 .name = "kmem.tcp.usage_in_bytes",
4873 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4874 .read_u64 = mem_cgroup_read_u64,
4877 .name = "kmem.tcp.failcnt",
4878 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4879 .write = mem_cgroup_reset,
4880 .read_u64 = mem_cgroup_read_u64,
4883 .name = "kmem.tcp.max_usage_in_bytes",
4884 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4885 .write = mem_cgroup_reset,
4886 .read_u64 = mem_cgroup_read_u64,
4888 { }, /* terminate */
4892 * Private memory cgroup IDR
4894 * Swap-out records and page cache shadow entries need to store memcg
4895 * references in constrained space, so we maintain an ID space that is
4896 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4897 * memory-controlled cgroups to 64k.
4899 * However, there usually are many references to the oflline CSS after
4900 * the cgroup has been destroyed, such as page cache or reclaimable
4901 * slab objects, that don't need to hang on to the ID. We want to keep
4902 * those dead CSS from occupying IDs, or we might quickly exhaust the
4903 * relatively small ID space and prevent the creation of new cgroups
4904 * even when there are much fewer than 64k cgroups - possibly none.
4906 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4907 * be freed and recycled when it's no longer needed, which is usually
4908 * when the CSS is offlined.
4910 * The only exception to that are records of swapped out tmpfs/shmem
4911 * pages that need to be attributed to live ancestors on swapin. But
4912 * those references are manageable from userspace.
4915 static DEFINE_IDR(mem_cgroup_idr);
4917 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4919 if (memcg->id.id > 0) {
4920 idr_remove(&mem_cgroup_idr, memcg->id.id);
4925 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
4928 refcount_add(n, &memcg->id.ref);
4931 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4933 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4934 mem_cgroup_id_remove(memcg);
4936 /* Memcg ID pins CSS */
4937 css_put(&memcg->css);
4941 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4943 mem_cgroup_id_put_many(memcg, 1);
4947 * mem_cgroup_from_id - look up a memcg from a memcg id
4948 * @id: the memcg id to look up
4950 * Caller must hold rcu_read_lock().
4952 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4954 WARN_ON_ONCE(!rcu_read_lock_held());
4955 return idr_find(&mem_cgroup_idr, id);
4958 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4960 struct mem_cgroup_per_node *pn;
4963 * This routine is called against possible nodes.
4964 * But it's BUG to call kmalloc() against offline node.
4966 * TODO: this routine can waste much memory for nodes which will
4967 * never be onlined. It's better to use memory hotplug callback
4970 if (!node_state(node, N_NORMAL_MEMORY))
4972 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4976 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
4977 if (!pn->lruvec_stat_local) {
4982 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4983 if (!pn->lruvec_stat_cpu) {
4984 free_percpu(pn->lruvec_stat_local);
4989 lruvec_init(&pn->lruvec);
4990 pn->usage_in_excess = 0;
4991 pn->on_tree = false;
4994 memcg->nodeinfo[node] = pn;
4998 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5000 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5005 free_percpu(pn->lruvec_stat_cpu);
5006 free_percpu(pn->lruvec_stat_local);
5010 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5015 free_mem_cgroup_per_node_info(memcg, node);
5016 free_percpu(memcg->vmstats_percpu);
5017 free_percpu(memcg->vmstats_local);
5021 static void mem_cgroup_free(struct mem_cgroup *memcg)
5023 memcg_wb_domain_exit(memcg);
5025 * Flush percpu vmstats and vmevents to guarantee the value correctness
5026 * on parent's and all ancestor levels.
5028 memcg_flush_percpu_vmstats(memcg);
5029 memcg_flush_percpu_vmevents(memcg);
5030 __mem_cgroup_free(memcg);
5033 static struct mem_cgroup *mem_cgroup_alloc(void)
5035 struct mem_cgroup *memcg;
5038 int __maybe_unused i;
5039 long error = -ENOMEM;
5041 size = sizeof(struct mem_cgroup);
5042 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5044 memcg = kzalloc(size, GFP_KERNEL);
5046 return ERR_PTR(error);
5048 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5049 1, MEM_CGROUP_ID_MAX,
5051 if (memcg->id.id < 0) {
5052 error = memcg->id.id;
5056 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
5057 if (!memcg->vmstats_local)
5060 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
5061 if (!memcg->vmstats_percpu)
5065 if (alloc_mem_cgroup_per_node_info(memcg, node))
5068 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5071 INIT_WORK(&memcg->high_work, high_work_func);
5072 INIT_LIST_HEAD(&memcg->oom_notify);
5073 mutex_init(&memcg->thresholds_lock);
5074 spin_lock_init(&memcg->move_lock);
5075 vmpressure_init(&memcg->vmpressure);
5076 INIT_LIST_HEAD(&memcg->event_list);
5077 spin_lock_init(&memcg->event_list_lock);
5078 memcg->socket_pressure = jiffies;
5079 #ifdef CONFIG_MEMCG_KMEM
5080 memcg->kmemcg_id = -1;
5082 #ifdef CONFIG_CGROUP_WRITEBACK
5083 INIT_LIST_HEAD(&memcg->cgwb_list);
5084 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5085 memcg->cgwb_frn[i].done =
5086 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5088 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5089 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5090 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5091 memcg->deferred_split_queue.split_queue_len = 0;
5093 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5096 mem_cgroup_id_remove(memcg);
5097 __mem_cgroup_free(memcg);
5098 return ERR_PTR(error);
5101 static struct cgroup_subsys_state * __ref
5102 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5104 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5105 struct mem_cgroup *memcg;
5106 long error = -ENOMEM;
5108 memcg = mem_cgroup_alloc();
5110 return ERR_CAST(memcg);
5112 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5113 memcg->soft_limit = PAGE_COUNTER_MAX;
5114 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5116 memcg->swappiness = mem_cgroup_swappiness(parent);
5117 memcg->oom_kill_disable = parent->oom_kill_disable;
5119 if (parent && parent->use_hierarchy) {
5120 memcg->use_hierarchy = true;
5121 page_counter_init(&memcg->memory, &parent->memory);
5122 page_counter_init(&memcg->swap, &parent->swap);
5123 page_counter_init(&memcg->memsw, &parent->memsw);
5124 page_counter_init(&memcg->kmem, &parent->kmem);
5125 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5127 page_counter_init(&memcg->memory, NULL);
5128 page_counter_init(&memcg->swap, NULL);
5129 page_counter_init(&memcg->memsw, NULL);
5130 page_counter_init(&memcg->kmem, NULL);
5131 page_counter_init(&memcg->tcpmem, NULL);
5133 * Deeper hierachy with use_hierarchy == false doesn't make
5134 * much sense so let cgroup subsystem know about this
5135 * unfortunate state in our controller.
5137 if (parent != root_mem_cgroup)
5138 memory_cgrp_subsys.broken_hierarchy = true;
5141 /* The following stuff does not apply to the root */
5143 #ifdef CONFIG_MEMCG_KMEM
5144 INIT_LIST_HEAD(&memcg->kmem_caches);
5146 root_mem_cgroup = memcg;
5150 error = memcg_online_kmem(memcg);
5154 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5155 static_branch_inc(&memcg_sockets_enabled_key);
5159 mem_cgroup_id_remove(memcg);
5160 mem_cgroup_free(memcg);
5161 return ERR_PTR(error);
5164 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5166 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5169 * A memcg must be visible for memcg_expand_shrinker_maps()
5170 * by the time the maps are allocated. So, we allocate maps
5171 * here, when for_each_mem_cgroup() can't skip it.
5173 if (memcg_alloc_shrinker_maps(memcg)) {
5174 mem_cgroup_id_remove(memcg);
5178 /* Online state pins memcg ID, memcg ID pins CSS */
5179 refcount_set(&memcg->id.ref, 1);
5184 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5186 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5187 struct mem_cgroup_event *event, *tmp;
5190 * Unregister events and notify userspace.
5191 * Notify userspace about cgroup removing only after rmdir of cgroup
5192 * directory to avoid race between userspace and kernelspace.
5194 spin_lock(&memcg->event_list_lock);
5195 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5196 list_del_init(&event->list);
5197 schedule_work(&event->remove);
5199 spin_unlock(&memcg->event_list_lock);
5201 page_counter_set_min(&memcg->memory, 0);
5202 page_counter_set_low(&memcg->memory, 0);
5204 memcg_offline_kmem(memcg);
5205 wb_memcg_offline(memcg);
5207 drain_all_stock(memcg);
5209 mem_cgroup_id_put(memcg);
5212 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5214 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5216 invalidate_reclaim_iterators(memcg);
5219 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5221 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5222 int __maybe_unused i;
5224 #ifdef CONFIG_CGROUP_WRITEBACK
5225 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5226 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5228 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5229 static_branch_dec(&memcg_sockets_enabled_key);
5231 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5232 static_branch_dec(&memcg_sockets_enabled_key);
5234 vmpressure_cleanup(&memcg->vmpressure);
5235 cancel_work_sync(&memcg->high_work);
5236 mem_cgroup_remove_from_trees(memcg);
5237 memcg_free_shrinker_maps(memcg);
5238 memcg_free_kmem(memcg);
5239 mem_cgroup_free(memcg);
5243 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5244 * @css: the target css
5246 * Reset the states of the mem_cgroup associated with @css. This is
5247 * invoked when the userland requests disabling on the default hierarchy
5248 * but the memcg is pinned through dependency. The memcg should stop
5249 * applying policies and should revert to the vanilla state as it may be
5250 * made visible again.
5252 * The current implementation only resets the essential configurations.
5253 * This needs to be expanded to cover all the visible parts.
5255 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5257 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5259 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5260 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5261 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
5262 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5263 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5264 page_counter_set_min(&memcg->memory, 0);
5265 page_counter_set_low(&memcg->memory, 0);
5266 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5267 memcg->soft_limit = PAGE_COUNTER_MAX;
5268 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5269 memcg_wb_domain_size_changed(memcg);
5273 /* Handlers for move charge at task migration. */
5274 static int mem_cgroup_do_precharge(unsigned long count)
5278 /* Try a single bulk charge without reclaim first, kswapd may wake */
5279 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5281 mc.precharge += count;
5285 /* Try charges one by one with reclaim, but do not retry */
5287 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5301 enum mc_target_type {
5308 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5309 unsigned long addr, pte_t ptent)
5311 struct page *page = vm_normal_page(vma, addr, ptent);
5313 if (!page || !page_mapped(page))
5315 if (PageAnon(page)) {
5316 if (!(mc.flags & MOVE_ANON))
5319 if (!(mc.flags & MOVE_FILE))
5322 if (!get_page_unless_zero(page))
5328 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5329 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5330 pte_t ptent, swp_entry_t *entry)
5332 struct page *page = NULL;
5333 swp_entry_t ent = pte_to_swp_entry(ptent);
5335 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
5339 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5340 * a device and because they are not accessible by CPU they are store
5341 * as special swap entry in the CPU page table.
5343 if (is_device_private_entry(ent)) {
5344 page = device_private_entry_to_page(ent);
5346 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5347 * a refcount of 1 when free (unlike normal page)
5349 if (!page_ref_add_unless(page, 1, 1))
5355 * Because lookup_swap_cache() updates some statistics counter,
5356 * we call find_get_page() with swapper_space directly.
5358 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5359 if (do_memsw_account())
5360 entry->val = ent.val;
5365 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5366 pte_t ptent, swp_entry_t *entry)
5372 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5373 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5375 struct page *page = NULL;
5376 struct address_space *mapping;
5379 if (!vma->vm_file) /* anonymous vma */
5381 if (!(mc.flags & MOVE_FILE))
5384 mapping = vma->vm_file->f_mapping;
5385 pgoff = linear_page_index(vma, addr);
5387 /* page is moved even if it's not RSS of this task(page-faulted). */
5389 /* shmem/tmpfs may report page out on swap: account for that too. */
5390 if (shmem_mapping(mapping)) {
5391 page = find_get_entry(mapping, pgoff);
5392 if (xa_is_value(page)) {
5393 swp_entry_t swp = radix_to_swp_entry(page);
5394 if (do_memsw_account())
5396 page = find_get_page(swap_address_space(swp),
5400 page = find_get_page(mapping, pgoff);
5402 page = find_get_page(mapping, pgoff);
5408 * mem_cgroup_move_account - move account of the page
5410 * @compound: charge the page as compound or small page
5411 * @from: mem_cgroup which the page is moved from.
5412 * @to: mem_cgroup which the page is moved to. @from != @to.
5414 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5416 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5419 static int mem_cgroup_move_account(struct page *page,
5421 struct mem_cgroup *from,
5422 struct mem_cgroup *to)
5424 struct lruvec *from_vec, *to_vec;
5425 struct pglist_data *pgdat;
5426 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5429 VM_BUG_ON(from == to);
5430 VM_BUG_ON_PAGE(PageLRU(page), page);
5431 VM_BUG_ON(compound && !PageTransHuge(page));
5434 * Prevent mem_cgroup_migrate() from looking at
5435 * page->mem_cgroup of its source page while we change it.
5438 if (!trylock_page(page))
5442 if (page->mem_cgroup != from)
5445 pgdat = page_pgdat(page);
5446 from_vec = mem_cgroup_lruvec(from, pgdat);
5447 to_vec = mem_cgroup_lruvec(to, pgdat);
5449 lock_page_memcg(page);
5451 if (PageAnon(page)) {
5452 if (page_mapped(page)) {
5453 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5454 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5457 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5458 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5460 if (PageSwapBacked(page)) {
5461 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5462 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5465 if (page_mapped(page)) {
5466 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5467 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5470 if (PageDirty(page)) {
5471 struct address_space *mapping = page_mapping(page);
5473 if (mapping_cap_account_dirty(mapping)) {
5474 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5476 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5482 if (PageWriteback(page)) {
5483 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5484 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5488 * All state has been migrated, let's switch to the new memcg.
5490 * It is safe to change page->mem_cgroup here because the page
5491 * is referenced, charged, isolated, and locked: we can't race
5492 * with (un)charging, migration, LRU putback, or anything else
5493 * that would rely on a stable page->mem_cgroup.
5495 * Note that lock_page_memcg is a memcg lock, not a page lock,
5496 * to save space. As soon as we switch page->mem_cgroup to a
5497 * new memcg that isn't locked, the above state can change
5498 * concurrently again. Make sure we're truly done with it.
5502 page->mem_cgroup = to; /* caller should have done css_get */
5504 __unlock_page_memcg(from);
5508 local_irq_disable();
5509 mem_cgroup_charge_statistics(to, page, nr_pages);
5510 memcg_check_events(to, page);
5511 mem_cgroup_charge_statistics(from, page, -nr_pages);
5512 memcg_check_events(from, page);
5521 * get_mctgt_type - get target type of moving charge
5522 * @vma: the vma the pte to be checked belongs
5523 * @addr: the address corresponding to the pte to be checked
5524 * @ptent: the pte to be checked
5525 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5528 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5529 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5530 * move charge. if @target is not NULL, the page is stored in target->page
5531 * with extra refcnt got(Callers should handle it).
5532 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5533 * target for charge migration. if @target is not NULL, the entry is stored
5535 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5536 * (so ZONE_DEVICE page and thus not on the lru).
5537 * For now we such page is charge like a regular page would be as for all
5538 * intent and purposes it is just special memory taking the place of a
5541 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5543 * Called with pte lock held.
5546 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5547 unsigned long addr, pte_t ptent, union mc_target *target)
5549 struct page *page = NULL;
5550 enum mc_target_type ret = MC_TARGET_NONE;
5551 swp_entry_t ent = { .val = 0 };
5553 if (pte_present(ptent))
5554 page = mc_handle_present_pte(vma, addr, ptent);
5555 else if (is_swap_pte(ptent))
5556 page = mc_handle_swap_pte(vma, ptent, &ent);
5557 else if (pte_none(ptent))
5558 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5560 if (!page && !ent.val)
5564 * Do only loose check w/o serialization.
5565 * mem_cgroup_move_account() checks the page is valid or
5566 * not under LRU exclusion.
5568 if (page->mem_cgroup == mc.from) {
5569 ret = MC_TARGET_PAGE;
5570 if (is_device_private_page(page))
5571 ret = MC_TARGET_DEVICE;
5573 target->page = page;
5575 if (!ret || !target)
5579 * There is a swap entry and a page doesn't exist or isn't charged.
5580 * But we cannot move a tail-page in a THP.
5582 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5583 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5584 ret = MC_TARGET_SWAP;
5591 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5593 * We don't consider PMD mapped swapping or file mapped pages because THP does
5594 * not support them for now.
5595 * Caller should make sure that pmd_trans_huge(pmd) is true.
5597 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5598 unsigned long addr, pmd_t pmd, union mc_target *target)
5600 struct page *page = NULL;
5601 enum mc_target_type ret = MC_TARGET_NONE;
5603 if (unlikely(is_swap_pmd(pmd))) {
5604 VM_BUG_ON(thp_migration_supported() &&
5605 !is_pmd_migration_entry(pmd));
5608 page = pmd_page(pmd);
5609 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5610 if (!(mc.flags & MOVE_ANON))
5612 if (page->mem_cgroup == mc.from) {
5613 ret = MC_TARGET_PAGE;
5616 target->page = page;
5622 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5623 unsigned long addr, pmd_t pmd, union mc_target *target)
5625 return MC_TARGET_NONE;
5629 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5630 unsigned long addr, unsigned long end,
5631 struct mm_walk *walk)
5633 struct vm_area_struct *vma = walk->vma;
5637 ptl = pmd_trans_huge_lock(pmd, vma);
5640 * Note their can not be MC_TARGET_DEVICE for now as we do not
5641 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5642 * this might change.
5644 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5645 mc.precharge += HPAGE_PMD_NR;
5650 if (pmd_trans_unstable(pmd))
5652 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5653 for (; addr != end; pte++, addr += PAGE_SIZE)
5654 if (get_mctgt_type(vma, addr, *pte, NULL))
5655 mc.precharge++; /* increment precharge temporarily */
5656 pte_unmap_unlock(pte - 1, ptl);
5662 static const struct mm_walk_ops precharge_walk_ops = {
5663 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5666 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5668 unsigned long precharge;
5670 down_read(&mm->mmap_sem);
5671 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5672 up_read(&mm->mmap_sem);
5674 precharge = mc.precharge;
5680 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5682 unsigned long precharge = mem_cgroup_count_precharge(mm);
5684 VM_BUG_ON(mc.moving_task);
5685 mc.moving_task = current;
5686 return mem_cgroup_do_precharge(precharge);
5689 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5690 static void __mem_cgroup_clear_mc(void)
5692 struct mem_cgroup *from = mc.from;
5693 struct mem_cgroup *to = mc.to;
5695 /* we must uncharge all the leftover precharges from mc.to */
5697 cancel_charge(mc.to, mc.precharge);
5701 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5702 * we must uncharge here.
5704 if (mc.moved_charge) {
5705 cancel_charge(mc.from, mc.moved_charge);
5706 mc.moved_charge = 0;
5708 /* we must fixup refcnts and charges */
5709 if (mc.moved_swap) {
5710 /* uncharge swap account from the old cgroup */
5711 if (!mem_cgroup_is_root(mc.from))
5712 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5714 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5717 * we charged both to->memory and to->memsw, so we
5718 * should uncharge to->memory.
5720 if (!mem_cgroup_is_root(mc.to))
5721 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5723 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5724 css_put_many(&mc.to->css, mc.moved_swap);
5728 memcg_oom_recover(from);
5729 memcg_oom_recover(to);
5730 wake_up_all(&mc.waitq);
5733 static void mem_cgroup_clear_mc(void)
5735 struct mm_struct *mm = mc.mm;
5738 * we must clear moving_task before waking up waiters at the end of
5741 mc.moving_task = NULL;
5742 __mem_cgroup_clear_mc();
5743 spin_lock(&mc.lock);
5747 spin_unlock(&mc.lock);
5752 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5754 struct cgroup_subsys_state *css;
5755 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5756 struct mem_cgroup *from;
5757 struct task_struct *leader, *p;
5758 struct mm_struct *mm;
5759 unsigned long move_flags;
5762 /* charge immigration isn't supported on the default hierarchy */
5763 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5767 * Multi-process migrations only happen on the default hierarchy
5768 * where charge immigration is not used. Perform charge
5769 * immigration if @tset contains a leader and whine if there are
5773 cgroup_taskset_for_each_leader(leader, css, tset) {
5776 memcg = mem_cgroup_from_css(css);
5782 * We are now commited to this value whatever it is. Changes in this
5783 * tunable will only affect upcoming migrations, not the current one.
5784 * So we need to save it, and keep it going.
5786 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5790 from = mem_cgroup_from_task(p);
5792 VM_BUG_ON(from == memcg);
5794 mm = get_task_mm(p);
5797 /* We move charges only when we move a owner of the mm */
5798 if (mm->owner == p) {
5801 VM_BUG_ON(mc.precharge);
5802 VM_BUG_ON(mc.moved_charge);
5803 VM_BUG_ON(mc.moved_swap);
5805 spin_lock(&mc.lock);
5809 mc.flags = move_flags;
5810 spin_unlock(&mc.lock);
5811 /* We set mc.moving_task later */
5813 ret = mem_cgroup_precharge_mc(mm);
5815 mem_cgroup_clear_mc();
5822 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5825 mem_cgroup_clear_mc();
5828 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5829 unsigned long addr, unsigned long end,
5830 struct mm_walk *walk)
5833 struct vm_area_struct *vma = walk->vma;
5836 enum mc_target_type target_type;
5837 union mc_target target;
5840 ptl = pmd_trans_huge_lock(pmd, vma);
5842 if (mc.precharge < HPAGE_PMD_NR) {
5846 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5847 if (target_type == MC_TARGET_PAGE) {
5849 if (!isolate_lru_page(page)) {
5850 if (!mem_cgroup_move_account(page, true,
5852 mc.precharge -= HPAGE_PMD_NR;
5853 mc.moved_charge += HPAGE_PMD_NR;
5855 putback_lru_page(page);
5858 } else if (target_type == MC_TARGET_DEVICE) {
5860 if (!mem_cgroup_move_account(page, true,
5862 mc.precharge -= HPAGE_PMD_NR;
5863 mc.moved_charge += HPAGE_PMD_NR;
5871 if (pmd_trans_unstable(pmd))
5874 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5875 for (; addr != end; addr += PAGE_SIZE) {
5876 pte_t ptent = *(pte++);
5877 bool device = false;
5883 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5884 case MC_TARGET_DEVICE:
5887 case MC_TARGET_PAGE:
5890 * We can have a part of the split pmd here. Moving it
5891 * can be done but it would be too convoluted so simply
5892 * ignore such a partial THP and keep it in original
5893 * memcg. There should be somebody mapping the head.
5895 if (PageTransCompound(page))
5897 if (!device && isolate_lru_page(page))
5899 if (!mem_cgroup_move_account(page, false,
5902 /* we uncharge from mc.from later. */
5906 putback_lru_page(page);
5907 put: /* get_mctgt_type() gets the page */
5910 case MC_TARGET_SWAP:
5912 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5914 /* we fixup refcnts and charges later. */
5922 pte_unmap_unlock(pte - 1, ptl);
5927 * We have consumed all precharges we got in can_attach().
5928 * We try charge one by one, but don't do any additional
5929 * charges to mc.to if we have failed in charge once in attach()
5932 ret = mem_cgroup_do_precharge(1);
5940 static const struct mm_walk_ops charge_walk_ops = {
5941 .pmd_entry = mem_cgroup_move_charge_pte_range,
5944 static void mem_cgroup_move_charge(void)
5946 lru_add_drain_all();
5948 * Signal lock_page_memcg() to take the memcg's move_lock
5949 * while we're moving its pages to another memcg. Then wait
5950 * for already started RCU-only updates to finish.
5952 atomic_inc(&mc.from->moving_account);
5955 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5957 * Someone who are holding the mmap_sem might be waiting in
5958 * waitq. So we cancel all extra charges, wake up all waiters,
5959 * and retry. Because we cancel precharges, we might not be able
5960 * to move enough charges, but moving charge is a best-effort
5961 * feature anyway, so it wouldn't be a big problem.
5963 __mem_cgroup_clear_mc();
5968 * When we have consumed all precharges and failed in doing
5969 * additional charge, the page walk just aborts.
5971 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
5974 up_read(&mc.mm->mmap_sem);
5975 atomic_dec(&mc.from->moving_account);
5978 static void mem_cgroup_move_task(void)
5981 mem_cgroup_move_charge();
5982 mem_cgroup_clear_mc();
5985 #else /* !CONFIG_MMU */
5986 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5990 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5993 static void mem_cgroup_move_task(void)
5999 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6000 * to verify whether we're attached to the default hierarchy on each mount
6003 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6006 * use_hierarchy is forced on the default hierarchy. cgroup core
6007 * guarantees that @root doesn't have any children, so turning it
6008 * on for the root memcg is enough.
6010 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6011 root_mem_cgroup->use_hierarchy = true;
6013 root_mem_cgroup->use_hierarchy = false;
6016 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6018 if (value == PAGE_COUNTER_MAX)
6019 seq_puts(m, "max\n");
6021 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6026 static u64 memory_current_read(struct cgroup_subsys_state *css,
6029 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6031 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6034 static int memory_min_show(struct seq_file *m, void *v)
6036 return seq_puts_memcg_tunable(m,
6037 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6040 static ssize_t memory_min_write(struct kernfs_open_file *of,
6041 char *buf, size_t nbytes, loff_t off)
6043 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6047 buf = strstrip(buf);
6048 err = page_counter_memparse(buf, "max", &min);
6052 page_counter_set_min(&memcg->memory, min);
6057 static int memory_low_show(struct seq_file *m, void *v)
6059 return seq_puts_memcg_tunable(m,
6060 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6063 static ssize_t memory_low_write(struct kernfs_open_file *of,
6064 char *buf, size_t nbytes, loff_t off)
6066 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6070 buf = strstrip(buf);
6071 err = page_counter_memparse(buf, "max", &low);
6075 page_counter_set_low(&memcg->memory, low);
6080 static int memory_high_show(struct seq_file *m, void *v)
6082 return seq_puts_memcg_tunable(m,
6083 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6086 static ssize_t memory_high_write(struct kernfs_open_file *of,
6087 char *buf, size_t nbytes, loff_t off)
6089 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6090 unsigned int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
6091 bool drained = false;
6095 buf = strstrip(buf);
6096 err = page_counter_memparse(buf, "max", &high);
6100 page_counter_set_high(&memcg->memory, high);
6103 unsigned long nr_pages = page_counter_read(&memcg->memory);
6104 unsigned long reclaimed;
6106 if (nr_pages <= high)
6109 if (signal_pending(current))
6113 drain_all_stock(memcg);
6118 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6121 if (!reclaimed && !nr_retries--)
6128 static int memory_max_show(struct seq_file *m, void *v)
6130 return seq_puts_memcg_tunable(m,
6131 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6134 static ssize_t memory_max_write(struct kernfs_open_file *of,
6135 char *buf, size_t nbytes, loff_t off)
6137 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6138 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
6139 bool drained = false;
6143 buf = strstrip(buf);
6144 err = page_counter_memparse(buf, "max", &max);
6148 xchg(&memcg->memory.max, max);
6151 unsigned long nr_pages = page_counter_read(&memcg->memory);
6153 if (nr_pages <= max)
6156 if (signal_pending(current))
6160 drain_all_stock(memcg);
6166 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6172 memcg_memory_event(memcg, MEMCG_OOM);
6173 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6177 memcg_wb_domain_size_changed(memcg);
6181 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6183 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6184 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6185 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6186 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6187 seq_printf(m, "oom_kill %lu\n",
6188 atomic_long_read(&events[MEMCG_OOM_KILL]));
6191 static int memory_events_show(struct seq_file *m, void *v)
6193 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6195 __memory_events_show(m, memcg->memory_events);
6199 static int memory_events_local_show(struct seq_file *m, void *v)
6201 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6203 __memory_events_show(m, memcg->memory_events_local);
6207 static int memory_stat_show(struct seq_file *m, void *v)
6209 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6212 buf = memory_stat_format(memcg);
6220 static int memory_oom_group_show(struct seq_file *m, void *v)
6222 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6224 seq_printf(m, "%d\n", memcg->oom_group);
6229 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6230 char *buf, size_t nbytes, loff_t off)
6232 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6235 buf = strstrip(buf);
6239 ret = kstrtoint(buf, 0, &oom_group);
6243 if (oom_group != 0 && oom_group != 1)
6246 memcg->oom_group = oom_group;
6251 static struct cftype memory_files[] = {
6254 .flags = CFTYPE_NOT_ON_ROOT,
6255 .read_u64 = memory_current_read,
6259 .flags = CFTYPE_NOT_ON_ROOT,
6260 .seq_show = memory_min_show,
6261 .write = memory_min_write,
6265 .flags = CFTYPE_NOT_ON_ROOT,
6266 .seq_show = memory_low_show,
6267 .write = memory_low_write,
6271 .flags = CFTYPE_NOT_ON_ROOT,
6272 .seq_show = memory_high_show,
6273 .write = memory_high_write,
6277 .flags = CFTYPE_NOT_ON_ROOT,
6278 .seq_show = memory_max_show,
6279 .write = memory_max_write,
6283 .flags = CFTYPE_NOT_ON_ROOT,
6284 .file_offset = offsetof(struct mem_cgroup, events_file),
6285 .seq_show = memory_events_show,
6288 .name = "events.local",
6289 .flags = CFTYPE_NOT_ON_ROOT,
6290 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6291 .seq_show = memory_events_local_show,
6295 .seq_show = memory_stat_show,
6298 .name = "oom.group",
6299 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6300 .seq_show = memory_oom_group_show,
6301 .write = memory_oom_group_write,
6306 struct cgroup_subsys memory_cgrp_subsys = {
6307 .css_alloc = mem_cgroup_css_alloc,
6308 .css_online = mem_cgroup_css_online,
6309 .css_offline = mem_cgroup_css_offline,
6310 .css_released = mem_cgroup_css_released,
6311 .css_free = mem_cgroup_css_free,
6312 .css_reset = mem_cgroup_css_reset,
6313 .can_attach = mem_cgroup_can_attach,
6314 .cancel_attach = mem_cgroup_cancel_attach,
6315 .post_attach = mem_cgroup_move_task,
6316 .bind = mem_cgroup_bind,
6317 .dfl_cftypes = memory_files,
6318 .legacy_cftypes = mem_cgroup_legacy_files,
6323 * This function calculates an individual cgroup's effective
6324 * protection which is derived from its own memory.min/low, its
6325 * parent's and siblings' settings, as well as the actual memory
6326 * distribution in the tree.
6328 * The following rules apply to the effective protection values:
6330 * 1. At the first level of reclaim, effective protection is equal to
6331 * the declared protection in memory.min and memory.low.
6333 * 2. To enable safe delegation of the protection configuration, at
6334 * subsequent levels the effective protection is capped to the
6335 * parent's effective protection.
6337 * 3. To make complex and dynamic subtrees easier to configure, the
6338 * user is allowed to overcommit the declared protection at a given
6339 * level. If that is the case, the parent's effective protection is
6340 * distributed to the children in proportion to how much protection
6341 * they have declared and how much of it they are utilizing.
6343 * This makes distribution proportional, but also work-conserving:
6344 * if one cgroup claims much more protection than it uses memory,
6345 * the unused remainder is available to its siblings.
6347 * 4. Conversely, when the declared protection is undercommitted at a
6348 * given level, the distribution of the larger parental protection
6349 * budget is NOT proportional. A cgroup's protection from a sibling
6350 * is capped to its own memory.min/low setting.
6352 * 5. However, to allow protecting recursive subtrees from each other
6353 * without having to declare each individual cgroup's fixed share
6354 * of the ancestor's claim to protection, any unutilized -
6355 * "floating" - protection from up the tree is distributed in
6356 * proportion to each cgroup's *usage*. This makes the protection
6357 * neutral wrt sibling cgroups and lets them compete freely over
6358 * the shared parental protection budget, but it protects the
6359 * subtree as a whole from neighboring subtrees.
6361 * Note that 4. and 5. are not in conflict: 4. is about protecting
6362 * against immediate siblings whereas 5. is about protecting against
6363 * neighboring subtrees.
6365 static unsigned long effective_protection(unsigned long usage,
6366 unsigned long parent_usage,
6367 unsigned long setting,
6368 unsigned long parent_effective,
6369 unsigned long siblings_protected)
6371 unsigned long protected;
6374 protected = min(usage, setting);
6376 * If all cgroups at this level combined claim and use more
6377 * protection then what the parent affords them, distribute
6378 * shares in proportion to utilization.
6380 * We are using actual utilization rather than the statically
6381 * claimed protection in order to be work-conserving: claimed
6382 * but unused protection is available to siblings that would
6383 * otherwise get a smaller chunk than what they claimed.
6385 if (siblings_protected > parent_effective)
6386 return protected * parent_effective / siblings_protected;
6389 * Ok, utilized protection of all children is within what the
6390 * parent affords them, so we know whatever this child claims
6391 * and utilizes is effectively protected.
6393 * If there is unprotected usage beyond this value, reclaim
6394 * will apply pressure in proportion to that amount.
6396 * If there is unutilized protection, the cgroup will be fully
6397 * shielded from reclaim, but we do return a smaller value for
6398 * protection than what the group could enjoy in theory. This
6399 * is okay. With the overcommit distribution above, effective
6400 * protection is always dependent on how memory is actually
6401 * consumed among the siblings anyway.
6406 * If the children aren't claiming (all of) the protection
6407 * afforded to them by the parent, distribute the remainder in
6408 * proportion to the (unprotected) memory of each cgroup. That
6409 * way, cgroups that aren't explicitly prioritized wrt each
6410 * other compete freely over the allowance, but they are
6411 * collectively protected from neighboring trees.
6413 * We're using unprotected memory for the weight so that if
6414 * some cgroups DO claim explicit protection, we don't protect
6415 * the same bytes twice.
6417 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6420 if (parent_effective > siblings_protected && usage > protected) {
6421 unsigned long unclaimed;
6423 unclaimed = parent_effective - siblings_protected;
6424 unclaimed *= usage - protected;
6425 unclaimed /= parent_usage - siblings_protected;
6434 * mem_cgroup_protected - check if memory consumption is in the normal range
6435 * @root: the top ancestor of the sub-tree being checked
6436 * @memcg: the memory cgroup to check
6438 * WARNING: This function is not stateless! It can only be used as part
6439 * of a top-down tree iteration, not for isolated queries.
6441 * Returns one of the following:
6442 * MEMCG_PROT_NONE: cgroup memory is not protected
6443 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6444 * an unprotected supply of reclaimable memory from other cgroups.
6445 * MEMCG_PROT_MIN: cgroup memory is protected
6447 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
6448 struct mem_cgroup *memcg)
6450 unsigned long usage, parent_usage;
6451 struct mem_cgroup *parent;
6453 if (mem_cgroup_disabled())
6454 return MEMCG_PROT_NONE;
6457 root = root_mem_cgroup;
6459 return MEMCG_PROT_NONE;
6461 usage = page_counter_read(&memcg->memory);
6463 return MEMCG_PROT_NONE;
6465 parent = parent_mem_cgroup(memcg);
6466 /* No parent means a non-hierarchical mode on v1 memcg */
6468 return MEMCG_PROT_NONE;
6470 if (parent == root) {
6471 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6472 memcg->memory.elow = memcg->memory.low;
6476 parent_usage = page_counter_read(&parent->memory);
6478 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6479 READ_ONCE(memcg->memory.min),
6480 READ_ONCE(parent->memory.emin),
6481 atomic_long_read(&parent->memory.children_min_usage)));
6483 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6484 memcg->memory.low, READ_ONCE(parent->memory.elow),
6485 atomic_long_read(&parent->memory.children_low_usage)));
6488 if (usage <= memcg->memory.emin)
6489 return MEMCG_PROT_MIN;
6490 else if (usage <= memcg->memory.elow)
6491 return MEMCG_PROT_LOW;
6493 return MEMCG_PROT_NONE;
6497 * mem_cgroup_try_charge - try charging a page
6498 * @page: page to charge
6499 * @mm: mm context of the victim
6500 * @gfp_mask: reclaim mode
6501 * @memcgp: charged memcg return
6503 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6504 * pages according to @gfp_mask if necessary.
6506 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6507 * Otherwise, an error code is returned.
6509 * After page->mapping has been set up, the caller must finalize the
6510 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6511 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6513 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6514 gfp_t gfp_mask, struct mem_cgroup **memcgp)
6516 unsigned int nr_pages = hpage_nr_pages(page);
6517 struct mem_cgroup *memcg = NULL;
6520 if (mem_cgroup_disabled())
6523 if (PageSwapCache(page)) {
6525 * Every swap fault against a single page tries to charge the
6526 * page, bail as early as possible. shmem_unuse() encounters
6527 * already charged pages, too. The USED bit is protected by
6528 * the page lock, which serializes swap cache removal, which
6529 * in turn serializes uncharging.
6531 VM_BUG_ON_PAGE(!PageLocked(page), page);
6532 if (compound_head(page)->mem_cgroup)
6535 if (do_swap_account) {
6536 swp_entry_t ent = { .val = page_private(page), };
6537 unsigned short id = lookup_swap_cgroup_id(ent);
6540 memcg = mem_cgroup_from_id(id);
6541 if (memcg && !css_tryget_online(&memcg->css))
6548 memcg = get_mem_cgroup_from_mm(mm);
6550 ret = try_charge(memcg, gfp_mask, nr_pages);
6552 css_put(&memcg->css);
6558 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6559 gfp_t gfp_mask, struct mem_cgroup **memcgp)
6563 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp);
6565 cgroup_throttle_swaprate(page, gfp_mask);
6570 * mem_cgroup_commit_charge - commit a page charge
6571 * @page: page to charge
6572 * @memcg: memcg to charge the page to
6573 * @lrucare: page might be on LRU already
6575 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6576 * after page->mapping has been set up. This must happen atomically
6577 * as part of the page instantiation, i.e. under the page table lock
6578 * for anonymous pages, under the page lock for page and swap cache.
6580 * In addition, the page must not be on the LRU during the commit, to
6581 * prevent racing with task migration. If it might be, use @lrucare.
6583 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6585 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6588 unsigned int nr_pages = hpage_nr_pages(page);
6590 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6592 if (mem_cgroup_disabled())
6595 * Swap faults will attempt to charge the same page multiple
6596 * times. But reuse_swap_page() might have removed the page
6597 * from swapcache already, so we can't check PageSwapCache().
6602 commit_charge(page, memcg, lrucare);
6604 local_irq_disable();
6605 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6606 memcg_check_events(memcg, page);
6609 if (do_memsw_account() && PageSwapCache(page)) {
6610 swp_entry_t entry = { .val = page_private(page) };
6612 * The swap entry might not get freed for a long time,
6613 * let's not wait for it. The page already received a
6614 * memory+swap charge, drop the swap entry duplicate.
6616 mem_cgroup_uncharge_swap(entry, nr_pages);
6621 * mem_cgroup_cancel_charge - cancel a page charge
6622 * @page: page to charge
6623 * @memcg: memcg to charge the page to
6625 * Cancel a charge transaction started by mem_cgroup_try_charge().
6627 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
6629 unsigned int nr_pages = hpage_nr_pages(page);
6631 if (mem_cgroup_disabled())
6634 * Swap faults will attempt to charge the same page multiple
6635 * times. But reuse_swap_page() might have removed the page
6636 * from swapcache already, so we can't check PageSwapCache().
6641 cancel_charge(memcg, nr_pages);
6645 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6646 * @page: page to charge
6647 * @mm: mm context of the victim
6648 * @gfp_mask: reclaim mode
6649 * @lrucare: page might be on the LRU already
6651 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6652 * pages according to @gfp_mask if necessary.
6654 * Returns 0 on success. Otherwise, an error code is returned.
6656 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask,
6659 struct mem_cgroup *memcg;
6662 ret = mem_cgroup_try_charge(page, mm, gfp_mask, &memcg);
6665 mem_cgroup_commit_charge(page, memcg, lrucare);
6669 struct uncharge_gather {
6670 struct mem_cgroup *memcg;
6671 unsigned long nr_pages;
6672 unsigned long pgpgout;
6673 unsigned long nr_kmem;
6674 unsigned long nr_huge;
6675 struct page *dummy_page;
6678 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6680 memset(ug, 0, sizeof(*ug));
6683 static void uncharge_batch(const struct uncharge_gather *ug)
6685 unsigned long flags;
6687 if (!mem_cgroup_is_root(ug->memcg)) {
6688 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6689 if (do_memsw_account())
6690 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6691 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6692 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6693 memcg_oom_recover(ug->memcg);
6696 local_irq_save(flags);
6697 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6698 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6699 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6700 memcg_check_events(ug->memcg, ug->dummy_page);
6701 local_irq_restore(flags);
6703 if (!mem_cgroup_is_root(ug->memcg))
6704 css_put_many(&ug->memcg->css, ug->nr_pages);
6707 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6709 unsigned long nr_pages;
6711 VM_BUG_ON_PAGE(PageLRU(page), page);
6713 if (!page->mem_cgroup)
6717 * Nobody should be changing or seriously looking at
6718 * page->mem_cgroup at this point, we have fully
6719 * exclusive access to the page.
6722 if (ug->memcg != page->mem_cgroup) {
6725 uncharge_gather_clear(ug);
6727 ug->memcg = page->mem_cgroup;
6730 nr_pages = compound_nr(page);
6731 ug->nr_pages += nr_pages;
6733 if (!PageKmemcg(page)) {
6734 if (PageTransHuge(page))
6735 ug->nr_huge += nr_pages;
6738 ug->nr_kmem += nr_pages;
6739 __ClearPageKmemcg(page);
6742 ug->dummy_page = page;
6743 page->mem_cgroup = NULL;
6746 static void uncharge_list(struct list_head *page_list)
6748 struct uncharge_gather ug;
6749 struct list_head *next;
6751 uncharge_gather_clear(&ug);
6754 * Note that the list can be a single page->lru; hence the
6755 * do-while loop instead of a simple list_for_each_entry().
6757 next = page_list->next;
6761 page = list_entry(next, struct page, lru);
6762 next = page->lru.next;
6764 uncharge_page(page, &ug);
6765 } while (next != page_list);
6768 uncharge_batch(&ug);
6772 * mem_cgroup_uncharge - uncharge a page
6773 * @page: page to uncharge
6775 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6776 * mem_cgroup_commit_charge().
6778 void mem_cgroup_uncharge(struct page *page)
6780 struct uncharge_gather ug;
6782 if (mem_cgroup_disabled())
6785 /* Don't touch page->lru of any random page, pre-check: */
6786 if (!page->mem_cgroup)
6789 uncharge_gather_clear(&ug);
6790 uncharge_page(page, &ug);
6791 uncharge_batch(&ug);
6795 * mem_cgroup_uncharge_list - uncharge a list of page
6796 * @page_list: list of pages to uncharge
6798 * Uncharge a list of pages previously charged with
6799 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6801 void mem_cgroup_uncharge_list(struct list_head *page_list)
6803 if (mem_cgroup_disabled())
6806 if (!list_empty(page_list))
6807 uncharge_list(page_list);
6811 * mem_cgroup_migrate - charge a page's replacement
6812 * @oldpage: currently circulating page
6813 * @newpage: replacement page
6815 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6816 * be uncharged upon free.
6818 * Both pages must be locked, @newpage->mapping must be set up.
6820 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6822 struct mem_cgroup *memcg;
6823 unsigned int nr_pages;
6824 unsigned long flags;
6826 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6827 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6828 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6829 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6832 if (mem_cgroup_disabled())
6835 /* Page cache replacement: new page already charged? */
6836 if (newpage->mem_cgroup)
6839 /* Swapcache readahead pages can get replaced before being charged */
6840 memcg = oldpage->mem_cgroup;
6844 /* Force-charge the new page. The old one will be freed soon */
6845 nr_pages = hpage_nr_pages(newpage);
6847 page_counter_charge(&memcg->memory, nr_pages);
6848 if (do_memsw_account())
6849 page_counter_charge(&memcg->memsw, nr_pages);
6850 css_get_many(&memcg->css, nr_pages);
6852 commit_charge(newpage, memcg, false);
6854 local_irq_save(flags);
6855 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6856 memcg_check_events(memcg, newpage);
6857 local_irq_restore(flags);
6860 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6861 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6863 void mem_cgroup_sk_alloc(struct sock *sk)
6865 struct mem_cgroup *memcg;
6867 if (!mem_cgroup_sockets_enabled)
6870 /* Do not associate the sock with unrelated interrupted task's memcg. */
6875 memcg = mem_cgroup_from_task(current);
6876 if (memcg == root_mem_cgroup)
6878 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6880 if (css_tryget(&memcg->css))
6881 sk->sk_memcg = memcg;
6886 void mem_cgroup_sk_free(struct sock *sk)
6889 css_put(&sk->sk_memcg->css);
6893 * mem_cgroup_charge_skmem - charge socket memory
6894 * @memcg: memcg to charge
6895 * @nr_pages: number of pages to charge
6897 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6898 * @memcg's configured limit, %false if the charge had to be forced.
6900 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6902 gfp_t gfp_mask = GFP_KERNEL;
6904 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6905 struct page_counter *fail;
6907 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6908 memcg->tcpmem_pressure = 0;
6911 page_counter_charge(&memcg->tcpmem, nr_pages);
6912 memcg->tcpmem_pressure = 1;
6916 /* Don't block in the packet receive path */
6918 gfp_mask = GFP_NOWAIT;
6920 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6922 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6925 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6930 * mem_cgroup_uncharge_skmem - uncharge socket memory
6931 * @memcg: memcg to uncharge
6932 * @nr_pages: number of pages to uncharge
6934 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6936 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6937 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6941 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6943 refill_stock(memcg, nr_pages);
6946 static int __init cgroup_memory(char *s)
6950 while ((token = strsep(&s, ",")) != NULL) {
6953 if (!strcmp(token, "nosocket"))
6954 cgroup_memory_nosocket = true;
6955 if (!strcmp(token, "nokmem"))
6956 cgroup_memory_nokmem = true;
6960 __setup("cgroup.memory=", cgroup_memory);
6963 * subsys_initcall() for memory controller.
6965 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6966 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6967 * basically everything that doesn't depend on a specific mem_cgroup structure
6968 * should be initialized from here.
6970 static int __init mem_cgroup_init(void)
6974 #ifdef CONFIG_MEMCG_KMEM
6976 * Kmem cache creation is mostly done with the slab_mutex held,
6977 * so use a workqueue with limited concurrency to avoid stalling
6978 * all worker threads in case lots of cgroups are created and
6979 * destroyed simultaneously.
6981 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6982 BUG_ON(!memcg_kmem_cache_wq);
6985 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6986 memcg_hotplug_cpu_dead);
6988 for_each_possible_cpu(cpu)
6989 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6992 for_each_node(node) {
6993 struct mem_cgroup_tree_per_node *rtpn;
6995 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6996 node_online(node) ? node : NUMA_NO_NODE);
6998 rtpn->rb_root = RB_ROOT;
6999 rtpn->rb_rightmost = NULL;
7000 spin_lock_init(&rtpn->lock);
7001 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7006 subsys_initcall(mem_cgroup_init);
7008 #ifdef CONFIG_MEMCG_SWAP
7009 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7011 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7013 * The root cgroup cannot be destroyed, so it's refcount must
7016 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7020 memcg = parent_mem_cgroup(memcg);
7022 memcg = root_mem_cgroup;
7028 * mem_cgroup_swapout - transfer a memsw charge to swap
7029 * @page: page whose memsw charge to transfer
7030 * @entry: swap entry to move the charge to
7032 * Transfer the memsw charge of @page to @entry.
7034 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7036 struct mem_cgroup *memcg, *swap_memcg;
7037 unsigned int nr_entries;
7038 unsigned short oldid;
7040 VM_BUG_ON_PAGE(PageLRU(page), page);
7041 VM_BUG_ON_PAGE(page_count(page), page);
7043 if (!do_memsw_account())
7046 memcg = page->mem_cgroup;
7048 /* Readahead page, never charged */
7053 * In case the memcg owning these pages has been offlined and doesn't
7054 * have an ID allocated to it anymore, charge the closest online
7055 * ancestor for the swap instead and transfer the memory+swap charge.
7057 swap_memcg = mem_cgroup_id_get_online(memcg);
7058 nr_entries = hpage_nr_pages(page);
7059 /* Get references for the tail pages, too */
7061 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7062 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7064 VM_BUG_ON_PAGE(oldid, page);
7065 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7067 page->mem_cgroup = NULL;
7069 if (!mem_cgroup_is_root(memcg))
7070 page_counter_uncharge(&memcg->memory, nr_entries);
7072 if (memcg != swap_memcg) {
7073 if (!mem_cgroup_is_root(swap_memcg))
7074 page_counter_charge(&swap_memcg->memsw, nr_entries);
7075 page_counter_uncharge(&memcg->memsw, nr_entries);
7079 * Interrupts should be disabled here because the caller holds the
7080 * i_pages lock which is taken with interrupts-off. It is
7081 * important here to have the interrupts disabled because it is the
7082 * only synchronisation we have for updating the per-CPU variables.
7084 VM_BUG_ON(!irqs_disabled());
7085 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7086 memcg_check_events(memcg, page);
7088 if (!mem_cgroup_is_root(memcg))
7089 css_put_many(&memcg->css, nr_entries);
7093 * mem_cgroup_try_charge_swap - try charging swap space for a page
7094 * @page: page being added to swap
7095 * @entry: swap entry to charge
7097 * Try to charge @page's memcg for the swap space at @entry.
7099 * Returns 0 on success, -ENOMEM on failure.
7101 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7103 unsigned int nr_pages = hpage_nr_pages(page);
7104 struct page_counter *counter;
7105 struct mem_cgroup *memcg;
7106 unsigned short oldid;
7108 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
7111 memcg = page->mem_cgroup;
7113 /* Readahead page, never charged */
7118 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7122 memcg = mem_cgroup_id_get_online(memcg);
7124 if (!mem_cgroup_is_root(memcg) &&
7125 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7126 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7127 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7128 mem_cgroup_id_put(memcg);
7132 /* Get references for the tail pages, too */
7134 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7135 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7136 VM_BUG_ON_PAGE(oldid, page);
7137 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7143 * mem_cgroup_uncharge_swap - uncharge swap space
7144 * @entry: swap entry to uncharge
7145 * @nr_pages: the amount of swap space to uncharge
7147 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7149 struct mem_cgroup *memcg;
7152 if (!do_swap_account)
7155 id = swap_cgroup_record(entry, 0, nr_pages);
7157 memcg = mem_cgroup_from_id(id);
7159 if (!mem_cgroup_is_root(memcg)) {
7160 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7161 page_counter_uncharge(&memcg->swap, nr_pages);
7163 page_counter_uncharge(&memcg->memsw, nr_pages);
7165 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7166 mem_cgroup_id_put_many(memcg, nr_pages);
7171 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7173 long nr_swap_pages = get_nr_swap_pages();
7175 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7176 return nr_swap_pages;
7177 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7178 nr_swap_pages = min_t(long, nr_swap_pages,
7179 READ_ONCE(memcg->swap.max) -
7180 page_counter_read(&memcg->swap));
7181 return nr_swap_pages;
7184 bool mem_cgroup_swap_full(struct page *page)
7186 struct mem_cgroup *memcg;
7188 VM_BUG_ON_PAGE(!PageLocked(page), page);
7192 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7195 memcg = page->mem_cgroup;
7199 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7200 unsigned long usage = page_counter_read(&memcg->swap);
7202 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7203 usage * 2 >= READ_ONCE(memcg->swap.max))
7210 /* for remember boot option*/
7211 #ifdef CONFIG_MEMCG_SWAP_ENABLED
7212 static int really_do_swap_account __initdata = 1;
7214 static int really_do_swap_account __initdata;
7217 static int __init enable_swap_account(char *s)
7219 if (!strcmp(s, "1"))
7220 really_do_swap_account = 1;
7221 else if (!strcmp(s, "0"))
7222 really_do_swap_account = 0;
7225 __setup("swapaccount=", enable_swap_account);
7227 static u64 swap_current_read(struct cgroup_subsys_state *css,
7230 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7232 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7235 static int swap_high_show(struct seq_file *m, void *v)
7237 return seq_puts_memcg_tunable(m,
7238 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7241 static ssize_t swap_high_write(struct kernfs_open_file *of,
7242 char *buf, size_t nbytes, loff_t off)
7244 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7248 buf = strstrip(buf);
7249 err = page_counter_memparse(buf, "max", &high);
7253 page_counter_set_high(&memcg->swap, high);
7258 static int swap_max_show(struct seq_file *m, void *v)
7260 return seq_puts_memcg_tunable(m,
7261 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7264 static ssize_t swap_max_write(struct kernfs_open_file *of,
7265 char *buf, size_t nbytes, loff_t off)
7267 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7271 buf = strstrip(buf);
7272 err = page_counter_memparse(buf, "max", &max);
7276 xchg(&memcg->swap.max, max);
7281 static int swap_events_show(struct seq_file *m, void *v)
7283 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7285 seq_printf(m, "high %lu\n",
7286 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7287 seq_printf(m, "max %lu\n",
7288 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7289 seq_printf(m, "fail %lu\n",
7290 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7295 static struct cftype swap_files[] = {
7297 .name = "swap.current",
7298 .flags = CFTYPE_NOT_ON_ROOT,
7299 .read_u64 = swap_current_read,
7302 .name = "swap.high",
7303 .flags = CFTYPE_NOT_ON_ROOT,
7304 .seq_show = swap_high_show,
7305 .write = swap_high_write,
7309 .flags = CFTYPE_NOT_ON_ROOT,
7310 .seq_show = swap_max_show,
7311 .write = swap_max_write,
7314 .name = "swap.events",
7315 .flags = CFTYPE_NOT_ON_ROOT,
7316 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7317 .seq_show = swap_events_show,
7322 static struct cftype memsw_cgroup_files[] = {
7324 .name = "memsw.usage_in_bytes",
7325 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7326 .read_u64 = mem_cgroup_read_u64,
7329 .name = "memsw.max_usage_in_bytes",
7330 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7331 .write = mem_cgroup_reset,
7332 .read_u64 = mem_cgroup_read_u64,
7335 .name = "memsw.limit_in_bytes",
7336 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7337 .write = mem_cgroup_write,
7338 .read_u64 = mem_cgroup_read_u64,
7341 .name = "memsw.failcnt",
7342 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7343 .write = mem_cgroup_reset,
7344 .read_u64 = mem_cgroup_read_u64,
7346 { }, /* terminate */
7349 static int __init mem_cgroup_swap_init(void)
7351 if (!mem_cgroup_disabled() && really_do_swap_account) {
7352 do_swap_account = 1;
7353 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
7355 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
7356 memsw_cgroup_files));
7360 subsys_initcall(mem_cgroup_swap_init);
7362 #endif /* CONFIG_MEMCG_SWAP */