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
24 * Per memcg lru locking
25 * Copyright (C) 2020 Alibaba, Inc, Alex Shi
28 #include <linux/page_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
31 #include <linux/pagewalk.h>
32 #include <linux/sched/mm.h>
33 #include <linux/shmem_fs.h>
34 #include <linux/hugetlb.h>
35 #include <linux/pagemap.h>
36 #include <linux/vm_event_item.h>
37 #include <linux/smp.h>
38 #include <linux/page-flags.h>
39 #include <linux/backing-dev.h>
40 #include <linux/bit_spinlock.h>
41 #include <linux/rcupdate.h>
42 #include <linux/limits.h>
43 #include <linux/export.h>
44 #include <linux/mutex.h>
45 #include <linux/rbtree.h>
46 #include <linux/slab.h>
47 #include <linux/swap.h>
48 #include <linux/swapops.h>
49 #include <linux/spinlock.h>
50 #include <linux/eventfd.h>
51 #include <linux/poll.h>
52 #include <linux/sort.h>
54 #include <linux/seq_file.h>
55 #include <linux/vmpressure.h>
56 #include <linux/mm_inline.h>
57 #include <linux/swap_cgroup.h>
58 #include <linux/cpu.h>
59 #include <linux/oom.h>
60 #include <linux/lockdep.h>
61 #include <linux/file.h>
62 #include <linux/tracehook.h>
63 #include <linux/psi.h>
64 #include <linux/seq_buf.h>
70 #include <linux/uaccess.h>
72 #include <trace/events/vmscan.h>
74 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
75 EXPORT_SYMBOL(memory_cgrp_subsys);
77 struct mem_cgroup *root_mem_cgroup __read_mostly;
79 /* Active memory cgroup to use from an interrupt context */
80 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
82 /* Socket memory accounting disabled? */
83 static bool cgroup_memory_nosocket;
85 /* Kernel memory accounting disabled? */
86 bool cgroup_memory_nokmem;
88 /* Whether the swap controller is active */
89 #ifdef CONFIG_MEMCG_SWAP
90 bool cgroup_memory_noswap __read_mostly;
92 #define cgroup_memory_noswap 1
95 #ifdef CONFIG_CGROUP_WRITEBACK
96 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
99 /* Whether legacy memory+swap accounting is active */
100 static bool do_memsw_account(void)
102 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
105 #define THRESHOLDS_EVENTS_TARGET 128
106 #define SOFTLIMIT_EVENTS_TARGET 1024
109 * Cgroups above their limits are maintained in a RB-Tree, independent of
110 * their hierarchy representation
113 struct mem_cgroup_tree_per_node {
114 struct rb_root rb_root;
115 struct rb_node *rb_rightmost;
119 struct mem_cgroup_tree {
120 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
123 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
126 struct mem_cgroup_eventfd_list {
127 struct list_head list;
128 struct eventfd_ctx *eventfd;
132 * cgroup_event represents events which userspace want to receive.
134 struct mem_cgroup_event {
136 * memcg which the event belongs to.
138 struct mem_cgroup *memcg;
140 * eventfd to signal userspace about the event.
142 struct eventfd_ctx *eventfd;
144 * Each of these stored in a list by the cgroup.
146 struct list_head list;
148 * register_event() callback will be used to add new userspace
149 * waiter for changes related to this event. Use eventfd_signal()
150 * on eventfd to send notification to userspace.
152 int (*register_event)(struct mem_cgroup *memcg,
153 struct eventfd_ctx *eventfd, const char *args);
155 * unregister_event() callback will be called when userspace closes
156 * the eventfd or on cgroup removing. This callback must be set,
157 * if you want provide notification functionality.
159 void (*unregister_event)(struct mem_cgroup *memcg,
160 struct eventfd_ctx *eventfd);
162 * All fields below needed to unregister event when
163 * userspace closes eventfd.
166 wait_queue_head_t *wqh;
167 wait_queue_entry_t wait;
168 struct work_struct remove;
171 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
172 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
174 /* Stuffs for move charges at task migration. */
176 * Types of charges to be moved.
178 #define MOVE_ANON 0x1U
179 #define MOVE_FILE 0x2U
180 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
182 /* "mc" and its members are protected by cgroup_mutex */
183 static struct move_charge_struct {
184 spinlock_t lock; /* for from, to */
185 struct mm_struct *mm;
186 struct mem_cgroup *from;
187 struct mem_cgroup *to;
189 unsigned long precharge;
190 unsigned long moved_charge;
191 unsigned long moved_swap;
192 struct task_struct *moving_task; /* a task moving charges */
193 wait_queue_head_t waitq; /* a waitq for other context */
195 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
196 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
200 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
201 * limit reclaim to prevent infinite loops, if they ever occur.
203 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
204 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
206 /* for encoding cft->private value on file */
215 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
216 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
217 #define MEMFILE_ATTR(val) ((val) & 0xffff)
218 /* Used for OOM notifier */
219 #define OOM_CONTROL (0)
222 * Iteration constructs for visiting all cgroups (under a tree). If
223 * loops are exited prematurely (break), mem_cgroup_iter_break() must
224 * be used for reference counting.
226 #define for_each_mem_cgroup_tree(iter, root) \
227 for (iter = mem_cgroup_iter(root, NULL, NULL); \
229 iter = mem_cgroup_iter(root, iter, NULL))
231 #define for_each_mem_cgroup(iter) \
232 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
234 iter = mem_cgroup_iter(NULL, iter, NULL))
236 static inline bool should_force_charge(void)
238 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
239 (current->flags & PF_EXITING);
242 /* Some nice accessors for the vmpressure. */
243 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
246 memcg = root_mem_cgroup;
247 return &memcg->vmpressure;
250 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
252 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
255 #ifdef CONFIG_MEMCG_KMEM
256 extern spinlock_t css_set_lock;
258 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
259 unsigned int nr_pages);
261 static void obj_cgroup_release(struct percpu_ref *ref)
263 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
264 unsigned int nr_bytes;
265 unsigned int nr_pages;
269 * At this point all allocated objects are freed, and
270 * objcg->nr_charged_bytes can't have an arbitrary byte value.
271 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
273 * The following sequence can lead to it:
274 * 1) CPU0: objcg == stock->cached_objcg
275 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
276 * PAGE_SIZE bytes are charged
277 * 3) CPU1: a process from another memcg is allocating something,
278 * the stock if flushed,
279 * objcg->nr_charged_bytes = PAGE_SIZE - 92
280 * 5) CPU0: we do release this object,
281 * 92 bytes are added to stock->nr_bytes
282 * 6) CPU0: stock is flushed,
283 * 92 bytes are added to objcg->nr_charged_bytes
285 * In the result, nr_charged_bytes == PAGE_SIZE.
286 * This page will be uncharged in obj_cgroup_release().
288 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
289 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
290 nr_pages = nr_bytes >> PAGE_SHIFT;
292 spin_lock_irqsave(&css_set_lock, flags);
294 obj_cgroup_uncharge_pages(objcg, nr_pages);
295 list_del(&objcg->list);
296 spin_unlock_irqrestore(&css_set_lock, flags);
298 percpu_ref_exit(ref);
299 kfree_rcu(objcg, rcu);
302 static struct obj_cgroup *obj_cgroup_alloc(void)
304 struct obj_cgroup *objcg;
307 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
311 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
317 INIT_LIST_HEAD(&objcg->list);
321 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
322 struct mem_cgroup *parent)
324 struct obj_cgroup *objcg, *iter;
326 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
328 spin_lock_irq(&css_set_lock);
330 /* 1) Ready to reparent active objcg. */
331 list_add(&objcg->list, &memcg->objcg_list);
332 /* 2) Reparent active objcg and already reparented objcgs to parent. */
333 list_for_each_entry(iter, &memcg->objcg_list, list)
334 WRITE_ONCE(iter->memcg, parent);
335 /* 3) Move already reparented objcgs to the parent's list */
336 list_splice(&memcg->objcg_list, &parent->objcg_list);
338 spin_unlock_irq(&css_set_lock);
340 percpu_ref_kill(&objcg->refcnt);
344 * This will be used as a shrinker list's index.
345 * The main reason for not using cgroup id for this:
346 * this works better in sparse environments, where we have a lot of memcgs,
347 * but only a few kmem-limited. Or also, if we have, for instance, 200
348 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
349 * 200 entry array for that.
351 * The current size of the caches array is stored in memcg_nr_cache_ids. It
352 * will double each time we have to increase it.
354 static DEFINE_IDA(memcg_cache_ida);
355 int memcg_nr_cache_ids;
357 /* Protects memcg_nr_cache_ids */
358 static DECLARE_RWSEM(memcg_cache_ids_sem);
360 void memcg_get_cache_ids(void)
362 down_read(&memcg_cache_ids_sem);
365 void memcg_put_cache_ids(void)
367 up_read(&memcg_cache_ids_sem);
371 * MIN_SIZE is different than 1, because we would like to avoid going through
372 * the alloc/free process all the time. In a small machine, 4 kmem-limited
373 * cgroups is a reasonable guess. In the future, it could be a parameter or
374 * tunable, but that is strictly not necessary.
376 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
377 * this constant directly from cgroup, but it is understandable that this is
378 * better kept as an internal representation in cgroup.c. In any case, the
379 * cgrp_id space is not getting any smaller, and we don't have to necessarily
380 * increase ours as well if it increases.
382 #define MEMCG_CACHES_MIN_SIZE 4
383 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
386 * A lot of the calls to the cache allocation functions are expected to be
387 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
388 * conditional to this static branch, we'll have to allow modules that does
389 * kmem_cache_alloc and the such to see this symbol as well
391 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
392 EXPORT_SYMBOL(memcg_kmem_enabled_key);
396 * mem_cgroup_css_from_page - css of the memcg associated with a page
397 * @page: page of interest
399 * If memcg is bound to the default hierarchy, css of the memcg associated
400 * with @page is returned. The returned css remains associated with @page
401 * until it is released.
403 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
406 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
408 struct mem_cgroup *memcg;
410 memcg = page_memcg(page);
412 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
413 memcg = root_mem_cgroup;
419 * page_cgroup_ino - return inode number of the memcg a page is charged to
422 * Look up the closest online ancestor of the memory cgroup @page is charged to
423 * and return its inode number or 0 if @page is not charged to any cgroup. It
424 * is safe to call this function without holding a reference to @page.
426 * Note, this function is inherently racy, because there is nothing to prevent
427 * the cgroup inode from getting torn down and potentially reallocated a moment
428 * after page_cgroup_ino() returns, so it only should be used by callers that
429 * do not care (such as procfs interfaces).
431 ino_t page_cgroup_ino(struct page *page)
433 struct mem_cgroup *memcg;
434 unsigned long ino = 0;
437 memcg = page_memcg_check(page);
439 while (memcg && !(memcg->css.flags & CSS_ONLINE))
440 memcg = parent_mem_cgroup(memcg);
442 ino = cgroup_ino(memcg->css.cgroup);
447 static struct mem_cgroup_per_node *
448 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
450 int nid = page_to_nid(page);
452 return memcg->nodeinfo[nid];
455 static struct mem_cgroup_tree_per_node *
456 soft_limit_tree_node(int nid)
458 return soft_limit_tree.rb_tree_per_node[nid];
461 static struct mem_cgroup_tree_per_node *
462 soft_limit_tree_from_page(struct page *page)
464 int nid = page_to_nid(page);
466 return soft_limit_tree.rb_tree_per_node[nid];
469 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
470 struct mem_cgroup_tree_per_node *mctz,
471 unsigned long new_usage_in_excess)
473 struct rb_node **p = &mctz->rb_root.rb_node;
474 struct rb_node *parent = NULL;
475 struct mem_cgroup_per_node *mz_node;
476 bool rightmost = true;
481 mz->usage_in_excess = new_usage_in_excess;
482 if (!mz->usage_in_excess)
486 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
488 if (mz->usage_in_excess < mz_node->usage_in_excess) {
497 mctz->rb_rightmost = &mz->tree_node;
499 rb_link_node(&mz->tree_node, parent, p);
500 rb_insert_color(&mz->tree_node, &mctz->rb_root);
504 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
505 struct mem_cgroup_tree_per_node *mctz)
510 if (&mz->tree_node == mctz->rb_rightmost)
511 mctz->rb_rightmost = rb_prev(&mz->tree_node);
513 rb_erase(&mz->tree_node, &mctz->rb_root);
517 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
518 struct mem_cgroup_tree_per_node *mctz)
522 spin_lock_irqsave(&mctz->lock, flags);
523 __mem_cgroup_remove_exceeded(mz, mctz);
524 spin_unlock_irqrestore(&mctz->lock, flags);
527 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
529 unsigned long nr_pages = page_counter_read(&memcg->memory);
530 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
531 unsigned long excess = 0;
533 if (nr_pages > soft_limit)
534 excess = nr_pages - soft_limit;
539 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
541 unsigned long excess;
542 struct mem_cgroup_per_node *mz;
543 struct mem_cgroup_tree_per_node *mctz;
545 mctz = soft_limit_tree_from_page(page);
549 * Necessary to update all ancestors when hierarchy is used.
550 * because their event counter is not touched.
552 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
553 mz = mem_cgroup_page_nodeinfo(memcg, page);
554 excess = soft_limit_excess(memcg);
556 * We have to update the tree if mz is on RB-tree or
557 * mem is over its softlimit.
559 if (excess || mz->on_tree) {
562 spin_lock_irqsave(&mctz->lock, flags);
563 /* if on-tree, remove it */
565 __mem_cgroup_remove_exceeded(mz, mctz);
567 * Insert again. mz->usage_in_excess will be updated.
568 * If excess is 0, no tree ops.
570 __mem_cgroup_insert_exceeded(mz, mctz, excess);
571 spin_unlock_irqrestore(&mctz->lock, flags);
576 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
578 struct mem_cgroup_tree_per_node *mctz;
579 struct mem_cgroup_per_node *mz;
583 mz = memcg->nodeinfo[nid];
584 mctz = soft_limit_tree_node(nid);
586 mem_cgroup_remove_exceeded(mz, mctz);
590 static struct mem_cgroup_per_node *
591 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
593 struct mem_cgroup_per_node *mz;
597 if (!mctz->rb_rightmost)
598 goto done; /* Nothing to reclaim from */
600 mz = rb_entry(mctz->rb_rightmost,
601 struct mem_cgroup_per_node, tree_node);
603 * Remove the node now but someone else can add it back,
604 * we will to add it back at the end of reclaim to its correct
605 * position in the tree.
607 __mem_cgroup_remove_exceeded(mz, mctz);
608 if (!soft_limit_excess(mz->memcg) ||
609 !css_tryget(&mz->memcg->css))
615 static struct mem_cgroup_per_node *
616 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
618 struct mem_cgroup_per_node *mz;
620 spin_lock_irq(&mctz->lock);
621 mz = __mem_cgroup_largest_soft_limit_node(mctz);
622 spin_unlock_irq(&mctz->lock);
627 * __mod_memcg_state - update cgroup memory statistics
628 * @memcg: the memory cgroup
629 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
630 * @val: delta to add to the counter, can be negative
632 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
634 if (mem_cgroup_disabled())
637 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
638 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
641 /* idx can be of type enum memcg_stat_item or node_stat_item. */
642 static unsigned long memcg_page_state(struct mem_cgroup *memcg, int idx)
644 long x = READ_ONCE(memcg->vmstats.state[idx]);
652 /* idx can be of type enum memcg_stat_item or node_stat_item. */
653 static unsigned long memcg_page_state_local(struct mem_cgroup *memcg, int idx)
658 for_each_possible_cpu(cpu)
659 x += per_cpu(memcg->vmstats_percpu->state[idx], cpu);
667 static struct mem_cgroup_per_node *
668 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
670 struct mem_cgroup *parent;
672 parent = parent_mem_cgroup(pn->memcg);
675 return parent->nodeinfo[nid];
678 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
681 struct mem_cgroup_per_node *pn;
682 struct mem_cgroup *memcg;
683 long x, threshold = MEMCG_CHARGE_BATCH;
685 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
689 __mod_memcg_state(memcg, idx, val);
692 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
694 if (vmstat_item_in_bytes(idx))
695 threshold <<= PAGE_SHIFT;
697 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
698 if (unlikely(abs(x) > threshold)) {
699 pg_data_t *pgdat = lruvec_pgdat(lruvec);
700 struct mem_cgroup_per_node *pi;
702 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
703 atomic_long_add(x, &pi->lruvec_stat[idx]);
706 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
710 * __mod_lruvec_state - update lruvec memory statistics
711 * @lruvec: the lruvec
712 * @idx: the stat item
713 * @val: delta to add to the counter, can be negative
715 * The lruvec is the intersection of the NUMA node and a cgroup. This
716 * function updates the all three counters that are affected by a
717 * change of state at this level: per-node, per-cgroup, per-lruvec.
719 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
723 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
725 /* Update memcg and lruvec */
726 if (!mem_cgroup_disabled())
727 __mod_memcg_lruvec_state(lruvec, idx, val);
730 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
733 struct page *head = compound_head(page); /* rmap on tail pages */
734 struct mem_cgroup *memcg;
735 pg_data_t *pgdat = page_pgdat(page);
736 struct lruvec *lruvec;
739 memcg = page_memcg(head);
740 /* Untracked pages have no memcg, no lruvec. Update only the node */
743 __mod_node_page_state(pgdat, idx, val);
747 lruvec = mem_cgroup_lruvec(memcg, pgdat);
748 __mod_lruvec_state(lruvec, idx, val);
751 EXPORT_SYMBOL(__mod_lruvec_page_state);
753 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
755 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
756 struct mem_cgroup *memcg;
757 struct lruvec *lruvec;
760 memcg = mem_cgroup_from_obj(p);
763 * Untracked pages have no memcg, no lruvec. Update only the
764 * node. If we reparent the slab objects to the root memcg,
765 * when we free the slab object, we need to update the per-memcg
766 * vmstats to keep it correct for the root memcg.
769 __mod_node_page_state(pgdat, idx, val);
771 lruvec = mem_cgroup_lruvec(memcg, pgdat);
772 __mod_lruvec_state(lruvec, idx, val);
778 * mod_objcg_mlstate() may be called with irq enabled, so
779 * mod_memcg_lruvec_state() should be used.
781 static inline void mod_objcg_mlstate(struct obj_cgroup *objcg,
782 struct pglist_data *pgdat,
783 enum node_stat_item idx, int nr)
785 struct mem_cgroup *memcg;
786 struct lruvec *lruvec;
789 memcg = obj_cgroup_memcg(objcg);
790 lruvec = mem_cgroup_lruvec(memcg, pgdat);
791 mod_memcg_lruvec_state(lruvec, idx, nr);
796 * __count_memcg_events - account VM events in a cgroup
797 * @memcg: the memory cgroup
798 * @idx: the event item
799 * @count: the number of events that occurred
801 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
804 if (mem_cgroup_disabled())
807 __this_cpu_add(memcg->vmstats_percpu->events[idx], count);
808 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
811 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
813 return READ_ONCE(memcg->vmstats.events[event]);
816 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
821 for_each_possible_cpu(cpu)
822 x += per_cpu(memcg->vmstats_percpu->events[event], cpu);
826 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
830 /* pagein of a big page is an event. So, ignore page size */
832 __count_memcg_events(memcg, PGPGIN, 1);
834 __count_memcg_events(memcg, PGPGOUT, 1);
835 nr_pages = -nr_pages; /* for event */
838 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
841 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
842 enum mem_cgroup_events_target target)
844 unsigned long val, next;
846 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
847 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
848 /* from time_after() in jiffies.h */
849 if ((long)(next - val) < 0) {
851 case MEM_CGROUP_TARGET_THRESH:
852 next = val + THRESHOLDS_EVENTS_TARGET;
854 case MEM_CGROUP_TARGET_SOFTLIMIT:
855 next = val + SOFTLIMIT_EVENTS_TARGET;
860 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
867 * Check events in order.
870 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
872 /* threshold event is triggered in finer grain than soft limit */
873 if (unlikely(mem_cgroup_event_ratelimit(memcg,
874 MEM_CGROUP_TARGET_THRESH))) {
877 do_softlimit = mem_cgroup_event_ratelimit(memcg,
878 MEM_CGROUP_TARGET_SOFTLIMIT);
879 mem_cgroup_threshold(memcg);
880 if (unlikely(do_softlimit))
881 mem_cgroup_update_tree(memcg, page);
885 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
888 * mm_update_next_owner() may clear mm->owner to NULL
889 * if it races with swapoff, page migration, etc.
890 * So this can be called with p == NULL.
895 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
897 EXPORT_SYMBOL(mem_cgroup_from_task);
900 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
901 * @mm: mm from which memcg should be extracted. It can be NULL.
903 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
904 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
907 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
909 struct mem_cgroup *memcg;
911 if (mem_cgroup_disabled())
915 * Page cache insertions can happen without an
916 * actual mm context, e.g. during disk probing
917 * on boot, loopback IO, acct() writes etc.
919 * No need to css_get on root memcg as the reference
920 * counting is disabled on the root level in the
921 * cgroup core. See CSS_NO_REF.
924 return root_mem_cgroup;
928 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
929 if (unlikely(!memcg))
930 memcg = root_mem_cgroup;
931 } while (!css_tryget(&memcg->css));
935 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
937 static __always_inline struct mem_cgroup *active_memcg(void)
940 return this_cpu_read(int_active_memcg);
942 return current->active_memcg;
945 static __always_inline bool memcg_kmem_bypass(void)
947 /* Allow remote memcg charging from any context. */
948 if (unlikely(active_memcg()))
951 /* Memcg to charge can't be determined. */
952 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
959 * mem_cgroup_iter - iterate over memory cgroup hierarchy
960 * @root: hierarchy root
961 * @prev: previously returned memcg, NULL on first invocation
962 * @reclaim: cookie for shared reclaim walks, NULL for full walks
964 * Returns references to children of the hierarchy below @root, or
965 * @root itself, or %NULL after a full round-trip.
967 * Caller must pass the return value in @prev on subsequent
968 * invocations for reference counting, or use mem_cgroup_iter_break()
969 * to cancel a hierarchy walk before the round-trip is complete.
971 * Reclaimers can specify a node in @reclaim to divide up the memcgs
972 * in the hierarchy among all concurrent reclaimers operating on the
975 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
976 struct mem_cgroup *prev,
977 struct mem_cgroup_reclaim_cookie *reclaim)
979 struct mem_cgroup_reclaim_iter *iter;
980 struct cgroup_subsys_state *css = NULL;
981 struct mem_cgroup *memcg = NULL;
982 struct mem_cgroup *pos = NULL;
984 if (mem_cgroup_disabled())
988 root = root_mem_cgroup;
990 if (prev && !reclaim)
996 struct mem_cgroup_per_node *mz;
998 mz = root->nodeinfo[reclaim->pgdat->node_id];
1001 if (prev && reclaim->generation != iter->generation)
1005 pos = READ_ONCE(iter->position);
1006 if (!pos || css_tryget(&pos->css))
1009 * css reference reached zero, so iter->position will
1010 * be cleared by ->css_released. However, we should not
1011 * rely on this happening soon, because ->css_released
1012 * is called from a work queue, and by busy-waiting we
1013 * might block it. So we clear iter->position right
1016 (void)cmpxchg(&iter->position, pos, NULL);
1024 css = css_next_descendant_pre(css, &root->css);
1027 * Reclaimers share the hierarchy walk, and a
1028 * new one might jump in right at the end of
1029 * the hierarchy - make sure they see at least
1030 * one group and restart from the beginning.
1038 * Verify the css and acquire a reference. The root
1039 * is provided by the caller, so we know it's alive
1040 * and kicking, and don't take an extra reference.
1042 memcg = mem_cgroup_from_css(css);
1044 if (css == &root->css)
1047 if (css_tryget(css))
1055 * The position could have already been updated by a competing
1056 * thread, so check that the value hasn't changed since we read
1057 * it to avoid reclaiming from the same cgroup twice.
1059 (void)cmpxchg(&iter->position, pos, memcg);
1067 reclaim->generation = iter->generation;
1072 if (prev && prev != root)
1073 css_put(&prev->css);
1079 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1080 * @root: hierarchy root
1081 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1083 void mem_cgroup_iter_break(struct mem_cgroup *root,
1084 struct mem_cgroup *prev)
1087 root = root_mem_cgroup;
1088 if (prev && prev != root)
1089 css_put(&prev->css);
1092 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1093 struct mem_cgroup *dead_memcg)
1095 struct mem_cgroup_reclaim_iter *iter;
1096 struct mem_cgroup_per_node *mz;
1099 for_each_node(nid) {
1100 mz = from->nodeinfo[nid];
1102 cmpxchg(&iter->position, dead_memcg, NULL);
1106 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1108 struct mem_cgroup *memcg = dead_memcg;
1109 struct mem_cgroup *last;
1112 __invalidate_reclaim_iterators(memcg, dead_memcg);
1114 } while ((memcg = parent_mem_cgroup(memcg)));
1117 * When cgruop1 non-hierarchy mode is used,
1118 * parent_mem_cgroup() does not walk all the way up to the
1119 * cgroup root (root_mem_cgroup). So we have to handle
1120 * dead_memcg from cgroup root separately.
1122 if (last != root_mem_cgroup)
1123 __invalidate_reclaim_iterators(root_mem_cgroup,
1128 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1129 * @memcg: hierarchy root
1130 * @fn: function to call for each task
1131 * @arg: argument passed to @fn
1133 * This function iterates over tasks attached to @memcg or to any of its
1134 * descendants and calls @fn for each task. If @fn returns a non-zero
1135 * value, the function breaks the iteration loop and returns the value.
1136 * Otherwise, it will iterate over all tasks and return 0.
1138 * This function must not be called for the root memory cgroup.
1140 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1141 int (*fn)(struct task_struct *, void *), void *arg)
1143 struct mem_cgroup *iter;
1146 BUG_ON(memcg == root_mem_cgroup);
1148 for_each_mem_cgroup_tree(iter, memcg) {
1149 struct css_task_iter it;
1150 struct task_struct *task;
1152 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1153 while (!ret && (task = css_task_iter_next(&it)))
1154 ret = fn(task, arg);
1155 css_task_iter_end(&it);
1157 mem_cgroup_iter_break(memcg, iter);
1164 #ifdef CONFIG_DEBUG_VM
1165 void lruvec_memcg_debug(struct lruvec *lruvec, struct page *page)
1167 struct mem_cgroup *memcg;
1169 if (mem_cgroup_disabled())
1172 memcg = page_memcg(page);
1175 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != root_mem_cgroup, page);
1177 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != memcg, page);
1182 * lock_page_lruvec - lock and return lruvec for a given page.
1185 * These functions are safe to use under any of the following conditions:
1188 * - lock_page_memcg()
1189 * - page->_refcount is zero
1191 struct lruvec *lock_page_lruvec(struct page *page)
1193 struct lruvec *lruvec;
1195 lruvec = mem_cgroup_page_lruvec(page);
1196 spin_lock(&lruvec->lru_lock);
1198 lruvec_memcg_debug(lruvec, page);
1203 struct lruvec *lock_page_lruvec_irq(struct page *page)
1205 struct lruvec *lruvec;
1207 lruvec = mem_cgroup_page_lruvec(page);
1208 spin_lock_irq(&lruvec->lru_lock);
1210 lruvec_memcg_debug(lruvec, page);
1215 struct lruvec *lock_page_lruvec_irqsave(struct page *page, unsigned long *flags)
1217 struct lruvec *lruvec;
1219 lruvec = mem_cgroup_page_lruvec(page);
1220 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1222 lruvec_memcg_debug(lruvec, page);
1228 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1229 * @lruvec: mem_cgroup per zone lru vector
1230 * @lru: index of lru list the page is sitting on
1231 * @zid: zone id of the accounted pages
1232 * @nr_pages: positive when adding or negative when removing
1234 * This function must be called under lru_lock, just before a page is added
1235 * to or just after a page is removed from an lru list (that ordering being
1236 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1238 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1239 int zid, int nr_pages)
1241 struct mem_cgroup_per_node *mz;
1242 unsigned long *lru_size;
1245 if (mem_cgroup_disabled())
1248 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1249 lru_size = &mz->lru_zone_size[zid][lru];
1252 *lru_size += nr_pages;
1255 if (WARN_ONCE(size < 0,
1256 "%s(%p, %d, %d): lru_size %ld\n",
1257 __func__, lruvec, lru, nr_pages, size)) {
1263 *lru_size += nr_pages;
1267 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1268 * @memcg: the memory cgroup
1270 * Returns the maximum amount of memory @mem can be charged with, in
1273 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1275 unsigned long margin = 0;
1276 unsigned long count;
1277 unsigned long limit;
1279 count = page_counter_read(&memcg->memory);
1280 limit = READ_ONCE(memcg->memory.max);
1282 margin = limit - count;
1284 if (do_memsw_account()) {
1285 count = page_counter_read(&memcg->memsw);
1286 limit = READ_ONCE(memcg->memsw.max);
1288 margin = min(margin, limit - count);
1297 * A routine for checking "mem" is under move_account() or not.
1299 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1300 * moving cgroups. This is for waiting at high-memory pressure
1303 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1305 struct mem_cgroup *from;
1306 struct mem_cgroup *to;
1309 * Unlike task_move routines, we access mc.to, mc.from not under
1310 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1312 spin_lock(&mc.lock);
1318 ret = mem_cgroup_is_descendant(from, memcg) ||
1319 mem_cgroup_is_descendant(to, memcg);
1321 spin_unlock(&mc.lock);
1325 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1327 if (mc.moving_task && current != mc.moving_task) {
1328 if (mem_cgroup_under_move(memcg)) {
1330 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1331 /* moving charge context might have finished. */
1334 finish_wait(&mc.waitq, &wait);
1341 struct memory_stat {
1346 static const struct memory_stat memory_stats[] = {
1347 { "anon", NR_ANON_MAPPED },
1348 { "file", NR_FILE_PAGES },
1349 { "kernel_stack", NR_KERNEL_STACK_KB },
1350 { "pagetables", NR_PAGETABLE },
1351 { "percpu", MEMCG_PERCPU_B },
1352 { "sock", MEMCG_SOCK },
1353 { "shmem", NR_SHMEM },
1354 { "file_mapped", NR_FILE_MAPPED },
1355 { "file_dirty", NR_FILE_DIRTY },
1356 { "file_writeback", NR_WRITEBACK },
1358 { "swapcached", NR_SWAPCACHE },
1360 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1361 { "anon_thp", NR_ANON_THPS },
1362 { "file_thp", NR_FILE_THPS },
1363 { "shmem_thp", NR_SHMEM_THPS },
1365 { "inactive_anon", NR_INACTIVE_ANON },
1366 { "active_anon", NR_ACTIVE_ANON },
1367 { "inactive_file", NR_INACTIVE_FILE },
1368 { "active_file", NR_ACTIVE_FILE },
1369 { "unevictable", NR_UNEVICTABLE },
1370 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B },
1371 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B },
1373 /* The memory events */
1374 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON },
1375 { "workingset_refault_file", WORKINGSET_REFAULT_FILE },
1376 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON },
1377 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE },
1378 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON },
1379 { "workingset_restore_file", WORKINGSET_RESTORE_FILE },
1380 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM },
1383 /* Translate stat items to the correct unit for memory.stat output */
1384 static int memcg_page_state_unit(int item)
1387 case MEMCG_PERCPU_B:
1388 case NR_SLAB_RECLAIMABLE_B:
1389 case NR_SLAB_UNRECLAIMABLE_B:
1390 case WORKINGSET_REFAULT_ANON:
1391 case WORKINGSET_REFAULT_FILE:
1392 case WORKINGSET_ACTIVATE_ANON:
1393 case WORKINGSET_ACTIVATE_FILE:
1394 case WORKINGSET_RESTORE_ANON:
1395 case WORKINGSET_RESTORE_FILE:
1396 case WORKINGSET_NODERECLAIM:
1398 case NR_KERNEL_STACK_KB:
1405 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1408 return memcg_page_state(memcg, item) * memcg_page_state_unit(item);
1411 static char *memory_stat_format(struct mem_cgroup *memcg)
1416 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1421 * Provide statistics on the state of the memory subsystem as
1422 * well as cumulative event counters that show past behavior.
1424 * This list is ordered following a combination of these gradients:
1425 * 1) generic big picture -> specifics and details
1426 * 2) reflecting userspace activity -> reflecting kernel heuristics
1428 * Current memory state:
1430 cgroup_rstat_flush(memcg->css.cgroup);
1432 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1435 size = memcg_page_state_output(memcg, memory_stats[i].idx);
1436 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1438 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1439 size += memcg_page_state_output(memcg,
1440 NR_SLAB_RECLAIMABLE_B);
1441 seq_buf_printf(&s, "slab %llu\n", size);
1445 /* Accumulated memory events */
1447 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1448 memcg_events(memcg, PGFAULT));
1449 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1450 memcg_events(memcg, PGMAJFAULT));
1451 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1452 memcg_events(memcg, PGREFILL));
1453 seq_buf_printf(&s, "pgscan %lu\n",
1454 memcg_events(memcg, PGSCAN_KSWAPD) +
1455 memcg_events(memcg, PGSCAN_DIRECT));
1456 seq_buf_printf(&s, "pgsteal %lu\n",
1457 memcg_events(memcg, PGSTEAL_KSWAPD) +
1458 memcg_events(memcg, PGSTEAL_DIRECT));
1459 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1460 memcg_events(memcg, PGACTIVATE));
1461 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1462 memcg_events(memcg, PGDEACTIVATE));
1463 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1464 memcg_events(memcg, PGLAZYFREE));
1465 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1466 memcg_events(memcg, PGLAZYFREED));
1468 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1469 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1470 memcg_events(memcg, THP_FAULT_ALLOC));
1471 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1472 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1473 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1475 /* The above should easily fit into one page */
1476 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1481 #define K(x) ((x) << (PAGE_SHIFT-10))
1483 * mem_cgroup_print_oom_context: Print OOM information relevant to
1484 * memory controller.
1485 * @memcg: The memory cgroup that went over limit
1486 * @p: Task that is going to be killed
1488 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1491 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1496 pr_cont(",oom_memcg=");
1497 pr_cont_cgroup_path(memcg->css.cgroup);
1499 pr_cont(",global_oom");
1501 pr_cont(",task_memcg=");
1502 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1508 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1509 * memory controller.
1510 * @memcg: The memory cgroup that went over limit
1512 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1516 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1517 K((u64)page_counter_read(&memcg->memory)),
1518 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1519 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1520 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1521 K((u64)page_counter_read(&memcg->swap)),
1522 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1524 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1525 K((u64)page_counter_read(&memcg->memsw)),
1526 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1527 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1528 K((u64)page_counter_read(&memcg->kmem)),
1529 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1532 pr_info("Memory cgroup stats for ");
1533 pr_cont_cgroup_path(memcg->css.cgroup);
1535 buf = memory_stat_format(memcg);
1543 * Return the memory (and swap, if configured) limit for a memcg.
1545 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1547 unsigned long max = READ_ONCE(memcg->memory.max);
1549 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1550 if (mem_cgroup_swappiness(memcg))
1551 max += min(READ_ONCE(memcg->swap.max),
1552 (unsigned long)total_swap_pages);
1554 if (mem_cgroup_swappiness(memcg)) {
1555 /* Calculate swap excess capacity from memsw limit */
1556 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1558 max += min(swap, (unsigned long)total_swap_pages);
1564 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1566 return page_counter_read(&memcg->memory);
1569 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1572 struct oom_control oc = {
1576 .gfp_mask = gfp_mask,
1581 if (mutex_lock_killable(&oom_lock))
1584 if (mem_cgroup_margin(memcg) >= (1 << order))
1588 * A few threads which were not waiting at mutex_lock_killable() can
1589 * fail to bail out. Therefore, check again after holding oom_lock.
1591 ret = should_force_charge() || out_of_memory(&oc);
1594 mutex_unlock(&oom_lock);
1598 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1601 unsigned long *total_scanned)
1603 struct mem_cgroup *victim = NULL;
1606 unsigned long excess;
1607 unsigned long nr_scanned;
1608 struct mem_cgroup_reclaim_cookie reclaim = {
1612 excess = soft_limit_excess(root_memcg);
1615 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1620 * If we have not been able to reclaim
1621 * anything, it might because there are
1622 * no reclaimable pages under this hierarchy
1627 * We want to do more targeted reclaim.
1628 * excess >> 2 is not to excessive so as to
1629 * reclaim too much, nor too less that we keep
1630 * coming back to reclaim from this cgroup
1632 if (total >= (excess >> 2) ||
1633 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1638 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1639 pgdat, &nr_scanned);
1640 *total_scanned += nr_scanned;
1641 if (!soft_limit_excess(root_memcg))
1644 mem_cgroup_iter_break(root_memcg, victim);
1648 #ifdef CONFIG_LOCKDEP
1649 static struct lockdep_map memcg_oom_lock_dep_map = {
1650 .name = "memcg_oom_lock",
1654 static DEFINE_SPINLOCK(memcg_oom_lock);
1657 * Check OOM-Killer is already running under our hierarchy.
1658 * If someone is running, return false.
1660 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1662 struct mem_cgroup *iter, *failed = NULL;
1664 spin_lock(&memcg_oom_lock);
1666 for_each_mem_cgroup_tree(iter, memcg) {
1667 if (iter->oom_lock) {
1669 * this subtree of our hierarchy is already locked
1670 * so we cannot give a lock.
1673 mem_cgroup_iter_break(memcg, iter);
1676 iter->oom_lock = true;
1681 * OK, we failed to lock the whole subtree so we have
1682 * to clean up what we set up to the failing subtree
1684 for_each_mem_cgroup_tree(iter, memcg) {
1685 if (iter == failed) {
1686 mem_cgroup_iter_break(memcg, iter);
1689 iter->oom_lock = false;
1692 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1694 spin_unlock(&memcg_oom_lock);
1699 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1701 struct mem_cgroup *iter;
1703 spin_lock(&memcg_oom_lock);
1704 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1705 for_each_mem_cgroup_tree(iter, memcg)
1706 iter->oom_lock = false;
1707 spin_unlock(&memcg_oom_lock);
1710 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1712 struct mem_cgroup *iter;
1714 spin_lock(&memcg_oom_lock);
1715 for_each_mem_cgroup_tree(iter, memcg)
1717 spin_unlock(&memcg_oom_lock);
1720 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1722 struct mem_cgroup *iter;
1725 * Be careful about under_oom underflows because a child memcg
1726 * could have been added after mem_cgroup_mark_under_oom.
1728 spin_lock(&memcg_oom_lock);
1729 for_each_mem_cgroup_tree(iter, memcg)
1730 if (iter->under_oom > 0)
1732 spin_unlock(&memcg_oom_lock);
1735 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1737 struct oom_wait_info {
1738 struct mem_cgroup *memcg;
1739 wait_queue_entry_t wait;
1742 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1743 unsigned mode, int sync, void *arg)
1745 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1746 struct mem_cgroup *oom_wait_memcg;
1747 struct oom_wait_info *oom_wait_info;
1749 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1750 oom_wait_memcg = oom_wait_info->memcg;
1752 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1753 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1755 return autoremove_wake_function(wait, mode, sync, arg);
1758 static void memcg_oom_recover(struct mem_cgroup *memcg)
1761 * For the following lockless ->under_oom test, the only required
1762 * guarantee is that it must see the state asserted by an OOM when
1763 * this function is called as a result of userland actions
1764 * triggered by the notification of the OOM. This is trivially
1765 * achieved by invoking mem_cgroup_mark_under_oom() before
1766 * triggering notification.
1768 if (memcg && memcg->under_oom)
1769 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1779 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1781 enum oom_status ret;
1784 if (order > PAGE_ALLOC_COSTLY_ORDER)
1787 memcg_memory_event(memcg, MEMCG_OOM);
1790 * We are in the middle of the charge context here, so we
1791 * don't want to block when potentially sitting on a callstack
1792 * that holds all kinds of filesystem and mm locks.
1794 * cgroup1 allows disabling the OOM killer and waiting for outside
1795 * handling until the charge can succeed; remember the context and put
1796 * the task to sleep at the end of the page fault when all locks are
1799 * On the other hand, in-kernel OOM killer allows for an async victim
1800 * memory reclaim (oom_reaper) and that means that we are not solely
1801 * relying on the oom victim to make a forward progress and we can
1802 * invoke the oom killer here.
1804 * Please note that mem_cgroup_out_of_memory might fail to find a
1805 * victim and then we have to bail out from the charge path.
1807 if (memcg->oom_kill_disable) {
1808 if (!current->in_user_fault)
1810 css_get(&memcg->css);
1811 current->memcg_in_oom = memcg;
1812 current->memcg_oom_gfp_mask = mask;
1813 current->memcg_oom_order = order;
1818 mem_cgroup_mark_under_oom(memcg);
1820 locked = mem_cgroup_oom_trylock(memcg);
1823 mem_cgroup_oom_notify(memcg);
1825 mem_cgroup_unmark_under_oom(memcg);
1826 if (mem_cgroup_out_of_memory(memcg, mask, order))
1832 mem_cgroup_oom_unlock(memcg);
1838 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1839 * @handle: actually kill/wait or just clean up the OOM state
1841 * This has to be called at the end of a page fault if the memcg OOM
1842 * handler was enabled.
1844 * Memcg supports userspace OOM handling where failed allocations must
1845 * sleep on a waitqueue until the userspace task resolves the
1846 * situation. Sleeping directly in the charge context with all kinds
1847 * of locks held is not a good idea, instead we remember an OOM state
1848 * in the task and mem_cgroup_oom_synchronize() has to be called at
1849 * the end of the page fault to complete the OOM handling.
1851 * Returns %true if an ongoing memcg OOM situation was detected and
1852 * completed, %false otherwise.
1854 bool mem_cgroup_oom_synchronize(bool handle)
1856 struct mem_cgroup *memcg = current->memcg_in_oom;
1857 struct oom_wait_info owait;
1860 /* OOM is global, do not handle */
1867 owait.memcg = memcg;
1868 owait.wait.flags = 0;
1869 owait.wait.func = memcg_oom_wake_function;
1870 owait.wait.private = current;
1871 INIT_LIST_HEAD(&owait.wait.entry);
1873 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1874 mem_cgroup_mark_under_oom(memcg);
1876 locked = mem_cgroup_oom_trylock(memcg);
1879 mem_cgroup_oom_notify(memcg);
1881 if (locked && !memcg->oom_kill_disable) {
1882 mem_cgroup_unmark_under_oom(memcg);
1883 finish_wait(&memcg_oom_waitq, &owait.wait);
1884 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1885 current->memcg_oom_order);
1888 mem_cgroup_unmark_under_oom(memcg);
1889 finish_wait(&memcg_oom_waitq, &owait.wait);
1893 mem_cgroup_oom_unlock(memcg);
1895 * There is no guarantee that an OOM-lock contender
1896 * sees the wakeups triggered by the OOM kill
1897 * uncharges. Wake any sleepers explicitly.
1899 memcg_oom_recover(memcg);
1902 current->memcg_in_oom = NULL;
1903 css_put(&memcg->css);
1908 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1909 * @victim: task to be killed by the OOM killer
1910 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1912 * Returns a pointer to a memory cgroup, which has to be cleaned up
1913 * by killing all belonging OOM-killable tasks.
1915 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1917 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1918 struct mem_cgroup *oom_domain)
1920 struct mem_cgroup *oom_group = NULL;
1921 struct mem_cgroup *memcg;
1923 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1927 oom_domain = root_mem_cgroup;
1931 memcg = mem_cgroup_from_task(victim);
1932 if (memcg == root_mem_cgroup)
1936 * If the victim task has been asynchronously moved to a different
1937 * memory cgroup, we might end up killing tasks outside oom_domain.
1938 * In this case it's better to ignore memory.group.oom.
1940 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
1944 * Traverse the memory cgroup hierarchy from the victim task's
1945 * cgroup up to the OOMing cgroup (or root) to find the
1946 * highest-level memory cgroup with oom.group set.
1948 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1949 if (memcg->oom_group)
1952 if (memcg == oom_domain)
1957 css_get(&oom_group->css);
1964 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1966 pr_info("Tasks in ");
1967 pr_cont_cgroup_path(memcg->css.cgroup);
1968 pr_cont(" are going to be killed due to memory.oom.group set\n");
1972 * lock_page_memcg - lock a page and memcg binding
1975 * This function protects unlocked LRU pages from being moved to
1978 * It ensures lifetime of the locked memcg. Caller is responsible
1979 * for the lifetime of the page.
1981 void lock_page_memcg(struct page *page)
1983 struct page *head = compound_head(page); /* rmap on tail pages */
1984 struct mem_cgroup *memcg;
1985 unsigned long flags;
1988 * The RCU lock is held throughout the transaction. The fast
1989 * path can get away without acquiring the memcg->move_lock
1990 * because page moving starts with an RCU grace period.
1994 if (mem_cgroup_disabled())
1997 memcg = page_memcg(head);
1998 if (unlikely(!memcg))
2001 #ifdef CONFIG_PROVE_LOCKING
2002 local_irq_save(flags);
2003 might_lock(&memcg->move_lock);
2004 local_irq_restore(flags);
2007 if (atomic_read(&memcg->moving_account) <= 0)
2010 spin_lock_irqsave(&memcg->move_lock, flags);
2011 if (memcg != page_memcg(head)) {
2012 spin_unlock_irqrestore(&memcg->move_lock, flags);
2017 * When charge migration first begins, we can have multiple
2018 * critical sections holding the fast-path RCU lock and one
2019 * holding the slowpath move_lock. Track the task who has the
2020 * move_lock for unlock_page_memcg().
2022 memcg->move_lock_task = current;
2023 memcg->move_lock_flags = flags;
2025 EXPORT_SYMBOL(lock_page_memcg);
2027 static void __unlock_page_memcg(struct mem_cgroup *memcg)
2029 if (memcg && memcg->move_lock_task == current) {
2030 unsigned long flags = memcg->move_lock_flags;
2032 memcg->move_lock_task = NULL;
2033 memcg->move_lock_flags = 0;
2035 spin_unlock_irqrestore(&memcg->move_lock, flags);
2042 * unlock_page_memcg - unlock a page and memcg binding
2045 void unlock_page_memcg(struct page *page)
2047 struct page *head = compound_head(page);
2049 __unlock_page_memcg(page_memcg(head));
2051 EXPORT_SYMBOL(unlock_page_memcg);
2054 #ifdef CONFIG_MEMCG_KMEM
2055 struct obj_cgroup *cached_objcg;
2056 struct pglist_data *cached_pgdat;
2057 unsigned int nr_bytes;
2058 int nr_slab_reclaimable_b;
2059 int nr_slab_unreclaimable_b;
2065 struct memcg_stock_pcp {
2066 struct mem_cgroup *cached; /* this never be root cgroup */
2067 unsigned int nr_pages;
2068 struct obj_stock task_obj;
2069 struct obj_stock irq_obj;
2071 struct work_struct work;
2072 unsigned long flags;
2073 #define FLUSHING_CACHED_CHARGE 0
2075 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2076 static DEFINE_MUTEX(percpu_charge_mutex);
2078 #ifdef CONFIG_MEMCG_KMEM
2079 static void drain_obj_stock(struct obj_stock *stock);
2080 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2081 struct mem_cgroup *root_memcg);
2084 static inline void drain_obj_stock(struct obj_stock *stock)
2087 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2088 struct mem_cgroup *root_memcg)
2095 * Most kmem_cache_alloc() calls are from user context. The irq disable/enable
2096 * sequence used in this case to access content from object stock is slow.
2097 * To optimize for user context access, there are now two object stocks for
2098 * task context and interrupt context access respectively.
2100 * The task context object stock can be accessed by disabling preemption only
2101 * which is cheap in non-preempt kernel. The interrupt context object stock
2102 * can only be accessed after disabling interrupt. User context code can
2103 * access interrupt object stock, but not vice versa.
2105 static inline struct obj_stock *get_obj_stock(unsigned long *pflags)
2107 struct memcg_stock_pcp *stock;
2109 if (likely(in_task())) {
2112 stock = this_cpu_ptr(&memcg_stock);
2113 return &stock->task_obj;
2116 local_irq_save(*pflags);
2117 stock = this_cpu_ptr(&memcg_stock);
2118 return &stock->irq_obj;
2121 static inline void put_obj_stock(unsigned long flags)
2123 if (likely(in_task()))
2126 local_irq_restore(flags);
2130 * consume_stock: Try to consume stocked charge on this cpu.
2131 * @memcg: memcg to consume from.
2132 * @nr_pages: how many pages to charge.
2134 * The charges will only happen if @memcg matches the current cpu's memcg
2135 * stock, and at least @nr_pages are available in that stock. Failure to
2136 * service an allocation will refill the stock.
2138 * returns true if successful, false otherwise.
2140 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2142 struct memcg_stock_pcp *stock;
2143 unsigned long flags;
2146 if (nr_pages > MEMCG_CHARGE_BATCH)
2149 local_irq_save(flags);
2151 stock = this_cpu_ptr(&memcg_stock);
2152 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2153 stock->nr_pages -= nr_pages;
2157 local_irq_restore(flags);
2163 * Returns stocks cached in percpu and reset cached information.
2165 static void drain_stock(struct memcg_stock_pcp *stock)
2167 struct mem_cgroup *old = stock->cached;
2172 if (stock->nr_pages) {
2173 page_counter_uncharge(&old->memory, stock->nr_pages);
2174 if (do_memsw_account())
2175 page_counter_uncharge(&old->memsw, stock->nr_pages);
2176 stock->nr_pages = 0;
2180 stock->cached = NULL;
2183 static void drain_local_stock(struct work_struct *dummy)
2185 struct memcg_stock_pcp *stock;
2186 unsigned long flags;
2189 * The only protection from memory hotplug vs. drain_stock races is
2190 * that we always operate on local CPU stock here with IRQ disabled
2192 local_irq_save(flags);
2194 stock = this_cpu_ptr(&memcg_stock);
2195 drain_obj_stock(&stock->irq_obj);
2197 drain_obj_stock(&stock->task_obj);
2199 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2201 local_irq_restore(flags);
2205 * Cache charges(val) to local per_cpu area.
2206 * This will be consumed by consume_stock() function, later.
2208 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2210 struct memcg_stock_pcp *stock;
2211 unsigned long flags;
2213 local_irq_save(flags);
2215 stock = this_cpu_ptr(&memcg_stock);
2216 if (stock->cached != memcg) { /* reset if necessary */
2218 css_get(&memcg->css);
2219 stock->cached = memcg;
2221 stock->nr_pages += nr_pages;
2223 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2226 local_irq_restore(flags);
2230 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2231 * of the hierarchy under it.
2233 static void drain_all_stock(struct mem_cgroup *root_memcg)
2237 /* If someone's already draining, avoid adding running more workers. */
2238 if (!mutex_trylock(&percpu_charge_mutex))
2241 * Notify other cpus that system-wide "drain" is running
2242 * We do not care about races with the cpu hotplug because cpu down
2243 * as well as workers from this path always operate on the local
2244 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2247 for_each_online_cpu(cpu) {
2248 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2249 struct mem_cgroup *memcg;
2253 memcg = stock->cached;
2254 if (memcg && stock->nr_pages &&
2255 mem_cgroup_is_descendant(memcg, root_memcg))
2257 if (obj_stock_flush_required(stock, root_memcg))
2262 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2264 drain_local_stock(&stock->work);
2266 schedule_work_on(cpu, &stock->work);
2270 mutex_unlock(&percpu_charge_mutex);
2273 static void memcg_flush_lruvec_page_state(struct mem_cgroup *memcg, int cpu)
2277 for_each_node(nid) {
2278 struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
2279 unsigned long stat[NR_VM_NODE_STAT_ITEMS];
2280 struct batched_lruvec_stat *lstatc;
2283 lstatc = per_cpu_ptr(pn->lruvec_stat_cpu, cpu);
2284 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
2285 stat[i] = lstatc->count[i];
2286 lstatc->count[i] = 0;
2290 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
2291 atomic_long_add(stat[i], &pn->lruvec_stat[i]);
2292 } while ((pn = parent_nodeinfo(pn, nid)));
2296 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2298 struct memcg_stock_pcp *stock;
2299 struct mem_cgroup *memcg;
2301 stock = &per_cpu(memcg_stock, cpu);
2304 for_each_mem_cgroup(memcg)
2305 memcg_flush_lruvec_page_state(memcg, cpu);
2310 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2311 unsigned int nr_pages,
2314 unsigned long nr_reclaimed = 0;
2317 unsigned long pflags;
2319 if (page_counter_read(&memcg->memory) <=
2320 READ_ONCE(memcg->memory.high))
2323 memcg_memory_event(memcg, MEMCG_HIGH);
2325 psi_memstall_enter(&pflags);
2326 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2328 psi_memstall_leave(&pflags);
2329 } while ((memcg = parent_mem_cgroup(memcg)) &&
2330 !mem_cgroup_is_root(memcg));
2332 return nr_reclaimed;
2335 static void high_work_func(struct work_struct *work)
2337 struct mem_cgroup *memcg;
2339 memcg = container_of(work, struct mem_cgroup, high_work);
2340 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2344 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2345 * enough to still cause a significant slowdown in most cases, while still
2346 * allowing diagnostics and tracing to proceed without becoming stuck.
2348 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2351 * When calculating the delay, we use these either side of the exponentiation to
2352 * maintain precision and scale to a reasonable number of jiffies (see the table
2355 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2356 * overage ratio to a delay.
2357 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2358 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2359 * to produce a reasonable delay curve.
2361 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2362 * reasonable delay curve compared to precision-adjusted overage, not
2363 * penalising heavily at first, but still making sure that growth beyond the
2364 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2365 * example, with a high of 100 megabytes:
2367 * +-------+------------------------+
2368 * | usage | time to allocate in ms |
2369 * +-------+------------------------+
2391 * +-------+------------------------+
2393 #define MEMCG_DELAY_PRECISION_SHIFT 20
2394 #define MEMCG_DELAY_SCALING_SHIFT 14
2396 static u64 calculate_overage(unsigned long usage, unsigned long high)
2404 * Prevent division by 0 in overage calculation by acting as if
2405 * it was a threshold of 1 page
2407 high = max(high, 1UL);
2409 overage = usage - high;
2410 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2411 return div64_u64(overage, high);
2414 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2416 u64 overage, max_overage = 0;
2419 overage = calculate_overage(page_counter_read(&memcg->memory),
2420 READ_ONCE(memcg->memory.high));
2421 max_overage = max(overage, max_overage);
2422 } while ((memcg = parent_mem_cgroup(memcg)) &&
2423 !mem_cgroup_is_root(memcg));
2428 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2430 u64 overage, max_overage = 0;
2433 overage = calculate_overage(page_counter_read(&memcg->swap),
2434 READ_ONCE(memcg->swap.high));
2436 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2437 max_overage = max(overage, max_overage);
2438 } while ((memcg = parent_mem_cgroup(memcg)) &&
2439 !mem_cgroup_is_root(memcg));
2445 * Get the number of jiffies that we should penalise a mischievous cgroup which
2446 * is exceeding its memory.high by checking both it and its ancestors.
2448 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2449 unsigned int nr_pages,
2452 unsigned long penalty_jiffies;
2458 * We use overage compared to memory.high to calculate the number of
2459 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2460 * fairly lenient on small overages, and increasingly harsh when the
2461 * memcg in question makes it clear that it has no intention of stopping
2462 * its crazy behaviour, so we exponentially increase the delay based on
2465 penalty_jiffies = max_overage * max_overage * HZ;
2466 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2467 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2470 * Factor in the task's own contribution to the overage, such that four
2471 * N-sized allocations are throttled approximately the same as one
2472 * 4N-sized allocation.
2474 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2475 * larger the current charge patch is than that.
2477 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2481 * Scheduled by try_charge() to be executed from the userland return path
2482 * and reclaims memory over the high limit.
2484 void mem_cgroup_handle_over_high(void)
2486 unsigned long penalty_jiffies;
2487 unsigned long pflags;
2488 unsigned long nr_reclaimed;
2489 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2490 int nr_retries = MAX_RECLAIM_RETRIES;
2491 struct mem_cgroup *memcg;
2492 bool in_retry = false;
2494 if (likely(!nr_pages))
2497 memcg = get_mem_cgroup_from_mm(current->mm);
2498 current->memcg_nr_pages_over_high = 0;
2502 * The allocating task should reclaim at least the batch size, but for
2503 * subsequent retries we only want to do what's necessary to prevent oom
2504 * or breaching resource isolation.
2506 * This is distinct from memory.max or page allocator behaviour because
2507 * memory.high is currently batched, whereas memory.max and the page
2508 * allocator run every time an allocation is made.
2510 nr_reclaimed = reclaim_high(memcg,
2511 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2515 * memory.high is breached and reclaim is unable to keep up. Throttle
2516 * allocators proactively to slow down excessive growth.
2518 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2519 mem_find_max_overage(memcg));
2521 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2522 swap_find_max_overage(memcg));
2525 * Clamp the max delay per usermode return so as to still keep the
2526 * application moving forwards and also permit diagnostics, albeit
2529 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2532 * Don't sleep if the amount of jiffies this memcg owes us is so low
2533 * that it's not even worth doing, in an attempt to be nice to those who
2534 * go only a small amount over their memory.high value and maybe haven't
2535 * been aggressively reclaimed enough yet.
2537 if (penalty_jiffies <= HZ / 100)
2541 * If reclaim is making forward progress but we're still over
2542 * memory.high, we want to encourage that rather than doing allocator
2545 if (nr_reclaimed || nr_retries--) {
2551 * If we exit early, we're guaranteed to die (since
2552 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2553 * need to account for any ill-begotten jiffies to pay them off later.
2555 psi_memstall_enter(&pflags);
2556 schedule_timeout_killable(penalty_jiffies);
2557 psi_memstall_leave(&pflags);
2560 css_put(&memcg->css);
2563 static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
2564 unsigned int nr_pages)
2566 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2567 int nr_retries = MAX_RECLAIM_RETRIES;
2568 struct mem_cgroup *mem_over_limit;
2569 struct page_counter *counter;
2570 enum oom_status oom_status;
2571 unsigned long nr_reclaimed;
2572 bool may_swap = true;
2573 bool drained = false;
2574 unsigned long pflags;
2577 if (consume_stock(memcg, nr_pages))
2580 if (!do_memsw_account() ||
2581 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2582 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2584 if (do_memsw_account())
2585 page_counter_uncharge(&memcg->memsw, batch);
2586 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2588 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2592 if (batch > nr_pages) {
2598 * Memcg doesn't have a dedicated reserve for atomic
2599 * allocations. But like the global atomic pool, we need to
2600 * put the burden of reclaim on regular allocation requests
2601 * and let these go through as privileged allocations.
2603 if (gfp_mask & __GFP_ATOMIC)
2607 * Unlike in global OOM situations, memcg is not in a physical
2608 * memory shortage. Allow dying and OOM-killed tasks to
2609 * bypass the last charges so that they can exit quickly and
2610 * free their memory.
2612 if (unlikely(should_force_charge()))
2616 * Prevent unbounded recursion when reclaim operations need to
2617 * allocate memory. This might exceed the limits temporarily,
2618 * but we prefer facilitating memory reclaim and getting back
2619 * under the limit over triggering OOM kills in these cases.
2621 if (unlikely(current->flags & PF_MEMALLOC))
2624 if (unlikely(task_in_memcg_oom(current)))
2627 if (!gfpflags_allow_blocking(gfp_mask))
2630 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2632 psi_memstall_enter(&pflags);
2633 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2634 gfp_mask, may_swap);
2635 psi_memstall_leave(&pflags);
2637 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2641 drain_all_stock(mem_over_limit);
2646 if (gfp_mask & __GFP_NORETRY)
2649 * Even though the limit is exceeded at this point, reclaim
2650 * may have been able to free some pages. Retry the charge
2651 * before killing the task.
2653 * Only for regular pages, though: huge pages are rather
2654 * unlikely to succeed so close to the limit, and we fall back
2655 * to regular pages anyway in case of failure.
2657 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2660 * At task move, charge accounts can be doubly counted. So, it's
2661 * better to wait until the end of task_move if something is going on.
2663 if (mem_cgroup_wait_acct_move(mem_over_limit))
2669 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2672 if (fatal_signal_pending(current))
2676 * keep retrying as long as the memcg oom killer is able to make
2677 * a forward progress or bypass the charge if the oom killer
2678 * couldn't make any progress.
2680 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2681 get_order(nr_pages * PAGE_SIZE));
2682 switch (oom_status) {
2684 nr_retries = MAX_RECLAIM_RETRIES;
2692 if (!(gfp_mask & __GFP_NOFAIL))
2696 * The allocation either can't fail or will lead to more memory
2697 * being freed very soon. Allow memory usage go over the limit
2698 * temporarily by force charging it.
2700 page_counter_charge(&memcg->memory, nr_pages);
2701 if (do_memsw_account())
2702 page_counter_charge(&memcg->memsw, nr_pages);
2707 if (batch > nr_pages)
2708 refill_stock(memcg, batch - nr_pages);
2711 * If the hierarchy is above the normal consumption range, schedule
2712 * reclaim on returning to userland. We can perform reclaim here
2713 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2714 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2715 * not recorded as it most likely matches current's and won't
2716 * change in the meantime. As high limit is checked again before
2717 * reclaim, the cost of mismatch is negligible.
2720 bool mem_high, swap_high;
2722 mem_high = page_counter_read(&memcg->memory) >
2723 READ_ONCE(memcg->memory.high);
2724 swap_high = page_counter_read(&memcg->swap) >
2725 READ_ONCE(memcg->swap.high);
2727 /* Don't bother a random interrupted task */
2728 if (in_interrupt()) {
2730 schedule_work(&memcg->high_work);
2736 if (mem_high || swap_high) {
2738 * The allocating tasks in this cgroup will need to do
2739 * reclaim or be throttled to prevent further growth
2740 * of the memory or swap footprints.
2742 * Target some best-effort fairness between the tasks,
2743 * and distribute reclaim work and delay penalties
2744 * based on how much each task is actually allocating.
2746 current->memcg_nr_pages_over_high += batch;
2747 set_notify_resume(current);
2750 } while ((memcg = parent_mem_cgroup(memcg)));
2755 static inline int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2756 unsigned int nr_pages)
2758 if (mem_cgroup_is_root(memcg))
2761 return try_charge_memcg(memcg, gfp_mask, nr_pages);
2764 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2765 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2767 if (mem_cgroup_is_root(memcg))
2770 page_counter_uncharge(&memcg->memory, nr_pages);
2771 if (do_memsw_account())
2772 page_counter_uncharge(&memcg->memsw, nr_pages);
2776 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2778 VM_BUG_ON_PAGE(page_memcg(page), page);
2780 * Any of the following ensures page's memcg stability:
2784 * - lock_page_memcg()
2785 * - exclusive reference
2787 page->memcg_data = (unsigned long)memcg;
2790 static struct mem_cgroup *get_mem_cgroup_from_objcg(struct obj_cgroup *objcg)
2792 struct mem_cgroup *memcg;
2796 memcg = obj_cgroup_memcg(objcg);
2797 if (unlikely(!css_tryget(&memcg->css)))
2804 #ifdef CONFIG_MEMCG_KMEM
2806 * The allocated objcg pointers array is not accounted directly.
2807 * Moreover, it should not come from DMA buffer and is not readily
2808 * reclaimable. So those GFP bits should be masked off.
2810 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | __GFP_ACCOUNT)
2812 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2813 gfp_t gfp, bool new_page)
2815 unsigned int objects = objs_per_slab_page(s, page);
2816 unsigned long memcg_data;
2819 gfp &= ~OBJCGS_CLEAR_MASK;
2820 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2825 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2828 * If the slab page is brand new and nobody can yet access
2829 * it's memcg_data, no synchronization is required and
2830 * memcg_data can be simply assigned.
2832 page->memcg_data = memcg_data;
2833 } else if (cmpxchg(&page->memcg_data, 0, memcg_data)) {
2835 * If the slab page is already in use, somebody can allocate
2836 * and assign obj_cgroups in parallel. In this case the existing
2837 * objcg vector should be reused.
2843 kmemleak_not_leak(vec);
2848 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2850 * A passed kernel object can be a slab object or a generic kernel page, so
2851 * different mechanisms for getting the memory cgroup pointer should be used.
2852 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2853 * can not know for sure how the kernel object is implemented.
2854 * mem_cgroup_from_obj() can be safely used in such cases.
2856 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2857 * cgroup_mutex, etc.
2859 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2863 if (mem_cgroup_disabled())
2866 page = virt_to_head_page(p);
2869 * Slab objects are accounted individually, not per-page.
2870 * Memcg membership data for each individual object is saved in
2871 * the page->obj_cgroups.
2873 if (page_objcgs_check(page)) {
2874 struct obj_cgroup *objcg;
2877 off = obj_to_index(page->slab_cache, page, p);
2878 objcg = page_objcgs(page)[off];
2880 return obj_cgroup_memcg(objcg);
2886 * page_memcg_check() is used here, because page_has_obj_cgroups()
2887 * check above could fail because the object cgroups vector wasn't set
2888 * at that moment, but it can be set concurrently.
2889 * page_memcg_check(page) will guarantee that a proper memory
2890 * cgroup pointer or NULL will be returned.
2892 return page_memcg_check(page);
2895 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2897 struct obj_cgroup *objcg = NULL;
2898 struct mem_cgroup *memcg;
2900 if (memcg_kmem_bypass())
2904 if (unlikely(active_memcg()))
2905 memcg = active_memcg();
2907 memcg = mem_cgroup_from_task(current);
2909 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2910 objcg = rcu_dereference(memcg->objcg);
2911 if (objcg && obj_cgroup_tryget(objcg))
2920 static int memcg_alloc_cache_id(void)
2925 id = ida_simple_get(&memcg_cache_ida,
2926 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2930 if (id < memcg_nr_cache_ids)
2934 * There's no space for the new id in memcg_caches arrays,
2935 * so we have to grow them.
2937 down_write(&memcg_cache_ids_sem);
2939 size = 2 * (id + 1);
2940 if (size < MEMCG_CACHES_MIN_SIZE)
2941 size = MEMCG_CACHES_MIN_SIZE;
2942 else if (size > MEMCG_CACHES_MAX_SIZE)
2943 size = MEMCG_CACHES_MAX_SIZE;
2945 err = memcg_update_all_list_lrus(size);
2947 memcg_nr_cache_ids = size;
2949 up_write(&memcg_cache_ids_sem);
2952 ida_simple_remove(&memcg_cache_ida, id);
2958 static void memcg_free_cache_id(int id)
2960 ida_simple_remove(&memcg_cache_ida, id);
2964 * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
2965 * @objcg: object cgroup to uncharge
2966 * @nr_pages: number of pages to uncharge
2968 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
2969 unsigned int nr_pages)
2971 struct mem_cgroup *memcg;
2973 memcg = get_mem_cgroup_from_objcg(objcg);
2975 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2976 page_counter_uncharge(&memcg->kmem, nr_pages);
2977 refill_stock(memcg, nr_pages);
2979 css_put(&memcg->css);
2983 * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
2984 * @objcg: object cgroup to charge
2985 * @gfp: reclaim mode
2986 * @nr_pages: number of pages to charge
2988 * Returns 0 on success, an error code on failure.
2990 static int obj_cgroup_charge_pages(struct obj_cgroup *objcg, gfp_t gfp,
2991 unsigned int nr_pages)
2993 struct page_counter *counter;
2994 struct mem_cgroup *memcg;
2997 memcg = get_mem_cgroup_from_objcg(objcg);
2999 ret = try_charge_memcg(memcg, gfp, nr_pages);
3003 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3004 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3007 * Enforce __GFP_NOFAIL allocation because callers are not
3008 * prepared to see failures and likely do not have any failure
3011 if (gfp & __GFP_NOFAIL) {
3012 page_counter_charge(&memcg->kmem, nr_pages);
3015 cancel_charge(memcg, nr_pages);
3019 css_put(&memcg->css);
3025 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3026 * @page: page to charge
3027 * @gfp: reclaim mode
3028 * @order: allocation order
3030 * Returns 0 on success, an error code on failure.
3032 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3034 struct obj_cgroup *objcg;
3037 objcg = get_obj_cgroup_from_current();
3039 ret = obj_cgroup_charge_pages(objcg, gfp, 1 << order);
3041 page->memcg_data = (unsigned long)objcg |
3045 obj_cgroup_put(objcg);
3051 * __memcg_kmem_uncharge_page: uncharge a kmem page
3052 * @page: page to uncharge
3053 * @order: allocation order
3055 void __memcg_kmem_uncharge_page(struct page *page, int order)
3057 struct obj_cgroup *objcg;
3058 unsigned int nr_pages = 1 << order;
3060 if (!PageMemcgKmem(page))
3063 objcg = __page_objcg(page);
3064 obj_cgroup_uncharge_pages(objcg, nr_pages);
3065 page->memcg_data = 0;
3066 obj_cgroup_put(objcg);
3069 void mod_objcg_state(struct obj_cgroup *objcg, struct pglist_data *pgdat,
3070 enum node_stat_item idx, int nr)
3072 unsigned long flags;
3073 struct obj_stock *stock = get_obj_stock(&flags);
3077 * Save vmstat data in stock and skip vmstat array update unless
3078 * accumulating over a page of vmstat data or when pgdat or idx
3081 if (stock->cached_objcg != objcg) {
3082 drain_obj_stock(stock);
3083 obj_cgroup_get(objcg);
3084 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3085 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3086 stock->cached_objcg = objcg;
3087 stock->cached_pgdat = pgdat;
3088 } else if (stock->cached_pgdat != pgdat) {
3089 /* Flush the existing cached vmstat data */
3090 if (stock->nr_slab_reclaimable_b) {
3091 mod_objcg_mlstate(objcg, pgdat, NR_SLAB_RECLAIMABLE_B,
3092 stock->nr_slab_reclaimable_b);
3093 stock->nr_slab_reclaimable_b = 0;
3095 if (stock->nr_slab_unreclaimable_b) {
3096 mod_objcg_mlstate(objcg, pgdat, NR_SLAB_UNRECLAIMABLE_B,
3097 stock->nr_slab_unreclaimable_b);
3098 stock->nr_slab_unreclaimable_b = 0;
3100 stock->cached_pgdat = pgdat;
3103 bytes = (idx == NR_SLAB_RECLAIMABLE_B) ? &stock->nr_slab_reclaimable_b
3104 : &stock->nr_slab_unreclaimable_b;
3106 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3107 * cached locally at least once before pushing it out.
3114 if (abs(*bytes) > PAGE_SIZE) {
3122 mod_objcg_mlstate(objcg, pgdat, idx, nr);
3124 put_obj_stock(flags);
3127 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3129 unsigned long flags;
3130 struct obj_stock *stock = get_obj_stock(&flags);
3133 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3134 stock->nr_bytes -= nr_bytes;
3138 put_obj_stock(flags);
3143 static void drain_obj_stock(struct obj_stock *stock)
3145 struct obj_cgroup *old = stock->cached_objcg;
3150 if (stock->nr_bytes) {
3151 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3152 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3155 obj_cgroup_uncharge_pages(old, nr_pages);
3158 * The leftover is flushed to the centralized per-memcg value.
3159 * On the next attempt to refill obj stock it will be moved
3160 * to a per-cpu stock (probably, on an other CPU), see
3161 * refill_obj_stock().
3163 * How often it's flushed is a trade-off between the memory
3164 * limit enforcement accuracy and potential CPU contention,
3165 * so it might be changed in the future.
3167 atomic_add(nr_bytes, &old->nr_charged_bytes);
3168 stock->nr_bytes = 0;
3172 * Flush the vmstat data in current stock
3174 if (stock->nr_slab_reclaimable_b || stock->nr_slab_unreclaimable_b) {
3175 if (stock->nr_slab_reclaimable_b) {
3176 mod_objcg_mlstate(old, stock->cached_pgdat,
3177 NR_SLAB_RECLAIMABLE_B,
3178 stock->nr_slab_reclaimable_b);
3179 stock->nr_slab_reclaimable_b = 0;
3181 if (stock->nr_slab_unreclaimable_b) {
3182 mod_objcg_mlstate(old, stock->cached_pgdat,
3183 NR_SLAB_UNRECLAIMABLE_B,
3184 stock->nr_slab_unreclaimable_b);
3185 stock->nr_slab_unreclaimable_b = 0;
3187 stock->cached_pgdat = NULL;
3190 obj_cgroup_put(old);
3191 stock->cached_objcg = NULL;
3194 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3195 struct mem_cgroup *root_memcg)
3197 struct mem_cgroup *memcg;
3199 if (in_task() && stock->task_obj.cached_objcg) {
3200 memcg = obj_cgroup_memcg(stock->task_obj.cached_objcg);
3201 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3204 if (stock->irq_obj.cached_objcg) {
3205 memcg = obj_cgroup_memcg(stock->irq_obj.cached_objcg);
3206 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3213 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes,
3214 bool allow_uncharge)
3216 unsigned long flags;
3217 struct obj_stock *stock = get_obj_stock(&flags);
3218 unsigned int nr_pages = 0;
3220 if (stock->cached_objcg != objcg) { /* reset if necessary */
3221 drain_obj_stock(stock);
3222 obj_cgroup_get(objcg);
3223 stock->cached_objcg = objcg;
3224 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3225 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3226 allow_uncharge = true; /* Allow uncharge when objcg changes */
3228 stock->nr_bytes += nr_bytes;
3230 if (allow_uncharge && (stock->nr_bytes > PAGE_SIZE)) {
3231 nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3232 stock->nr_bytes &= (PAGE_SIZE - 1);
3235 put_obj_stock(flags);
3238 obj_cgroup_uncharge_pages(objcg, nr_pages);
3241 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3243 unsigned int nr_pages, nr_bytes;
3246 if (consume_obj_stock(objcg, size))
3250 * In theory, objcg->nr_charged_bytes can have enough
3251 * pre-charged bytes to satisfy the allocation. However,
3252 * flushing objcg->nr_charged_bytes requires two atomic
3253 * operations, and objcg->nr_charged_bytes can't be big.
3254 * The shared objcg->nr_charged_bytes can also become a
3255 * performance bottleneck if all tasks of the same memcg are
3256 * trying to update it. So it's better to ignore it and try
3257 * grab some new pages. The stock's nr_bytes will be flushed to
3258 * objcg->nr_charged_bytes later on when objcg changes.
3260 * The stock's nr_bytes may contain enough pre-charged bytes
3261 * to allow one less page from being charged, but we can't rely
3262 * on the pre-charged bytes not being changed outside of
3263 * consume_obj_stock() or refill_obj_stock(). So ignore those
3264 * pre-charged bytes as well when charging pages. To avoid a
3265 * page uncharge right after a page charge, we set the
3266 * allow_uncharge flag to false when calling refill_obj_stock()
3267 * to temporarily allow the pre-charged bytes to exceed the page
3268 * size limit. The maximum reachable value of the pre-charged
3269 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3272 nr_pages = size >> PAGE_SHIFT;
3273 nr_bytes = size & (PAGE_SIZE - 1);
3278 ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3279 if (!ret && nr_bytes)
3280 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes, false);
3285 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3287 refill_obj_stock(objcg, size, true);
3290 #endif /* CONFIG_MEMCG_KMEM */
3293 * Because page_memcg(head) is not set on tails, set it now.
3295 void split_page_memcg(struct page *head, unsigned int nr)
3297 struct mem_cgroup *memcg = page_memcg(head);
3300 if (mem_cgroup_disabled() || !memcg)
3303 for (i = 1; i < nr; i++)
3304 head[i].memcg_data = head->memcg_data;
3306 if (PageMemcgKmem(head))
3307 obj_cgroup_get_many(__page_objcg(head), nr - 1);
3309 css_get_many(&memcg->css, nr - 1);
3312 #ifdef CONFIG_MEMCG_SWAP
3314 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3315 * @entry: swap entry to be moved
3316 * @from: mem_cgroup which the entry is moved from
3317 * @to: mem_cgroup which the entry is moved to
3319 * It succeeds only when the swap_cgroup's record for this entry is the same
3320 * as the mem_cgroup's id of @from.
3322 * Returns 0 on success, -EINVAL on failure.
3324 * The caller must have charged to @to, IOW, called page_counter_charge() about
3325 * both res and memsw, and called css_get().
3327 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3328 struct mem_cgroup *from, struct mem_cgroup *to)
3330 unsigned short old_id, new_id;
3332 old_id = mem_cgroup_id(from);
3333 new_id = mem_cgroup_id(to);
3335 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3336 mod_memcg_state(from, MEMCG_SWAP, -1);
3337 mod_memcg_state(to, MEMCG_SWAP, 1);
3343 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3344 struct mem_cgroup *from, struct mem_cgroup *to)
3350 static DEFINE_MUTEX(memcg_max_mutex);
3352 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3353 unsigned long max, bool memsw)
3355 bool enlarge = false;
3356 bool drained = false;
3358 bool limits_invariant;
3359 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3362 if (signal_pending(current)) {
3367 mutex_lock(&memcg_max_mutex);
3369 * Make sure that the new limit (memsw or memory limit) doesn't
3370 * break our basic invariant rule memory.max <= memsw.max.
3372 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3373 max <= memcg->memsw.max;
3374 if (!limits_invariant) {
3375 mutex_unlock(&memcg_max_mutex);
3379 if (max > counter->max)
3381 ret = page_counter_set_max(counter, max);
3382 mutex_unlock(&memcg_max_mutex);
3388 drain_all_stock(memcg);
3393 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3394 GFP_KERNEL, !memsw)) {
3400 if (!ret && enlarge)
3401 memcg_oom_recover(memcg);
3406 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3408 unsigned long *total_scanned)
3410 unsigned long nr_reclaimed = 0;
3411 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3412 unsigned long reclaimed;
3414 struct mem_cgroup_tree_per_node *mctz;
3415 unsigned long excess;
3416 unsigned long nr_scanned;
3421 mctz = soft_limit_tree_node(pgdat->node_id);
3424 * Do not even bother to check the largest node if the root
3425 * is empty. Do it lockless to prevent lock bouncing. Races
3426 * are acceptable as soft limit is best effort anyway.
3428 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3432 * This loop can run a while, specially if mem_cgroup's continuously
3433 * keep exceeding their soft limit and putting the system under
3440 mz = mem_cgroup_largest_soft_limit_node(mctz);
3445 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3446 gfp_mask, &nr_scanned);
3447 nr_reclaimed += reclaimed;
3448 *total_scanned += nr_scanned;
3449 spin_lock_irq(&mctz->lock);
3450 __mem_cgroup_remove_exceeded(mz, mctz);
3453 * If we failed to reclaim anything from this memory cgroup
3454 * it is time to move on to the next cgroup
3458 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3460 excess = soft_limit_excess(mz->memcg);
3462 * One school of thought says that we should not add
3463 * back the node to the tree if reclaim returns 0.
3464 * But our reclaim could return 0, simply because due
3465 * to priority we are exposing a smaller subset of
3466 * memory to reclaim from. Consider this as a longer
3469 /* If excess == 0, no tree ops */
3470 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3471 spin_unlock_irq(&mctz->lock);
3472 css_put(&mz->memcg->css);
3475 * Could not reclaim anything and there are no more
3476 * mem cgroups to try or we seem to be looping without
3477 * reclaiming anything.
3479 if (!nr_reclaimed &&
3481 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3483 } while (!nr_reclaimed);
3485 css_put(&next_mz->memcg->css);
3486 return nr_reclaimed;
3490 * Reclaims as many pages from the given memcg as possible.
3492 * Caller is responsible for holding css reference for memcg.
3494 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3496 int nr_retries = MAX_RECLAIM_RETRIES;
3498 /* we call try-to-free pages for make this cgroup empty */
3499 lru_add_drain_all();
3501 drain_all_stock(memcg);
3503 /* try to free all pages in this cgroup */
3504 while (nr_retries && page_counter_read(&memcg->memory)) {
3507 if (signal_pending(current))
3510 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3514 /* maybe some writeback is necessary */
3515 congestion_wait(BLK_RW_ASYNC, HZ/10);
3523 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3524 char *buf, size_t nbytes,
3527 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3529 if (mem_cgroup_is_root(memcg))
3531 return mem_cgroup_force_empty(memcg) ?: nbytes;
3534 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3540 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3541 struct cftype *cft, u64 val)
3546 pr_warn_once("Non-hierarchical mode is deprecated. "
3547 "Please report your usecase to linux-mm@kvack.org if you "
3548 "depend on this functionality.\n");
3553 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3557 if (mem_cgroup_is_root(memcg)) {
3558 cgroup_rstat_flush(memcg->css.cgroup);
3559 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3560 memcg_page_state(memcg, NR_ANON_MAPPED);
3562 val += memcg_page_state(memcg, MEMCG_SWAP);
3565 val = page_counter_read(&memcg->memory);
3567 val = page_counter_read(&memcg->memsw);
3580 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3583 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3584 struct page_counter *counter;
3586 switch (MEMFILE_TYPE(cft->private)) {
3588 counter = &memcg->memory;
3591 counter = &memcg->memsw;
3594 counter = &memcg->kmem;
3597 counter = &memcg->tcpmem;
3603 switch (MEMFILE_ATTR(cft->private)) {
3605 if (counter == &memcg->memory)
3606 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3607 if (counter == &memcg->memsw)
3608 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3609 return (u64)page_counter_read(counter) * PAGE_SIZE;
3611 return (u64)counter->max * PAGE_SIZE;
3613 return (u64)counter->watermark * PAGE_SIZE;
3615 return counter->failcnt;
3616 case RES_SOFT_LIMIT:
3617 return (u64)memcg->soft_limit * PAGE_SIZE;
3623 #ifdef CONFIG_MEMCG_KMEM
3624 static int memcg_online_kmem(struct mem_cgroup *memcg)
3626 struct obj_cgroup *objcg;
3629 if (cgroup_memory_nokmem)
3632 BUG_ON(memcg->kmemcg_id >= 0);
3633 BUG_ON(memcg->kmem_state);
3635 memcg_id = memcg_alloc_cache_id();
3639 objcg = obj_cgroup_alloc();
3641 memcg_free_cache_id(memcg_id);
3644 objcg->memcg = memcg;
3645 rcu_assign_pointer(memcg->objcg, objcg);
3647 static_branch_enable(&memcg_kmem_enabled_key);
3649 memcg->kmemcg_id = memcg_id;
3650 memcg->kmem_state = KMEM_ONLINE;
3655 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3657 struct cgroup_subsys_state *css;
3658 struct mem_cgroup *parent, *child;
3661 if (memcg->kmem_state != KMEM_ONLINE)
3664 memcg->kmem_state = KMEM_ALLOCATED;
3666 parent = parent_mem_cgroup(memcg);
3668 parent = root_mem_cgroup;
3670 memcg_reparent_objcgs(memcg, parent);
3672 kmemcg_id = memcg->kmemcg_id;
3673 BUG_ON(kmemcg_id < 0);
3676 * Change kmemcg_id of this cgroup and all its descendants to the
3677 * parent's id, and then move all entries from this cgroup's list_lrus
3678 * to ones of the parent. After we have finished, all list_lrus
3679 * corresponding to this cgroup are guaranteed to remain empty. The
3680 * ordering is imposed by list_lru_node->lock taken by
3681 * memcg_drain_all_list_lrus().
3683 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3684 css_for_each_descendant_pre(css, &memcg->css) {
3685 child = mem_cgroup_from_css(css);
3686 BUG_ON(child->kmemcg_id != kmemcg_id);
3687 child->kmemcg_id = parent->kmemcg_id;
3691 memcg_drain_all_list_lrus(kmemcg_id, parent);
3693 memcg_free_cache_id(kmemcg_id);
3696 static void memcg_free_kmem(struct mem_cgroup *memcg)
3698 /* css_alloc() failed, offlining didn't happen */
3699 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3700 memcg_offline_kmem(memcg);
3703 static int memcg_online_kmem(struct mem_cgroup *memcg)
3707 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3710 static void memcg_free_kmem(struct mem_cgroup *memcg)
3713 #endif /* CONFIG_MEMCG_KMEM */
3715 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3720 mutex_lock(&memcg_max_mutex);
3721 ret = page_counter_set_max(&memcg->kmem, max);
3722 mutex_unlock(&memcg_max_mutex);
3726 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3730 mutex_lock(&memcg_max_mutex);
3732 ret = page_counter_set_max(&memcg->tcpmem, max);
3736 if (!memcg->tcpmem_active) {
3738 * The active flag needs to be written after the static_key
3739 * update. This is what guarantees that the socket activation
3740 * function is the last one to run. See mem_cgroup_sk_alloc()
3741 * for details, and note that we don't mark any socket as
3742 * belonging to this memcg until that flag is up.
3744 * We need to do this, because static_keys will span multiple
3745 * sites, but we can't control their order. If we mark a socket
3746 * as accounted, but the accounting functions are not patched in
3747 * yet, we'll lose accounting.
3749 * We never race with the readers in mem_cgroup_sk_alloc(),
3750 * because when this value change, the code to process it is not
3753 static_branch_inc(&memcg_sockets_enabled_key);
3754 memcg->tcpmem_active = true;
3757 mutex_unlock(&memcg_max_mutex);
3762 * The user of this function is...
3765 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3766 char *buf, size_t nbytes, loff_t off)
3768 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3769 unsigned long nr_pages;
3772 buf = strstrip(buf);
3773 ret = page_counter_memparse(buf, "-1", &nr_pages);
3777 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3779 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3783 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3785 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3788 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3791 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3792 "Please report your usecase to linux-mm@kvack.org if you "
3793 "depend on this functionality.\n");
3794 ret = memcg_update_kmem_max(memcg, nr_pages);
3797 ret = memcg_update_tcp_max(memcg, nr_pages);
3801 case RES_SOFT_LIMIT:
3802 memcg->soft_limit = nr_pages;
3806 return ret ?: nbytes;
3809 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3810 size_t nbytes, loff_t off)
3812 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3813 struct page_counter *counter;
3815 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3817 counter = &memcg->memory;
3820 counter = &memcg->memsw;
3823 counter = &memcg->kmem;
3826 counter = &memcg->tcpmem;
3832 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3834 page_counter_reset_watermark(counter);
3837 counter->failcnt = 0;
3846 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3849 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3853 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3854 struct cftype *cft, u64 val)
3856 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3858 if (val & ~MOVE_MASK)
3862 * No kind of locking is needed in here, because ->can_attach() will
3863 * check this value once in the beginning of the process, and then carry
3864 * on with stale data. This means that changes to this value will only
3865 * affect task migrations starting after the change.
3867 memcg->move_charge_at_immigrate = val;
3871 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3872 struct cftype *cft, u64 val)
3880 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3881 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3882 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3884 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3885 int nid, unsigned int lru_mask, bool tree)
3887 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3888 unsigned long nr = 0;
3891 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3894 if (!(BIT(lru) & lru_mask))
3897 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3899 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3904 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3905 unsigned int lru_mask,
3908 unsigned long nr = 0;
3912 if (!(BIT(lru) & lru_mask))
3915 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3917 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3922 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3926 unsigned int lru_mask;
3929 static const struct numa_stat stats[] = {
3930 { "total", LRU_ALL },
3931 { "file", LRU_ALL_FILE },
3932 { "anon", LRU_ALL_ANON },
3933 { "unevictable", BIT(LRU_UNEVICTABLE) },
3935 const struct numa_stat *stat;
3937 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3939 cgroup_rstat_flush(memcg->css.cgroup);
3941 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3942 seq_printf(m, "%s=%lu", stat->name,
3943 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3945 for_each_node_state(nid, N_MEMORY)
3946 seq_printf(m, " N%d=%lu", nid,
3947 mem_cgroup_node_nr_lru_pages(memcg, nid,
3948 stat->lru_mask, false));
3952 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3954 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3955 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3957 for_each_node_state(nid, N_MEMORY)
3958 seq_printf(m, " N%d=%lu", nid,
3959 mem_cgroup_node_nr_lru_pages(memcg, nid,
3960 stat->lru_mask, true));
3966 #endif /* CONFIG_NUMA */
3968 static const unsigned int memcg1_stats[] = {
3971 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3981 static const char *const memcg1_stat_names[] = {
3984 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3994 /* Universal VM events cgroup1 shows, original sort order */
3995 static const unsigned int memcg1_events[] = {
4002 static int memcg_stat_show(struct seq_file *m, void *v)
4004 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4005 unsigned long memory, memsw;
4006 struct mem_cgroup *mi;
4009 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4011 cgroup_rstat_flush(memcg->css.cgroup);
4013 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4016 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4018 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4019 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4022 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4023 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4024 memcg_events_local(memcg, memcg1_events[i]));
4026 for (i = 0; i < NR_LRU_LISTS; i++)
4027 seq_printf(m, "%s %lu\n", lru_list_name(i),
4028 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4031 /* Hierarchical information */
4032 memory = memsw = PAGE_COUNTER_MAX;
4033 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4034 memory = min(memory, READ_ONCE(mi->memory.max));
4035 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4037 seq_printf(m, "hierarchical_memory_limit %llu\n",
4038 (u64)memory * PAGE_SIZE);
4039 if (do_memsw_account())
4040 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4041 (u64)memsw * PAGE_SIZE);
4043 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4046 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4048 nr = memcg_page_state(memcg, memcg1_stats[i]);
4049 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4050 (u64)nr * PAGE_SIZE);
4053 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4054 seq_printf(m, "total_%s %llu\n",
4055 vm_event_name(memcg1_events[i]),
4056 (u64)memcg_events(memcg, memcg1_events[i]));
4058 for (i = 0; i < NR_LRU_LISTS; i++)
4059 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4060 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4063 #ifdef CONFIG_DEBUG_VM
4066 struct mem_cgroup_per_node *mz;
4067 unsigned long anon_cost = 0;
4068 unsigned long file_cost = 0;
4070 for_each_online_pgdat(pgdat) {
4071 mz = memcg->nodeinfo[pgdat->node_id];
4073 anon_cost += mz->lruvec.anon_cost;
4074 file_cost += mz->lruvec.file_cost;
4076 seq_printf(m, "anon_cost %lu\n", anon_cost);
4077 seq_printf(m, "file_cost %lu\n", file_cost);
4084 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4087 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4089 return mem_cgroup_swappiness(memcg);
4092 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4093 struct cftype *cft, u64 val)
4095 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4100 if (!mem_cgroup_is_root(memcg))
4101 memcg->swappiness = val;
4103 vm_swappiness = val;
4108 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4110 struct mem_cgroup_threshold_ary *t;
4111 unsigned long usage;
4116 t = rcu_dereference(memcg->thresholds.primary);
4118 t = rcu_dereference(memcg->memsw_thresholds.primary);
4123 usage = mem_cgroup_usage(memcg, swap);
4126 * current_threshold points to threshold just below or equal to usage.
4127 * If it's not true, a threshold was crossed after last
4128 * call of __mem_cgroup_threshold().
4130 i = t->current_threshold;
4133 * Iterate backward over array of thresholds starting from
4134 * current_threshold and check if a threshold is crossed.
4135 * If none of thresholds below usage is crossed, we read
4136 * only one element of the array here.
4138 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4139 eventfd_signal(t->entries[i].eventfd, 1);
4141 /* i = current_threshold + 1 */
4145 * Iterate forward over array of thresholds starting from
4146 * current_threshold+1 and check if a threshold is crossed.
4147 * If none of thresholds above usage is crossed, we read
4148 * only one element of the array here.
4150 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4151 eventfd_signal(t->entries[i].eventfd, 1);
4153 /* Update current_threshold */
4154 t->current_threshold = i - 1;
4159 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4162 __mem_cgroup_threshold(memcg, false);
4163 if (do_memsw_account())
4164 __mem_cgroup_threshold(memcg, true);
4166 memcg = parent_mem_cgroup(memcg);
4170 static int compare_thresholds(const void *a, const void *b)
4172 const struct mem_cgroup_threshold *_a = a;
4173 const struct mem_cgroup_threshold *_b = b;
4175 if (_a->threshold > _b->threshold)
4178 if (_a->threshold < _b->threshold)
4184 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4186 struct mem_cgroup_eventfd_list *ev;
4188 spin_lock(&memcg_oom_lock);
4190 list_for_each_entry(ev, &memcg->oom_notify, list)
4191 eventfd_signal(ev->eventfd, 1);
4193 spin_unlock(&memcg_oom_lock);
4197 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4199 struct mem_cgroup *iter;
4201 for_each_mem_cgroup_tree(iter, memcg)
4202 mem_cgroup_oom_notify_cb(iter);
4205 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4206 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4208 struct mem_cgroup_thresholds *thresholds;
4209 struct mem_cgroup_threshold_ary *new;
4210 unsigned long threshold;
4211 unsigned long usage;
4214 ret = page_counter_memparse(args, "-1", &threshold);
4218 mutex_lock(&memcg->thresholds_lock);
4221 thresholds = &memcg->thresholds;
4222 usage = mem_cgroup_usage(memcg, false);
4223 } else if (type == _MEMSWAP) {
4224 thresholds = &memcg->memsw_thresholds;
4225 usage = mem_cgroup_usage(memcg, true);
4229 /* Check if a threshold crossed before adding a new one */
4230 if (thresholds->primary)
4231 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4233 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4235 /* Allocate memory for new array of thresholds */
4236 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4243 /* Copy thresholds (if any) to new array */
4244 if (thresholds->primary)
4245 memcpy(new->entries, thresholds->primary->entries,
4246 flex_array_size(new, entries, size - 1));
4248 /* Add new threshold */
4249 new->entries[size - 1].eventfd = eventfd;
4250 new->entries[size - 1].threshold = threshold;
4252 /* Sort thresholds. Registering of new threshold isn't time-critical */
4253 sort(new->entries, size, sizeof(*new->entries),
4254 compare_thresholds, NULL);
4256 /* Find current threshold */
4257 new->current_threshold = -1;
4258 for (i = 0; i < size; i++) {
4259 if (new->entries[i].threshold <= usage) {
4261 * new->current_threshold will not be used until
4262 * rcu_assign_pointer(), so it's safe to increment
4265 ++new->current_threshold;
4270 /* Free old spare buffer and save old primary buffer as spare */
4271 kfree(thresholds->spare);
4272 thresholds->spare = thresholds->primary;
4274 rcu_assign_pointer(thresholds->primary, new);
4276 /* To be sure that nobody uses thresholds */
4280 mutex_unlock(&memcg->thresholds_lock);
4285 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4286 struct eventfd_ctx *eventfd, const char *args)
4288 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4291 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4292 struct eventfd_ctx *eventfd, const char *args)
4294 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4297 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4298 struct eventfd_ctx *eventfd, enum res_type type)
4300 struct mem_cgroup_thresholds *thresholds;
4301 struct mem_cgroup_threshold_ary *new;
4302 unsigned long usage;
4303 int i, j, size, entries;
4305 mutex_lock(&memcg->thresholds_lock);
4308 thresholds = &memcg->thresholds;
4309 usage = mem_cgroup_usage(memcg, false);
4310 } else if (type == _MEMSWAP) {
4311 thresholds = &memcg->memsw_thresholds;
4312 usage = mem_cgroup_usage(memcg, true);
4316 if (!thresholds->primary)
4319 /* Check if a threshold crossed before removing */
4320 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4322 /* Calculate new number of threshold */
4324 for (i = 0; i < thresholds->primary->size; i++) {
4325 if (thresholds->primary->entries[i].eventfd != eventfd)
4331 new = thresholds->spare;
4333 /* If no items related to eventfd have been cleared, nothing to do */
4337 /* Set thresholds array to NULL if we don't have thresholds */
4346 /* Copy thresholds and find current threshold */
4347 new->current_threshold = -1;
4348 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4349 if (thresholds->primary->entries[i].eventfd == eventfd)
4352 new->entries[j] = thresholds->primary->entries[i];
4353 if (new->entries[j].threshold <= usage) {
4355 * new->current_threshold will not be used
4356 * until rcu_assign_pointer(), so it's safe to increment
4359 ++new->current_threshold;
4365 /* Swap primary and spare array */
4366 thresholds->spare = thresholds->primary;
4368 rcu_assign_pointer(thresholds->primary, new);
4370 /* To be sure that nobody uses thresholds */
4373 /* If all events are unregistered, free the spare array */
4375 kfree(thresholds->spare);
4376 thresholds->spare = NULL;
4379 mutex_unlock(&memcg->thresholds_lock);
4382 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4383 struct eventfd_ctx *eventfd)
4385 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4388 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4389 struct eventfd_ctx *eventfd)
4391 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4394 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4395 struct eventfd_ctx *eventfd, const char *args)
4397 struct mem_cgroup_eventfd_list *event;
4399 event = kmalloc(sizeof(*event), GFP_KERNEL);
4403 spin_lock(&memcg_oom_lock);
4405 event->eventfd = eventfd;
4406 list_add(&event->list, &memcg->oom_notify);
4408 /* already in OOM ? */
4409 if (memcg->under_oom)
4410 eventfd_signal(eventfd, 1);
4411 spin_unlock(&memcg_oom_lock);
4416 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4417 struct eventfd_ctx *eventfd)
4419 struct mem_cgroup_eventfd_list *ev, *tmp;
4421 spin_lock(&memcg_oom_lock);
4423 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4424 if (ev->eventfd == eventfd) {
4425 list_del(&ev->list);
4430 spin_unlock(&memcg_oom_lock);
4433 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4435 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4437 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4438 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4439 seq_printf(sf, "oom_kill %lu\n",
4440 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4444 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4445 struct cftype *cft, u64 val)
4447 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4449 /* cannot set to root cgroup and only 0 and 1 are allowed */
4450 if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4453 memcg->oom_kill_disable = val;
4455 memcg_oom_recover(memcg);
4460 #ifdef CONFIG_CGROUP_WRITEBACK
4462 #include <trace/events/writeback.h>
4464 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4466 return wb_domain_init(&memcg->cgwb_domain, gfp);
4469 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4471 wb_domain_exit(&memcg->cgwb_domain);
4474 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4476 wb_domain_size_changed(&memcg->cgwb_domain);
4479 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4481 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4483 if (!memcg->css.parent)
4486 return &memcg->cgwb_domain;
4490 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4491 * @wb: bdi_writeback in question
4492 * @pfilepages: out parameter for number of file pages
4493 * @pheadroom: out parameter for number of allocatable pages according to memcg
4494 * @pdirty: out parameter for number of dirty pages
4495 * @pwriteback: out parameter for number of pages under writeback
4497 * Determine the numbers of file, headroom, dirty, and writeback pages in
4498 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4499 * is a bit more involved.
4501 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4502 * headroom is calculated as the lowest headroom of itself and the
4503 * ancestors. Note that this doesn't consider the actual amount of
4504 * available memory in the system. The caller should further cap
4505 * *@pheadroom accordingly.
4507 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4508 unsigned long *pheadroom, unsigned long *pdirty,
4509 unsigned long *pwriteback)
4511 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4512 struct mem_cgroup *parent;
4514 cgroup_rstat_flush_irqsafe(memcg->css.cgroup);
4516 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4517 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4518 *pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4519 memcg_page_state(memcg, NR_ACTIVE_FILE);
4521 *pheadroom = PAGE_COUNTER_MAX;
4522 while ((parent = parent_mem_cgroup(memcg))) {
4523 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4524 READ_ONCE(memcg->memory.high));
4525 unsigned long used = page_counter_read(&memcg->memory);
4527 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4533 * Foreign dirty flushing
4535 * There's an inherent mismatch between memcg and writeback. The former
4536 * tracks ownership per-page while the latter per-inode. This was a
4537 * deliberate design decision because honoring per-page ownership in the
4538 * writeback path is complicated, may lead to higher CPU and IO overheads
4539 * and deemed unnecessary given that write-sharing an inode across
4540 * different cgroups isn't a common use-case.
4542 * Combined with inode majority-writer ownership switching, this works well
4543 * enough in most cases but there are some pathological cases. For
4544 * example, let's say there are two cgroups A and B which keep writing to
4545 * different but confined parts of the same inode. B owns the inode and
4546 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4547 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4548 * triggering background writeback. A will be slowed down without a way to
4549 * make writeback of the dirty pages happen.
4551 * Conditions like the above can lead to a cgroup getting repeatedly and
4552 * severely throttled after making some progress after each
4553 * dirty_expire_interval while the underlying IO device is almost
4556 * Solving this problem completely requires matching the ownership tracking
4557 * granularities between memcg and writeback in either direction. However,
4558 * the more egregious behaviors can be avoided by simply remembering the
4559 * most recent foreign dirtying events and initiating remote flushes on
4560 * them when local writeback isn't enough to keep the memory clean enough.
4562 * The following two functions implement such mechanism. When a foreign
4563 * page - a page whose memcg and writeback ownerships don't match - is
4564 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4565 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4566 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4567 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4568 * foreign bdi_writebacks which haven't expired. Both the numbers of
4569 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4570 * limited to MEMCG_CGWB_FRN_CNT.
4572 * The mechanism only remembers IDs and doesn't hold any object references.
4573 * As being wrong occasionally doesn't matter, updates and accesses to the
4574 * records are lockless and racy.
4576 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4577 struct bdi_writeback *wb)
4579 struct mem_cgroup *memcg = page_memcg(page);
4580 struct memcg_cgwb_frn *frn;
4581 u64 now = get_jiffies_64();
4582 u64 oldest_at = now;
4586 trace_track_foreign_dirty(page, wb);
4589 * Pick the slot to use. If there is already a slot for @wb, keep
4590 * using it. If not replace the oldest one which isn't being
4593 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4594 frn = &memcg->cgwb_frn[i];
4595 if (frn->bdi_id == wb->bdi->id &&
4596 frn->memcg_id == wb->memcg_css->id)
4598 if (time_before64(frn->at, oldest_at) &&
4599 atomic_read(&frn->done.cnt) == 1) {
4601 oldest_at = frn->at;
4605 if (i < MEMCG_CGWB_FRN_CNT) {
4607 * Re-using an existing one. Update timestamp lazily to
4608 * avoid making the cacheline hot. We want them to be
4609 * reasonably up-to-date and significantly shorter than
4610 * dirty_expire_interval as that's what expires the record.
4611 * Use the shorter of 1s and dirty_expire_interval / 8.
4613 unsigned long update_intv =
4614 min_t(unsigned long, HZ,
4615 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4617 if (time_before64(frn->at, now - update_intv))
4619 } else if (oldest >= 0) {
4620 /* replace the oldest free one */
4621 frn = &memcg->cgwb_frn[oldest];
4622 frn->bdi_id = wb->bdi->id;
4623 frn->memcg_id = wb->memcg_css->id;
4628 /* issue foreign writeback flushes for recorded foreign dirtying events */
4629 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4631 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4632 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4633 u64 now = jiffies_64;
4636 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4637 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4640 * If the record is older than dirty_expire_interval,
4641 * writeback on it has already started. No need to kick it
4642 * off again. Also, don't start a new one if there's
4643 * already one in flight.
4645 if (time_after64(frn->at, now - intv) &&
4646 atomic_read(&frn->done.cnt) == 1) {
4648 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4649 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4650 WB_REASON_FOREIGN_FLUSH,
4656 #else /* CONFIG_CGROUP_WRITEBACK */
4658 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4663 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4667 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4671 #endif /* CONFIG_CGROUP_WRITEBACK */
4674 * DO NOT USE IN NEW FILES.
4676 * "cgroup.event_control" implementation.
4678 * This is way over-engineered. It tries to support fully configurable
4679 * events for each user. Such level of flexibility is completely
4680 * unnecessary especially in the light of the planned unified hierarchy.
4682 * Please deprecate this and replace with something simpler if at all
4687 * Unregister event and free resources.
4689 * Gets called from workqueue.
4691 static void memcg_event_remove(struct work_struct *work)
4693 struct mem_cgroup_event *event =
4694 container_of(work, struct mem_cgroup_event, remove);
4695 struct mem_cgroup *memcg = event->memcg;
4697 remove_wait_queue(event->wqh, &event->wait);
4699 event->unregister_event(memcg, event->eventfd);
4701 /* Notify userspace the event is going away. */
4702 eventfd_signal(event->eventfd, 1);
4704 eventfd_ctx_put(event->eventfd);
4706 css_put(&memcg->css);
4710 * Gets called on EPOLLHUP on eventfd when user closes it.
4712 * Called with wqh->lock held and interrupts disabled.
4714 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4715 int sync, void *key)
4717 struct mem_cgroup_event *event =
4718 container_of(wait, struct mem_cgroup_event, wait);
4719 struct mem_cgroup *memcg = event->memcg;
4720 __poll_t flags = key_to_poll(key);
4722 if (flags & EPOLLHUP) {
4724 * If the event has been detached at cgroup removal, we
4725 * can simply return knowing the other side will cleanup
4728 * We can't race against event freeing since the other
4729 * side will require wqh->lock via remove_wait_queue(),
4732 spin_lock(&memcg->event_list_lock);
4733 if (!list_empty(&event->list)) {
4734 list_del_init(&event->list);
4736 * We are in atomic context, but cgroup_event_remove()
4737 * may sleep, so we have to call it in workqueue.
4739 schedule_work(&event->remove);
4741 spin_unlock(&memcg->event_list_lock);
4747 static void memcg_event_ptable_queue_proc(struct file *file,
4748 wait_queue_head_t *wqh, poll_table *pt)
4750 struct mem_cgroup_event *event =
4751 container_of(pt, struct mem_cgroup_event, pt);
4754 add_wait_queue(wqh, &event->wait);
4758 * DO NOT USE IN NEW FILES.
4760 * Parse input and register new cgroup event handler.
4762 * Input must be in format '<event_fd> <control_fd> <args>'.
4763 * Interpretation of args is defined by control file implementation.
4765 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4766 char *buf, size_t nbytes, loff_t off)
4768 struct cgroup_subsys_state *css = of_css(of);
4769 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4770 struct mem_cgroup_event *event;
4771 struct cgroup_subsys_state *cfile_css;
4772 unsigned int efd, cfd;
4779 buf = strstrip(buf);
4781 efd = simple_strtoul(buf, &endp, 10);
4786 cfd = simple_strtoul(buf, &endp, 10);
4787 if ((*endp != ' ') && (*endp != '\0'))
4791 event = kzalloc(sizeof(*event), GFP_KERNEL);
4795 event->memcg = memcg;
4796 INIT_LIST_HEAD(&event->list);
4797 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4798 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4799 INIT_WORK(&event->remove, memcg_event_remove);
4807 event->eventfd = eventfd_ctx_fileget(efile.file);
4808 if (IS_ERR(event->eventfd)) {
4809 ret = PTR_ERR(event->eventfd);
4816 goto out_put_eventfd;
4819 /* the process need read permission on control file */
4820 /* AV: shouldn't we check that it's been opened for read instead? */
4821 ret = file_permission(cfile.file, MAY_READ);
4826 * Determine the event callbacks and set them in @event. This used
4827 * to be done via struct cftype but cgroup core no longer knows
4828 * about these events. The following is crude but the whole thing
4829 * is for compatibility anyway.
4831 * DO NOT ADD NEW FILES.
4833 name = cfile.file->f_path.dentry->d_name.name;
4835 if (!strcmp(name, "memory.usage_in_bytes")) {
4836 event->register_event = mem_cgroup_usage_register_event;
4837 event->unregister_event = mem_cgroup_usage_unregister_event;
4838 } else if (!strcmp(name, "memory.oom_control")) {
4839 event->register_event = mem_cgroup_oom_register_event;
4840 event->unregister_event = mem_cgroup_oom_unregister_event;
4841 } else if (!strcmp(name, "memory.pressure_level")) {
4842 event->register_event = vmpressure_register_event;
4843 event->unregister_event = vmpressure_unregister_event;
4844 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4845 event->register_event = memsw_cgroup_usage_register_event;
4846 event->unregister_event = memsw_cgroup_usage_unregister_event;
4853 * Verify @cfile should belong to @css. Also, remaining events are
4854 * automatically removed on cgroup destruction but the removal is
4855 * asynchronous, so take an extra ref on @css.
4857 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4858 &memory_cgrp_subsys);
4860 if (IS_ERR(cfile_css))
4862 if (cfile_css != css) {
4867 ret = event->register_event(memcg, event->eventfd, buf);
4871 vfs_poll(efile.file, &event->pt);
4873 spin_lock(&memcg->event_list_lock);
4874 list_add(&event->list, &memcg->event_list);
4875 spin_unlock(&memcg->event_list_lock);
4887 eventfd_ctx_put(event->eventfd);
4896 static struct cftype mem_cgroup_legacy_files[] = {
4898 .name = "usage_in_bytes",
4899 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4900 .read_u64 = mem_cgroup_read_u64,
4903 .name = "max_usage_in_bytes",
4904 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4905 .write = mem_cgroup_reset,
4906 .read_u64 = mem_cgroup_read_u64,
4909 .name = "limit_in_bytes",
4910 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4911 .write = mem_cgroup_write,
4912 .read_u64 = mem_cgroup_read_u64,
4915 .name = "soft_limit_in_bytes",
4916 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4917 .write = mem_cgroup_write,
4918 .read_u64 = mem_cgroup_read_u64,
4922 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4923 .write = mem_cgroup_reset,
4924 .read_u64 = mem_cgroup_read_u64,
4928 .seq_show = memcg_stat_show,
4931 .name = "force_empty",
4932 .write = mem_cgroup_force_empty_write,
4935 .name = "use_hierarchy",
4936 .write_u64 = mem_cgroup_hierarchy_write,
4937 .read_u64 = mem_cgroup_hierarchy_read,
4940 .name = "cgroup.event_control", /* XXX: for compat */
4941 .write = memcg_write_event_control,
4942 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4945 .name = "swappiness",
4946 .read_u64 = mem_cgroup_swappiness_read,
4947 .write_u64 = mem_cgroup_swappiness_write,
4950 .name = "move_charge_at_immigrate",
4951 .read_u64 = mem_cgroup_move_charge_read,
4952 .write_u64 = mem_cgroup_move_charge_write,
4955 .name = "oom_control",
4956 .seq_show = mem_cgroup_oom_control_read,
4957 .write_u64 = mem_cgroup_oom_control_write,
4958 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4961 .name = "pressure_level",
4965 .name = "numa_stat",
4966 .seq_show = memcg_numa_stat_show,
4970 .name = "kmem.limit_in_bytes",
4971 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4972 .write = mem_cgroup_write,
4973 .read_u64 = mem_cgroup_read_u64,
4976 .name = "kmem.usage_in_bytes",
4977 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4978 .read_u64 = mem_cgroup_read_u64,
4981 .name = "kmem.failcnt",
4982 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4983 .write = mem_cgroup_reset,
4984 .read_u64 = mem_cgroup_read_u64,
4987 .name = "kmem.max_usage_in_bytes",
4988 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4989 .write = mem_cgroup_reset,
4990 .read_u64 = mem_cgroup_read_u64,
4992 #if defined(CONFIG_MEMCG_KMEM) && \
4993 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4995 .name = "kmem.slabinfo",
4996 .seq_show = memcg_slab_show,
5000 .name = "kmem.tcp.limit_in_bytes",
5001 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5002 .write = mem_cgroup_write,
5003 .read_u64 = mem_cgroup_read_u64,
5006 .name = "kmem.tcp.usage_in_bytes",
5007 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5008 .read_u64 = mem_cgroup_read_u64,
5011 .name = "kmem.tcp.failcnt",
5012 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5013 .write = mem_cgroup_reset,
5014 .read_u64 = mem_cgroup_read_u64,
5017 .name = "kmem.tcp.max_usage_in_bytes",
5018 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5019 .write = mem_cgroup_reset,
5020 .read_u64 = mem_cgroup_read_u64,
5022 { }, /* terminate */
5026 * Private memory cgroup IDR
5028 * Swap-out records and page cache shadow entries need to store memcg
5029 * references in constrained space, so we maintain an ID space that is
5030 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5031 * memory-controlled cgroups to 64k.
5033 * However, there usually are many references to the offline CSS after
5034 * the cgroup has been destroyed, such as page cache or reclaimable
5035 * slab objects, that don't need to hang on to the ID. We want to keep
5036 * those dead CSS from occupying IDs, or we might quickly exhaust the
5037 * relatively small ID space and prevent the creation of new cgroups
5038 * even when there are much fewer than 64k cgroups - possibly none.
5040 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5041 * be freed and recycled when it's no longer needed, which is usually
5042 * when the CSS is offlined.
5044 * The only exception to that are records of swapped out tmpfs/shmem
5045 * pages that need to be attributed to live ancestors on swapin. But
5046 * those references are manageable from userspace.
5049 static DEFINE_IDR(mem_cgroup_idr);
5051 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5053 if (memcg->id.id > 0) {
5054 idr_remove(&mem_cgroup_idr, memcg->id.id);
5059 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5062 refcount_add(n, &memcg->id.ref);
5065 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5067 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5068 mem_cgroup_id_remove(memcg);
5070 /* Memcg ID pins CSS */
5071 css_put(&memcg->css);
5075 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5077 mem_cgroup_id_put_many(memcg, 1);
5081 * mem_cgroup_from_id - look up a memcg from a memcg id
5082 * @id: the memcg id to look up
5084 * Caller must hold rcu_read_lock().
5086 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5088 WARN_ON_ONCE(!rcu_read_lock_held());
5089 return idr_find(&mem_cgroup_idr, id);
5092 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5094 struct mem_cgroup_per_node *pn;
5097 * This routine is called against possible nodes.
5098 * But it's BUG to call kmalloc() against offline node.
5100 * TODO: this routine can waste much memory for nodes which will
5101 * never be onlined. It's better to use memory hotplug callback
5104 if (!node_state(node, N_NORMAL_MEMORY))
5106 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5110 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5111 GFP_KERNEL_ACCOUNT);
5112 if (!pn->lruvec_stat_local) {
5117 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct batched_lruvec_stat,
5118 GFP_KERNEL_ACCOUNT);
5119 if (!pn->lruvec_stat_cpu) {
5120 free_percpu(pn->lruvec_stat_local);
5125 lruvec_init(&pn->lruvec);
5126 pn->usage_in_excess = 0;
5127 pn->on_tree = false;
5130 memcg->nodeinfo[node] = pn;
5134 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5136 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5141 free_percpu(pn->lruvec_stat_cpu);
5142 free_percpu(pn->lruvec_stat_local);
5146 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5151 free_mem_cgroup_per_node_info(memcg, node);
5152 free_percpu(memcg->vmstats_percpu);
5156 static void mem_cgroup_free(struct mem_cgroup *memcg)
5160 memcg_wb_domain_exit(memcg);
5162 * Flush percpu lruvec stats to guarantee the value
5163 * correctness on parent's and all ancestor levels.
5165 for_each_online_cpu(cpu)
5166 memcg_flush_lruvec_page_state(memcg, cpu);
5167 __mem_cgroup_free(memcg);
5170 static struct mem_cgroup *mem_cgroup_alloc(void)
5172 struct mem_cgroup *memcg;
5175 int __maybe_unused i;
5176 long error = -ENOMEM;
5178 size = sizeof(struct mem_cgroup);
5179 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5181 memcg = kzalloc(size, GFP_KERNEL);
5183 return ERR_PTR(error);
5185 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5186 1, MEM_CGROUP_ID_MAX,
5188 if (memcg->id.id < 0) {
5189 error = memcg->id.id;
5193 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5194 GFP_KERNEL_ACCOUNT);
5195 if (!memcg->vmstats_percpu)
5199 if (alloc_mem_cgroup_per_node_info(memcg, node))
5202 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5205 INIT_WORK(&memcg->high_work, high_work_func);
5206 INIT_LIST_HEAD(&memcg->oom_notify);
5207 mutex_init(&memcg->thresholds_lock);
5208 spin_lock_init(&memcg->move_lock);
5209 vmpressure_init(&memcg->vmpressure);
5210 INIT_LIST_HEAD(&memcg->event_list);
5211 spin_lock_init(&memcg->event_list_lock);
5212 memcg->socket_pressure = jiffies;
5213 #ifdef CONFIG_MEMCG_KMEM
5214 memcg->kmemcg_id = -1;
5215 INIT_LIST_HEAD(&memcg->objcg_list);
5217 #ifdef CONFIG_CGROUP_WRITEBACK
5218 INIT_LIST_HEAD(&memcg->cgwb_list);
5219 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5220 memcg->cgwb_frn[i].done =
5221 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5223 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5224 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5225 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5226 memcg->deferred_split_queue.split_queue_len = 0;
5228 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5231 mem_cgroup_id_remove(memcg);
5232 __mem_cgroup_free(memcg);
5233 return ERR_PTR(error);
5236 static struct cgroup_subsys_state * __ref
5237 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5239 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5240 struct mem_cgroup *memcg, *old_memcg;
5241 long error = -ENOMEM;
5243 old_memcg = set_active_memcg(parent);
5244 memcg = mem_cgroup_alloc();
5245 set_active_memcg(old_memcg);
5247 return ERR_CAST(memcg);
5249 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5250 memcg->soft_limit = PAGE_COUNTER_MAX;
5251 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5253 memcg->swappiness = mem_cgroup_swappiness(parent);
5254 memcg->oom_kill_disable = parent->oom_kill_disable;
5256 page_counter_init(&memcg->memory, &parent->memory);
5257 page_counter_init(&memcg->swap, &parent->swap);
5258 page_counter_init(&memcg->kmem, &parent->kmem);
5259 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5261 page_counter_init(&memcg->memory, NULL);
5262 page_counter_init(&memcg->swap, NULL);
5263 page_counter_init(&memcg->kmem, NULL);
5264 page_counter_init(&memcg->tcpmem, NULL);
5266 root_mem_cgroup = memcg;
5270 /* The following stuff does not apply to the root */
5271 error = memcg_online_kmem(memcg);
5275 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5276 static_branch_inc(&memcg_sockets_enabled_key);
5280 mem_cgroup_id_remove(memcg);
5281 mem_cgroup_free(memcg);
5282 return ERR_PTR(error);
5285 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5287 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5290 * A memcg must be visible for expand_shrinker_info()
5291 * by the time the maps are allocated. So, we allocate maps
5292 * here, when for_each_mem_cgroup() can't skip it.
5294 if (alloc_shrinker_info(memcg)) {
5295 mem_cgroup_id_remove(memcg);
5299 /* Online state pins memcg ID, memcg ID pins CSS */
5300 refcount_set(&memcg->id.ref, 1);
5305 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5307 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5308 struct mem_cgroup_event *event, *tmp;
5311 * Unregister events and notify userspace.
5312 * Notify userspace about cgroup removing only after rmdir of cgroup
5313 * directory to avoid race between userspace and kernelspace.
5315 spin_lock(&memcg->event_list_lock);
5316 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5317 list_del_init(&event->list);
5318 schedule_work(&event->remove);
5320 spin_unlock(&memcg->event_list_lock);
5322 page_counter_set_min(&memcg->memory, 0);
5323 page_counter_set_low(&memcg->memory, 0);
5325 memcg_offline_kmem(memcg);
5326 reparent_shrinker_deferred(memcg);
5327 wb_memcg_offline(memcg);
5329 drain_all_stock(memcg);
5331 mem_cgroup_id_put(memcg);
5334 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5336 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5338 invalidate_reclaim_iterators(memcg);
5341 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5343 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5344 int __maybe_unused i;
5346 #ifdef CONFIG_CGROUP_WRITEBACK
5347 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5348 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5350 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5351 static_branch_dec(&memcg_sockets_enabled_key);
5353 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5354 static_branch_dec(&memcg_sockets_enabled_key);
5356 vmpressure_cleanup(&memcg->vmpressure);
5357 cancel_work_sync(&memcg->high_work);
5358 mem_cgroup_remove_from_trees(memcg);
5359 free_shrinker_info(memcg);
5360 memcg_free_kmem(memcg);
5361 mem_cgroup_free(memcg);
5365 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5366 * @css: the target css
5368 * Reset the states of the mem_cgroup associated with @css. This is
5369 * invoked when the userland requests disabling on the default hierarchy
5370 * but the memcg is pinned through dependency. The memcg should stop
5371 * applying policies and should revert to the vanilla state as it may be
5372 * made visible again.
5374 * The current implementation only resets the essential configurations.
5375 * This needs to be expanded to cover all the visible parts.
5377 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5379 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5381 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5382 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5383 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5384 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5385 page_counter_set_min(&memcg->memory, 0);
5386 page_counter_set_low(&memcg->memory, 0);
5387 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5388 memcg->soft_limit = PAGE_COUNTER_MAX;
5389 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5390 memcg_wb_domain_size_changed(memcg);
5393 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5395 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5396 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5397 struct memcg_vmstats_percpu *statc;
5401 statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5403 for (i = 0; i < MEMCG_NR_STAT; i++) {
5405 * Collect the aggregated propagation counts of groups
5406 * below us. We're in a per-cpu loop here and this is
5407 * a global counter, so the first cycle will get them.
5409 delta = memcg->vmstats.state_pending[i];
5411 memcg->vmstats.state_pending[i] = 0;
5413 /* Add CPU changes on this level since the last flush */
5414 v = READ_ONCE(statc->state[i]);
5415 if (v != statc->state_prev[i]) {
5416 delta += v - statc->state_prev[i];
5417 statc->state_prev[i] = v;
5423 /* Aggregate counts on this level and propagate upwards */
5424 memcg->vmstats.state[i] += delta;
5426 parent->vmstats.state_pending[i] += delta;
5429 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
5430 delta = memcg->vmstats.events_pending[i];
5432 memcg->vmstats.events_pending[i] = 0;
5434 v = READ_ONCE(statc->events[i]);
5435 if (v != statc->events_prev[i]) {
5436 delta += v - statc->events_prev[i];
5437 statc->events_prev[i] = v;
5443 memcg->vmstats.events[i] += delta;
5445 parent->vmstats.events_pending[i] += delta;
5450 /* Handlers for move charge at task migration. */
5451 static int mem_cgroup_do_precharge(unsigned long count)
5455 /* Try a single bulk charge without reclaim first, kswapd may wake */
5456 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5458 mc.precharge += count;
5462 /* Try charges one by one with reclaim, but do not retry */
5464 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5478 enum mc_target_type {
5485 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5486 unsigned long addr, pte_t ptent)
5488 struct page *page = vm_normal_page(vma, addr, ptent);
5490 if (!page || !page_mapped(page))
5492 if (PageAnon(page)) {
5493 if (!(mc.flags & MOVE_ANON))
5496 if (!(mc.flags & MOVE_FILE))
5499 if (!get_page_unless_zero(page))
5505 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5506 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5507 pte_t ptent, swp_entry_t *entry)
5509 struct page *page = NULL;
5510 swp_entry_t ent = pte_to_swp_entry(ptent);
5512 if (!(mc.flags & MOVE_ANON))
5516 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5517 * a device and because they are not accessible by CPU they are store
5518 * as special swap entry in the CPU page table.
5520 if (is_device_private_entry(ent)) {
5521 page = device_private_entry_to_page(ent);
5523 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5524 * a refcount of 1 when free (unlike normal page)
5526 if (!page_ref_add_unless(page, 1, 1))
5531 if (non_swap_entry(ent))
5535 * Because lookup_swap_cache() updates some statistics counter,
5536 * we call find_get_page() with swapper_space directly.
5538 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5539 entry->val = ent.val;
5544 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5545 pte_t ptent, swp_entry_t *entry)
5551 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5552 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5554 if (!vma->vm_file) /* anonymous vma */
5556 if (!(mc.flags & MOVE_FILE))
5559 /* page is moved even if it's not RSS of this task(page-faulted). */
5560 /* shmem/tmpfs may report page out on swap: account for that too. */
5561 return find_get_incore_page(vma->vm_file->f_mapping,
5562 linear_page_index(vma, addr));
5566 * mem_cgroup_move_account - move account of the page
5568 * @compound: charge the page as compound or small page
5569 * @from: mem_cgroup which the page is moved from.
5570 * @to: mem_cgroup which the page is moved to. @from != @to.
5572 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5574 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5577 static int mem_cgroup_move_account(struct page *page,
5579 struct mem_cgroup *from,
5580 struct mem_cgroup *to)
5582 struct lruvec *from_vec, *to_vec;
5583 struct pglist_data *pgdat;
5584 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5587 VM_BUG_ON(from == to);
5588 VM_BUG_ON_PAGE(PageLRU(page), page);
5589 VM_BUG_ON(compound && !PageTransHuge(page));
5592 * Prevent mem_cgroup_migrate() from looking at
5593 * page's memory cgroup of its source page while we change it.
5596 if (!trylock_page(page))
5600 if (page_memcg(page) != from)
5603 pgdat = page_pgdat(page);
5604 from_vec = mem_cgroup_lruvec(from, pgdat);
5605 to_vec = mem_cgroup_lruvec(to, pgdat);
5607 lock_page_memcg(page);
5609 if (PageAnon(page)) {
5610 if (page_mapped(page)) {
5611 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5612 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5613 if (PageTransHuge(page)) {
5614 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5616 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5621 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5622 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5624 if (PageSwapBacked(page)) {
5625 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5626 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5629 if (page_mapped(page)) {
5630 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5631 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5634 if (PageDirty(page)) {
5635 struct address_space *mapping = page_mapping(page);
5637 if (mapping_can_writeback(mapping)) {
5638 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5640 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5646 if (PageWriteback(page)) {
5647 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5648 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5652 * All state has been migrated, let's switch to the new memcg.
5654 * It is safe to change page's memcg here because the page
5655 * is referenced, charged, isolated, and locked: we can't race
5656 * with (un)charging, migration, LRU putback, or anything else
5657 * that would rely on a stable page's memory cgroup.
5659 * Note that lock_page_memcg is a memcg lock, not a page lock,
5660 * to save space. As soon as we switch page's memory cgroup to a
5661 * new memcg that isn't locked, the above state can change
5662 * concurrently again. Make sure we're truly done with it.
5667 css_put(&from->css);
5669 page->memcg_data = (unsigned long)to;
5671 __unlock_page_memcg(from);
5675 local_irq_disable();
5676 mem_cgroup_charge_statistics(to, page, nr_pages);
5677 memcg_check_events(to, page);
5678 mem_cgroup_charge_statistics(from, page, -nr_pages);
5679 memcg_check_events(from, page);
5688 * get_mctgt_type - get target type of moving charge
5689 * @vma: the vma the pte to be checked belongs
5690 * @addr: the address corresponding to the pte to be checked
5691 * @ptent: the pte to be checked
5692 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5695 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5696 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5697 * move charge. if @target is not NULL, the page is stored in target->page
5698 * with extra refcnt got(Callers should handle it).
5699 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5700 * target for charge migration. if @target is not NULL, the entry is stored
5702 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5703 * (so ZONE_DEVICE page and thus not on the lru).
5704 * For now we such page is charge like a regular page would be as for all
5705 * intent and purposes it is just special memory taking the place of a
5708 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5710 * Called with pte lock held.
5713 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5714 unsigned long addr, pte_t ptent, union mc_target *target)
5716 struct page *page = NULL;
5717 enum mc_target_type ret = MC_TARGET_NONE;
5718 swp_entry_t ent = { .val = 0 };
5720 if (pte_present(ptent))
5721 page = mc_handle_present_pte(vma, addr, ptent);
5722 else if (is_swap_pte(ptent))
5723 page = mc_handle_swap_pte(vma, ptent, &ent);
5724 else if (pte_none(ptent))
5725 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5727 if (!page && !ent.val)
5731 * Do only loose check w/o serialization.
5732 * mem_cgroup_move_account() checks the page is valid or
5733 * not under LRU exclusion.
5735 if (page_memcg(page) == mc.from) {
5736 ret = MC_TARGET_PAGE;
5737 if (is_device_private_page(page))
5738 ret = MC_TARGET_DEVICE;
5740 target->page = page;
5742 if (!ret || !target)
5746 * There is a swap entry and a page doesn't exist or isn't charged.
5747 * But we cannot move a tail-page in a THP.
5749 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5750 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5751 ret = MC_TARGET_SWAP;
5758 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5760 * We don't consider PMD mapped swapping or file mapped pages because THP does
5761 * not support them for now.
5762 * Caller should make sure that pmd_trans_huge(pmd) is true.
5764 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5765 unsigned long addr, pmd_t pmd, union mc_target *target)
5767 struct page *page = NULL;
5768 enum mc_target_type ret = MC_TARGET_NONE;
5770 if (unlikely(is_swap_pmd(pmd))) {
5771 VM_BUG_ON(thp_migration_supported() &&
5772 !is_pmd_migration_entry(pmd));
5775 page = pmd_page(pmd);
5776 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5777 if (!(mc.flags & MOVE_ANON))
5779 if (page_memcg(page) == mc.from) {
5780 ret = MC_TARGET_PAGE;
5783 target->page = page;
5789 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5790 unsigned long addr, pmd_t pmd, union mc_target *target)
5792 return MC_TARGET_NONE;
5796 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5797 unsigned long addr, unsigned long end,
5798 struct mm_walk *walk)
5800 struct vm_area_struct *vma = walk->vma;
5804 ptl = pmd_trans_huge_lock(pmd, vma);
5807 * Note their can not be MC_TARGET_DEVICE for now as we do not
5808 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5809 * this might change.
5811 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5812 mc.precharge += HPAGE_PMD_NR;
5817 if (pmd_trans_unstable(pmd))
5819 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5820 for (; addr != end; pte++, addr += PAGE_SIZE)
5821 if (get_mctgt_type(vma, addr, *pte, NULL))
5822 mc.precharge++; /* increment precharge temporarily */
5823 pte_unmap_unlock(pte - 1, ptl);
5829 static const struct mm_walk_ops precharge_walk_ops = {
5830 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5833 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5835 unsigned long precharge;
5838 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5839 mmap_read_unlock(mm);
5841 precharge = mc.precharge;
5847 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5849 unsigned long precharge = mem_cgroup_count_precharge(mm);
5851 VM_BUG_ON(mc.moving_task);
5852 mc.moving_task = current;
5853 return mem_cgroup_do_precharge(precharge);
5856 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5857 static void __mem_cgroup_clear_mc(void)
5859 struct mem_cgroup *from = mc.from;
5860 struct mem_cgroup *to = mc.to;
5862 /* we must uncharge all the leftover precharges from mc.to */
5864 cancel_charge(mc.to, mc.precharge);
5868 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5869 * we must uncharge here.
5871 if (mc.moved_charge) {
5872 cancel_charge(mc.from, mc.moved_charge);
5873 mc.moved_charge = 0;
5875 /* we must fixup refcnts and charges */
5876 if (mc.moved_swap) {
5877 /* uncharge swap account from the old cgroup */
5878 if (!mem_cgroup_is_root(mc.from))
5879 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5881 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5884 * we charged both to->memory and to->memsw, so we
5885 * should uncharge to->memory.
5887 if (!mem_cgroup_is_root(mc.to))
5888 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5892 memcg_oom_recover(from);
5893 memcg_oom_recover(to);
5894 wake_up_all(&mc.waitq);
5897 static void mem_cgroup_clear_mc(void)
5899 struct mm_struct *mm = mc.mm;
5902 * we must clear moving_task before waking up waiters at the end of
5905 mc.moving_task = NULL;
5906 __mem_cgroup_clear_mc();
5907 spin_lock(&mc.lock);
5911 spin_unlock(&mc.lock);
5916 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5918 struct cgroup_subsys_state *css;
5919 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5920 struct mem_cgroup *from;
5921 struct task_struct *leader, *p;
5922 struct mm_struct *mm;
5923 unsigned long move_flags;
5926 /* charge immigration isn't supported on the default hierarchy */
5927 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5931 * Multi-process migrations only happen on the default hierarchy
5932 * where charge immigration is not used. Perform charge
5933 * immigration if @tset contains a leader and whine if there are
5937 cgroup_taskset_for_each_leader(leader, css, tset) {
5940 memcg = mem_cgroup_from_css(css);
5946 * We are now committed to this value whatever it is. Changes in this
5947 * tunable will only affect upcoming migrations, not the current one.
5948 * So we need to save it, and keep it going.
5950 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5954 from = mem_cgroup_from_task(p);
5956 VM_BUG_ON(from == memcg);
5958 mm = get_task_mm(p);
5961 /* We move charges only when we move a owner of the mm */
5962 if (mm->owner == p) {
5965 VM_BUG_ON(mc.precharge);
5966 VM_BUG_ON(mc.moved_charge);
5967 VM_BUG_ON(mc.moved_swap);
5969 spin_lock(&mc.lock);
5973 mc.flags = move_flags;
5974 spin_unlock(&mc.lock);
5975 /* We set mc.moving_task later */
5977 ret = mem_cgroup_precharge_mc(mm);
5979 mem_cgroup_clear_mc();
5986 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5989 mem_cgroup_clear_mc();
5992 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5993 unsigned long addr, unsigned long end,
5994 struct mm_walk *walk)
5997 struct vm_area_struct *vma = walk->vma;
6000 enum mc_target_type target_type;
6001 union mc_target target;
6004 ptl = pmd_trans_huge_lock(pmd, vma);
6006 if (mc.precharge < HPAGE_PMD_NR) {
6010 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6011 if (target_type == MC_TARGET_PAGE) {
6013 if (!isolate_lru_page(page)) {
6014 if (!mem_cgroup_move_account(page, true,
6016 mc.precharge -= HPAGE_PMD_NR;
6017 mc.moved_charge += HPAGE_PMD_NR;
6019 putback_lru_page(page);
6022 } else if (target_type == MC_TARGET_DEVICE) {
6024 if (!mem_cgroup_move_account(page, true,
6026 mc.precharge -= HPAGE_PMD_NR;
6027 mc.moved_charge += HPAGE_PMD_NR;
6035 if (pmd_trans_unstable(pmd))
6038 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6039 for (; addr != end; addr += PAGE_SIZE) {
6040 pte_t ptent = *(pte++);
6041 bool device = false;
6047 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6048 case MC_TARGET_DEVICE:
6051 case MC_TARGET_PAGE:
6054 * We can have a part of the split pmd here. Moving it
6055 * can be done but it would be too convoluted so simply
6056 * ignore such a partial THP and keep it in original
6057 * memcg. There should be somebody mapping the head.
6059 if (PageTransCompound(page))
6061 if (!device && isolate_lru_page(page))
6063 if (!mem_cgroup_move_account(page, false,
6066 /* we uncharge from mc.from later. */
6070 putback_lru_page(page);
6071 put: /* get_mctgt_type() gets the page */
6074 case MC_TARGET_SWAP:
6076 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6078 mem_cgroup_id_get_many(mc.to, 1);
6079 /* we fixup other refcnts and charges later. */
6087 pte_unmap_unlock(pte - 1, ptl);
6092 * We have consumed all precharges we got in can_attach().
6093 * We try charge one by one, but don't do any additional
6094 * charges to mc.to if we have failed in charge once in attach()
6097 ret = mem_cgroup_do_precharge(1);
6105 static const struct mm_walk_ops charge_walk_ops = {
6106 .pmd_entry = mem_cgroup_move_charge_pte_range,
6109 static void mem_cgroup_move_charge(void)
6111 lru_add_drain_all();
6113 * Signal lock_page_memcg() to take the memcg's move_lock
6114 * while we're moving its pages to another memcg. Then wait
6115 * for already started RCU-only updates to finish.
6117 atomic_inc(&mc.from->moving_account);
6120 if (unlikely(!mmap_read_trylock(mc.mm))) {
6122 * Someone who are holding the mmap_lock might be waiting in
6123 * waitq. So we cancel all extra charges, wake up all waiters,
6124 * and retry. Because we cancel precharges, we might not be able
6125 * to move enough charges, but moving charge is a best-effort
6126 * feature anyway, so it wouldn't be a big problem.
6128 __mem_cgroup_clear_mc();
6133 * When we have consumed all precharges and failed in doing
6134 * additional charge, the page walk just aborts.
6136 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6139 mmap_read_unlock(mc.mm);
6140 atomic_dec(&mc.from->moving_account);
6143 static void mem_cgroup_move_task(void)
6146 mem_cgroup_move_charge();
6147 mem_cgroup_clear_mc();
6150 #else /* !CONFIG_MMU */
6151 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6155 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6158 static void mem_cgroup_move_task(void)
6163 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6165 if (value == PAGE_COUNTER_MAX)
6166 seq_puts(m, "max\n");
6168 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6173 static u64 memory_current_read(struct cgroup_subsys_state *css,
6176 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6178 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6181 static int memory_min_show(struct seq_file *m, void *v)
6183 return seq_puts_memcg_tunable(m,
6184 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6187 static ssize_t memory_min_write(struct kernfs_open_file *of,
6188 char *buf, size_t nbytes, loff_t off)
6190 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6194 buf = strstrip(buf);
6195 err = page_counter_memparse(buf, "max", &min);
6199 page_counter_set_min(&memcg->memory, min);
6204 static int memory_low_show(struct seq_file *m, void *v)
6206 return seq_puts_memcg_tunable(m,
6207 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6210 static ssize_t memory_low_write(struct kernfs_open_file *of,
6211 char *buf, size_t nbytes, loff_t off)
6213 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6217 buf = strstrip(buf);
6218 err = page_counter_memparse(buf, "max", &low);
6222 page_counter_set_low(&memcg->memory, low);
6227 static int memory_high_show(struct seq_file *m, void *v)
6229 return seq_puts_memcg_tunable(m,
6230 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6233 static ssize_t memory_high_write(struct kernfs_open_file *of,
6234 char *buf, size_t nbytes, loff_t off)
6236 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6237 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6238 bool drained = false;
6242 buf = strstrip(buf);
6243 err = page_counter_memparse(buf, "max", &high);
6247 page_counter_set_high(&memcg->memory, high);
6250 unsigned long nr_pages = page_counter_read(&memcg->memory);
6251 unsigned long reclaimed;
6253 if (nr_pages <= high)
6256 if (signal_pending(current))
6260 drain_all_stock(memcg);
6265 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6268 if (!reclaimed && !nr_retries--)
6272 memcg_wb_domain_size_changed(memcg);
6276 static int memory_max_show(struct seq_file *m, void *v)
6278 return seq_puts_memcg_tunable(m,
6279 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6282 static ssize_t memory_max_write(struct kernfs_open_file *of,
6283 char *buf, size_t nbytes, loff_t off)
6285 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6286 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6287 bool drained = false;
6291 buf = strstrip(buf);
6292 err = page_counter_memparse(buf, "max", &max);
6296 xchg(&memcg->memory.max, max);
6299 unsigned long nr_pages = page_counter_read(&memcg->memory);
6301 if (nr_pages <= max)
6304 if (signal_pending(current))
6308 drain_all_stock(memcg);
6314 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6320 memcg_memory_event(memcg, MEMCG_OOM);
6321 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6325 memcg_wb_domain_size_changed(memcg);
6329 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6331 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6332 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6333 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6334 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6335 seq_printf(m, "oom_kill %lu\n",
6336 atomic_long_read(&events[MEMCG_OOM_KILL]));
6339 static int memory_events_show(struct seq_file *m, void *v)
6341 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6343 __memory_events_show(m, memcg->memory_events);
6347 static int memory_events_local_show(struct seq_file *m, void *v)
6349 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6351 __memory_events_show(m, memcg->memory_events_local);
6355 static int memory_stat_show(struct seq_file *m, void *v)
6357 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6360 buf = memory_stat_format(memcg);
6369 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6372 return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6375 static int memory_numa_stat_show(struct seq_file *m, void *v)
6378 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6380 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6383 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6386 seq_printf(m, "%s", memory_stats[i].name);
6387 for_each_node_state(nid, N_MEMORY) {
6389 struct lruvec *lruvec;
6391 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6392 size = lruvec_page_state_output(lruvec,
6393 memory_stats[i].idx);
6394 seq_printf(m, " N%d=%llu", nid, size);
6403 static int memory_oom_group_show(struct seq_file *m, void *v)
6405 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6407 seq_printf(m, "%d\n", memcg->oom_group);
6412 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6413 char *buf, size_t nbytes, loff_t off)
6415 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6418 buf = strstrip(buf);
6422 ret = kstrtoint(buf, 0, &oom_group);
6426 if (oom_group != 0 && oom_group != 1)
6429 memcg->oom_group = oom_group;
6434 static struct cftype memory_files[] = {
6437 .flags = CFTYPE_NOT_ON_ROOT,
6438 .read_u64 = memory_current_read,
6442 .flags = CFTYPE_NOT_ON_ROOT,
6443 .seq_show = memory_min_show,
6444 .write = memory_min_write,
6448 .flags = CFTYPE_NOT_ON_ROOT,
6449 .seq_show = memory_low_show,
6450 .write = memory_low_write,
6454 .flags = CFTYPE_NOT_ON_ROOT,
6455 .seq_show = memory_high_show,
6456 .write = memory_high_write,
6460 .flags = CFTYPE_NOT_ON_ROOT,
6461 .seq_show = memory_max_show,
6462 .write = memory_max_write,
6466 .flags = CFTYPE_NOT_ON_ROOT,
6467 .file_offset = offsetof(struct mem_cgroup, events_file),
6468 .seq_show = memory_events_show,
6471 .name = "events.local",
6472 .flags = CFTYPE_NOT_ON_ROOT,
6473 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6474 .seq_show = memory_events_local_show,
6478 .seq_show = memory_stat_show,
6482 .name = "numa_stat",
6483 .seq_show = memory_numa_stat_show,
6487 .name = "oom.group",
6488 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6489 .seq_show = memory_oom_group_show,
6490 .write = memory_oom_group_write,
6495 struct cgroup_subsys memory_cgrp_subsys = {
6496 .css_alloc = mem_cgroup_css_alloc,
6497 .css_online = mem_cgroup_css_online,
6498 .css_offline = mem_cgroup_css_offline,
6499 .css_released = mem_cgroup_css_released,
6500 .css_free = mem_cgroup_css_free,
6501 .css_reset = mem_cgroup_css_reset,
6502 .css_rstat_flush = mem_cgroup_css_rstat_flush,
6503 .can_attach = mem_cgroup_can_attach,
6504 .cancel_attach = mem_cgroup_cancel_attach,
6505 .post_attach = mem_cgroup_move_task,
6506 .dfl_cftypes = memory_files,
6507 .legacy_cftypes = mem_cgroup_legacy_files,
6512 * This function calculates an individual cgroup's effective
6513 * protection which is derived from its own memory.min/low, its
6514 * parent's and siblings' settings, as well as the actual memory
6515 * distribution in the tree.
6517 * The following rules apply to the effective protection values:
6519 * 1. At the first level of reclaim, effective protection is equal to
6520 * the declared protection in memory.min and memory.low.
6522 * 2. To enable safe delegation of the protection configuration, at
6523 * subsequent levels the effective protection is capped to the
6524 * parent's effective protection.
6526 * 3. To make complex and dynamic subtrees easier to configure, the
6527 * user is allowed to overcommit the declared protection at a given
6528 * level. If that is the case, the parent's effective protection is
6529 * distributed to the children in proportion to how much protection
6530 * they have declared and how much of it they are utilizing.
6532 * This makes distribution proportional, but also work-conserving:
6533 * if one cgroup claims much more protection than it uses memory,
6534 * the unused remainder is available to its siblings.
6536 * 4. Conversely, when the declared protection is undercommitted at a
6537 * given level, the distribution of the larger parental protection
6538 * budget is NOT proportional. A cgroup's protection from a sibling
6539 * is capped to its own memory.min/low setting.
6541 * 5. However, to allow protecting recursive subtrees from each other
6542 * without having to declare each individual cgroup's fixed share
6543 * of the ancestor's claim to protection, any unutilized -
6544 * "floating" - protection from up the tree is distributed in
6545 * proportion to each cgroup's *usage*. This makes the protection
6546 * neutral wrt sibling cgroups and lets them compete freely over
6547 * the shared parental protection budget, but it protects the
6548 * subtree as a whole from neighboring subtrees.
6550 * Note that 4. and 5. are not in conflict: 4. is about protecting
6551 * against immediate siblings whereas 5. is about protecting against
6552 * neighboring subtrees.
6554 static unsigned long effective_protection(unsigned long usage,
6555 unsigned long parent_usage,
6556 unsigned long setting,
6557 unsigned long parent_effective,
6558 unsigned long siblings_protected)
6560 unsigned long protected;
6563 protected = min(usage, setting);
6565 * If all cgroups at this level combined claim and use more
6566 * protection then what the parent affords them, distribute
6567 * shares in proportion to utilization.
6569 * We are using actual utilization rather than the statically
6570 * claimed protection in order to be work-conserving: claimed
6571 * but unused protection is available to siblings that would
6572 * otherwise get a smaller chunk than what they claimed.
6574 if (siblings_protected > parent_effective)
6575 return protected * parent_effective / siblings_protected;
6578 * Ok, utilized protection of all children is within what the
6579 * parent affords them, so we know whatever this child claims
6580 * and utilizes is effectively protected.
6582 * If there is unprotected usage beyond this value, reclaim
6583 * will apply pressure in proportion to that amount.
6585 * If there is unutilized protection, the cgroup will be fully
6586 * shielded from reclaim, but we do return a smaller value for
6587 * protection than what the group could enjoy in theory. This
6588 * is okay. With the overcommit distribution above, effective
6589 * protection is always dependent on how memory is actually
6590 * consumed among the siblings anyway.
6595 * If the children aren't claiming (all of) the protection
6596 * afforded to them by the parent, distribute the remainder in
6597 * proportion to the (unprotected) memory of each cgroup. That
6598 * way, cgroups that aren't explicitly prioritized wrt each
6599 * other compete freely over the allowance, but they are
6600 * collectively protected from neighboring trees.
6602 * We're using unprotected memory for the weight so that if
6603 * some cgroups DO claim explicit protection, we don't protect
6604 * the same bytes twice.
6606 * Check both usage and parent_usage against the respective
6607 * protected values. One should imply the other, but they
6608 * aren't read atomically - make sure the division is sane.
6610 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6612 if (parent_effective > siblings_protected &&
6613 parent_usage > siblings_protected &&
6614 usage > protected) {
6615 unsigned long unclaimed;
6617 unclaimed = parent_effective - siblings_protected;
6618 unclaimed *= usage - protected;
6619 unclaimed /= parent_usage - siblings_protected;
6628 * mem_cgroup_protected - check if memory consumption is in the normal range
6629 * @root: the top ancestor of the sub-tree being checked
6630 * @memcg: the memory cgroup to check
6632 * WARNING: This function is not stateless! It can only be used as part
6633 * of a top-down tree iteration, not for isolated queries.
6635 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6636 struct mem_cgroup *memcg)
6638 unsigned long usage, parent_usage;
6639 struct mem_cgroup *parent;
6641 if (mem_cgroup_disabled())
6645 root = root_mem_cgroup;
6648 * Effective values of the reclaim targets are ignored so they
6649 * can be stale. Have a look at mem_cgroup_protection for more
6651 * TODO: calculation should be more robust so that we do not need
6652 * that special casing.
6657 usage = page_counter_read(&memcg->memory);
6661 parent = parent_mem_cgroup(memcg);
6662 /* No parent means a non-hierarchical mode on v1 memcg */
6666 if (parent == root) {
6667 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6668 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6672 parent_usage = page_counter_read(&parent->memory);
6674 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6675 READ_ONCE(memcg->memory.min),
6676 READ_ONCE(parent->memory.emin),
6677 atomic_long_read(&parent->memory.children_min_usage)));
6679 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6680 READ_ONCE(memcg->memory.low),
6681 READ_ONCE(parent->memory.elow),
6682 atomic_long_read(&parent->memory.children_low_usage)));
6685 static int __mem_cgroup_charge(struct page *page, struct mem_cgroup *memcg,
6688 unsigned int nr_pages = thp_nr_pages(page);
6691 ret = try_charge(memcg, gfp, nr_pages);
6695 css_get(&memcg->css);
6696 commit_charge(page, memcg);
6698 local_irq_disable();
6699 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6700 memcg_check_events(memcg, page);
6707 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6708 * @page: page to charge
6709 * @mm: mm context of the victim
6710 * @gfp_mask: reclaim mode
6712 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6713 * pages according to @gfp_mask if necessary.
6715 * Do not use this for pages allocated for swapin.
6717 * Returns 0 on success. Otherwise, an error code is returned.
6719 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6721 struct mem_cgroup *memcg;
6724 if (mem_cgroup_disabled())
6727 memcg = get_mem_cgroup_from_mm(mm);
6728 ret = __mem_cgroup_charge(page, memcg, gfp_mask);
6729 css_put(&memcg->css);
6735 * mem_cgroup_swapin_charge_page - charge a newly allocated page for swapin
6736 * @page: page to charge
6737 * @mm: mm context of the victim
6738 * @gfp: reclaim mode
6739 * @entry: swap entry for which the page is allocated
6741 * This function charges a page allocated for swapin. Please call this before
6742 * adding the page to the swapcache.
6744 * Returns 0 on success. Otherwise, an error code is returned.
6746 int mem_cgroup_swapin_charge_page(struct page *page, struct mm_struct *mm,
6747 gfp_t gfp, swp_entry_t entry)
6749 struct mem_cgroup *memcg;
6753 if (mem_cgroup_disabled())
6756 id = lookup_swap_cgroup_id(entry);
6758 memcg = mem_cgroup_from_id(id);
6759 if (!memcg || !css_tryget_online(&memcg->css))
6760 memcg = get_mem_cgroup_from_mm(mm);
6763 ret = __mem_cgroup_charge(page, memcg, gfp);
6765 css_put(&memcg->css);
6770 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6771 * @entry: swap entry for which the page is charged
6773 * Call this function after successfully adding the charged page to swapcache.
6775 * Note: This function assumes the page for which swap slot is being uncharged
6778 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
6781 * Cgroup1's unified memory+swap counter has been charged with the
6782 * new swapcache page, finish the transfer by uncharging the swap
6783 * slot. The swap slot would also get uncharged when it dies, but
6784 * it can stick around indefinitely and we'd count the page twice
6787 * Cgroup2 has separate resource counters for memory and swap,
6788 * so this is a non-issue here. Memory and swap charge lifetimes
6789 * correspond 1:1 to page and swap slot lifetimes: we charge the
6790 * page to memory here, and uncharge swap when the slot is freed.
6792 if (!mem_cgroup_disabled() && do_memsw_account()) {
6794 * The swap entry might not get freed for a long time,
6795 * let's not wait for it. The page already received a
6796 * memory+swap charge, drop the swap entry duplicate.
6798 mem_cgroup_uncharge_swap(entry, 1);
6802 struct uncharge_gather {
6803 struct mem_cgroup *memcg;
6804 unsigned long nr_memory;
6805 unsigned long pgpgout;
6806 unsigned long nr_kmem;
6807 struct page *dummy_page;
6810 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6812 memset(ug, 0, sizeof(*ug));
6815 static void uncharge_batch(const struct uncharge_gather *ug)
6817 unsigned long flags;
6819 if (ug->nr_memory) {
6820 page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
6821 if (do_memsw_account())
6822 page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
6823 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6824 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6825 memcg_oom_recover(ug->memcg);
6828 local_irq_save(flags);
6829 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6830 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
6831 memcg_check_events(ug->memcg, ug->dummy_page);
6832 local_irq_restore(flags);
6834 /* drop reference from uncharge_page */
6835 css_put(&ug->memcg->css);
6838 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6840 unsigned long nr_pages;
6841 struct mem_cgroup *memcg;
6842 struct obj_cgroup *objcg;
6843 bool use_objcg = PageMemcgKmem(page);
6845 VM_BUG_ON_PAGE(PageLRU(page), page);
6848 * Nobody should be changing or seriously looking at
6849 * page memcg or objcg at this point, we have fully
6850 * exclusive access to the page.
6853 objcg = __page_objcg(page);
6855 * This get matches the put at the end of the function and
6856 * kmem pages do not hold memcg references anymore.
6858 memcg = get_mem_cgroup_from_objcg(objcg);
6860 memcg = __page_memcg(page);
6866 if (ug->memcg != memcg) {
6869 uncharge_gather_clear(ug);
6872 ug->dummy_page = page;
6874 /* pairs with css_put in uncharge_batch */
6875 css_get(&memcg->css);
6878 nr_pages = compound_nr(page);
6881 ug->nr_memory += nr_pages;
6882 ug->nr_kmem += nr_pages;
6884 page->memcg_data = 0;
6885 obj_cgroup_put(objcg);
6887 /* LRU pages aren't accounted at the root level */
6888 if (!mem_cgroup_is_root(memcg))
6889 ug->nr_memory += nr_pages;
6892 page->memcg_data = 0;
6895 css_put(&memcg->css);
6899 * mem_cgroup_uncharge - uncharge a page
6900 * @page: page to uncharge
6902 * Uncharge a page previously charged with mem_cgroup_charge().
6904 void mem_cgroup_uncharge(struct page *page)
6906 struct uncharge_gather ug;
6908 if (mem_cgroup_disabled())
6911 /* Don't touch page->lru of any random page, pre-check: */
6912 if (!page_memcg(page))
6915 uncharge_gather_clear(&ug);
6916 uncharge_page(page, &ug);
6917 uncharge_batch(&ug);
6921 * mem_cgroup_uncharge_list - uncharge a list of page
6922 * @page_list: list of pages to uncharge
6924 * Uncharge a list of pages previously charged with
6925 * mem_cgroup_charge().
6927 void mem_cgroup_uncharge_list(struct list_head *page_list)
6929 struct uncharge_gather ug;
6932 if (mem_cgroup_disabled())
6935 uncharge_gather_clear(&ug);
6936 list_for_each_entry(page, page_list, lru)
6937 uncharge_page(page, &ug);
6939 uncharge_batch(&ug);
6943 * mem_cgroup_migrate - charge a page's replacement
6944 * @oldpage: currently circulating page
6945 * @newpage: replacement page
6947 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6948 * be uncharged upon free.
6950 * Both pages must be locked, @newpage->mapping must be set up.
6952 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6954 struct mem_cgroup *memcg;
6955 unsigned int nr_pages;
6956 unsigned long flags;
6958 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6959 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6960 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6961 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6964 if (mem_cgroup_disabled())
6967 /* Page cache replacement: new page already charged? */
6968 if (page_memcg(newpage))
6971 memcg = page_memcg(oldpage);
6972 VM_WARN_ON_ONCE_PAGE(!memcg, oldpage);
6976 /* Force-charge the new page. The old one will be freed soon */
6977 nr_pages = thp_nr_pages(newpage);
6979 if (!mem_cgroup_is_root(memcg)) {
6980 page_counter_charge(&memcg->memory, nr_pages);
6981 if (do_memsw_account())
6982 page_counter_charge(&memcg->memsw, nr_pages);
6985 css_get(&memcg->css);
6986 commit_charge(newpage, memcg);
6988 local_irq_save(flags);
6989 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6990 memcg_check_events(memcg, newpage);
6991 local_irq_restore(flags);
6994 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6995 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6997 void mem_cgroup_sk_alloc(struct sock *sk)
6999 struct mem_cgroup *memcg;
7001 if (!mem_cgroup_sockets_enabled)
7004 /* Do not associate the sock with unrelated interrupted task's memcg. */
7009 memcg = mem_cgroup_from_task(current);
7010 if (memcg == root_mem_cgroup)
7012 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7014 if (css_tryget(&memcg->css))
7015 sk->sk_memcg = memcg;
7020 void mem_cgroup_sk_free(struct sock *sk)
7023 css_put(&sk->sk_memcg->css);
7027 * mem_cgroup_charge_skmem - charge socket memory
7028 * @memcg: memcg to charge
7029 * @nr_pages: number of pages to charge
7031 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7032 * @memcg's configured limit, %false if the charge had to be forced.
7034 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7036 gfp_t gfp_mask = GFP_KERNEL;
7038 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7039 struct page_counter *fail;
7041 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7042 memcg->tcpmem_pressure = 0;
7045 page_counter_charge(&memcg->tcpmem, nr_pages);
7046 memcg->tcpmem_pressure = 1;
7050 /* Don't block in the packet receive path */
7052 gfp_mask = GFP_NOWAIT;
7054 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7056 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7059 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7064 * mem_cgroup_uncharge_skmem - uncharge socket memory
7065 * @memcg: memcg to uncharge
7066 * @nr_pages: number of pages to uncharge
7068 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7070 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7071 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7075 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7077 refill_stock(memcg, nr_pages);
7080 static int __init cgroup_memory(char *s)
7084 while ((token = strsep(&s, ",")) != NULL) {
7087 if (!strcmp(token, "nosocket"))
7088 cgroup_memory_nosocket = true;
7089 if (!strcmp(token, "nokmem"))
7090 cgroup_memory_nokmem = true;
7094 __setup("cgroup.memory=", cgroup_memory);
7097 * subsys_initcall() for memory controller.
7099 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7100 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7101 * basically everything that doesn't depend on a specific mem_cgroup structure
7102 * should be initialized from here.
7104 static int __init mem_cgroup_init(void)
7109 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7110 * used for per-memcg-per-cpu caching of per-node statistics. In order
7111 * to work fine, we should make sure that the overfill threshold can't
7112 * exceed S32_MAX / PAGE_SIZE.
7114 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7116 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7117 memcg_hotplug_cpu_dead);
7119 for_each_possible_cpu(cpu)
7120 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7123 for_each_node(node) {
7124 struct mem_cgroup_tree_per_node *rtpn;
7126 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7127 node_online(node) ? node : NUMA_NO_NODE);
7129 rtpn->rb_root = RB_ROOT;
7130 rtpn->rb_rightmost = NULL;
7131 spin_lock_init(&rtpn->lock);
7132 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7137 subsys_initcall(mem_cgroup_init);
7139 #ifdef CONFIG_MEMCG_SWAP
7140 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7142 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7144 * The root cgroup cannot be destroyed, so it's refcount must
7147 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7151 memcg = parent_mem_cgroup(memcg);
7153 memcg = root_mem_cgroup;
7159 * mem_cgroup_swapout - transfer a memsw charge to swap
7160 * @page: page whose memsw charge to transfer
7161 * @entry: swap entry to move the charge to
7163 * Transfer the memsw charge of @page to @entry.
7165 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7167 struct mem_cgroup *memcg, *swap_memcg;
7168 unsigned int nr_entries;
7169 unsigned short oldid;
7171 VM_BUG_ON_PAGE(PageLRU(page), page);
7172 VM_BUG_ON_PAGE(page_count(page), page);
7174 if (mem_cgroup_disabled())
7177 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7180 memcg = page_memcg(page);
7182 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7187 * In case the memcg owning these pages has been offlined and doesn't
7188 * have an ID allocated to it anymore, charge the closest online
7189 * ancestor for the swap instead and transfer the memory+swap charge.
7191 swap_memcg = mem_cgroup_id_get_online(memcg);
7192 nr_entries = thp_nr_pages(page);
7193 /* Get references for the tail pages, too */
7195 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7196 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7198 VM_BUG_ON_PAGE(oldid, page);
7199 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7201 page->memcg_data = 0;
7203 if (!mem_cgroup_is_root(memcg))
7204 page_counter_uncharge(&memcg->memory, nr_entries);
7206 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7207 if (!mem_cgroup_is_root(swap_memcg))
7208 page_counter_charge(&swap_memcg->memsw, nr_entries);
7209 page_counter_uncharge(&memcg->memsw, nr_entries);
7213 * Interrupts should be disabled here because the caller holds the
7214 * i_pages lock which is taken with interrupts-off. It is
7215 * important here to have the interrupts disabled because it is the
7216 * only synchronisation we have for updating the per-CPU variables.
7218 VM_BUG_ON(!irqs_disabled());
7219 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7220 memcg_check_events(memcg, page);
7222 css_put(&memcg->css);
7226 * mem_cgroup_try_charge_swap - try charging swap space for a page
7227 * @page: page being added to swap
7228 * @entry: swap entry to charge
7230 * Try to charge @page's memcg for the swap space at @entry.
7232 * Returns 0 on success, -ENOMEM on failure.
7234 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7236 unsigned int nr_pages = thp_nr_pages(page);
7237 struct page_counter *counter;
7238 struct mem_cgroup *memcg;
7239 unsigned short oldid;
7241 if (mem_cgroup_disabled())
7244 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7247 memcg = page_memcg(page);
7249 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7254 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7258 memcg = mem_cgroup_id_get_online(memcg);
7260 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7261 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7262 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7263 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7264 mem_cgroup_id_put(memcg);
7268 /* Get references for the tail pages, too */
7270 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7271 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7272 VM_BUG_ON_PAGE(oldid, page);
7273 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7279 * mem_cgroup_uncharge_swap - uncharge swap space
7280 * @entry: swap entry to uncharge
7281 * @nr_pages: the amount of swap space to uncharge
7283 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7285 struct mem_cgroup *memcg;
7288 id = swap_cgroup_record(entry, 0, nr_pages);
7290 memcg = mem_cgroup_from_id(id);
7292 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7293 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7294 page_counter_uncharge(&memcg->swap, nr_pages);
7296 page_counter_uncharge(&memcg->memsw, nr_pages);
7298 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7299 mem_cgroup_id_put_many(memcg, nr_pages);
7304 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7306 long nr_swap_pages = get_nr_swap_pages();
7308 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7309 return nr_swap_pages;
7310 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7311 nr_swap_pages = min_t(long, nr_swap_pages,
7312 READ_ONCE(memcg->swap.max) -
7313 page_counter_read(&memcg->swap));
7314 return nr_swap_pages;
7317 bool mem_cgroup_swap_full(struct page *page)
7319 struct mem_cgroup *memcg;
7321 VM_BUG_ON_PAGE(!PageLocked(page), page);
7325 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7328 memcg = page_memcg(page);
7332 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7333 unsigned long usage = page_counter_read(&memcg->swap);
7335 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7336 usage * 2 >= READ_ONCE(memcg->swap.max))
7343 static int __init setup_swap_account(char *s)
7345 if (!strcmp(s, "1"))
7346 cgroup_memory_noswap = false;
7347 else if (!strcmp(s, "0"))
7348 cgroup_memory_noswap = true;
7351 __setup("swapaccount=", setup_swap_account);
7353 static u64 swap_current_read(struct cgroup_subsys_state *css,
7356 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7358 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7361 static int swap_high_show(struct seq_file *m, void *v)
7363 return seq_puts_memcg_tunable(m,
7364 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7367 static ssize_t swap_high_write(struct kernfs_open_file *of,
7368 char *buf, size_t nbytes, loff_t off)
7370 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7374 buf = strstrip(buf);
7375 err = page_counter_memparse(buf, "max", &high);
7379 page_counter_set_high(&memcg->swap, high);
7384 static int swap_max_show(struct seq_file *m, void *v)
7386 return seq_puts_memcg_tunable(m,
7387 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7390 static ssize_t swap_max_write(struct kernfs_open_file *of,
7391 char *buf, size_t nbytes, loff_t off)
7393 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7397 buf = strstrip(buf);
7398 err = page_counter_memparse(buf, "max", &max);
7402 xchg(&memcg->swap.max, max);
7407 static int swap_events_show(struct seq_file *m, void *v)
7409 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7411 seq_printf(m, "high %lu\n",
7412 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7413 seq_printf(m, "max %lu\n",
7414 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7415 seq_printf(m, "fail %lu\n",
7416 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7421 static struct cftype swap_files[] = {
7423 .name = "swap.current",
7424 .flags = CFTYPE_NOT_ON_ROOT,
7425 .read_u64 = swap_current_read,
7428 .name = "swap.high",
7429 .flags = CFTYPE_NOT_ON_ROOT,
7430 .seq_show = swap_high_show,
7431 .write = swap_high_write,
7435 .flags = CFTYPE_NOT_ON_ROOT,
7436 .seq_show = swap_max_show,
7437 .write = swap_max_write,
7440 .name = "swap.events",
7441 .flags = CFTYPE_NOT_ON_ROOT,
7442 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7443 .seq_show = swap_events_show,
7448 static struct cftype memsw_files[] = {
7450 .name = "memsw.usage_in_bytes",
7451 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7452 .read_u64 = mem_cgroup_read_u64,
7455 .name = "memsw.max_usage_in_bytes",
7456 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7457 .write = mem_cgroup_reset,
7458 .read_u64 = mem_cgroup_read_u64,
7461 .name = "memsw.limit_in_bytes",
7462 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7463 .write = mem_cgroup_write,
7464 .read_u64 = mem_cgroup_read_u64,
7467 .name = "memsw.failcnt",
7468 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7469 .write = mem_cgroup_reset,
7470 .read_u64 = mem_cgroup_read_u64,
7472 { }, /* terminate */
7476 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7477 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7478 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7479 * boot parameter. This may result in premature OOPS inside
7480 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7482 static int __init mem_cgroup_swap_init(void)
7484 /* No memory control -> no swap control */
7485 if (mem_cgroup_disabled())
7486 cgroup_memory_noswap = true;
7488 if (cgroup_memory_noswap)
7491 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7492 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7496 core_initcall(mem_cgroup_swap_init);
7498 #endif /* CONFIG_MEMCG_SWAP */