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);
81 EXPORT_PER_CPU_SYMBOL_GPL(int_active_memcg);
83 /* Socket memory accounting disabled? */
84 static bool cgroup_memory_nosocket __ro_after_init;
86 /* Kernel memory accounting disabled? */
87 bool cgroup_memory_nokmem __ro_after_init;
89 /* Whether the swap controller is active */
90 #ifdef CONFIG_MEMCG_SWAP
91 bool cgroup_memory_noswap __ro_after_init;
93 #define cgroup_memory_noswap 1
96 #ifdef CONFIG_CGROUP_WRITEBACK
97 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
100 /* Whether legacy memory+swap accounting is active */
101 static bool do_memsw_account(void)
103 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
106 #define THRESHOLDS_EVENTS_TARGET 128
107 #define SOFTLIMIT_EVENTS_TARGET 1024
110 * Cgroups above their limits are maintained in a RB-Tree, independent of
111 * their hierarchy representation
114 struct mem_cgroup_tree_per_node {
115 struct rb_root rb_root;
116 struct rb_node *rb_rightmost;
120 struct mem_cgroup_tree {
121 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
124 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
127 struct mem_cgroup_eventfd_list {
128 struct list_head list;
129 struct eventfd_ctx *eventfd;
133 * cgroup_event represents events which userspace want to receive.
135 struct mem_cgroup_event {
137 * memcg which the event belongs to.
139 struct mem_cgroup *memcg;
141 * eventfd to signal userspace about the event.
143 struct eventfd_ctx *eventfd;
145 * Each of these stored in a list by the cgroup.
147 struct list_head list;
149 * register_event() callback will be used to add new userspace
150 * waiter for changes related to this event. Use eventfd_signal()
151 * on eventfd to send notification to userspace.
153 int (*register_event)(struct mem_cgroup *memcg,
154 struct eventfd_ctx *eventfd, const char *args);
156 * unregister_event() callback will be called when userspace closes
157 * the eventfd or on cgroup removing. This callback must be set,
158 * if you want provide notification functionality.
160 void (*unregister_event)(struct mem_cgroup *memcg,
161 struct eventfd_ctx *eventfd);
163 * All fields below needed to unregister event when
164 * userspace closes eventfd.
167 wait_queue_head_t *wqh;
168 wait_queue_entry_t wait;
169 struct work_struct remove;
172 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
173 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
175 /* Stuffs for move charges at task migration. */
177 * Types of charges to be moved.
179 #define MOVE_ANON 0x1U
180 #define MOVE_FILE 0x2U
181 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
183 /* "mc" and its members are protected by cgroup_mutex */
184 static struct move_charge_struct {
185 spinlock_t lock; /* for from, to */
186 struct mm_struct *mm;
187 struct mem_cgroup *from;
188 struct mem_cgroup *to;
190 unsigned long precharge;
191 unsigned long moved_charge;
192 unsigned long moved_swap;
193 struct task_struct *moving_task; /* a task moving charges */
194 wait_queue_head_t waitq; /* a waitq for other context */
196 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
197 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
201 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
202 * limit reclaim to prevent infinite loops, if they ever occur.
204 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
205 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
207 /* for encoding cft->private value on file */
216 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
217 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
218 #define MEMFILE_ATTR(val) ((val) & 0xffff)
219 /* Used for OOM notifier */
220 #define OOM_CONTROL (0)
223 * Iteration constructs for visiting all cgroups (under a tree). If
224 * loops are exited prematurely (break), mem_cgroup_iter_break() must
225 * be used for reference counting.
227 #define for_each_mem_cgroup_tree(iter, root) \
228 for (iter = mem_cgroup_iter(root, NULL, NULL); \
230 iter = mem_cgroup_iter(root, iter, NULL))
232 #define for_each_mem_cgroup(iter) \
233 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
235 iter = mem_cgroup_iter(NULL, iter, NULL))
237 static inline bool should_force_charge(void)
239 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
240 (current->flags & PF_EXITING);
243 /* Some nice accessors for the vmpressure. */
244 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
247 memcg = root_mem_cgroup;
248 return &memcg->vmpressure;
251 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
253 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
256 #ifdef CONFIG_MEMCG_KMEM
257 extern spinlock_t css_set_lock;
259 bool mem_cgroup_kmem_disabled(void)
261 return cgroup_memory_nokmem;
264 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
265 unsigned int nr_pages);
267 static void obj_cgroup_release(struct percpu_ref *ref)
269 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
270 unsigned int nr_bytes;
271 unsigned int nr_pages;
275 * At this point all allocated objects are freed, and
276 * objcg->nr_charged_bytes can't have an arbitrary byte value.
277 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
279 * The following sequence can lead to it:
280 * 1) CPU0: objcg == stock->cached_objcg
281 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
282 * PAGE_SIZE bytes are charged
283 * 3) CPU1: a process from another memcg is allocating something,
284 * the stock if flushed,
285 * objcg->nr_charged_bytes = PAGE_SIZE - 92
286 * 5) CPU0: we do release this object,
287 * 92 bytes are added to stock->nr_bytes
288 * 6) CPU0: stock is flushed,
289 * 92 bytes are added to objcg->nr_charged_bytes
291 * In the result, nr_charged_bytes == PAGE_SIZE.
292 * This page will be uncharged in obj_cgroup_release().
294 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
295 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
296 nr_pages = nr_bytes >> PAGE_SHIFT;
299 obj_cgroup_uncharge_pages(objcg, nr_pages);
301 spin_lock_irqsave(&css_set_lock, flags);
302 list_del(&objcg->list);
303 spin_unlock_irqrestore(&css_set_lock, flags);
305 percpu_ref_exit(ref);
306 kfree_rcu(objcg, rcu);
309 static struct obj_cgroup *obj_cgroup_alloc(void)
311 struct obj_cgroup *objcg;
314 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
318 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
324 INIT_LIST_HEAD(&objcg->list);
328 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
329 struct mem_cgroup *parent)
331 struct obj_cgroup *objcg, *iter;
333 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
335 spin_lock_irq(&css_set_lock);
337 /* 1) Ready to reparent active objcg. */
338 list_add(&objcg->list, &memcg->objcg_list);
339 /* 2) Reparent active objcg and already reparented objcgs to parent. */
340 list_for_each_entry(iter, &memcg->objcg_list, list)
341 WRITE_ONCE(iter->memcg, parent);
342 /* 3) Move already reparented objcgs to the parent's list */
343 list_splice(&memcg->objcg_list, &parent->objcg_list);
345 spin_unlock_irq(&css_set_lock);
347 percpu_ref_kill(&objcg->refcnt);
351 * This will be used as a shrinker list's index.
352 * The main reason for not using cgroup id for this:
353 * this works better in sparse environments, where we have a lot of memcgs,
354 * but only a few kmem-limited. Or also, if we have, for instance, 200
355 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
356 * 200 entry array for that.
358 * The current size of the caches array is stored in memcg_nr_cache_ids. It
359 * will double each time we have to increase it.
361 static DEFINE_IDA(memcg_cache_ida);
362 int memcg_nr_cache_ids;
364 /* Protects memcg_nr_cache_ids */
365 static DECLARE_RWSEM(memcg_cache_ids_sem);
367 void memcg_get_cache_ids(void)
369 down_read(&memcg_cache_ids_sem);
372 void memcg_put_cache_ids(void)
374 up_read(&memcg_cache_ids_sem);
378 * MIN_SIZE is different than 1, because we would like to avoid going through
379 * the alloc/free process all the time. In a small machine, 4 kmem-limited
380 * cgroups is a reasonable guess. In the future, it could be a parameter or
381 * tunable, but that is strictly not necessary.
383 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
384 * this constant directly from cgroup, but it is understandable that this is
385 * better kept as an internal representation in cgroup.c. In any case, the
386 * cgrp_id space is not getting any smaller, and we don't have to necessarily
387 * increase ours as well if it increases.
389 #define MEMCG_CACHES_MIN_SIZE 4
390 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
393 * A lot of the calls to the cache allocation functions are expected to be
394 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
395 * conditional to this static branch, we'll have to allow modules that does
396 * kmem_cache_alloc and the such to see this symbol as well
398 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
399 EXPORT_SYMBOL(memcg_kmem_enabled_key);
403 * mem_cgroup_css_from_page - css of the memcg associated with a page
404 * @page: page of interest
406 * If memcg is bound to the default hierarchy, css of the memcg associated
407 * with @page is returned. The returned css remains associated with @page
408 * until it is released.
410 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
413 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
415 struct mem_cgroup *memcg;
417 memcg = page_memcg(page);
419 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
420 memcg = root_mem_cgroup;
426 * page_cgroup_ino - return inode number of the memcg a page is charged to
429 * Look up the closest online ancestor of the memory cgroup @page is charged to
430 * and return its inode number or 0 if @page is not charged to any cgroup. It
431 * is safe to call this function without holding a reference to @page.
433 * Note, this function is inherently racy, because there is nothing to prevent
434 * the cgroup inode from getting torn down and potentially reallocated a moment
435 * after page_cgroup_ino() returns, so it only should be used by callers that
436 * do not care (such as procfs interfaces).
438 ino_t page_cgroup_ino(struct page *page)
440 struct mem_cgroup *memcg;
441 unsigned long ino = 0;
444 memcg = page_memcg_check(page);
446 while (memcg && !(memcg->css.flags & CSS_ONLINE))
447 memcg = parent_mem_cgroup(memcg);
449 ino = cgroup_ino(memcg->css.cgroup);
454 static struct mem_cgroup_per_node *
455 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
457 int nid = page_to_nid(page);
459 return memcg->nodeinfo[nid];
462 static struct mem_cgroup_tree_per_node *
463 soft_limit_tree_node(int nid)
465 return soft_limit_tree.rb_tree_per_node[nid];
468 static struct mem_cgroup_tree_per_node *
469 soft_limit_tree_from_page(struct page *page)
471 int nid = page_to_nid(page);
473 return soft_limit_tree.rb_tree_per_node[nid];
476 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
477 struct mem_cgroup_tree_per_node *mctz,
478 unsigned long new_usage_in_excess)
480 struct rb_node **p = &mctz->rb_root.rb_node;
481 struct rb_node *parent = NULL;
482 struct mem_cgroup_per_node *mz_node;
483 bool rightmost = true;
488 mz->usage_in_excess = new_usage_in_excess;
489 if (!mz->usage_in_excess)
493 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
495 if (mz->usage_in_excess < mz_node->usage_in_excess) {
504 mctz->rb_rightmost = &mz->tree_node;
506 rb_link_node(&mz->tree_node, parent, p);
507 rb_insert_color(&mz->tree_node, &mctz->rb_root);
511 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
512 struct mem_cgroup_tree_per_node *mctz)
517 if (&mz->tree_node == mctz->rb_rightmost)
518 mctz->rb_rightmost = rb_prev(&mz->tree_node);
520 rb_erase(&mz->tree_node, &mctz->rb_root);
524 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
525 struct mem_cgroup_tree_per_node *mctz)
529 spin_lock_irqsave(&mctz->lock, flags);
530 __mem_cgroup_remove_exceeded(mz, mctz);
531 spin_unlock_irqrestore(&mctz->lock, flags);
534 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
536 unsigned long nr_pages = page_counter_read(&memcg->memory);
537 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
538 unsigned long excess = 0;
540 if (nr_pages > soft_limit)
541 excess = nr_pages - soft_limit;
546 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
548 unsigned long excess;
549 struct mem_cgroup_per_node *mz;
550 struct mem_cgroup_tree_per_node *mctz;
552 mctz = soft_limit_tree_from_page(page);
556 * Necessary to update all ancestors when hierarchy is used.
557 * because their event counter is not touched.
559 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
560 mz = mem_cgroup_page_nodeinfo(memcg, page);
561 excess = soft_limit_excess(memcg);
563 * We have to update the tree if mz is on RB-tree or
564 * mem is over its softlimit.
566 if (excess || mz->on_tree) {
569 spin_lock_irqsave(&mctz->lock, flags);
570 /* if on-tree, remove it */
572 __mem_cgroup_remove_exceeded(mz, mctz);
574 * Insert again. mz->usage_in_excess will be updated.
575 * If excess is 0, no tree ops.
577 __mem_cgroup_insert_exceeded(mz, mctz, excess);
578 spin_unlock_irqrestore(&mctz->lock, flags);
583 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
585 struct mem_cgroup_tree_per_node *mctz;
586 struct mem_cgroup_per_node *mz;
590 mz = memcg->nodeinfo[nid];
591 mctz = soft_limit_tree_node(nid);
593 mem_cgroup_remove_exceeded(mz, mctz);
597 static struct mem_cgroup_per_node *
598 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
600 struct mem_cgroup_per_node *mz;
604 if (!mctz->rb_rightmost)
605 goto done; /* Nothing to reclaim from */
607 mz = rb_entry(mctz->rb_rightmost,
608 struct mem_cgroup_per_node, tree_node);
610 * Remove the node now but someone else can add it back,
611 * we will to add it back at the end of reclaim to its correct
612 * position in the tree.
614 __mem_cgroup_remove_exceeded(mz, mctz);
615 if (!soft_limit_excess(mz->memcg) ||
616 !css_tryget(&mz->memcg->css))
622 static struct mem_cgroup_per_node *
623 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
625 struct mem_cgroup_per_node *mz;
627 spin_lock_irq(&mctz->lock);
628 mz = __mem_cgroup_largest_soft_limit_node(mctz);
629 spin_unlock_irq(&mctz->lock);
634 * __mod_memcg_state - update cgroup memory statistics
635 * @memcg: the memory cgroup
636 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
637 * @val: delta to add to the counter, can be negative
639 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
641 if (mem_cgroup_disabled())
644 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
645 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
648 /* idx can be of type enum memcg_stat_item or node_stat_item. */
649 static unsigned long memcg_page_state(struct mem_cgroup *memcg, int idx)
651 long x = READ_ONCE(memcg->vmstats.state[idx]);
659 /* idx can be of type enum memcg_stat_item or node_stat_item. */
660 static unsigned long memcg_page_state_local(struct mem_cgroup *memcg, int idx)
665 for_each_possible_cpu(cpu)
666 x += per_cpu(memcg->vmstats_percpu->state[idx], cpu);
674 static struct mem_cgroup_per_node *
675 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
677 struct mem_cgroup *parent;
679 parent = parent_mem_cgroup(pn->memcg);
682 return parent->nodeinfo[nid];
685 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
688 struct mem_cgroup_per_node *pn;
689 struct mem_cgroup *memcg;
690 long x, threshold = MEMCG_CHARGE_BATCH;
692 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
696 __mod_memcg_state(memcg, idx, val);
699 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
701 if (vmstat_item_in_bytes(idx))
702 threshold <<= PAGE_SHIFT;
704 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
705 if (unlikely(abs(x) > threshold)) {
706 pg_data_t *pgdat = lruvec_pgdat(lruvec);
707 struct mem_cgroup_per_node *pi;
709 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
710 atomic_long_add(x, &pi->lruvec_stat[idx]);
713 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
717 * __mod_lruvec_state - update lruvec memory statistics
718 * @lruvec: the lruvec
719 * @idx: the stat item
720 * @val: delta to add to the counter, can be negative
722 * The lruvec is the intersection of the NUMA node and a cgroup. This
723 * function updates the all three counters that are affected by a
724 * change of state at this level: per-node, per-cgroup, per-lruvec.
726 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
730 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
732 /* Update memcg and lruvec */
733 if (!mem_cgroup_disabled())
734 __mod_memcg_lruvec_state(lruvec, idx, val);
737 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
740 struct page *head = compound_head(page); /* rmap on tail pages */
741 struct mem_cgroup *memcg;
742 pg_data_t *pgdat = page_pgdat(page);
743 struct lruvec *lruvec;
746 memcg = page_memcg(head);
747 /* Untracked pages have no memcg, no lruvec. Update only the node */
750 __mod_node_page_state(pgdat, idx, val);
754 lruvec = mem_cgroup_lruvec(memcg, pgdat);
755 __mod_lruvec_state(lruvec, idx, val);
758 EXPORT_SYMBOL(__mod_lruvec_page_state);
760 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
762 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
763 struct mem_cgroup *memcg;
764 struct lruvec *lruvec;
767 memcg = mem_cgroup_from_obj(p);
770 * Untracked pages have no memcg, no lruvec. Update only the
771 * node. If we reparent the slab objects to the root memcg,
772 * when we free the slab object, we need to update the per-memcg
773 * vmstats to keep it correct for the root memcg.
776 __mod_node_page_state(pgdat, idx, val);
778 lruvec = mem_cgroup_lruvec(memcg, pgdat);
779 __mod_lruvec_state(lruvec, idx, val);
785 * mod_objcg_mlstate() may be called with irq enabled, so
786 * mod_memcg_lruvec_state() should be used.
788 static inline void mod_objcg_mlstate(struct obj_cgroup *objcg,
789 struct pglist_data *pgdat,
790 enum node_stat_item idx, int nr)
792 struct mem_cgroup *memcg;
793 struct lruvec *lruvec;
796 memcg = obj_cgroup_memcg(objcg);
797 lruvec = mem_cgroup_lruvec(memcg, pgdat);
798 mod_memcg_lruvec_state(lruvec, idx, nr);
803 * __count_memcg_events - account VM events in a cgroup
804 * @memcg: the memory cgroup
805 * @idx: the event item
806 * @count: the number of events that occurred
808 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
811 if (mem_cgroup_disabled())
814 __this_cpu_add(memcg->vmstats_percpu->events[idx], count);
815 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
818 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
820 return READ_ONCE(memcg->vmstats.events[event]);
823 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
828 for_each_possible_cpu(cpu)
829 x += per_cpu(memcg->vmstats_percpu->events[event], cpu);
833 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
837 /* pagein of a big page is an event. So, ignore page size */
839 __count_memcg_events(memcg, PGPGIN, 1);
841 __count_memcg_events(memcg, PGPGOUT, 1);
842 nr_pages = -nr_pages; /* for event */
845 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
848 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
849 enum mem_cgroup_events_target target)
851 unsigned long val, next;
853 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
854 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
855 /* from time_after() in jiffies.h */
856 if ((long)(next - val) < 0) {
858 case MEM_CGROUP_TARGET_THRESH:
859 next = val + THRESHOLDS_EVENTS_TARGET;
861 case MEM_CGROUP_TARGET_SOFTLIMIT:
862 next = val + SOFTLIMIT_EVENTS_TARGET;
867 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
874 * Check events in order.
877 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
879 /* threshold event is triggered in finer grain than soft limit */
880 if (unlikely(mem_cgroup_event_ratelimit(memcg,
881 MEM_CGROUP_TARGET_THRESH))) {
884 do_softlimit = mem_cgroup_event_ratelimit(memcg,
885 MEM_CGROUP_TARGET_SOFTLIMIT);
886 mem_cgroup_threshold(memcg);
887 if (unlikely(do_softlimit))
888 mem_cgroup_update_tree(memcg, page);
892 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
895 * mm_update_next_owner() may clear mm->owner to NULL
896 * if it races with swapoff, page migration, etc.
897 * So this can be called with p == NULL.
902 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
904 EXPORT_SYMBOL(mem_cgroup_from_task);
906 static __always_inline struct mem_cgroup *active_memcg(void)
909 return this_cpu_read(int_active_memcg);
911 return current->active_memcg;
915 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
916 * @mm: mm from which memcg should be extracted. It can be NULL.
918 * Obtain a reference on mm->memcg and returns it if successful. If mm
919 * is NULL, then the memcg is chosen as follows:
920 * 1) The active memcg, if set.
921 * 2) current->mm->memcg, if available
923 * If mem_cgroup is disabled, NULL is returned.
925 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
927 struct mem_cgroup *memcg;
929 if (mem_cgroup_disabled())
933 * Page cache insertions can happen without an
934 * actual mm context, e.g. during disk probing
935 * on boot, loopback IO, acct() writes etc.
937 * No need to css_get on root memcg as the reference
938 * counting is disabled on the root level in the
939 * cgroup core. See CSS_NO_REF.
942 memcg = active_memcg();
943 if (unlikely(memcg)) {
944 /* remote memcg must hold a ref */
945 css_get(&memcg->css);
950 return root_mem_cgroup;
955 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
956 if (unlikely(!memcg))
957 memcg = root_mem_cgroup;
958 } while (!css_tryget(&memcg->css));
962 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
964 static __always_inline bool memcg_kmem_bypass(void)
966 /* Allow remote memcg charging from any context. */
967 if (unlikely(active_memcg()))
970 /* Memcg to charge can't be determined. */
971 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
978 * mem_cgroup_iter - iterate over memory cgroup hierarchy
979 * @root: hierarchy root
980 * @prev: previously returned memcg, NULL on first invocation
981 * @reclaim: cookie for shared reclaim walks, NULL for full walks
983 * Returns references to children of the hierarchy below @root, or
984 * @root itself, or %NULL after a full round-trip.
986 * Caller must pass the return value in @prev on subsequent
987 * invocations for reference counting, or use mem_cgroup_iter_break()
988 * to cancel a hierarchy walk before the round-trip is complete.
990 * Reclaimers can specify a node in @reclaim to divide up the memcgs
991 * in the hierarchy among all concurrent reclaimers operating on the
994 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
995 struct mem_cgroup *prev,
996 struct mem_cgroup_reclaim_cookie *reclaim)
998 struct mem_cgroup_reclaim_iter *iter;
999 struct cgroup_subsys_state *css = NULL;
1000 struct mem_cgroup *memcg = NULL;
1001 struct mem_cgroup *pos = NULL;
1003 if (mem_cgroup_disabled())
1007 root = root_mem_cgroup;
1009 if (prev && !reclaim)
1015 struct mem_cgroup_per_node *mz;
1017 mz = root->nodeinfo[reclaim->pgdat->node_id];
1020 if (prev && reclaim->generation != iter->generation)
1024 pos = READ_ONCE(iter->position);
1025 if (!pos || css_tryget(&pos->css))
1028 * css reference reached zero, so iter->position will
1029 * be cleared by ->css_released. However, we should not
1030 * rely on this happening soon, because ->css_released
1031 * is called from a work queue, and by busy-waiting we
1032 * might block it. So we clear iter->position right
1035 (void)cmpxchg(&iter->position, pos, NULL);
1043 css = css_next_descendant_pre(css, &root->css);
1046 * Reclaimers share the hierarchy walk, and a
1047 * new one might jump in right at the end of
1048 * the hierarchy - make sure they see at least
1049 * one group and restart from the beginning.
1057 * Verify the css and acquire a reference. The root
1058 * is provided by the caller, so we know it's alive
1059 * and kicking, and don't take an extra reference.
1061 memcg = mem_cgroup_from_css(css);
1063 if (css == &root->css)
1066 if (css_tryget(css))
1074 * The position could have already been updated by a competing
1075 * thread, so check that the value hasn't changed since we read
1076 * it to avoid reclaiming from the same cgroup twice.
1078 (void)cmpxchg(&iter->position, pos, memcg);
1086 reclaim->generation = iter->generation;
1091 if (prev && prev != root)
1092 css_put(&prev->css);
1098 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1099 * @root: hierarchy root
1100 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1102 void mem_cgroup_iter_break(struct mem_cgroup *root,
1103 struct mem_cgroup *prev)
1106 root = root_mem_cgroup;
1107 if (prev && prev != root)
1108 css_put(&prev->css);
1111 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1112 struct mem_cgroup *dead_memcg)
1114 struct mem_cgroup_reclaim_iter *iter;
1115 struct mem_cgroup_per_node *mz;
1118 for_each_node(nid) {
1119 mz = from->nodeinfo[nid];
1121 cmpxchg(&iter->position, dead_memcg, NULL);
1125 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1127 struct mem_cgroup *memcg = dead_memcg;
1128 struct mem_cgroup *last;
1131 __invalidate_reclaim_iterators(memcg, dead_memcg);
1133 } while ((memcg = parent_mem_cgroup(memcg)));
1136 * When cgruop1 non-hierarchy mode is used,
1137 * parent_mem_cgroup() does not walk all the way up to the
1138 * cgroup root (root_mem_cgroup). So we have to handle
1139 * dead_memcg from cgroup root separately.
1141 if (last != root_mem_cgroup)
1142 __invalidate_reclaim_iterators(root_mem_cgroup,
1147 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1148 * @memcg: hierarchy root
1149 * @fn: function to call for each task
1150 * @arg: argument passed to @fn
1152 * This function iterates over tasks attached to @memcg or to any of its
1153 * descendants and calls @fn for each task. If @fn returns a non-zero
1154 * value, the function breaks the iteration loop and returns the value.
1155 * Otherwise, it will iterate over all tasks and return 0.
1157 * This function must not be called for the root memory cgroup.
1159 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1160 int (*fn)(struct task_struct *, void *), void *arg)
1162 struct mem_cgroup *iter;
1165 BUG_ON(memcg == root_mem_cgroup);
1167 for_each_mem_cgroup_tree(iter, memcg) {
1168 struct css_task_iter it;
1169 struct task_struct *task;
1171 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1172 while (!ret && (task = css_task_iter_next(&it)))
1173 ret = fn(task, arg);
1174 css_task_iter_end(&it);
1176 mem_cgroup_iter_break(memcg, iter);
1183 #ifdef CONFIG_DEBUG_VM
1184 void lruvec_memcg_debug(struct lruvec *lruvec, struct page *page)
1186 struct mem_cgroup *memcg;
1188 if (mem_cgroup_disabled())
1191 memcg = page_memcg(page);
1194 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != root_mem_cgroup, page);
1196 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != memcg, page);
1201 * lock_page_lruvec - lock and return lruvec for a given page.
1204 * These functions are safe to use under any of the following conditions:
1207 * - lock_page_memcg()
1208 * - page->_refcount is zero
1210 struct lruvec *lock_page_lruvec(struct page *page)
1212 struct lruvec *lruvec;
1214 lruvec = mem_cgroup_page_lruvec(page);
1215 spin_lock(&lruvec->lru_lock);
1217 lruvec_memcg_debug(lruvec, page);
1222 struct lruvec *lock_page_lruvec_irq(struct page *page)
1224 struct lruvec *lruvec;
1226 lruvec = mem_cgroup_page_lruvec(page);
1227 spin_lock_irq(&lruvec->lru_lock);
1229 lruvec_memcg_debug(lruvec, page);
1234 struct lruvec *lock_page_lruvec_irqsave(struct page *page, unsigned long *flags)
1236 struct lruvec *lruvec;
1238 lruvec = mem_cgroup_page_lruvec(page);
1239 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1241 lruvec_memcg_debug(lruvec, page);
1247 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1248 * @lruvec: mem_cgroup per zone lru vector
1249 * @lru: index of lru list the page is sitting on
1250 * @zid: zone id of the accounted pages
1251 * @nr_pages: positive when adding or negative when removing
1253 * This function must be called under lru_lock, just before a page is added
1254 * to or just after a page is removed from an lru list (that ordering being
1255 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1257 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1258 int zid, int nr_pages)
1260 struct mem_cgroup_per_node *mz;
1261 unsigned long *lru_size;
1264 if (mem_cgroup_disabled())
1267 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1268 lru_size = &mz->lru_zone_size[zid][lru];
1271 *lru_size += nr_pages;
1274 if (WARN_ONCE(size < 0,
1275 "%s(%p, %d, %d): lru_size %ld\n",
1276 __func__, lruvec, lru, nr_pages, size)) {
1282 *lru_size += nr_pages;
1286 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1287 * @memcg: the memory cgroup
1289 * Returns the maximum amount of memory @mem can be charged with, in
1292 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1294 unsigned long margin = 0;
1295 unsigned long count;
1296 unsigned long limit;
1298 count = page_counter_read(&memcg->memory);
1299 limit = READ_ONCE(memcg->memory.max);
1301 margin = limit - count;
1303 if (do_memsw_account()) {
1304 count = page_counter_read(&memcg->memsw);
1305 limit = READ_ONCE(memcg->memsw.max);
1307 margin = min(margin, limit - count);
1316 * A routine for checking "mem" is under move_account() or not.
1318 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1319 * moving cgroups. This is for waiting at high-memory pressure
1322 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1324 struct mem_cgroup *from;
1325 struct mem_cgroup *to;
1328 * Unlike task_move routines, we access mc.to, mc.from not under
1329 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1331 spin_lock(&mc.lock);
1337 ret = mem_cgroup_is_descendant(from, memcg) ||
1338 mem_cgroup_is_descendant(to, memcg);
1340 spin_unlock(&mc.lock);
1344 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1346 if (mc.moving_task && current != mc.moving_task) {
1347 if (mem_cgroup_under_move(memcg)) {
1349 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1350 /* moving charge context might have finished. */
1353 finish_wait(&mc.waitq, &wait);
1360 struct memory_stat {
1365 static const struct memory_stat memory_stats[] = {
1366 { "anon", NR_ANON_MAPPED },
1367 { "file", NR_FILE_PAGES },
1368 { "kernel_stack", NR_KERNEL_STACK_KB },
1369 { "pagetables", NR_PAGETABLE },
1370 { "percpu", MEMCG_PERCPU_B },
1371 { "sock", MEMCG_SOCK },
1372 { "shmem", NR_SHMEM },
1373 { "file_mapped", NR_FILE_MAPPED },
1374 { "file_dirty", NR_FILE_DIRTY },
1375 { "file_writeback", NR_WRITEBACK },
1377 { "swapcached", NR_SWAPCACHE },
1379 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1380 { "anon_thp", NR_ANON_THPS },
1381 { "file_thp", NR_FILE_THPS },
1382 { "shmem_thp", NR_SHMEM_THPS },
1384 { "inactive_anon", NR_INACTIVE_ANON },
1385 { "active_anon", NR_ACTIVE_ANON },
1386 { "inactive_file", NR_INACTIVE_FILE },
1387 { "active_file", NR_ACTIVE_FILE },
1388 { "unevictable", NR_UNEVICTABLE },
1389 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B },
1390 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B },
1392 /* The memory events */
1393 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON },
1394 { "workingset_refault_file", WORKINGSET_REFAULT_FILE },
1395 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON },
1396 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE },
1397 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON },
1398 { "workingset_restore_file", WORKINGSET_RESTORE_FILE },
1399 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM },
1402 /* Translate stat items to the correct unit for memory.stat output */
1403 static int memcg_page_state_unit(int item)
1406 case MEMCG_PERCPU_B:
1407 case NR_SLAB_RECLAIMABLE_B:
1408 case NR_SLAB_UNRECLAIMABLE_B:
1409 case WORKINGSET_REFAULT_ANON:
1410 case WORKINGSET_REFAULT_FILE:
1411 case WORKINGSET_ACTIVATE_ANON:
1412 case WORKINGSET_ACTIVATE_FILE:
1413 case WORKINGSET_RESTORE_ANON:
1414 case WORKINGSET_RESTORE_FILE:
1415 case WORKINGSET_NODERECLAIM:
1417 case NR_KERNEL_STACK_KB:
1424 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1427 return memcg_page_state(memcg, item) * memcg_page_state_unit(item);
1430 static char *memory_stat_format(struct mem_cgroup *memcg)
1435 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1440 * Provide statistics on the state of the memory subsystem as
1441 * well as cumulative event counters that show past behavior.
1443 * This list is ordered following a combination of these gradients:
1444 * 1) generic big picture -> specifics and details
1445 * 2) reflecting userspace activity -> reflecting kernel heuristics
1447 * Current memory state:
1449 cgroup_rstat_flush(memcg->css.cgroup);
1451 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1454 size = memcg_page_state_output(memcg, memory_stats[i].idx);
1455 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1457 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1458 size += memcg_page_state_output(memcg,
1459 NR_SLAB_RECLAIMABLE_B);
1460 seq_buf_printf(&s, "slab %llu\n", size);
1464 /* Accumulated memory events */
1466 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1467 memcg_events(memcg, PGFAULT));
1468 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1469 memcg_events(memcg, PGMAJFAULT));
1470 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1471 memcg_events(memcg, PGREFILL));
1472 seq_buf_printf(&s, "pgscan %lu\n",
1473 memcg_events(memcg, PGSCAN_KSWAPD) +
1474 memcg_events(memcg, PGSCAN_DIRECT));
1475 seq_buf_printf(&s, "pgsteal %lu\n",
1476 memcg_events(memcg, PGSTEAL_KSWAPD) +
1477 memcg_events(memcg, PGSTEAL_DIRECT));
1478 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1479 memcg_events(memcg, PGACTIVATE));
1480 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1481 memcg_events(memcg, PGDEACTIVATE));
1482 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1483 memcg_events(memcg, PGLAZYFREE));
1484 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1485 memcg_events(memcg, PGLAZYFREED));
1487 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1488 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1489 memcg_events(memcg, THP_FAULT_ALLOC));
1490 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1491 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1492 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1494 /* The above should easily fit into one page */
1495 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1500 #define K(x) ((x) << (PAGE_SHIFT-10))
1502 * mem_cgroup_print_oom_context: Print OOM information relevant to
1503 * memory controller.
1504 * @memcg: The memory cgroup that went over limit
1505 * @p: Task that is going to be killed
1507 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1510 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1515 pr_cont(",oom_memcg=");
1516 pr_cont_cgroup_path(memcg->css.cgroup);
1518 pr_cont(",global_oom");
1520 pr_cont(",task_memcg=");
1521 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1527 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1528 * memory controller.
1529 * @memcg: The memory cgroup that went over limit
1531 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1535 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1536 K((u64)page_counter_read(&memcg->memory)),
1537 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1538 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1539 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1540 K((u64)page_counter_read(&memcg->swap)),
1541 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1543 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1544 K((u64)page_counter_read(&memcg->memsw)),
1545 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1546 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1547 K((u64)page_counter_read(&memcg->kmem)),
1548 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1551 pr_info("Memory cgroup stats for ");
1552 pr_cont_cgroup_path(memcg->css.cgroup);
1554 buf = memory_stat_format(memcg);
1562 * Return the memory (and swap, if configured) limit for a memcg.
1564 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1566 unsigned long max = READ_ONCE(memcg->memory.max);
1568 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1569 if (mem_cgroup_swappiness(memcg))
1570 max += min(READ_ONCE(memcg->swap.max),
1571 (unsigned long)total_swap_pages);
1573 if (mem_cgroup_swappiness(memcg)) {
1574 /* Calculate swap excess capacity from memsw limit */
1575 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1577 max += min(swap, (unsigned long)total_swap_pages);
1583 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1585 return page_counter_read(&memcg->memory);
1588 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1591 struct oom_control oc = {
1595 .gfp_mask = gfp_mask,
1600 if (mutex_lock_killable(&oom_lock))
1603 if (mem_cgroup_margin(memcg) >= (1 << order))
1607 * A few threads which were not waiting at mutex_lock_killable() can
1608 * fail to bail out. Therefore, check again after holding oom_lock.
1610 ret = should_force_charge() || out_of_memory(&oc);
1613 mutex_unlock(&oom_lock);
1617 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1620 unsigned long *total_scanned)
1622 struct mem_cgroup *victim = NULL;
1625 unsigned long excess;
1626 unsigned long nr_scanned;
1627 struct mem_cgroup_reclaim_cookie reclaim = {
1631 excess = soft_limit_excess(root_memcg);
1634 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1639 * If we have not been able to reclaim
1640 * anything, it might because there are
1641 * no reclaimable pages under this hierarchy
1646 * We want to do more targeted reclaim.
1647 * excess >> 2 is not to excessive so as to
1648 * reclaim too much, nor too less that we keep
1649 * coming back to reclaim from this cgroup
1651 if (total >= (excess >> 2) ||
1652 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1657 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1658 pgdat, &nr_scanned);
1659 *total_scanned += nr_scanned;
1660 if (!soft_limit_excess(root_memcg))
1663 mem_cgroup_iter_break(root_memcg, victim);
1667 #ifdef CONFIG_LOCKDEP
1668 static struct lockdep_map memcg_oom_lock_dep_map = {
1669 .name = "memcg_oom_lock",
1673 static DEFINE_SPINLOCK(memcg_oom_lock);
1676 * Check OOM-Killer is already running under our hierarchy.
1677 * If someone is running, return false.
1679 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1681 struct mem_cgroup *iter, *failed = NULL;
1683 spin_lock(&memcg_oom_lock);
1685 for_each_mem_cgroup_tree(iter, memcg) {
1686 if (iter->oom_lock) {
1688 * this subtree of our hierarchy is already locked
1689 * so we cannot give a lock.
1692 mem_cgroup_iter_break(memcg, iter);
1695 iter->oom_lock = true;
1700 * OK, we failed to lock the whole subtree so we have
1701 * to clean up what we set up to the failing subtree
1703 for_each_mem_cgroup_tree(iter, memcg) {
1704 if (iter == failed) {
1705 mem_cgroup_iter_break(memcg, iter);
1708 iter->oom_lock = false;
1711 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1713 spin_unlock(&memcg_oom_lock);
1718 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1720 struct mem_cgroup *iter;
1722 spin_lock(&memcg_oom_lock);
1723 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1724 for_each_mem_cgroup_tree(iter, memcg)
1725 iter->oom_lock = false;
1726 spin_unlock(&memcg_oom_lock);
1729 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1731 struct mem_cgroup *iter;
1733 spin_lock(&memcg_oom_lock);
1734 for_each_mem_cgroup_tree(iter, memcg)
1736 spin_unlock(&memcg_oom_lock);
1739 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1741 struct mem_cgroup *iter;
1744 * Be careful about under_oom underflows because a child memcg
1745 * could have been added after mem_cgroup_mark_under_oom.
1747 spin_lock(&memcg_oom_lock);
1748 for_each_mem_cgroup_tree(iter, memcg)
1749 if (iter->under_oom > 0)
1751 spin_unlock(&memcg_oom_lock);
1754 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1756 struct oom_wait_info {
1757 struct mem_cgroup *memcg;
1758 wait_queue_entry_t wait;
1761 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1762 unsigned mode, int sync, void *arg)
1764 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1765 struct mem_cgroup *oom_wait_memcg;
1766 struct oom_wait_info *oom_wait_info;
1768 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1769 oom_wait_memcg = oom_wait_info->memcg;
1771 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1772 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1774 return autoremove_wake_function(wait, mode, sync, arg);
1777 static void memcg_oom_recover(struct mem_cgroup *memcg)
1780 * For the following lockless ->under_oom test, the only required
1781 * guarantee is that it must see the state asserted by an OOM when
1782 * this function is called as a result of userland actions
1783 * triggered by the notification of the OOM. This is trivially
1784 * achieved by invoking mem_cgroup_mark_under_oom() before
1785 * triggering notification.
1787 if (memcg && memcg->under_oom)
1788 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1798 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1800 enum oom_status ret;
1803 if (order > PAGE_ALLOC_COSTLY_ORDER)
1806 memcg_memory_event(memcg, MEMCG_OOM);
1809 * We are in the middle of the charge context here, so we
1810 * don't want to block when potentially sitting on a callstack
1811 * that holds all kinds of filesystem and mm locks.
1813 * cgroup1 allows disabling the OOM killer and waiting for outside
1814 * handling until the charge can succeed; remember the context and put
1815 * the task to sleep at the end of the page fault when all locks are
1818 * On the other hand, in-kernel OOM killer allows for an async victim
1819 * memory reclaim (oom_reaper) and that means that we are not solely
1820 * relying on the oom victim to make a forward progress and we can
1821 * invoke the oom killer here.
1823 * Please note that mem_cgroup_out_of_memory might fail to find a
1824 * victim and then we have to bail out from the charge path.
1826 if (memcg->oom_kill_disable) {
1827 if (!current->in_user_fault)
1829 css_get(&memcg->css);
1830 current->memcg_in_oom = memcg;
1831 current->memcg_oom_gfp_mask = mask;
1832 current->memcg_oom_order = order;
1837 mem_cgroup_mark_under_oom(memcg);
1839 locked = mem_cgroup_oom_trylock(memcg);
1842 mem_cgroup_oom_notify(memcg);
1844 mem_cgroup_unmark_under_oom(memcg);
1845 if (mem_cgroup_out_of_memory(memcg, mask, order))
1851 mem_cgroup_oom_unlock(memcg);
1857 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1858 * @handle: actually kill/wait or just clean up the OOM state
1860 * This has to be called at the end of a page fault if the memcg OOM
1861 * handler was enabled.
1863 * Memcg supports userspace OOM handling where failed allocations must
1864 * sleep on a waitqueue until the userspace task resolves the
1865 * situation. Sleeping directly in the charge context with all kinds
1866 * of locks held is not a good idea, instead we remember an OOM state
1867 * in the task and mem_cgroup_oom_synchronize() has to be called at
1868 * the end of the page fault to complete the OOM handling.
1870 * Returns %true if an ongoing memcg OOM situation was detected and
1871 * completed, %false otherwise.
1873 bool mem_cgroup_oom_synchronize(bool handle)
1875 struct mem_cgroup *memcg = current->memcg_in_oom;
1876 struct oom_wait_info owait;
1879 /* OOM is global, do not handle */
1886 owait.memcg = memcg;
1887 owait.wait.flags = 0;
1888 owait.wait.func = memcg_oom_wake_function;
1889 owait.wait.private = current;
1890 INIT_LIST_HEAD(&owait.wait.entry);
1892 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1893 mem_cgroup_mark_under_oom(memcg);
1895 locked = mem_cgroup_oom_trylock(memcg);
1898 mem_cgroup_oom_notify(memcg);
1900 if (locked && !memcg->oom_kill_disable) {
1901 mem_cgroup_unmark_under_oom(memcg);
1902 finish_wait(&memcg_oom_waitq, &owait.wait);
1903 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1904 current->memcg_oom_order);
1907 mem_cgroup_unmark_under_oom(memcg);
1908 finish_wait(&memcg_oom_waitq, &owait.wait);
1912 mem_cgroup_oom_unlock(memcg);
1914 * There is no guarantee that an OOM-lock contender
1915 * sees the wakeups triggered by the OOM kill
1916 * uncharges. Wake any sleepers explicitly.
1918 memcg_oom_recover(memcg);
1921 current->memcg_in_oom = NULL;
1922 css_put(&memcg->css);
1927 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1928 * @victim: task to be killed by the OOM killer
1929 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1931 * Returns a pointer to a memory cgroup, which has to be cleaned up
1932 * by killing all belonging OOM-killable tasks.
1934 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1936 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1937 struct mem_cgroup *oom_domain)
1939 struct mem_cgroup *oom_group = NULL;
1940 struct mem_cgroup *memcg;
1942 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1946 oom_domain = root_mem_cgroup;
1950 memcg = mem_cgroup_from_task(victim);
1951 if (memcg == root_mem_cgroup)
1955 * If the victim task has been asynchronously moved to a different
1956 * memory cgroup, we might end up killing tasks outside oom_domain.
1957 * In this case it's better to ignore memory.group.oom.
1959 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
1963 * Traverse the memory cgroup hierarchy from the victim task's
1964 * cgroup up to the OOMing cgroup (or root) to find the
1965 * highest-level memory cgroup with oom.group set.
1967 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1968 if (memcg->oom_group)
1971 if (memcg == oom_domain)
1976 css_get(&oom_group->css);
1983 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1985 pr_info("Tasks in ");
1986 pr_cont_cgroup_path(memcg->css.cgroup);
1987 pr_cont(" are going to be killed due to memory.oom.group set\n");
1991 * lock_page_memcg - lock a page and memcg binding
1994 * This function protects unlocked LRU pages from being moved to
1997 * It ensures lifetime of the locked memcg. Caller is responsible
1998 * for the lifetime of the page.
2000 void lock_page_memcg(struct page *page)
2002 struct page *head = compound_head(page); /* rmap on tail pages */
2003 struct mem_cgroup *memcg;
2004 unsigned long flags;
2007 * The RCU lock is held throughout the transaction. The fast
2008 * path can get away without acquiring the memcg->move_lock
2009 * because page moving starts with an RCU grace period.
2013 if (mem_cgroup_disabled())
2016 memcg = page_memcg(head);
2017 if (unlikely(!memcg))
2020 #ifdef CONFIG_PROVE_LOCKING
2021 local_irq_save(flags);
2022 might_lock(&memcg->move_lock);
2023 local_irq_restore(flags);
2026 if (atomic_read(&memcg->moving_account) <= 0)
2029 spin_lock_irqsave(&memcg->move_lock, flags);
2030 if (memcg != page_memcg(head)) {
2031 spin_unlock_irqrestore(&memcg->move_lock, flags);
2036 * When charge migration first begins, we can have multiple
2037 * critical sections holding the fast-path RCU lock and one
2038 * holding the slowpath move_lock. Track the task who has the
2039 * move_lock for unlock_page_memcg().
2041 memcg->move_lock_task = current;
2042 memcg->move_lock_flags = flags;
2044 EXPORT_SYMBOL(lock_page_memcg);
2046 static void __unlock_page_memcg(struct mem_cgroup *memcg)
2048 if (memcg && memcg->move_lock_task == current) {
2049 unsigned long flags = memcg->move_lock_flags;
2051 memcg->move_lock_task = NULL;
2052 memcg->move_lock_flags = 0;
2054 spin_unlock_irqrestore(&memcg->move_lock, flags);
2061 * unlock_page_memcg - unlock a page and memcg binding
2064 void unlock_page_memcg(struct page *page)
2066 struct page *head = compound_head(page);
2068 __unlock_page_memcg(page_memcg(head));
2070 EXPORT_SYMBOL(unlock_page_memcg);
2073 #ifdef CONFIG_MEMCG_KMEM
2074 struct obj_cgroup *cached_objcg;
2075 struct pglist_data *cached_pgdat;
2076 unsigned int nr_bytes;
2077 int nr_slab_reclaimable_b;
2078 int nr_slab_unreclaimable_b;
2084 struct memcg_stock_pcp {
2085 struct mem_cgroup *cached; /* this never be root cgroup */
2086 unsigned int nr_pages;
2087 struct obj_stock task_obj;
2088 struct obj_stock irq_obj;
2090 struct work_struct work;
2091 unsigned long flags;
2092 #define FLUSHING_CACHED_CHARGE 0
2094 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2095 static DEFINE_MUTEX(percpu_charge_mutex);
2097 #ifdef CONFIG_MEMCG_KMEM
2098 static void drain_obj_stock(struct obj_stock *stock);
2099 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2100 struct mem_cgroup *root_memcg);
2103 static inline void drain_obj_stock(struct obj_stock *stock)
2106 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2107 struct mem_cgroup *root_memcg)
2114 * Most kmem_cache_alloc() calls are from user context. The irq disable/enable
2115 * sequence used in this case to access content from object stock is slow.
2116 * To optimize for user context access, there are now two object stocks for
2117 * task context and interrupt context access respectively.
2119 * The task context object stock can be accessed by disabling preemption only
2120 * which is cheap in non-preempt kernel. The interrupt context object stock
2121 * can only be accessed after disabling interrupt. User context code can
2122 * access interrupt object stock, but not vice versa.
2124 static inline struct obj_stock *get_obj_stock(unsigned long *pflags)
2126 struct memcg_stock_pcp *stock;
2128 if (likely(in_task())) {
2131 stock = this_cpu_ptr(&memcg_stock);
2132 return &stock->task_obj;
2135 local_irq_save(*pflags);
2136 stock = this_cpu_ptr(&memcg_stock);
2137 return &stock->irq_obj;
2140 static inline void put_obj_stock(unsigned long flags)
2142 if (likely(in_task()))
2145 local_irq_restore(flags);
2149 * consume_stock: Try to consume stocked charge on this cpu.
2150 * @memcg: memcg to consume from.
2151 * @nr_pages: how many pages to charge.
2153 * The charges will only happen if @memcg matches the current cpu's memcg
2154 * stock, and at least @nr_pages are available in that stock. Failure to
2155 * service an allocation will refill the stock.
2157 * returns true if successful, false otherwise.
2159 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2161 struct memcg_stock_pcp *stock;
2162 unsigned long flags;
2165 if (nr_pages > MEMCG_CHARGE_BATCH)
2168 local_irq_save(flags);
2170 stock = this_cpu_ptr(&memcg_stock);
2171 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2172 stock->nr_pages -= nr_pages;
2176 local_irq_restore(flags);
2182 * Returns stocks cached in percpu and reset cached information.
2184 static void drain_stock(struct memcg_stock_pcp *stock)
2186 struct mem_cgroup *old = stock->cached;
2191 if (stock->nr_pages) {
2192 page_counter_uncharge(&old->memory, stock->nr_pages);
2193 if (do_memsw_account())
2194 page_counter_uncharge(&old->memsw, stock->nr_pages);
2195 stock->nr_pages = 0;
2199 stock->cached = NULL;
2202 static void drain_local_stock(struct work_struct *dummy)
2204 struct memcg_stock_pcp *stock;
2205 unsigned long flags;
2208 * The only protection from memory hotplug vs. drain_stock races is
2209 * that we always operate on local CPU stock here with IRQ disabled
2211 local_irq_save(flags);
2213 stock = this_cpu_ptr(&memcg_stock);
2214 drain_obj_stock(&stock->irq_obj);
2216 drain_obj_stock(&stock->task_obj);
2218 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2220 local_irq_restore(flags);
2224 * Cache charges(val) to local per_cpu area.
2225 * This will be consumed by consume_stock() function, later.
2227 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2229 struct memcg_stock_pcp *stock;
2230 unsigned long flags;
2232 local_irq_save(flags);
2234 stock = this_cpu_ptr(&memcg_stock);
2235 if (stock->cached != memcg) { /* reset if necessary */
2237 css_get(&memcg->css);
2238 stock->cached = memcg;
2240 stock->nr_pages += nr_pages;
2242 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2245 local_irq_restore(flags);
2249 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2250 * of the hierarchy under it.
2252 static void drain_all_stock(struct mem_cgroup *root_memcg)
2256 /* If someone's already draining, avoid adding running more workers. */
2257 if (!mutex_trylock(&percpu_charge_mutex))
2260 * Notify other cpus that system-wide "drain" is running
2261 * We do not care about races with the cpu hotplug because cpu down
2262 * as well as workers from this path always operate on the local
2263 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2266 for_each_online_cpu(cpu) {
2267 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2268 struct mem_cgroup *memcg;
2272 memcg = stock->cached;
2273 if (memcg && stock->nr_pages &&
2274 mem_cgroup_is_descendant(memcg, root_memcg))
2276 if (obj_stock_flush_required(stock, root_memcg))
2281 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2283 drain_local_stock(&stock->work);
2285 schedule_work_on(cpu, &stock->work);
2289 mutex_unlock(&percpu_charge_mutex);
2292 static void memcg_flush_lruvec_page_state(struct mem_cgroup *memcg, int cpu)
2296 for_each_node(nid) {
2297 struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
2298 unsigned long stat[NR_VM_NODE_STAT_ITEMS];
2299 struct batched_lruvec_stat *lstatc;
2302 lstatc = per_cpu_ptr(pn->lruvec_stat_cpu, cpu);
2303 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
2304 stat[i] = lstatc->count[i];
2305 lstatc->count[i] = 0;
2309 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
2310 atomic_long_add(stat[i], &pn->lruvec_stat[i]);
2311 } while ((pn = parent_nodeinfo(pn, nid)));
2315 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2317 struct memcg_stock_pcp *stock;
2318 struct mem_cgroup *memcg;
2320 stock = &per_cpu(memcg_stock, cpu);
2323 for_each_mem_cgroup(memcg)
2324 memcg_flush_lruvec_page_state(memcg, cpu);
2329 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2330 unsigned int nr_pages,
2333 unsigned long nr_reclaimed = 0;
2336 unsigned long pflags;
2338 if (page_counter_read(&memcg->memory) <=
2339 READ_ONCE(memcg->memory.high))
2342 memcg_memory_event(memcg, MEMCG_HIGH);
2344 psi_memstall_enter(&pflags);
2345 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2347 psi_memstall_leave(&pflags);
2348 } while ((memcg = parent_mem_cgroup(memcg)) &&
2349 !mem_cgroup_is_root(memcg));
2351 return nr_reclaimed;
2354 static void high_work_func(struct work_struct *work)
2356 struct mem_cgroup *memcg;
2358 memcg = container_of(work, struct mem_cgroup, high_work);
2359 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2363 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2364 * enough to still cause a significant slowdown in most cases, while still
2365 * allowing diagnostics and tracing to proceed without becoming stuck.
2367 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2370 * When calculating the delay, we use these either side of the exponentiation to
2371 * maintain precision and scale to a reasonable number of jiffies (see the table
2374 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2375 * overage ratio to a delay.
2376 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2377 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2378 * to produce a reasonable delay curve.
2380 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2381 * reasonable delay curve compared to precision-adjusted overage, not
2382 * penalising heavily at first, but still making sure that growth beyond the
2383 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2384 * example, with a high of 100 megabytes:
2386 * +-------+------------------------+
2387 * | usage | time to allocate in ms |
2388 * +-------+------------------------+
2410 * +-------+------------------------+
2412 #define MEMCG_DELAY_PRECISION_SHIFT 20
2413 #define MEMCG_DELAY_SCALING_SHIFT 14
2415 static u64 calculate_overage(unsigned long usage, unsigned long high)
2423 * Prevent division by 0 in overage calculation by acting as if
2424 * it was a threshold of 1 page
2426 high = max(high, 1UL);
2428 overage = usage - high;
2429 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2430 return div64_u64(overage, high);
2433 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2435 u64 overage, max_overage = 0;
2438 overage = calculate_overage(page_counter_read(&memcg->memory),
2439 READ_ONCE(memcg->memory.high));
2440 max_overage = max(overage, max_overage);
2441 } while ((memcg = parent_mem_cgroup(memcg)) &&
2442 !mem_cgroup_is_root(memcg));
2447 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2449 u64 overage, max_overage = 0;
2452 overage = calculate_overage(page_counter_read(&memcg->swap),
2453 READ_ONCE(memcg->swap.high));
2455 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2456 max_overage = max(overage, max_overage);
2457 } while ((memcg = parent_mem_cgroup(memcg)) &&
2458 !mem_cgroup_is_root(memcg));
2464 * Get the number of jiffies that we should penalise a mischievous cgroup which
2465 * is exceeding its memory.high by checking both it and its ancestors.
2467 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2468 unsigned int nr_pages,
2471 unsigned long penalty_jiffies;
2477 * We use overage compared to memory.high to calculate the number of
2478 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2479 * fairly lenient on small overages, and increasingly harsh when the
2480 * memcg in question makes it clear that it has no intention of stopping
2481 * its crazy behaviour, so we exponentially increase the delay based on
2484 penalty_jiffies = max_overage * max_overage * HZ;
2485 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2486 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2489 * Factor in the task's own contribution to the overage, such that four
2490 * N-sized allocations are throttled approximately the same as one
2491 * 4N-sized allocation.
2493 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2494 * larger the current charge patch is than that.
2496 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2500 * Scheduled by try_charge() to be executed from the userland return path
2501 * and reclaims memory over the high limit.
2503 void mem_cgroup_handle_over_high(void)
2505 unsigned long penalty_jiffies;
2506 unsigned long pflags;
2507 unsigned long nr_reclaimed;
2508 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2509 int nr_retries = MAX_RECLAIM_RETRIES;
2510 struct mem_cgroup *memcg;
2511 bool in_retry = false;
2513 if (likely(!nr_pages))
2516 memcg = get_mem_cgroup_from_mm(current->mm);
2517 current->memcg_nr_pages_over_high = 0;
2521 * The allocating task should reclaim at least the batch size, but for
2522 * subsequent retries we only want to do what's necessary to prevent oom
2523 * or breaching resource isolation.
2525 * This is distinct from memory.max or page allocator behaviour because
2526 * memory.high is currently batched, whereas memory.max and the page
2527 * allocator run every time an allocation is made.
2529 nr_reclaimed = reclaim_high(memcg,
2530 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2534 * memory.high is breached and reclaim is unable to keep up. Throttle
2535 * allocators proactively to slow down excessive growth.
2537 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2538 mem_find_max_overage(memcg));
2540 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2541 swap_find_max_overage(memcg));
2544 * Clamp the max delay per usermode return so as to still keep the
2545 * application moving forwards and also permit diagnostics, albeit
2548 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2551 * Don't sleep if the amount of jiffies this memcg owes us is so low
2552 * that it's not even worth doing, in an attempt to be nice to those who
2553 * go only a small amount over their memory.high value and maybe haven't
2554 * been aggressively reclaimed enough yet.
2556 if (penalty_jiffies <= HZ / 100)
2560 * If reclaim is making forward progress but we're still over
2561 * memory.high, we want to encourage that rather than doing allocator
2564 if (nr_reclaimed || nr_retries--) {
2570 * If we exit early, we're guaranteed to die (since
2571 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2572 * need to account for any ill-begotten jiffies to pay them off later.
2574 psi_memstall_enter(&pflags);
2575 schedule_timeout_killable(penalty_jiffies);
2576 psi_memstall_leave(&pflags);
2579 css_put(&memcg->css);
2582 static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
2583 unsigned int nr_pages)
2585 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2586 int nr_retries = MAX_RECLAIM_RETRIES;
2587 struct mem_cgroup *mem_over_limit;
2588 struct page_counter *counter;
2589 enum oom_status oom_status;
2590 unsigned long nr_reclaimed;
2591 bool may_swap = true;
2592 bool drained = false;
2593 unsigned long pflags;
2596 if (consume_stock(memcg, nr_pages))
2599 if (!do_memsw_account() ||
2600 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2601 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2603 if (do_memsw_account())
2604 page_counter_uncharge(&memcg->memsw, batch);
2605 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2607 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2611 if (batch > nr_pages) {
2617 * Memcg doesn't have a dedicated reserve for atomic
2618 * allocations. But like the global atomic pool, we need to
2619 * put the burden of reclaim on regular allocation requests
2620 * and let these go through as privileged allocations.
2622 if (gfp_mask & __GFP_ATOMIC)
2626 * Unlike in global OOM situations, memcg is not in a physical
2627 * memory shortage. Allow dying and OOM-killed tasks to
2628 * bypass the last charges so that they can exit quickly and
2629 * free their memory.
2631 if (unlikely(should_force_charge()))
2635 * Prevent unbounded recursion when reclaim operations need to
2636 * allocate memory. This might exceed the limits temporarily,
2637 * but we prefer facilitating memory reclaim and getting back
2638 * under the limit over triggering OOM kills in these cases.
2640 if (unlikely(current->flags & PF_MEMALLOC))
2643 if (unlikely(task_in_memcg_oom(current)))
2646 if (!gfpflags_allow_blocking(gfp_mask))
2649 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2651 psi_memstall_enter(&pflags);
2652 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2653 gfp_mask, may_swap);
2654 psi_memstall_leave(&pflags);
2656 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2660 drain_all_stock(mem_over_limit);
2665 if (gfp_mask & __GFP_NORETRY)
2668 * Even though the limit is exceeded at this point, reclaim
2669 * may have been able to free some pages. Retry the charge
2670 * before killing the task.
2672 * Only for regular pages, though: huge pages are rather
2673 * unlikely to succeed so close to the limit, and we fall back
2674 * to regular pages anyway in case of failure.
2676 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2679 * At task move, charge accounts can be doubly counted. So, it's
2680 * better to wait until the end of task_move if something is going on.
2682 if (mem_cgroup_wait_acct_move(mem_over_limit))
2688 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2691 if (fatal_signal_pending(current))
2695 * keep retrying as long as the memcg oom killer is able to make
2696 * a forward progress or bypass the charge if the oom killer
2697 * couldn't make any progress.
2699 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2700 get_order(nr_pages * PAGE_SIZE));
2701 switch (oom_status) {
2703 nr_retries = MAX_RECLAIM_RETRIES;
2711 if (!(gfp_mask & __GFP_NOFAIL))
2715 * The allocation either can't fail or will lead to more memory
2716 * being freed very soon. Allow memory usage go over the limit
2717 * temporarily by force charging it.
2719 page_counter_charge(&memcg->memory, nr_pages);
2720 if (do_memsw_account())
2721 page_counter_charge(&memcg->memsw, nr_pages);
2726 if (batch > nr_pages)
2727 refill_stock(memcg, batch - nr_pages);
2730 * If the hierarchy is above the normal consumption range, schedule
2731 * reclaim on returning to userland. We can perform reclaim here
2732 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2733 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2734 * not recorded as it most likely matches current's and won't
2735 * change in the meantime. As high limit is checked again before
2736 * reclaim, the cost of mismatch is negligible.
2739 bool mem_high, swap_high;
2741 mem_high = page_counter_read(&memcg->memory) >
2742 READ_ONCE(memcg->memory.high);
2743 swap_high = page_counter_read(&memcg->swap) >
2744 READ_ONCE(memcg->swap.high);
2746 /* Don't bother a random interrupted task */
2747 if (in_interrupt()) {
2749 schedule_work(&memcg->high_work);
2755 if (mem_high || swap_high) {
2757 * The allocating tasks in this cgroup will need to do
2758 * reclaim or be throttled to prevent further growth
2759 * of the memory or swap footprints.
2761 * Target some best-effort fairness between the tasks,
2762 * and distribute reclaim work and delay penalties
2763 * based on how much each task is actually allocating.
2765 current->memcg_nr_pages_over_high += batch;
2766 set_notify_resume(current);
2769 } while ((memcg = parent_mem_cgroup(memcg)));
2774 static inline int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2775 unsigned int nr_pages)
2777 if (mem_cgroup_is_root(memcg))
2780 return try_charge_memcg(memcg, gfp_mask, nr_pages);
2783 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2784 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2786 if (mem_cgroup_is_root(memcg))
2789 page_counter_uncharge(&memcg->memory, nr_pages);
2790 if (do_memsw_account())
2791 page_counter_uncharge(&memcg->memsw, nr_pages);
2795 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2797 VM_BUG_ON_PAGE(page_memcg(page), page);
2799 * Any of the following ensures page's memcg stability:
2803 * - lock_page_memcg()
2804 * - exclusive reference
2806 page->memcg_data = (unsigned long)memcg;
2809 static struct mem_cgroup *get_mem_cgroup_from_objcg(struct obj_cgroup *objcg)
2811 struct mem_cgroup *memcg;
2815 memcg = obj_cgroup_memcg(objcg);
2816 if (unlikely(!css_tryget(&memcg->css)))
2823 #ifdef CONFIG_MEMCG_KMEM
2825 * The allocated objcg pointers array is not accounted directly.
2826 * Moreover, it should not come from DMA buffer and is not readily
2827 * reclaimable. So those GFP bits should be masked off.
2829 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | __GFP_ACCOUNT)
2831 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2832 gfp_t gfp, bool new_page)
2834 unsigned int objects = objs_per_slab_page(s, page);
2835 unsigned long memcg_data;
2838 gfp &= ~OBJCGS_CLEAR_MASK;
2839 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2844 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2847 * If the slab page is brand new and nobody can yet access
2848 * it's memcg_data, no synchronization is required and
2849 * memcg_data can be simply assigned.
2851 page->memcg_data = memcg_data;
2852 } else if (cmpxchg(&page->memcg_data, 0, memcg_data)) {
2854 * If the slab page is already in use, somebody can allocate
2855 * and assign obj_cgroups in parallel. In this case the existing
2856 * objcg vector should be reused.
2862 kmemleak_not_leak(vec);
2867 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2869 * A passed kernel object can be a slab object or a generic kernel page, so
2870 * different mechanisms for getting the memory cgroup pointer should be used.
2871 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2872 * can not know for sure how the kernel object is implemented.
2873 * mem_cgroup_from_obj() can be safely used in such cases.
2875 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2876 * cgroup_mutex, etc.
2878 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2882 if (mem_cgroup_disabled())
2885 page = virt_to_head_page(p);
2888 * Slab objects are accounted individually, not per-page.
2889 * Memcg membership data for each individual object is saved in
2890 * the page->obj_cgroups.
2892 if (page_objcgs_check(page)) {
2893 struct obj_cgroup *objcg;
2896 off = obj_to_index(page->slab_cache, page, p);
2897 objcg = page_objcgs(page)[off];
2899 return obj_cgroup_memcg(objcg);
2905 * page_memcg_check() is used here, because page_has_obj_cgroups()
2906 * check above could fail because the object cgroups vector wasn't set
2907 * at that moment, but it can be set concurrently.
2908 * page_memcg_check(page) will guarantee that a proper memory
2909 * cgroup pointer or NULL will be returned.
2911 return page_memcg_check(page);
2914 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2916 struct obj_cgroup *objcg = NULL;
2917 struct mem_cgroup *memcg;
2919 if (memcg_kmem_bypass())
2923 if (unlikely(active_memcg()))
2924 memcg = active_memcg();
2926 memcg = mem_cgroup_from_task(current);
2928 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2929 objcg = rcu_dereference(memcg->objcg);
2930 if (objcg && obj_cgroup_tryget(objcg))
2939 static int memcg_alloc_cache_id(void)
2944 id = ida_simple_get(&memcg_cache_ida,
2945 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2949 if (id < memcg_nr_cache_ids)
2953 * There's no space for the new id in memcg_caches arrays,
2954 * so we have to grow them.
2956 down_write(&memcg_cache_ids_sem);
2958 size = 2 * (id + 1);
2959 if (size < MEMCG_CACHES_MIN_SIZE)
2960 size = MEMCG_CACHES_MIN_SIZE;
2961 else if (size > MEMCG_CACHES_MAX_SIZE)
2962 size = MEMCG_CACHES_MAX_SIZE;
2964 err = memcg_update_all_list_lrus(size);
2966 memcg_nr_cache_ids = size;
2968 up_write(&memcg_cache_ids_sem);
2971 ida_simple_remove(&memcg_cache_ida, id);
2977 static void memcg_free_cache_id(int id)
2979 ida_simple_remove(&memcg_cache_ida, id);
2983 * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
2984 * @objcg: object cgroup to uncharge
2985 * @nr_pages: number of pages to uncharge
2987 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
2988 unsigned int nr_pages)
2990 struct mem_cgroup *memcg;
2992 memcg = get_mem_cgroup_from_objcg(objcg);
2994 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2995 page_counter_uncharge(&memcg->kmem, nr_pages);
2996 refill_stock(memcg, nr_pages);
2998 css_put(&memcg->css);
3002 * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
3003 * @objcg: object cgroup to charge
3004 * @gfp: reclaim mode
3005 * @nr_pages: number of pages to charge
3007 * Returns 0 on success, an error code on failure.
3009 static int obj_cgroup_charge_pages(struct obj_cgroup *objcg, gfp_t gfp,
3010 unsigned int nr_pages)
3012 struct page_counter *counter;
3013 struct mem_cgroup *memcg;
3016 memcg = get_mem_cgroup_from_objcg(objcg);
3018 ret = try_charge_memcg(memcg, gfp, nr_pages);
3022 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3023 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3026 * Enforce __GFP_NOFAIL allocation because callers are not
3027 * prepared to see failures and likely do not have any failure
3030 if (gfp & __GFP_NOFAIL) {
3031 page_counter_charge(&memcg->kmem, nr_pages);
3034 cancel_charge(memcg, nr_pages);
3038 css_put(&memcg->css);
3044 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3045 * @page: page to charge
3046 * @gfp: reclaim mode
3047 * @order: allocation order
3049 * Returns 0 on success, an error code on failure.
3051 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3053 struct obj_cgroup *objcg;
3056 objcg = get_obj_cgroup_from_current();
3058 ret = obj_cgroup_charge_pages(objcg, gfp, 1 << order);
3060 page->memcg_data = (unsigned long)objcg |
3064 obj_cgroup_put(objcg);
3070 * __memcg_kmem_uncharge_page: uncharge a kmem page
3071 * @page: page to uncharge
3072 * @order: allocation order
3074 void __memcg_kmem_uncharge_page(struct page *page, int order)
3076 struct obj_cgroup *objcg;
3077 unsigned int nr_pages = 1 << order;
3079 if (!PageMemcgKmem(page))
3082 objcg = __page_objcg(page);
3083 obj_cgroup_uncharge_pages(objcg, nr_pages);
3084 page->memcg_data = 0;
3085 obj_cgroup_put(objcg);
3088 void mod_objcg_state(struct obj_cgroup *objcg, struct pglist_data *pgdat,
3089 enum node_stat_item idx, int nr)
3091 unsigned long flags;
3092 struct obj_stock *stock = get_obj_stock(&flags);
3096 * Save vmstat data in stock and skip vmstat array update unless
3097 * accumulating over a page of vmstat data or when pgdat or idx
3100 if (stock->cached_objcg != objcg) {
3101 drain_obj_stock(stock);
3102 obj_cgroup_get(objcg);
3103 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3104 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3105 stock->cached_objcg = objcg;
3106 stock->cached_pgdat = pgdat;
3107 } else if (stock->cached_pgdat != pgdat) {
3108 /* Flush the existing cached vmstat data */
3109 struct pglist_data *oldpg = stock->cached_pgdat;
3111 if (stock->nr_slab_reclaimable_b) {
3112 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_RECLAIMABLE_B,
3113 stock->nr_slab_reclaimable_b);
3114 stock->nr_slab_reclaimable_b = 0;
3116 if (stock->nr_slab_unreclaimable_b) {
3117 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_UNRECLAIMABLE_B,
3118 stock->nr_slab_unreclaimable_b);
3119 stock->nr_slab_unreclaimable_b = 0;
3121 stock->cached_pgdat = pgdat;
3124 bytes = (idx == NR_SLAB_RECLAIMABLE_B) ? &stock->nr_slab_reclaimable_b
3125 : &stock->nr_slab_unreclaimable_b;
3127 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3128 * cached locally at least once before pushing it out.
3135 if (abs(*bytes) > PAGE_SIZE) {
3143 mod_objcg_mlstate(objcg, pgdat, idx, nr);
3145 put_obj_stock(flags);
3148 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3150 unsigned long flags;
3151 struct obj_stock *stock = get_obj_stock(&flags);
3154 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3155 stock->nr_bytes -= nr_bytes;
3159 put_obj_stock(flags);
3164 static void drain_obj_stock(struct obj_stock *stock)
3166 struct obj_cgroup *old = stock->cached_objcg;
3171 if (stock->nr_bytes) {
3172 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3173 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3176 obj_cgroup_uncharge_pages(old, nr_pages);
3179 * The leftover is flushed to the centralized per-memcg value.
3180 * On the next attempt to refill obj stock it will be moved
3181 * to a per-cpu stock (probably, on an other CPU), see
3182 * refill_obj_stock().
3184 * How often it's flushed is a trade-off between the memory
3185 * limit enforcement accuracy and potential CPU contention,
3186 * so it might be changed in the future.
3188 atomic_add(nr_bytes, &old->nr_charged_bytes);
3189 stock->nr_bytes = 0;
3193 * Flush the vmstat data in current stock
3195 if (stock->nr_slab_reclaimable_b || stock->nr_slab_unreclaimable_b) {
3196 if (stock->nr_slab_reclaimable_b) {
3197 mod_objcg_mlstate(old, stock->cached_pgdat,
3198 NR_SLAB_RECLAIMABLE_B,
3199 stock->nr_slab_reclaimable_b);
3200 stock->nr_slab_reclaimable_b = 0;
3202 if (stock->nr_slab_unreclaimable_b) {
3203 mod_objcg_mlstate(old, stock->cached_pgdat,
3204 NR_SLAB_UNRECLAIMABLE_B,
3205 stock->nr_slab_unreclaimable_b);
3206 stock->nr_slab_unreclaimable_b = 0;
3208 stock->cached_pgdat = NULL;
3211 obj_cgroup_put(old);
3212 stock->cached_objcg = NULL;
3215 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3216 struct mem_cgroup *root_memcg)
3218 struct mem_cgroup *memcg;
3220 if (in_task() && stock->task_obj.cached_objcg) {
3221 memcg = obj_cgroup_memcg(stock->task_obj.cached_objcg);
3222 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3225 if (stock->irq_obj.cached_objcg) {
3226 memcg = obj_cgroup_memcg(stock->irq_obj.cached_objcg);
3227 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3234 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes,
3235 bool allow_uncharge)
3237 unsigned long flags;
3238 struct obj_stock *stock = get_obj_stock(&flags);
3239 unsigned int nr_pages = 0;
3241 if (stock->cached_objcg != objcg) { /* reset if necessary */
3242 drain_obj_stock(stock);
3243 obj_cgroup_get(objcg);
3244 stock->cached_objcg = objcg;
3245 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3246 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3247 allow_uncharge = true; /* Allow uncharge when objcg changes */
3249 stock->nr_bytes += nr_bytes;
3251 if (allow_uncharge && (stock->nr_bytes > PAGE_SIZE)) {
3252 nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3253 stock->nr_bytes &= (PAGE_SIZE - 1);
3256 put_obj_stock(flags);
3259 obj_cgroup_uncharge_pages(objcg, nr_pages);
3262 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3264 unsigned int nr_pages, nr_bytes;
3267 if (consume_obj_stock(objcg, size))
3271 * In theory, objcg->nr_charged_bytes can have enough
3272 * pre-charged bytes to satisfy the allocation. However,
3273 * flushing objcg->nr_charged_bytes requires two atomic
3274 * operations, and objcg->nr_charged_bytes can't be big.
3275 * The shared objcg->nr_charged_bytes can also become a
3276 * performance bottleneck if all tasks of the same memcg are
3277 * trying to update it. So it's better to ignore it and try
3278 * grab some new pages. The stock's nr_bytes will be flushed to
3279 * objcg->nr_charged_bytes later on when objcg changes.
3281 * The stock's nr_bytes may contain enough pre-charged bytes
3282 * to allow one less page from being charged, but we can't rely
3283 * on the pre-charged bytes not being changed outside of
3284 * consume_obj_stock() or refill_obj_stock(). So ignore those
3285 * pre-charged bytes as well when charging pages. To avoid a
3286 * page uncharge right after a page charge, we set the
3287 * allow_uncharge flag to false when calling refill_obj_stock()
3288 * to temporarily allow the pre-charged bytes to exceed the page
3289 * size limit. The maximum reachable value of the pre-charged
3290 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3293 nr_pages = size >> PAGE_SHIFT;
3294 nr_bytes = size & (PAGE_SIZE - 1);
3299 ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3300 if (!ret && nr_bytes)
3301 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes, false);
3306 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3308 refill_obj_stock(objcg, size, true);
3311 #endif /* CONFIG_MEMCG_KMEM */
3314 * Because page_memcg(head) is not set on tails, set it now.
3316 void split_page_memcg(struct page *head, unsigned int nr)
3318 struct mem_cgroup *memcg = page_memcg(head);
3321 if (mem_cgroup_disabled() || !memcg)
3324 for (i = 1; i < nr; i++)
3325 head[i].memcg_data = head->memcg_data;
3327 if (PageMemcgKmem(head))
3328 obj_cgroup_get_many(__page_objcg(head), nr - 1);
3330 css_get_many(&memcg->css, nr - 1);
3333 #ifdef CONFIG_MEMCG_SWAP
3335 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3336 * @entry: swap entry to be moved
3337 * @from: mem_cgroup which the entry is moved from
3338 * @to: mem_cgroup which the entry is moved to
3340 * It succeeds only when the swap_cgroup's record for this entry is the same
3341 * as the mem_cgroup's id of @from.
3343 * Returns 0 on success, -EINVAL on failure.
3345 * The caller must have charged to @to, IOW, called page_counter_charge() about
3346 * both res and memsw, and called css_get().
3348 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3349 struct mem_cgroup *from, struct mem_cgroup *to)
3351 unsigned short old_id, new_id;
3353 old_id = mem_cgroup_id(from);
3354 new_id = mem_cgroup_id(to);
3356 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3357 mod_memcg_state(from, MEMCG_SWAP, -1);
3358 mod_memcg_state(to, MEMCG_SWAP, 1);
3364 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3365 struct mem_cgroup *from, struct mem_cgroup *to)
3371 static DEFINE_MUTEX(memcg_max_mutex);
3373 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3374 unsigned long max, bool memsw)
3376 bool enlarge = false;
3377 bool drained = false;
3379 bool limits_invariant;
3380 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3383 if (signal_pending(current)) {
3388 mutex_lock(&memcg_max_mutex);
3390 * Make sure that the new limit (memsw or memory limit) doesn't
3391 * break our basic invariant rule memory.max <= memsw.max.
3393 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3394 max <= memcg->memsw.max;
3395 if (!limits_invariant) {
3396 mutex_unlock(&memcg_max_mutex);
3400 if (max > counter->max)
3402 ret = page_counter_set_max(counter, max);
3403 mutex_unlock(&memcg_max_mutex);
3409 drain_all_stock(memcg);
3414 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3415 GFP_KERNEL, !memsw)) {
3421 if (!ret && enlarge)
3422 memcg_oom_recover(memcg);
3427 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3429 unsigned long *total_scanned)
3431 unsigned long nr_reclaimed = 0;
3432 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3433 unsigned long reclaimed;
3435 struct mem_cgroup_tree_per_node *mctz;
3436 unsigned long excess;
3437 unsigned long nr_scanned;
3442 mctz = soft_limit_tree_node(pgdat->node_id);
3445 * Do not even bother to check the largest node if the root
3446 * is empty. Do it lockless to prevent lock bouncing. Races
3447 * are acceptable as soft limit is best effort anyway.
3449 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3453 * This loop can run a while, specially if mem_cgroup's continuously
3454 * keep exceeding their soft limit and putting the system under
3461 mz = mem_cgroup_largest_soft_limit_node(mctz);
3466 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3467 gfp_mask, &nr_scanned);
3468 nr_reclaimed += reclaimed;
3469 *total_scanned += nr_scanned;
3470 spin_lock_irq(&mctz->lock);
3471 __mem_cgroup_remove_exceeded(mz, mctz);
3474 * If we failed to reclaim anything from this memory cgroup
3475 * it is time to move on to the next cgroup
3479 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3481 excess = soft_limit_excess(mz->memcg);
3483 * One school of thought says that we should not add
3484 * back the node to the tree if reclaim returns 0.
3485 * But our reclaim could return 0, simply because due
3486 * to priority we are exposing a smaller subset of
3487 * memory to reclaim from. Consider this as a longer
3490 /* If excess == 0, no tree ops */
3491 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3492 spin_unlock_irq(&mctz->lock);
3493 css_put(&mz->memcg->css);
3496 * Could not reclaim anything and there are no more
3497 * mem cgroups to try or we seem to be looping without
3498 * reclaiming anything.
3500 if (!nr_reclaimed &&
3502 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3504 } while (!nr_reclaimed);
3506 css_put(&next_mz->memcg->css);
3507 return nr_reclaimed;
3511 * Reclaims as many pages from the given memcg as possible.
3513 * Caller is responsible for holding css reference for memcg.
3515 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3517 int nr_retries = MAX_RECLAIM_RETRIES;
3519 /* we call try-to-free pages for make this cgroup empty */
3520 lru_add_drain_all();
3522 drain_all_stock(memcg);
3524 /* try to free all pages in this cgroup */
3525 while (nr_retries && page_counter_read(&memcg->memory)) {
3528 if (signal_pending(current))
3531 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3535 /* maybe some writeback is necessary */
3536 congestion_wait(BLK_RW_ASYNC, HZ/10);
3544 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3545 char *buf, size_t nbytes,
3548 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3550 if (mem_cgroup_is_root(memcg))
3552 return mem_cgroup_force_empty(memcg) ?: nbytes;
3555 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3561 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3562 struct cftype *cft, u64 val)
3567 pr_warn_once("Non-hierarchical mode is deprecated. "
3568 "Please report your usecase to linux-mm@kvack.org if you "
3569 "depend on this functionality.\n");
3574 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3578 if (mem_cgroup_is_root(memcg)) {
3579 /* mem_cgroup_threshold() calls here from irqsafe context */
3580 cgroup_rstat_flush_irqsafe(memcg->css.cgroup);
3581 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3582 memcg_page_state(memcg, NR_ANON_MAPPED);
3584 val += memcg_page_state(memcg, MEMCG_SWAP);
3587 val = page_counter_read(&memcg->memory);
3589 val = page_counter_read(&memcg->memsw);
3602 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3605 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3606 struct page_counter *counter;
3608 switch (MEMFILE_TYPE(cft->private)) {
3610 counter = &memcg->memory;
3613 counter = &memcg->memsw;
3616 counter = &memcg->kmem;
3619 counter = &memcg->tcpmem;
3625 switch (MEMFILE_ATTR(cft->private)) {
3627 if (counter == &memcg->memory)
3628 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3629 if (counter == &memcg->memsw)
3630 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3631 return (u64)page_counter_read(counter) * PAGE_SIZE;
3633 return (u64)counter->max * PAGE_SIZE;
3635 return (u64)counter->watermark * PAGE_SIZE;
3637 return counter->failcnt;
3638 case RES_SOFT_LIMIT:
3639 return (u64)memcg->soft_limit * PAGE_SIZE;
3645 #ifdef CONFIG_MEMCG_KMEM
3646 static int memcg_online_kmem(struct mem_cgroup *memcg)
3648 struct obj_cgroup *objcg;
3651 if (cgroup_memory_nokmem)
3654 BUG_ON(memcg->kmemcg_id >= 0);
3655 BUG_ON(memcg->kmem_state);
3657 memcg_id = memcg_alloc_cache_id();
3661 objcg = obj_cgroup_alloc();
3663 memcg_free_cache_id(memcg_id);
3666 objcg->memcg = memcg;
3667 rcu_assign_pointer(memcg->objcg, objcg);
3669 static_branch_enable(&memcg_kmem_enabled_key);
3671 memcg->kmemcg_id = memcg_id;
3672 memcg->kmem_state = KMEM_ONLINE;
3677 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3679 struct cgroup_subsys_state *css;
3680 struct mem_cgroup *parent, *child;
3683 if (memcg->kmem_state != KMEM_ONLINE)
3686 memcg->kmem_state = KMEM_ALLOCATED;
3688 parent = parent_mem_cgroup(memcg);
3690 parent = root_mem_cgroup;
3692 memcg_reparent_objcgs(memcg, parent);
3694 kmemcg_id = memcg->kmemcg_id;
3695 BUG_ON(kmemcg_id < 0);
3698 * Change kmemcg_id of this cgroup and all its descendants to the
3699 * parent's id, and then move all entries from this cgroup's list_lrus
3700 * to ones of the parent. After we have finished, all list_lrus
3701 * corresponding to this cgroup are guaranteed to remain empty. The
3702 * ordering is imposed by list_lru_node->lock taken by
3703 * memcg_drain_all_list_lrus().
3705 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3706 css_for_each_descendant_pre(css, &memcg->css) {
3707 child = mem_cgroup_from_css(css);
3708 BUG_ON(child->kmemcg_id != kmemcg_id);
3709 child->kmemcg_id = parent->kmemcg_id;
3713 memcg_drain_all_list_lrus(kmemcg_id, parent);
3715 memcg_free_cache_id(kmemcg_id);
3718 static void memcg_free_kmem(struct mem_cgroup *memcg)
3720 /* css_alloc() failed, offlining didn't happen */
3721 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3722 memcg_offline_kmem(memcg);
3725 static int memcg_online_kmem(struct mem_cgroup *memcg)
3729 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3732 static void memcg_free_kmem(struct mem_cgroup *memcg)
3735 #endif /* CONFIG_MEMCG_KMEM */
3737 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3742 mutex_lock(&memcg_max_mutex);
3743 ret = page_counter_set_max(&memcg->kmem, max);
3744 mutex_unlock(&memcg_max_mutex);
3748 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3752 mutex_lock(&memcg_max_mutex);
3754 ret = page_counter_set_max(&memcg->tcpmem, max);
3758 if (!memcg->tcpmem_active) {
3760 * The active flag needs to be written after the static_key
3761 * update. This is what guarantees that the socket activation
3762 * function is the last one to run. See mem_cgroup_sk_alloc()
3763 * for details, and note that we don't mark any socket as
3764 * belonging to this memcg until that flag is up.
3766 * We need to do this, because static_keys will span multiple
3767 * sites, but we can't control their order. If we mark a socket
3768 * as accounted, but the accounting functions are not patched in
3769 * yet, we'll lose accounting.
3771 * We never race with the readers in mem_cgroup_sk_alloc(),
3772 * because when this value change, the code to process it is not
3775 static_branch_inc(&memcg_sockets_enabled_key);
3776 memcg->tcpmem_active = true;
3779 mutex_unlock(&memcg_max_mutex);
3784 * The user of this function is...
3787 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3788 char *buf, size_t nbytes, loff_t off)
3790 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3791 unsigned long nr_pages;
3794 buf = strstrip(buf);
3795 ret = page_counter_memparse(buf, "-1", &nr_pages);
3799 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3801 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3805 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3807 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3810 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3813 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3814 "Please report your usecase to linux-mm@kvack.org if you "
3815 "depend on this functionality.\n");
3816 ret = memcg_update_kmem_max(memcg, nr_pages);
3819 ret = memcg_update_tcp_max(memcg, nr_pages);
3823 case RES_SOFT_LIMIT:
3824 memcg->soft_limit = nr_pages;
3828 return ret ?: nbytes;
3831 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3832 size_t nbytes, loff_t off)
3834 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3835 struct page_counter *counter;
3837 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3839 counter = &memcg->memory;
3842 counter = &memcg->memsw;
3845 counter = &memcg->kmem;
3848 counter = &memcg->tcpmem;
3854 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3856 page_counter_reset_watermark(counter);
3859 counter->failcnt = 0;
3868 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3871 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3875 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3876 struct cftype *cft, u64 val)
3878 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3880 if (val & ~MOVE_MASK)
3884 * No kind of locking is needed in here, because ->can_attach() will
3885 * check this value once in the beginning of the process, and then carry
3886 * on with stale data. This means that changes to this value will only
3887 * affect task migrations starting after the change.
3889 memcg->move_charge_at_immigrate = val;
3893 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3894 struct cftype *cft, u64 val)
3902 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3903 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3904 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3906 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3907 int nid, unsigned int lru_mask, bool tree)
3909 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3910 unsigned long nr = 0;
3913 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3916 if (!(BIT(lru) & lru_mask))
3919 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3921 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3926 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3927 unsigned int lru_mask,
3930 unsigned long nr = 0;
3934 if (!(BIT(lru) & lru_mask))
3937 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3939 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3944 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3948 unsigned int lru_mask;
3951 static const struct numa_stat stats[] = {
3952 { "total", LRU_ALL },
3953 { "file", LRU_ALL_FILE },
3954 { "anon", LRU_ALL_ANON },
3955 { "unevictable", BIT(LRU_UNEVICTABLE) },
3957 const struct numa_stat *stat;
3959 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3961 cgroup_rstat_flush(memcg->css.cgroup);
3963 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3964 seq_printf(m, "%s=%lu", stat->name,
3965 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3967 for_each_node_state(nid, N_MEMORY)
3968 seq_printf(m, " N%d=%lu", nid,
3969 mem_cgroup_node_nr_lru_pages(memcg, nid,
3970 stat->lru_mask, false));
3974 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3976 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3977 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3979 for_each_node_state(nid, N_MEMORY)
3980 seq_printf(m, " N%d=%lu", nid,
3981 mem_cgroup_node_nr_lru_pages(memcg, nid,
3982 stat->lru_mask, true));
3988 #endif /* CONFIG_NUMA */
3990 static const unsigned int memcg1_stats[] = {
3993 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4003 static const char *const memcg1_stat_names[] = {
4006 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4016 /* Universal VM events cgroup1 shows, original sort order */
4017 static const unsigned int memcg1_events[] = {
4024 static int memcg_stat_show(struct seq_file *m, void *v)
4026 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4027 unsigned long memory, memsw;
4028 struct mem_cgroup *mi;
4031 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4033 cgroup_rstat_flush(memcg->css.cgroup);
4035 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4038 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4040 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4041 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4044 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4045 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4046 memcg_events_local(memcg, memcg1_events[i]));
4048 for (i = 0; i < NR_LRU_LISTS; i++)
4049 seq_printf(m, "%s %lu\n", lru_list_name(i),
4050 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4053 /* Hierarchical information */
4054 memory = memsw = PAGE_COUNTER_MAX;
4055 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4056 memory = min(memory, READ_ONCE(mi->memory.max));
4057 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4059 seq_printf(m, "hierarchical_memory_limit %llu\n",
4060 (u64)memory * PAGE_SIZE);
4061 if (do_memsw_account())
4062 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4063 (u64)memsw * PAGE_SIZE);
4065 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4068 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4070 nr = memcg_page_state(memcg, memcg1_stats[i]);
4071 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4072 (u64)nr * PAGE_SIZE);
4075 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4076 seq_printf(m, "total_%s %llu\n",
4077 vm_event_name(memcg1_events[i]),
4078 (u64)memcg_events(memcg, memcg1_events[i]));
4080 for (i = 0; i < NR_LRU_LISTS; i++)
4081 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4082 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4085 #ifdef CONFIG_DEBUG_VM
4088 struct mem_cgroup_per_node *mz;
4089 unsigned long anon_cost = 0;
4090 unsigned long file_cost = 0;
4092 for_each_online_pgdat(pgdat) {
4093 mz = memcg->nodeinfo[pgdat->node_id];
4095 anon_cost += mz->lruvec.anon_cost;
4096 file_cost += mz->lruvec.file_cost;
4098 seq_printf(m, "anon_cost %lu\n", anon_cost);
4099 seq_printf(m, "file_cost %lu\n", file_cost);
4106 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4109 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4111 return mem_cgroup_swappiness(memcg);
4114 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4115 struct cftype *cft, u64 val)
4117 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4122 if (!mem_cgroup_is_root(memcg))
4123 memcg->swappiness = val;
4125 vm_swappiness = val;
4130 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4132 struct mem_cgroup_threshold_ary *t;
4133 unsigned long usage;
4138 t = rcu_dereference(memcg->thresholds.primary);
4140 t = rcu_dereference(memcg->memsw_thresholds.primary);
4145 usage = mem_cgroup_usage(memcg, swap);
4148 * current_threshold points to threshold just below or equal to usage.
4149 * If it's not true, a threshold was crossed after last
4150 * call of __mem_cgroup_threshold().
4152 i = t->current_threshold;
4155 * Iterate backward over array of thresholds starting from
4156 * current_threshold and check if a threshold is crossed.
4157 * If none of thresholds below usage is crossed, we read
4158 * only one element of the array here.
4160 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4161 eventfd_signal(t->entries[i].eventfd, 1);
4163 /* i = current_threshold + 1 */
4167 * Iterate forward over array of thresholds starting from
4168 * current_threshold+1 and check if a threshold is crossed.
4169 * If none of thresholds above usage is crossed, we read
4170 * only one element of the array here.
4172 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4173 eventfd_signal(t->entries[i].eventfd, 1);
4175 /* Update current_threshold */
4176 t->current_threshold = i - 1;
4181 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4184 __mem_cgroup_threshold(memcg, false);
4185 if (do_memsw_account())
4186 __mem_cgroup_threshold(memcg, true);
4188 memcg = parent_mem_cgroup(memcg);
4192 static int compare_thresholds(const void *a, const void *b)
4194 const struct mem_cgroup_threshold *_a = a;
4195 const struct mem_cgroup_threshold *_b = b;
4197 if (_a->threshold > _b->threshold)
4200 if (_a->threshold < _b->threshold)
4206 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4208 struct mem_cgroup_eventfd_list *ev;
4210 spin_lock(&memcg_oom_lock);
4212 list_for_each_entry(ev, &memcg->oom_notify, list)
4213 eventfd_signal(ev->eventfd, 1);
4215 spin_unlock(&memcg_oom_lock);
4219 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4221 struct mem_cgroup *iter;
4223 for_each_mem_cgroup_tree(iter, memcg)
4224 mem_cgroup_oom_notify_cb(iter);
4227 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4228 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4230 struct mem_cgroup_thresholds *thresholds;
4231 struct mem_cgroup_threshold_ary *new;
4232 unsigned long threshold;
4233 unsigned long usage;
4236 ret = page_counter_memparse(args, "-1", &threshold);
4240 mutex_lock(&memcg->thresholds_lock);
4243 thresholds = &memcg->thresholds;
4244 usage = mem_cgroup_usage(memcg, false);
4245 } else if (type == _MEMSWAP) {
4246 thresholds = &memcg->memsw_thresholds;
4247 usage = mem_cgroup_usage(memcg, true);
4251 /* Check if a threshold crossed before adding a new one */
4252 if (thresholds->primary)
4253 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4255 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4257 /* Allocate memory for new array of thresholds */
4258 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4265 /* Copy thresholds (if any) to new array */
4266 if (thresholds->primary)
4267 memcpy(new->entries, thresholds->primary->entries,
4268 flex_array_size(new, entries, size - 1));
4270 /* Add new threshold */
4271 new->entries[size - 1].eventfd = eventfd;
4272 new->entries[size - 1].threshold = threshold;
4274 /* Sort thresholds. Registering of new threshold isn't time-critical */
4275 sort(new->entries, size, sizeof(*new->entries),
4276 compare_thresholds, NULL);
4278 /* Find current threshold */
4279 new->current_threshold = -1;
4280 for (i = 0; i < size; i++) {
4281 if (new->entries[i].threshold <= usage) {
4283 * new->current_threshold will not be used until
4284 * rcu_assign_pointer(), so it's safe to increment
4287 ++new->current_threshold;
4292 /* Free old spare buffer and save old primary buffer as spare */
4293 kfree(thresholds->spare);
4294 thresholds->spare = thresholds->primary;
4296 rcu_assign_pointer(thresholds->primary, new);
4298 /* To be sure that nobody uses thresholds */
4302 mutex_unlock(&memcg->thresholds_lock);
4307 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4308 struct eventfd_ctx *eventfd, const char *args)
4310 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4313 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4314 struct eventfd_ctx *eventfd, const char *args)
4316 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4319 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4320 struct eventfd_ctx *eventfd, enum res_type type)
4322 struct mem_cgroup_thresholds *thresholds;
4323 struct mem_cgroup_threshold_ary *new;
4324 unsigned long usage;
4325 int i, j, size, entries;
4327 mutex_lock(&memcg->thresholds_lock);
4330 thresholds = &memcg->thresholds;
4331 usage = mem_cgroup_usage(memcg, false);
4332 } else if (type == _MEMSWAP) {
4333 thresholds = &memcg->memsw_thresholds;
4334 usage = mem_cgroup_usage(memcg, true);
4338 if (!thresholds->primary)
4341 /* Check if a threshold crossed before removing */
4342 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4344 /* Calculate new number of threshold */
4346 for (i = 0; i < thresholds->primary->size; i++) {
4347 if (thresholds->primary->entries[i].eventfd != eventfd)
4353 new = thresholds->spare;
4355 /* If no items related to eventfd have been cleared, nothing to do */
4359 /* Set thresholds array to NULL if we don't have thresholds */
4368 /* Copy thresholds and find current threshold */
4369 new->current_threshold = -1;
4370 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4371 if (thresholds->primary->entries[i].eventfd == eventfd)
4374 new->entries[j] = thresholds->primary->entries[i];
4375 if (new->entries[j].threshold <= usage) {
4377 * new->current_threshold will not be used
4378 * until rcu_assign_pointer(), so it's safe to increment
4381 ++new->current_threshold;
4387 /* Swap primary and spare array */
4388 thresholds->spare = thresholds->primary;
4390 rcu_assign_pointer(thresholds->primary, new);
4392 /* To be sure that nobody uses thresholds */
4395 /* If all events are unregistered, free the spare array */
4397 kfree(thresholds->spare);
4398 thresholds->spare = NULL;
4401 mutex_unlock(&memcg->thresholds_lock);
4404 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4405 struct eventfd_ctx *eventfd)
4407 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4410 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4411 struct eventfd_ctx *eventfd)
4413 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4416 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4417 struct eventfd_ctx *eventfd, const char *args)
4419 struct mem_cgroup_eventfd_list *event;
4421 event = kmalloc(sizeof(*event), GFP_KERNEL);
4425 spin_lock(&memcg_oom_lock);
4427 event->eventfd = eventfd;
4428 list_add(&event->list, &memcg->oom_notify);
4430 /* already in OOM ? */
4431 if (memcg->under_oom)
4432 eventfd_signal(eventfd, 1);
4433 spin_unlock(&memcg_oom_lock);
4438 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4439 struct eventfd_ctx *eventfd)
4441 struct mem_cgroup_eventfd_list *ev, *tmp;
4443 spin_lock(&memcg_oom_lock);
4445 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4446 if (ev->eventfd == eventfd) {
4447 list_del(&ev->list);
4452 spin_unlock(&memcg_oom_lock);
4455 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4457 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4459 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4460 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4461 seq_printf(sf, "oom_kill %lu\n",
4462 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4466 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4467 struct cftype *cft, u64 val)
4469 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4471 /* cannot set to root cgroup and only 0 and 1 are allowed */
4472 if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4475 memcg->oom_kill_disable = val;
4477 memcg_oom_recover(memcg);
4482 #ifdef CONFIG_CGROUP_WRITEBACK
4484 #include <trace/events/writeback.h>
4486 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4488 return wb_domain_init(&memcg->cgwb_domain, gfp);
4491 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4493 wb_domain_exit(&memcg->cgwb_domain);
4496 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4498 wb_domain_size_changed(&memcg->cgwb_domain);
4501 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4503 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4505 if (!memcg->css.parent)
4508 return &memcg->cgwb_domain;
4512 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4513 * @wb: bdi_writeback in question
4514 * @pfilepages: out parameter for number of file pages
4515 * @pheadroom: out parameter for number of allocatable pages according to memcg
4516 * @pdirty: out parameter for number of dirty pages
4517 * @pwriteback: out parameter for number of pages under writeback
4519 * Determine the numbers of file, headroom, dirty, and writeback pages in
4520 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4521 * is a bit more involved.
4523 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4524 * headroom is calculated as the lowest headroom of itself and the
4525 * ancestors. Note that this doesn't consider the actual amount of
4526 * available memory in the system. The caller should further cap
4527 * *@pheadroom accordingly.
4529 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4530 unsigned long *pheadroom, unsigned long *pdirty,
4531 unsigned long *pwriteback)
4533 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4534 struct mem_cgroup *parent;
4536 cgroup_rstat_flush_irqsafe(memcg->css.cgroup);
4538 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4539 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4540 *pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4541 memcg_page_state(memcg, NR_ACTIVE_FILE);
4543 *pheadroom = PAGE_COUNTER_MAX;
4544 while ((parent = parent_mem_cgroup(memcg))) {
4545 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4546 READ_ONCE(memcg->memory.high));
4547 unsigned long used = page_counter_read(&memcg->memory);
4549 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4555 * Foreign dirty flushing
4557 * There's an inherent mismatch between memcg and writeback. The former
4558 * tracks ownership per-page while the latter per-inode. This was a
4559 * deliberate design decision because honoring per-page ownership in the
4560 * writeback path is complicated, may lead to higher CPU and IO overheads
4561 * and deemed unnecessary given that write-sharing an inode across
4562 * different cgroups isn't a common use-case.
4564 * Combined with inode majority-writer ownership switching, this works well
4565 * enough in most cases but there are some pathological cases. For
4566 * example, let's say there are two cgroups A and B which keep writing to
4567 * different but confined parts of the same inode. B owns the inode and
4568 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4569 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4570 * triggering background writeback. A will be slowed down without a way to
4571 * make writeback of the dirty pages happen.
4573 * Conditions like the above can lead to a cgroup getting repeatedly and
4574 * severely throttled after making some progress after each
4575 * dirty_expire_interval while the underlying IO device is almost
4578 * Solving this problem completely requires matching the ownership tracking
4579 * granularities between memcg and writeback in either direction. However,
4580 * the more egregious behaviors can be avoided by simply remembering the
4581 * most recent foreign dirtying events and initiating remote flushes on
4582 * them when local writeback isn't enough to keep the memory clean enough.
4584 * The following two functions implement such mechanism. When a foreign
4585 * page - a page whose memcg and writeback ownerships don't match - is
4586 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4587 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4588 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4589 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4590 * foreign bdi_writebacks which haven't expired. Both the numbers of
4591 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4592 * limited to MEMCG_CGWB_FRN_CNT.
4594 * The mechanism only remembers IDs and doesn't hold any object references.
4595 * As being wrong occasionally doesn't matter, updates and accesses to the
4596 * records are lockless and racy.
4598 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4599 struct bdi_writeback *wb)
4601 struct mem_cgroup *memcg = page_memcg(page);
4602 struct memcg_cgwb_frn *frn;
4603 u64 now = get_jiffies_64();
4604 u64 oldest_at = now;
4608 trace_track_foreign_dirty(page, wb);
4611 * Pick the slot to use. If there is already a slot for @wb, keep
4612 * using it. If not replace the oldest one which isn't being
4615 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4616 frn = &memcg->cgwb_frn[i];
4617 if (frn->bdi_id == wb->bdi->id &&
4618 frn->memcg_id == wb->memcg_css->id)
4620 if (time_before64(frn->at, oldest_at) &&
4621 atomic_read(&frn->done.cnt) == 1) {
4623 oldest_at = frn->at;
4627 if (i < MEMCG_CGWB_FRN_CNT) {
4629 * Re-using an existing one. Update timestamp lazily to
4630 * avoid making the cacheline hot. We want them to be
4631 * reasonably up-to-date and significantly shorter than
4632 * dirty_expire_interval as that's what expires the record.
4633 * Use the shorter of 1s and dirty_expire_interval / 8.
4635 unsigned long update_intv =
4636 min_t(unsigned long, HZ,
4637 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4639 if (time_before64(frn->at, now - update_intv))
4641 } else if (oldest >= 0) {
4642 /* replace the oldest free one */
4643 frn = &memcg->cgwb_frn[oldest];
4644 frn->bdi_id = wb->bdi->id;
4645 frn->memcg_id = wb->memcg_css->id;
4650 /* issue foreign writeback flushes for recorded foreign dirtying events */
4651 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4653 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4654 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4655 u64 now = jiffies_64;
4658 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4659 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4662 * If the record is older than dirty_expire_interval,
4663 * writeback on it has already started. No need to kick it
4664 * off again. Also, don't start a new one if there's
4665 * already one in flight.
4667 if (time_after64(frn->at, now - intv) &&
4668 atomic_read(&frn->done.cnt) == 1) {
4670 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4671 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4672 WB_REASON_FOREIGN_FLUSH,
4678 #else /* CONFIG_CGROUP_WRITEBACK */
4680 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4685 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4689 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4693 #endif /* CONFIG_CGROUP_WRITEBACK */
4696 * DO NOT USE IN NEW FILES.
4698 * "cgroup.event_control" implementation.
4700 * This is way over-engineered. It tries to support fully configurable
4701 * events for each user. Such level of flexibility is completely
4702 * unnecessary especially in the light of the planned unified hierarchy.
4704 * Please deprecate this and replace with something simpler if at all
4709 * Unregister event and free resources.
4711 * Gets called from workqueue.
4713 static void memcg_event_remove(struct work_struct *work)
4715 struct mem_cgroup_event *event =
4716 container_of(work, struct mem_cgroup_event, remove);
4717 struct mem_cgroup *memcg = event->memcg;
4719 remove_wait_queue(event->wqh, &event->wait);
4721 event->unregister_event(memcg, event->eventfd);
4723 /* Notify userspace the event is going away. */
4724 eventfd_signal(event->eventfd, 1);
4726 eventfd_ctx_put(event->eventfd);
4728 css_put(&memcg->css);
4732 * Gets called on EPOLLHUP on eventfd when user closes it.
4734 * Called with wqh->lock held and interrupts disabled.
4736 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4737 int sync, void *key)
4739 struct mem_cgroup_event *event =
4740 container_of(wait, struct mem_cgroup_event, wait);
4741 struct mem_cgroup *memcg = event->memcg;
4742 __poll_t flags = key_to_poll(key);
4744 if (flags & EPOLLHUP) {
4746 * If the event has been detached at cgroup removal, we
4747 * can simply return knowing the other side will cleanup
4750 * We can't race against event freeing since the other
4751 * side will require wqh->lock via remove_wait_queue(),
4754 spin_lock(&memcg->event_list_lock);
4755 if (!list_empty(&event->list)) {
4756 list_del_init(&event->list);
4758 * We are in atomic context, but cgroup_event_remove()
4759 * may sleep, so we have to call it in workqueue.
4761 schedule_work(&event->remove);
4763 spin_unlock(&memcg->event_list_lock);
4769 static void memcg_event_ptable_queue_proc(struct file *file,
4770 wait_queue_head_t *wqh, poll_table *pt)
4772 struct mem_cgroup_event *event =
4773 container_of(pt, struct mem_cgroup_event, pt);
4776 add_wait_queue(wqh, &event->wait);
4780 * DO NOT USE IN NEW FILES.
4782 * Parse input and register new cgroup event handler.
4784 * Input must be in format '<event_fd> <control_fd> <args>'.
4785 * Interpretation of args is defined by control file implementation.
4787 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4788 char *buf, size_t nbytes, loff_t off)
4790 struct cgroup_subsys_state *css = of_css(of);
4791 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4792 struct mem_cgroup_event *event;
4793 struct cgroup_subsys_state *cfile_css;
4794 unsigned int efd, cfd;
4801 buf = strstrip(buf);
4803 efd = simple_strtoul(buf, &endp, 10);
4808 cfd = simple_strtoul(buf, &endp, 10);
4809 if ((*endp != ' ') && (*endp != '\0'))
4813 event = kzalloc(sizeof(*event), GFP_KERNEL);
4817 event->memcg = memcg;
4818 INIT_LIST_HEAD(&event->list);
4819 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4820 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4821 INIT_WORK(&event->remove, memcg_event_remove);
4829 event->eventfd = eventfd_ctx_fileget(efile.file);
4830 if (IS_ERR(event->eventfd)) {
4831 ret = PTR_ERR(event->eventfd);
4838 goto out_put_eventfd;
4841 /* the process need read permission on control file */
4842 /* AV: shouldn't we check that it's been opened for read instead? */
4843 ret = file_permission(cfile.file, MAY_READ);
4848 * Determine the event callbacks and set them in @event. This used
4849 * to be done via struct cftype but cgroup core no longer knows
4850 * about these events. The following is crude but the whole thing
4851 * is for compatibility anyway.
4853 * DO NOT ADD NEW FILES.
4855 name = cfile.file->f_path.dentry->d_name.name;
4857 if (!strcmp(name, "memory.usage_in_bytes")) {
4858 event->register_event = mem_cgroup_usage_register_event;
4859 event->unregister_event = mem_cgroup_usage_unregister_event;
4860 } else if (!strcmp(name, "memory.oom_control")) {
4861 event->register_event = mem_cgroup_oom_register_event;
4862 event->unregister_event = mem_cgroup_oom_unregister_event;
4863 } else if (!strcmp(name, "memory.pressure_level")) {
4864 event->register_event = vmpressure_register_event;
4865 event->unregister_event = vmpressure_unregister_event;
4866 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4867 event->register_event = memsw_cgroup_usage_register_event;
4868 event->unregister_event = memsw_cgroup_usage_unregister_event;
4875 * Verify @cfile should belong to @css. Also, remaining events are
4876 * automatically removed on cgroup destruction but the removal is
4877 * asynchronous, so take an extra ref on @css.
4879 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4880 &memory_cgrp_subsys);
4882 if (IS_ERR(cfile_css))
4884 if (cfile_css != css) {
4889 ret = event->register_event(memcg, event->eventfd, buf);
4893 vfs_poll(efile.file, &event->pt);
4895 spin_lock(&memcg->event_list_lock);
4896 list_add(&event->list, &memcg->event_list);
4897 spin_unlock(&memcg->event_list_lock);
4909 eventfd_ctx_put(event->eventfd);
4918 static struct cftype mem_cgroup_legacy_files[] = {
4920 .name = "usage_in_bytes",
4921 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4922 .read_u64 = mem_cgroup_read_u64,
4925 .name = "max_usage_in_bytes",
4926 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4927 .write = mem_cgroup_reset,
4928 .read_u64 = mem_cgroup_read_u64,
4931 .name = "limit_in_bytes",
4932 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4933 .write = mem_cgroup_write,
4934 .read_u64 = mem_cgroup_read_u64,
4937 .name = "soft_limit_in_bytes",
4938 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4939 .write = mem_cgroup_write,
4940 .read_u64 = mem_cgroup_read_u64,
4944 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4945 .write = mem_cgroup_reset,
4946 .read_u64 = mem_cgroup_read_u64,
4950 .seq_show = memcg_stat_show,
4953 .name = "force_empty",
4954 .write = mem_cgroup_force_empty_write,
4957 .name = "use_hierarchy",
4958 .write_u64 = mem_cgroup_hierarchy_write,
4959 .read_u64 = mem_cgroup_hierarchy_read,
4962 .name = "cgroup.event_control", /* XXX: for compat */
4963 .write = memcg_write_event_control,
4964 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4967 .name = "swappiness",
4968 .read_u64 = mem_cgroup_swappiness_read,
4969 .write_u64 = mem_cgroup_swappiness_write,
4972 .name = "move_charge_at_immigrate",
4973 .read_u64 = mem_cgroup_move_charge_read,
4974 .write_u64 = mem_cgroup_move_charge_write,
4977 .name = "oom_control",
4978 .seq_show = mem_cgroup_oom_control_read,
4979 .write_u64 = mem_cgroup_oom_control_write,
4980 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4983 .name = "pressure_level",
4987 .name = "numa_stat",
4988 .seq_show = memcg_numa_stat_show,
4992 .name = "kmem.limit_in_bytes",
4993 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4994 .write = mem_cgroup_write,
4995 .read_u64 = mem_cgroup_read_u64,
4998 .name = "kmem.usage_in_bytes",
4999 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5000 .read_u64 = mem_cgroup_read_u64,
5003 .name = "kmem.failcnt",
5004 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5005 .write = mem_cgroup_reset,
5006 .read_u64 = mem_cgroup_read_u64,
5009 .name = "kmem.max_usage_in_bytes",
5010 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5011 .write = mem_cgroup_reset,
5012 .read_u64 = mem_cgroup_read_u64,
5014 #if defined(CONFIG_MEMCG_KMEM) && \
5015 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5017 .name = "kmem.slabinfo",
5018 .seq_show = memcg_slab_show,
5022 .name = "kmem.tcp.limit_in_bytes",
5023 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5024 .write = mem_cgroup_write,
5025 .read_u64 = mem_cgroup_read_u64,
5028 .name = "kmem.tcp.usage_in_bytes",
5029 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5030 .read_u64 = mem_cgroup_read_u64,
5033 .name = "kmem.tcp.failcnt",
5034 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5035 .write = mem_cgroup_reset,
5036 .read_u64 = mem_cgroup_read_u64,
5039 .name = "kmem.tcp.max_usage_in_bytes",
5040 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5041 .write = mem_cgroup_reset,
5042 .read_u64 = mem_cgroup_read_u64,
5044 { }, /* terminate */
5048 * Private memory cgroup IDR
5050 * Swap-out records and page cache shadow entries need to store memcg
5051 * references in constrained space, so we maintain an ID space that is
5052 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5053 * memory-controlled cgroups to 64k.
5055 * However, there usually are many references to the offline CSS after
5056 * the cgroup has been destroyed, such as page cache or reclaimable
5057 * slab objects, that don't need to hang on to the ID. We want to keep
5058 * those dead CSS from occupying IDs, or we might quickly exhaust the
5059 * relatively small ID space and prevent the creation of new cgroups
5060 * even when there are much fewer than 64k cgroups - possibly none.
5062 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5063 * be freed and recycled when it's no longer needed, which is usually
5064 * when the CSS is offlined.
5066 * The only exception to that are records of swapped out tmpfs/shmem
5067 * pages that need to be attributed to live ancestors on swapin. But
5068 * those references are manageable from userspace.
5071 static DEFINE_IDR(mem_cgroup_idr);
5073 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5075 if (memcg->id.id > 0) {
5076 idr_remove(&mem_cgroup_idr, memcg->id.id);
5081 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5084 refcount_add(n, &memcg->id.ref);
5087 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5089 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5090 mem_cgroup_id_remove(memcg);
5092 /* Memcg ID pins CSS */
5093 css_put(&memcg->css);
5097 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5099 mem_cgroup_id_put_many(memcg, 1);
5103 * mem_cgroup_from_id - look up a memcg from a memcg id
5104 * @id: the memcg id to look up
5106 * Caller must hold rcu_read_lock().
5108 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5110 WARN_ON_ONCE(!rcu_read_lock_held());
5111 return idr_find(&mem_cgroup_idr, id);
5114 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5116 struct mem_cgroup_per_node *pn;
5119 * This routine is called against possible nodes.
5120 * But it's BUG to call kmalloc() against offline node.
5122 * TODO: this routine can waste much memory for nodes which will
5123 * never be onlined. It's better to use memory hotplug callback
5126 if (!node_state(node, N_NORMAL_MEMORY))
5128 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5132 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5133 GFP_KERNEL_ACCOUNT);
5134 if (!pn->lruvec_stat_local) {
5139 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct batched_lruvec_stat,
5140 GFP_KERNEL_ACCOUNT);
5141 if (!pn->lruvec_stat_cpu) {
5142 free_percpu(pn->lruvec_stat_local);
5147 lruvec_init(&pn->lruvec);
5148 pn->usage_in_excess = 0;
5149 pn->on_tree = false;
5152 memcg->nodeinfo[node] = pn;
5156 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5158 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5163 free_percpu(pn->lruvec_stat_cpu);
5164 free_percpu(pn->lruvec_stat_local);
5168 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5173 free_mem_cgroup_per_node_info(memcg, node);
5174 free_percpu(memcg->vmstats_percpu);
5178 static void mem_cgroup_free(struct mem_cgroup *memcg)
5182 memcg_wb_domain_exit(memcg);
5184 * Flush percpu lruvec stats to guarantee the value
5185 * correctness on parent's and all ancestor levels.
5187 for_each_online_cpu(cpu)
5188 memcg_flush_lruvec_page_state(memcg, cpu);
5189 __mem_cgroup_free(memcg);
5192 static struct mem_cgroup *mem_cgroup_alloc(void)
5194 struct mem_cgroup *memcg;
5197 int __maybe_unused i;
5198 long error = -ENOMEM;
5200 size = sizeof(struct mem_cgroup);
5201 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5203 memcg = kzalloc(size, GFP_KERNEL);
5205 return ERR_PTR(error);
5207 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5208 1, MEM_CGROUP_ID_MAX,
5210 if (memcg->id.id < 0) {
5211 error = memcg->id.id;
5215 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5216 GFP_KERNEL_ACCOUNT);
5217 if (!memcg->vmstats_percpu)
5221 if (alloc_mem_cgroup_per_node_info(memcg, node))
5224 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5227 INIT_WORK(&memcg->high_work, high_work_func);
5228 INIT_LIST_HEAD(&memcg->oom_notify);
5229 mutex_init(&memcg->thresholds_lock);
5230 spin_lock_init(&memcg->move_lock);
5231 vmpressure_init(&memcg->vmpressure);
5232 INIT_LIST_HEAD(&memcg->event_list);
5233 spin_lock_init(&memcg->event_list_lock);
5234 memcg->socket_pressure = jiffies;
5235 #ifdef CONFIG_MEMCG_KMEM
5236 memcg->kmemcg_id = -1;
5237 INIT_LIST_HEAD(&memcg->objcg_list);
5239 #ifdef CONFIG_CGROUP_WRITEBACK
5240 INIT_LIST_HEAD(&memcg->cgwb_list);
5241 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5242 memcg->cgwb_frn[i].done =
5243 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5245 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5246 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5247 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5248 memcg->deferred_split_queue.split_queue_len = 0;
5250 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5253 mem_cgroup_id_remove(memcg);
5254 __mem_cgroup_free(memcg);
5255 return ERR_PTR(error);
5258 static struct cgroup_subsys_state * __ref
5259 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5261 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5262 struct mem_cgroup *memcg, *old_memcg;
5263 long error = -ENOMEM;
5265 old_memcg = set_active_memcg(parent);
5266 memcg = mem_cgroup_alloc();
5267 set_active_memcg(old_memcg);
5269 return ERR_CAST(memcg);
5271 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5272 memcg->soft_limit = PAGE_COUNTER_MAX;
5273 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5275 memcg->swappiness = mem_cgroup_swappiness(parent);
5276 memcg->oom_kill_disable = parent->oom_kill_disable;
5278 page_counter_init(&memcg->memory, &parent->memory);
5279 page_counter_init(&memcg->swap, &parent->swap);
5280 page_counter_init(&memcg->kmem, &parent->kmem);
5281 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5283 page_counter_init(&memcg->memory, NULL);
5284 page_counter_init(&memcg->swap, NULL);
5285 page_counter_init(&memcg->kmem, NULL);
5286 page_counter_init(&memcg->tcpmem, NULL);
5288 root_mem_cgroup = memcg;
5292 /* The following stuff does not apply to the root */
5293 error = memcg_online_kmem(memcg);
5297 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5298 static_branch_inc(&memcg_sockets_enabled_key);
5302 mem_cgroup_id_remove(memcg);
5303 mem_cgroup_free(memcg);
5304 return ERR_PTR(error);
5307 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5309 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5312 * A memcg must be visible for expand_shrinker_info()
5313 * by the time the maps are allocated. So, we allocate maps
5314 * here, when for_each_mem_cgroup() can't skip it.
5316 if (alloc_shrinker_info(memcg)) {
5317 mem_cgroup_id_remove(memcg);
5321 /* Online state pins memcg ID, memcg ID pins CSS */
5322 refcount_set(&memcg->id.ref, 1);
5327 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5329 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5330 struct mem_cgroup_event *event, *tmp;
5333 * Unregister events and notify userspace.
5334 * Notify userspace about cgroup removing only after rmdir of cgroup
5335 * directory to avoid race between userspace and kernelspace.
5337 spin_lock(&memcg->event_list_lock);
5338 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5339 list_del_init(&event->list);
5340 schedule_work(&event->remove);
5342 spin_unlock(&memcg->event_list_lock);
5344 page_counter_set_min(&memcg->memory, 0);
5345 page_counter_set_low(&memcg->memory, 0);
5347 memcg_offline_kmem(memcg);
5348 reparent_shrinker_deferred(memcg);
5349 wb_memcg_offline(memcg);
5351 drain_all_stock(memcg);
5353 mem_cgroup_id_put(memcg);
5356 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5358 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5360 invalidate_reclaim_iterators(memcg);
5363 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5365 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5366 int __maybe_unused i;
5368 #ifdef CONFIG_CGROUP_WRITEBACK
5369 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5370 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5372 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5373 static_branch_dec(&memcg_sockets_enabled_key);
5375 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5376 static_branch_dec(&memcg_sockets_enabled_key);
5378 vmpressure_cleanup(&memcg->vmpressure);
5379 cancel_work_sync(&memcg->high_work);
5380 mem_cgroup_remove_from_trees(memcg);
5381 free_shrinker_info(memcg);
5382 memcg_free_kmem(memcg);
5383 mem_cgroup_free(memcg);
5387 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5388 * @css: the target css
5390 * Reset the states of the mem_cgroup associated with @css. This is
5391 * invoked when the userland requests disabling on the default hierarchy
5392 * but the memcg is pinned through dependency. The memcg should stop
5393 * applying policies and should revert to the vanilla state as it may be
5394 * made visible again.
5396 * The current implementation only resets the essential configurations.
5397 * This needs to be expanded to cover all the visible parts.
5399 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5401 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5403 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5404 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5405 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5406 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5407 page_counter_set_min(&memcg->memory, 0);
5408 page_counter_set_low(&memcg->memory, 0);
5409 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5410 memcg->soft_limit = PAGE_COUNTER_MAX;
5411 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5412 memcg_wb_domain_size_changed(memcg);
5415 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5417 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5418 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5419 struct memcg_vmstats_percpu *statc;
5423 statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5425 for (i = 0; i < MEMCG_NR_STAT; i++) {
5427 * Collect the aggregated propagation counts of groups
5428 * below us. We're in a per-cpu loop here and this is
5429 * a global counter, so the first cycle will get them.
5431 delta = memcg->vmstats.state_pending[i];
5433 memcg->vmstats.state_pending[i] = 0;
5435 /* Add CPU changes on this level since the last flush */
5436 v = READ_ONCE(statc->state[i]);
5437 if (v != statc->state_prev[i]) {
5438 delta += v - statc->state_prev[i];
5439 statc->state_prev[i] = v;
5445 /* Aggregate counts on this level and propagate upwards */
5446 memcg->vmstats.state[i] += delta;
5448 parent->vmstats.state_pending[i] += delta;
5451 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
5452 delta = memcg->vmstats.events_pending[i];
5454 memcg->vmstats.events_pending[i] = 0;
5456 v = READ_ONCE(statc->events[i]);
5457 if (v != statc->events_prev[i]) {
5458 delta += v - statc->events_prev[i];
5459 statc->events_prev[i] = v;
5465 memcg->vmstats.events[i] += delta;
5467 parent->vmstats.events_pending[i] += delta;
5472 /* Handlers for move charge at task migration. */
5473 static int mem_cgroup_do_precharge(unsigned long count)
5477 /* Try a single bulk charge without reclaim first, kswapd may wake */
5478 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5480 mc.precharge += count;
5484 /* Try charges one by one with reclaim, but do not retry */
5486 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5500 enum mc_target_type {
5507 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5508 unsigned long addr, pte_t ptent)
5510 struct page *page = vm_normal_page(vma, addr, ptent);
5512 if (!page || !page_mapped(page))
5514 if (PageAnon(page)) {
5515 if (!(mc.flags & MOVE_ANON))
5518 if (!(mc.flags & MOVE_FILE))
5521 if (!get_page_unless_zero(page))
5527 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5528 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5529 pte_t ptent, swp_entry_t *entry)
5531 struct page *page = NULL;
5532 swp_entry_t ent = pte_to_swp_entry(ptent);
5534 if (!(mc.flags & MOVE_ANON))
5538 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5539 * a device and because they are not accessible by CPU they are store
5540 * as special swap entry in the CPU page table.
5542 if (is_device_private_entry(ent)) {
5543 page = pfn_swap_entry_to_page(ent);
5545 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5546 * a refcount of 1 when free (unlike normal page)
5548 if (!page_ref_add_unless(page, 1, 1))
5553 if (non_swap_entry(ent))
5557 * Because lookup_swap_cache() updates some statistics counter,
5558 * we call find_get_page() with swapper_space directly.
5560 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5561 entry->val = ent.val;
5566 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5567 pte_t ptent, swp_entry_t *entry)
5573 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5574 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5576 if (!vma->vm_file) /* anonymous vma */
5578 if (!(mc.flags & MOVE_FILE))
5581 /* page is moved even if it's not RSS of this task(page-faulted). */
5582 /* shmem/tmpfs may report page out on swap: account for that too. */
5583 return find_get_incore_page(vma->vm_file->f_mapping,
5584 linear_page_index(vma, addr));
5588 * mem_cgroup_move_account - move account of the page
5590 * @compound: charge the page as compound or small page
5591 * @from: mem_cgroup which the page is moved from.
5592 * @to: mem_cgroup which the page is moved to. @from != @to.
5594 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5596 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5599 static int mem_cgroup_move_account(struct page *page,
5601 struct mem_cgroup *from,
5602 struct mem_cgroup *to)
5604 struct lruvec *from_vec, *to_vec;
5605 struct pglist_data *pgdat;
5606 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5609 VM_BUG_ON(from == to);
5610 VM_BUG_ON_PAGE(PageLRU(page), page);
5611 VM_BUG_ON(compound && !PageTransHuge(page));
5614 * Prevent mem_cgroup_migrate() from looking at
5615 * page's memory cgroup of its source page while we change it.
5618 if (!trylock_page(page))
5622 if (page_memcg(page) != from)
5625 pgdat = page_pgdat(page);
5626 from_vec = mem_cgroup_lruvec(from, pgdat);
5627 to_vec = mem_cgroup_lruvec(to, pgdat);
5629 lock_page_memcg(page);
5631 if (PageAnon(page)) {
5632 if (page_mapped(page)) {
5633 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5634 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5635 if (PageTransHuge(page)) {
5636 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5638 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5643 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5644 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5646 if (PageSwapBacked(page)) {
5647 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5648 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5651 if (page_mapped(page)) {
5652 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5653 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5656 if (PageDirty(page)) {
5657 struct address_space *mapping = page_mapping(page);
5659 if (mapping_can_writeback(mapping)) {
5660 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5662 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5668 if (PageWriteback(page)) {
5669 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5670 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5674 * All state has been migrated, let's switch to the new memcg.
5676 * It is safe to change page's memcg here because the page
5677 * is referenced, charged, isolated, and locked: we can't race
5678 * with (un)charging, migration, LRU putback, or anything else
5679 * that would rely on a stable page's memory cgroup.
5681 * Note that lock_page_memcg is a memcg lock, not a page lock,
5682 * to save space. As soon as we switch page's memory cgroup to a
5683 * new memcg that isn't locked, the above state can change
5684 * concurrently again. Make sure we're truly done with it.
5689 css_put(&from->css);
5691 page->memcg_data = (unsigned long)to;
5693 __unlock_page_memcg(from);
5697 local_irq_disable();
5698 mem_cgroup_charge_statistics(to, page, nr_pages);
5699 memcg_check_events(to, page);
5700 mem_cgroup_charge_statistics(from, page, -nr_pages);
5701 memcg_check_events(from, page);
5710 * get_mctgt_type - get target type of moving charge
5711 * @vma: the vma the pte to be checked belongs
5712 * @addr: the address corresponding to the pte to be checked
5713 * @ptent: the pte to be checked
5714 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5717 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5718 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5719 * move charge. if @target is not NULL, the page is stored in target->page
5720 * with extra refcnt got(Callers should handle it).
5721 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5722 * target for charge migration. if @target is not NULL, the entry is stored
5724 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5725 * (so ZONE_DEVICE page and thus not on the lru).
5726 * For now we such page is charge like a regular page would be as for all
5727 * intent and purposes it is just special memory taking the place of a
5730 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5732 * Called with pte lock held.
5735 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5736 unsigned long addr, pte_t ptent, union mc_target *target)
5738 struct page *page = NULL;
5739 enum mc_target_type ret = MC_TARGET_NONE;
5740 swp_entry_t ent = { .val = 0 };
5742 if (pte_present(ptent))
5743 page = mc_handle_present_pte(vma, addr, ptent);
5744 else if (is_swap_pte(ptent))
5745 page = mc_handle_swap_pte(vma, ptent, &ent);
5746 else if (pte_none(ptent))
5747 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5749 if (!page && !ent.val)
5753 * Do only loose check w/o serialization.
5754 * mem_cgroup_move_account() checks the page is valid or
5755 * not under LRU exclusion.
5757 if (page_memcg(page) == mc.from) {
5758 ret = MC_TARGET_PAGE;
5759 if (is_device_private_page(page))
5760 ret = MC_TARGET_DEVICE;
5762 target->page = page;
5764 if (!ret || !target)
5768 * There is a swap entry and a page doesn't exist or isn't charged.
5769 * But we cannot move a tail-page in a THP.
5771 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5772 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5773 ret = MC_TARGET_SWAP;
5780 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5782 * We don't consider PMD mapped swapping or file mapped pages because THP does
5783 * not support them for now.
5784 * Caller should make sure that pmd_trans_huge(pmd) is true.
5786 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5787 unsigned long addr, pmd_t pmd, union mc_target *target)
5789 struct page *page = NULL;
5790 enum mc_target_type ret = MC_TARGET_NONE;
5792 if (unlikely(is_swap_pmd(pmd))) {
5793 VM_BUG_ON(thp_migration_supported() &&
5794 !is_pmd_migration_entry(pmd));
5797 page = pmd_page(pmd);
5798 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5799 if (!(mc.flags & MOVE_ANON))
5801 if (page_memcg(page) == mc.from) {
5802 ret = MC_TARGET_PAGE;
5805 target->page = page;
5811 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5812 unsigned long addr, pmd_t pmd, union mc_target *target)
5814 return MC_TARGET_NONE;
5818 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5819 unsigned long addr, unsigned long end,
5820 struct mm_walk *walk)
5822 struct vm_area_struct *vma = walk->vma;
5826 ptl = pmd_trans_huge_lock(pmd, vma);
5829 * Note their can not be MC_TARGET_DEVICE for now as we do not
5830 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5831 * this might change.
5833 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5834 mc.precharge += HPAGE_PMD_NR;
5839 if (pmd_trans_unstable(pmd))
5841 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5842 for (; addr != end; pte++, addr += PAGE_SIZE)
5843 if (get_mctgt_type(vma, addr, *pte, NULL))
5844 mc.precharge++; /* increment precharge temporarily */
5845 pte_unmap_unlock(pte - 1, ptl);
5851 static const struct mm_walk_ops precharge_walk_ops = {
5852 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5855 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5857 unsigned long precharge;
5860 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5861 mmap_read_unlock(mm);
5863 precharge = mc.precharge;
5869 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5871 unsigned long precharge = mem_cgroup_count_precharge(mm);
5873 VM_BUG_ON(mc.moving_task);
5874 mc.moving_task = current;
5875 return mem_cgroup_do_precharge(precharge);
5878 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5879 static void __mem_cgroup_clear_mc(void)
5881 struct mem_cgroup *from = mc.from;
5882 struct mem_cgroup *to = mc.to;
5884 /* we must uncharge all the leftover precharges from mc.to */
5886 cancel_charge(mc.to, mc.precharge);
5890 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5891 * we must uncharge here.
5893 if (mc.moved_charge) {
5894 cancel_charge(mc.from, mc.moved_charge);
5895 mc.moved_charge = 0;
5897 /* we must fixup refcnts and charges */
5898 if (mc.moved_swap) {
5899 /* uncharge swap account from the old cgroup */
5900 if (!mem_cgroup_is_root(mc.from))
5901 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5903 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5906 * we charged both to->memory and to->memsw, so we
5907 * should uncharge to->memory.
5909 if (!mem_cgroup_is_root(mc.to))
5910 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5914 memcg_oom_recover(from);
5915 memcg_oom_recover(to);
5916 wake_up_all(&mc.waitq);
5919 static void mem_cgroup_clear_mc(void)
5921 struct mm_struct *mm = mc.mm;
5924 * we must clear moving_task before waking up waiters at the end of
5927 mc.moving_task = NULL;
5928 __mem_cgroup_clear_mc();
5929 spin_lock(&mc.lock);
5933 spin_unlock(&mc.lock);
5938 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5940 struct cgroup_subsys_state *css;
5941 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5942 struct mem_cgroup *from;
5943 struct task_struct *leader, *p;
5944 struct mm_struct *mm;
5945 unsigned long move_flags;
5948 /* charge immigration isn't supported on the default hierarchy */
5949 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5953 * Multi-process migrations only happen on the default hierarchy
5954 * where charge immigration is not used. Perform charge
5955 * immigration if @tset contains a leader and whine if there are
5959 cgroup_taskset_for_each_leader(leader, css, tset) {
5962 memcg = mem_cgroup_from_css(css);
5968 * We are now committed to this value whatever it is. Changes in this
5969 * tunable will only affect upcoming migrations, not the current one.
5970 * So we need to save it, and keep it going.
5972 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5976 from = mem_cgroup_from_task(p);
5978 VM_BUG_ON(from == memcg);
5980 mm = get_task_mm(p);
5983 /* We move charges only when we move a owner of the mm */
5984 if (mm->owner == p) {
5987 VM_BUG_ON(mc.precharge);
5988 VM_BUG_ON(mc.moved_charge);
5989 VM_BUG_ON(mc.moved_swap);
5991 spin_lock(&mc.lock);
5995 mc.flags = move_flags;
5996 spin_unlock(&mc.lock);
5997 /* We set mc.moving_task later */
5999 ret = mem_cgroup_precharge_mc(mm);
6001 mem_cgroup_clear_mc();
6008 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6011 mem_cgroup_clear_mc();
6014 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6015 unsigned long addr, unsigned long end,
6016 struct mm_walk *walk)
6019 struct vm_area_struct *vma = walk->vma;
6022 enum mc_target_type target_type;
6023 union mc_target target;
6026 ptl = pmd_trans_huge_lock(pmd, vma);
6028 if (mc.precharge < HPAGE_PMD_NR) {
6032 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6033 if (target_type == MC_TARGET_PAGE) {
6035 if (!isolate_lru_page(page)) {
6036 if (!mem_cgroup_move_account(page, true,
6038 mc.precharge -= HPAGE_PMD_NR;
6039 mc.moved_charge += HPAGE_PMD_NR;
6041 putback_lru_page(page);
6044 } else if (target_type == MC_TARGET_DEVICE) {
6046 if (!mem_cgroup_move_account(page, true,
6048 mc.precharge -= HPAGE_PMD_NR;
6049 mc.moved_charge += HPAGE_PMD_NR;
6057 if (pmd_trans_unstable(pmd))
6060 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6061 for (; addr != end; addr += PAGE_SIZE) {
6062 pte_t ptent = *(pte++);
6063 bool device = false;
6069 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6070 case MC_TARGET_DEVICE:
6073 case MC_TARGET_PAGE:
6076 * We can have a part of the split pmd here. Moving it
6077 * can be done but it would be too convoluted so simply
6078 * ignore such a partial THP and keep it in original
6079 * memcg. There should be somebody mapping the head.
6081 if (PageTransCompound(page))
6083 if (!device && isolate_lru_page(page))
6085 if (!mem_cgroup_move_account(page, false,
6088 /* we uncharge from mc.from later. */
6092 putback_lru_page(page);
6093 put: /* get_mctgt_type() gets the page */
6096 case MC_TARGET_SWAP:
6098 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6100 mem_cgroup_id_get_many(mc.to, 1);
6101 /* we fixup other refcnts and charges later. */
6109 pte_unmap_unlock(pte - 1, ptl);
6114 * We have consumed all precharges we got in can_attach().
6115 * We try charge one by one, but don't do any additional
6116 * charges to mc.to if we have failed in charge once in attach()
6119 ret = mem_cgroup_do_precharge(1);
6127 static const struct mm_walk_ops charge_walk_ops = {
6128 .pmd_entry = mem_cgroup_move_charge_pte_range,
6131 static void mem_cgroup_move_charge(void)
6133 lru_add_drain_all();
6135 * Signal lock_page_memcg() to take the memcg's move_lock
6136 * while we're moving its pages to another memcg. Then wait
6137 * for already started RCU-only updates to finish.
6139 atomic_inc(&mc.from->moving_account);
6142 if (unlikely(!mmap_read_trylock(mc.mm))) {
6144 * Someone who are holding the mmap_lock might be waiting in
6145 * waitq. So we cancel all extra charges, wake up all waiters,
6146 * and retry. Because we cancel precharges, we might not be able
6147 * to move enough charges, but moving charge is a best-effort
6148 * feature anyway, so it wouldn't be a big problem.
6150 __mem_cgroup_clear_mc();
6155 * When we have consumed all precharges and failed in doing
6156 * additional charge, the page walk just aborts.
6158 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6161 mmap_read_unlock(mc.mm);
6162 atomic_dec(&mc.from->moving_account);
6165 static void mem_cgroup_move_task(void)
6168 mem_cgroup_move_charge();
6169 mem_cgroup_clear_mc();
6172 #else /* !CONFIG_MMU */
6173 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6177 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6180 static void mem_cgroup_move_task(void)
6185 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6187 if (value == PAGE_COUNTER_MAX)
6188 seq_puts(m, "max\n");
6190 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6195 static u64 memory_current_read(struct cgroup_subsys_state *css,
6198 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6200 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6203 static int memory_min_show(struct seq_file *m, void *v)
6205 return seq_puts_memcg_tunable(m,
6206 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6209 static ssize_t memory_min_write(struct kernfs_open_file *of,
6210 char *buf, size_t nbytes, loff_t off)
6212 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6216 buf = strstrip(buf);
6217 err = page_counter_memparse(buf, "max", &min);
6221 page_counter_set_min(&memcg->memory, min);
6226 static int memory_low_show(struct seq_file *m, void *v)
6228 return seq_puts_memcg_tunable(m,
6229 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6232 static ssize_t memory_low_write(struct kernfs_open_file *of,
6233 char *buf, size_t nbytes, loff_t off)
6235 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6239 buf = strstrip(buf);
6240 err = page_counter_memparse(buf, "max", &low);
6244 page_counter_set_low(&memcg->memory, low);
6249 static int memory_high_show(struct seq_file *m, void *v)
6251 return seq_puts_memcg_tunable(m,
6252 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6255 static ssize_t memory_high_write(struct kernfs_open_file *of,
6256 char *buf, size_t nbytes, loff_t off)
6258 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6259 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6260 bool drained = false;
6264 buf = strstrip(buf);
6265 err = page_counter_memparse(buf, "max", &high);
6269 page_counter_set_high(&memcg->memory, high);
6272 unsigned long nr_pages = page_counter_read(&memcg->memory);
6273 unsigned long reclaimed;
6275 if (nr_pages <= high)
6278 if (signal_pending(current))
6282 drain_all_stock(memcg);
6287 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6290 if (!reclaimed && !nr_retries--)
6294 memcg_wb_domain_size_changed(memcg);
6298 static int memory_max_show(struct seq_file *m, void *v)
6300 return seq_puts_memcg_tunable(m,
6301 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6304 static ssize_t memory_max_write(struct kernfs_open_file *of,
6305 char *buf, size_t nbytes, loff_t off)
6307 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6308 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6309 bool drained = false;
6313 buf = strstrip(buf);
6314 err = page_counter_memparse(buf, "max", &max);
6318 xchg(&memcg->memory.max, max);
6321 unsigned long nr_pages = page_counter_read(&memcg->memory);
6323 if (nr_pages <= max)
6326 if (signal_pending(current))
6330 drain_all_stock(memcg);
6336 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6342 memcg_memory_event(memcg, MEMCG_OOM);
6343 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6347 memcg_wb_domain_size_changed(memcg);
6351 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6353 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6354 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6355 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6356 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6357 seq_printf(m, "oom_kill %lu\n",
6358 atomic_long_read(&events[MEMCG_OOM_KILL]));
6361 static int memory_events_show(struct seq_file *m, void *v)
6363 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6365 __memory_events_show(m, memcg->memory_events);
6369 static int memory_events_local_show(struct seq_file *m, void *v)
6371 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6373 __memory_events_show(m, memcg->memory_events_local);
6377 static int memory_stat_show(struct seq_file *m, void *v)
6379 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6382 buf = memory_stat_format(memcg);
6391 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6394 return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6397 static int memory_numa_stat_show(struct seq_file *m, void *v)
6400 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6402 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6405 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6408 seq_printf(m, "%s", memory_stats[i].name);
6409 for_each_node_state(nid, N_MEMORY) {
6411 struct lruvec *lruvec;
6413 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6414 size = lruvec_page_state_output(lruvec,
6415 memory_stats[i].idx);
6416 seq_printf(m, " N%d=%llu", nid, size);
6425 static int memory_oom_group_show(struct seq_file *m, void *v)
6427 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6429 seq_printf(m, "%d\n", memcg->oom_group);
6434 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6435 char *buf, size_t nbytes, loff_t off)
6437 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6440 buf = strstrip(buf);
6444 ret = kstrtoint(buf, 0, &oom_group);
6448 if (oom_group != 0 && oom_group != 1)
6451 memcg->oom_group = oom_group;
6456 static struct cftype memory_files[] = {
6459 .flags = CFTYPE_NOT_ON_ROOT,
6460 .read_u64 = memory_current_read,
6464 .flags = CFTYPE_NOT_ON_ROOT,
6465 .seq_show = memory_min_show,
6466 .write = memory_min_write,
6470 .flags = CFTYPE_NOT_ON_ROOT,
6471 .seq_show = memory_low_show,
6472 .write = memory_low_write,
6476 .flags = CFTYPE_NOT_ON_ROOT,
6477 .seq_show = memory_high_show,
6478 .write = memory_high_write,
6482 .flags = CFTYPE_NOT_ON_ROOT,
6483 .seq_show = memory_max_show,
6484 .write = memory_max_write,
6488 .flags = CFTYPE_NOT_ON_ROOT,
6489 .file_offset = offsetof(struct mem_cgroup, events_file),
6490 .seq_show = memory_events_show,
6493 .name = "events.local",
6494 .flags = CFTYPE_NOT_ON_ROOT,
6495 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6496 .seq_show = memory_events_local_show,
6500 .seq_show = memory_stat_show,
6504 .name = "numa_stat",
6505 .seq_show = memory_numa_stat_show,
6509 .name = "oom.group",
6510 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6511 .seq_show = memory_oom_group_show,
6512 .write = memory_oom_group_write,
6517 struct cgroup_subsys memory_cgrp_subsys = {
6518 .css_alloc = mem_cgroup_css_alloc,
6519 .css_online = mem_cgroup_css_online,
6520 .css_offline = mem_cgroup_css_offline,
6521 .css_released = mem_cgroup_css_released,
6522 .css_free = mem_cgroup_css_free,
6523 .css_reset = mem_cgroup_css_reset,
6524 .css_rstat_flush = mem_cgroup_css_rstat_flush,
6525 .can_attach = mem_cgroup_can_attach,
6526 .cancel_attach = mem_cgroup_cancel_attach,
6527 .post_attach = mem_cgroup_move_task,
6528 .dfl_cftypes = memory_files,
6529 .legacy_cftypes = mem_cgroup_legacy_files,
6534 * This function calculates an individual cgroup's effective
6535 * protection which is derived from its own memory.min/low, its
6536 * parent's and siblings' settings, as well as the actual memory
6537 * distribution in the tree.
6539 * The following rules apply to the effective protection values:
6541 * 1. At the first level of reclaim, effective protection is equal to
6542 * the declared protection in memory.min and memory.low.
6544 * 2. To enable safe delegation of the protection configuration, at
6545 * subsequent levels the effective protection is capped to the
6546 * parent's effective protection.
6548 * 3. To make complex and dynamic subtrees easier to configure, the
6549 * user is allowed to overcommit the declared protection at a given
6550 * level. If that is the case, the parent's effective protection is
6551 * distributed to the children in proportion to how much protection
6552 * they have declared and how much of it they are utilizing.
6554 * This makes distribution proportional, but also work-conserving:
6555 * if one cgroup claims much more protection than it uses memory,
6556 * the unused remainder is available to its siblings.
6558 * 4. Conversely, when the declared protection is undercommitted at a
6559 * given level, the distribution of the larger parental protection
6560 * budget is NOT proportional. A cgroup's protection from a sibling
6561 * is capped to its own memory.min/low setting.
6563 * 5. However, to allow protecting recursive subtrees from each other
6564 * without having to declare each individual cgroup's fixed share
6565 * of the ancestor's claim to protection, any unutilized -
6566 * "floating" - protection from up the tree is distributed in
6567 * proportion to each cgroup's *usage*. This makes the protection
6568 * neutral wrt sibling cgroups and lets them compete freely over
6569 * the shared parental protection budget, but it protects the
6570 * subtree as a whole from neighboring subtrees.
6572 * Note that 4. and 5. are not in conflict: 4. is about protecting
6573 * against immediate siblings whereas 5. is about protecting against
6574 * neighboring subtrees.
6576 static unsigned long effective_protection(unsigned long usage,
6577 unsigned long parent_usage,
6578 unsigned long setting,
6579 unsigned long parent_effective,
6580 unsigned long siblings_protected)
6582 unsigned long protected;
6585 protected = min(usage, setting);
6587 * If all cgroups at this level combined claim and use more
6588 * protection then what the parent affords them, distribute
6589 * shares in proportion to utilization.
6591 * We are using actual utilization rather than the statically
6592 * claimed protection in order to be work-conserving: claimed
6593 * but unused protection is available to siblings that would
6594 * otherwise get a smaller chunk than what they claimed.
6596 if (siblings_protected > parent_effective)
6597 return protected * parent_effective / siblings_protected;
6600 * Ok, utilized protection of all children is within what the
6601 * parent affords them, so we know whatever this child claims
6602 * and utilizes is effectively protected.
6604 * If there is unprotected usage beyond this value, reclaim
6605 * will apply pressure in proportion to that amount.
6607 * If there is unutilized protection, the cgroup will be fully
6608 * shielded from reclaim, but we do return a smaller value for
6609 * protection than what the group could enjoy in theory. This
6610 * is okay. With the overcommit distribution above, effective
6611 * protection is always dependent on how memory is actually
6612 * consumed among the siblings anyway.
6617 * If the children aren't claiming (all of) the protection
6618 * afforded to them by the parent, distribute the remainder in
6619 * proportion to the (unprotected) memory of each cgroup. That
6620 * way, cgroups that aren't explicitly prioritized wrt each
6621 * other compete freely over the allowance, but they are
6622 * collectively protected from neighboring trees.
6624 * We're using unprotected memory for the weight so that if
6625 * some cgroups DO claim explicit protection, we don't protect
6626 * the same bytes twice.
6628 * Check both usage and parent_usage against the respective
6629 * protected values. One should imply the other, but they
6630 * aren't read atomically - make sure the division is sane.
6632 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6634 if (parent_effective > siblings_protected &&
6635 parent_usage > siblings_protected &&
6636 usage > protected) {
6637 unsigned long unclaimed;
6639 unclaimed = parent_effective - siblings_protected;
6640 unclaimed *= usage - protected;
6641 unclaimed /= parent_usage - siblings_protected;
6650 * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
6651 * @root: the top ancestor of the sub-tree being checked
6652 * @memcg: the memory cgroup to check
6654 * WARNING: This function is not stateless! It can only be used as part
6655 * of a top-down tree iteration, not for isolated queries.
6657 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6658 struct mem_cgroup *memcg)
6660 unsigned long usage, parent_usage;
6661 struct mem_cgroup *parent;
6663 if (mem_cgroup_disabled())
6667 root = root_mem_cgroup;
6670 * Effective values of the reclaim targets are ignored so they
6671 * can be stale. Have a look at mem_cgroup_protection for more
6673 * TODO: calculation should be more robust so that we do not need
6674 * that special casing.
6679 usage = page_counter_read(&memcg->memory);
6683 parent = parent_mem_cgroup(memcg);
6684 /* No parent means a non-hierarchical mode on v1 memcg */
6688 if (parent == root) {
6689 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6690 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6694 parent_usage = page_counter_read(&parent->memory);
6696 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6697 READ_ONCE(memcg->memory.min),
6698 READ_ONCE(parent->memory.emin),
6699 atomic_long_read(&parent->memory.children_min_usage)));
6701 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6702 READ_ONCE(memcg->memory.low),
6703 READ_ONCE(parent->memory.elow),
6704 atomic_long_read(&parent->memory.children_low_usage)));
6707 static int __mem_cgroup_charge(struct page *page, struct mem_cgroup *memcg,
6710 unsigned int nr_pages = thp_nr_pages(page);
6713 ret = try_charge(memcg, gfp, nr_pages);
6717 css_get(&memcg->css);
6718 commit_charge(page, memcg);
6720 local_irq_disable();
6721 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6722 memcg_check_events(memcg, page);
6729 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6730 * @page: page to charge
6731 * @mm: mm context of the victim
6732 * @gfp_mask: reclaim mode
6734 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6735 * pages according to @gfp_mask if necessary. if @mm is NULL, try to
6736 * charge to the active memcg.
6738 * Do not use this for pages allocated for swapin.
6740 * Returns 0 on success. Otherwise, an error code is returned.
6742 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6744 struct mem_cgroup *memcg;
6747 if (mem_cgroup_disabled())
6750 memcg = get_mem_cgroup_from_mm(mm);
6751 ret = __mem_cgroup_charge(page, memcg, gfp_mask);
6752 css_put(&memcg->css);
6758 * mem_cgroup_swapin_charge_page - charge a newly allocated page for swapin
6759 * @page: page to charge
6760 * @mm: mm context of the victim
6761 * @gfp: reclaim mode
6762 * @entry: swap entry for which the page is allocated
6764 * This function charges a page allocated for swapin. Please call this before
6765 * adding the page to the swapcache.
6767 * Returns 0 on success. Otherwise, an error code is returned.
6769 int mem_cgroup_swapin_charge_page(struct page *page, struct mm_struct *mm,
6770 gfp_t gfp, swp_entry_t entry)
6772 struct mem_cgroup *memcg;
6776 if (mem_cgroup_disabled())
6779 id = lookup_swap_cgroup_id(entry);
6781 memcg = mem_cgroup_from_id(id);
6782 if (!memcg || !css_tryget_online(&memcg->css))
6783 memcg = get_mem_cgroup_from_mm(mm);
6786 ret = __mem_cgroup_charge(page, memcg, gfp);
6788 css_put(&memcg->css);
6793 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6794 * @entry: swap entry for which the page is charged
6796 * Call this function after successfully adding the charged page to swapcache.
6798 * Note: This function assumes the page for which swap slot is being uncharged
6801 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
6804 * Cgroup1's unified memory+swap counter has been charged with the
6805 * new swapcache page, finish the transfer by uncharging the swap
6806 * slot. The swap slot would also get uncharged when it dies, but
6807 * it can stick around indefinitely and we'd count the page twice
6810 * Cgroup2 has separate resource counters for memory and swap,
6811 * so this is a non-issue here. Memory and swap charge lifetimes
6812 * correspond 1:1 to page and swap slot lifetimes: we charge the
6813 * page to memory here, and uncharge swap when the slot is freed.
6815 if (!mem_cgroup_disabled() && do_memsw_account()) {
6817 * The swap entry might not get freed for a long time,
6818 * let's not wait for it. The page already received a
6819 * memory+swap charge, drop the swap entry duplicate.
6821 mem_cgroup_uncharge_swap(entry, 1);
6825 struct uncharge_gather {
6826 struct mem_cgroup *memcg;
6827 unsigned long nr_memory;
6828 unsigned long pgpgout;
6829 unsigned long nr_kmem;
6830 struct page *dummy_page;
6833 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6835 memset(ug, 0, sizeof(*ug));
6838 static void uncharge_batch(const struct uncharge_gather *ug)
6840 unsigned long flags;
6842 if (ug->nr_memory) {
6843 page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
6844 if (do_memsw_account())
6845 page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
6846 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6847 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6848 memcg_oom_recover(ug->memcg);
6851 local_irq_save(flags);
6852 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6853 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
6854 memcg_check_events(ug->memcg, ug->dummy_page);
6855 local_irq_restore(flags);
6857 /* drop reference from uncharge_page */
6858 css_put(&ug->memcg->css);
6861 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6863 unsigned long nr_pages;
6864 struct mem_cgroup *memcg;
6865 struct obj_cgroup *objcg;
6866 bool use_objcg = PageMemcgKmem(page);
6868 VM_BUG_ON_PAGE(PageLRU(page), page);
6871 * Nobody should be changing or seriously looking at
6872 * page memcg or objcg at this point, we have fully
6873 * exclusive access to the page.
6876 objcg = __page_objcg(page);
6878 * This get matches the put at the end of the function and
6879 * kmem pages do not hold memcg references anymore.
6881 memcg = get_mem_cgroup_from_objcg(objcg);
6883 memcg = __page_memcg(page);
6889 if (ug->memcg != memcg) {
6892 uncharge_gather_clear(ug);
6895 ug->dummy_page = page;
6897 /* pairs with css_put in uncharge_batch */
6898 css_get(&memcg->css);
6901 nr_pages = compound_nr(page);
6904 ug->nr_memory += nr_pages;
6905 ug->nr_kmem += nr_pages;
6907 page->memcg_data = 0;
6908 obj_cgroup_put(objcg);
6910 /* LRU pages aren't accounted at the root level */
6911 if (!mem_cgroup_is_root(memcg))
6912 ug->nr_memory += nr_pages;
6915 page->memcg_data = 0;
6918 css_put(&memcg->css);
6922 * mem_cgroup_uncharge - uncharge a page
6923 * @page: page to uncharge
6925 * Uncharge a page previously charged with mem_cgroup_charge().
6927 void mem_cgroup_uncharge(struct page *page)
6929 struct uncharge_gather ug;
6931 if (mem_cgroup_disabled())
6934 /* Don't touch page->lru of any random page, pre-check: */
6935 if (!page_memcg(page))
6938 uncharge_gather_clear(&ug);
6939 uncharge_page(page, &ug);
6940 uncharge_batch(&ug);
6944 * mem_cgroup_uncharge_list - uncharge a list of page
6945 * @page_list: list of pages to uncharge
6947 * Uncharge a list of pages previously charged with
6948 * mem_cgroup_charge().
6950 void mem_cgroup_uncharge_list(struct list_head *page_list)
6952 struct uncharge_gather ug;
6955 if (mem_cgroup_disabled())
6958 uncharge_gather_clear(&ug);
6959 list_for_each_entry(page, page_list, lru)
6960 uncharge_page(page, &ug);
6962 uncharge_batch(&ug);
6966 * mem_cgroup_migrate - charge a page's replacement
6967 * @oldpage: currently circulating page
6968 * @newpage: replacement page
6970 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6971 * be uncharged upon free.
6973 * Both pages must be locked, @newpage->mapping must be set up.
6975 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6977 struct mem_cgroup *memcg;
6978 unsigned int nr_pages;
6979 unsigned long flags;
6981 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6982 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6983 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6984 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6987 if (mem_cgroup_disabled())
6990 /* Page cache replacement: new page already charged? */
6991 if (page_memcg(newpage))
6994 memcg = page_memcg(oldpage);
6995 VM_WARN_ON_ONCE_PAGE(!memcg, oldpage);
6999 /* Force-charge the new page. The old one will be freed soon */
7000 nr_pages = thp_nr_pages(newpage);
7002 if (!mem_cgroup_is_root(memcg)) {
7003 page_counter_charge(&memcg->memory, nr_pages);
7004 if (do_memsw_account())
7005 page_counter_charge(&memcg->memsw, nr_pages);
7008 css_get(&memcg->css);
7009 commit_charge(newpage, memcg);
7011 local_irq_save(flags);
7012 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
7013 memcg_check_events(memcg, newpage);
7014 local_irq_restore(flags);
7017 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7018 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7020 void mem_cgroup_sk_alloc(struct sock *sk)
7022 struct mem_cgroup *memcg;
7024 if (!mem_cgroup_sockets_enabled)
7027 /* Do not associate the sock with unrelated interrupted task's memcg. */
7032 memcg = mem_cgroup_from_task(current);
7033 if (memcg == root_mem_cgroup)
7035 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7037 if (css_tryget(&memcg->css))
7038 sk->sk_memcg = memcg;
7043 void mem_cgroup_sk_free(struct sock *sk)
7046 css_put(&sk->sk_memcg->css);
7050 * mem_cgroup_charge_skmem - charge socket memory
7051 * @memcg: memcg to charge
7052 * @nr_pages: number of pages to charge
7054 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7055 * @memcg's configured limit, %false if the charge had to be forced.
7057 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7059 gfp_t gfp_mask = GFP_KERNEL;
7061 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7062 struct page_counter *fail;
7064 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7065 memcg->tcpmem_pressure = 0;
7068 page_counter_charge(&memcg->tcpmem, nr_pages);
7069 memcg->tcpmem_pressure = 1;
7073 /* Don't block in the packet receive path */
7075 gfp_mask = GFP_NOWAIT;
7077 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7079 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7082 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7087 * mem_cgroup_uncharge_skmem - uncharge socket memory
7088 * @memcg: memcg to uncharge
7089 * @nr_pages: number of pages to uncharge
7091 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7093 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7094 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7098 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7100 refill_stock(memcg, nr_pages);
7103 static int __init cgroup_memory(char *s)
7107 while ((token = strsep(&s, ",")) != NULL) {
7110 if (!strcmp(token, "nosocket"))
7111 cgroup_memory_nosocket = true;
7112 if (!strcmp(token, "nokmem"))
7113 cgroup_memory_nokmem = true;
7117 __setup("cgroup.memory=", cgroup_memory);
7120 * subsys_initcall() for memory controller.
7122 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7123 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7124 * basically everything that doesn't depend on a specific mem_cgroup structure
7125 * should be initialized from here.
7127 static int __init mem_cgroup_init(void)
7132 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7133 * used for per-memcg-per-cpu caching of per-node statistics. In order
7134 * to work fine, we should make sure that the overfill threshold can't
7135 * exceed S32_MAX / PAGE_SIZE.
7137 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7139 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7140 memcg_hotplug_cpu_dead);
7142 for_each_possible_cpu(cpu)
7143 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7146 for_each_node(node) {
7147 struct mem_cgroup_tree_per_node *rtpn;
7149 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7150 node_online(node) ? node : NUMA_NO_NODE);
7152 rtpn->rb_root = RB_ROOT;
7153 rtpn->rb_rightmost = NULL;
7154 spin_lock_init(&rtpn->lock);
7155 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7160 subsys_initcall(mem_cgroup_init);
7162 #ifdef CONFIG_MEMCG_SWAP
7163 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7165 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7167 * The root cgroup cannot be destroyed, so it's refcount must
7170 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7174 memcg = parent_mem_cgroup(memcg);
7176 memcg = root_mem_cgroup;
7182 * mem_cgroup_swapout - transfer a memsw charge to swap
7183 * @page: page whose memsw charge to transfer
7184 * @entry: swap entry to move the charge to
7186 * Transfer the memsw charge of @page to @entry.
7188 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7190 struct mem_cgroup *memcg, *swap_memcg;
7191 unsigned int nr_entries;
7192 unsigned short oldid;
7194 VM_BUG_ON_PAGE(PageLRU(page), page);
7195 VM_BUG_ON_PAGE(page_count(page), page);
7197 if (mem_cgroup_disabled())
7200 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7203 memcg = page_memcg(page);
7205 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7210 * In case the memcg owning these pages has been offlined and doesn't
7211 * have an ID allocated to it anymore, charge the closest online
7212 * ancestor for the swap instead and transfer the memory+swap charge.
7214 swap_memcg = mem_cgroup_id_get_online(memcg);
7215 nr_entries = thp_nr_pages(page);
7216 /* Get references for the tail pages, too */
7218 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7219 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7221 VM_BUG_ON_PAGE(oldid, page);
7222 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7224 page->memcg_data = 0;
7226 if (!mem_cgroup_is_root(memcg))
7227 page_counter_uncharge(&memcg->memory, nr_entries);
7229 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7230 if (!mem_cgroup_is_root(swap_memcg))
7231 page_counter_charge(&swap_memcg->memsw, nr_entries);
7232 page_counter_uncharge(&memcg->memsw, nr_entries);
7236 * Interrupts should be disabled here because the caller holds the
7237 * i_pages lock which is taken with interrupts-off. It is
7238 * important here to have the interrupts disabled because it is the
7239 * only synchronisation we have for updating the per-CPU variables.
7241 VM_BUG_ON(!irqs_disabled());
7242 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7243 memcg_check_events(memcg, page);
7245 css_put(&memcg->css);
7249 * mem_cgroup_try_charge_swap - try charging swap space for a page
7250 * @page: page being added to swap
7251 * @entry: swap entry to charge
7253 * Try to charge @page's memcg for the swap space at @entry.
7255 * Returns 0 on success, -ENOMEM on failure.
7257 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7259 unsigned int nr_pages = thp_nr_pages(page);
7260 struct page_counter *counter;
7261 struct mem_cgroup *memcg;
7262 unsigned short oldid;
7264 if (mem_cgroup_disabled())
7267 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7270 memcg = page_memcg(page);
7272 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7277 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7281 memcg = mem_cgroup_id_get_online(memcg);
7283 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7284 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7285 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7286 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7287 mem_cgroup_id_put(memcg);
7291 /* Get references for the tail pages, too */
7293 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7294 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7295 VM_BUG_ON_PAGE(oldid, page);
7296 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7302 * mem_cgroup_uncharge_swap - uncharge swap space
7303 * @entry: swap entry to uncharge
7304 * @nr_pages: the amount of swap space to uncharge
7306 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7308 struct mem_cgroup *memcg;
7311 id = swap_cgroup_record(entry, 0, nr_pages);
7313 memcg = mem_cgroup_from_id(id);
7315 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7316 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7317 page_counter_uncharge(&memcg->swap, nr_pages);
7319 page_counter_uncharge(&memcg->memsw, nr_pages);
7321 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7322 mem_cgroup_id_put_many(memcg, nr_pages);
7327 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7329 long nr_swap_pages = get_nr_swap_pages();
7331 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7332 return nr_swap_pages;
7333 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7334 nr_swap_pages = min_t(long, nr_swap_pages,
7335 READ_ONCE(memcg->swap.max) -
7336 page_counter_read(&memcg->swap));
7337 return nr_swap_pages;
7340 bool mem_cgroup_swap_full(struct page *page)
7342 struct mem_cgroup *memcg;
7344 VM_BUG_ON_PAGE(!PageLocked(page), page);
7348 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7351 memcg = page_memcg(page);
7355 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7356 unsigned long usage = page_counter_read(&memcg->swap);
7358 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7359 usage * 2 >= READ_ONCE(memcg->swap.max))
7366 static int __init setup_swap_account(char *s)
7368 if (!strcmp(s, "1"))
7369 cgroup_memory_noswap = false;
7370 else if (!strcmp(s, "0"))
7371 cgroup_memory_noswap = true;
7374 __setup("swapaccount=", setup_swap_account);
7376 static u64 swap_current_read(struct cgroup_subsys_state *css,
7379 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7381 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7384 static int swap_high_show(struct seq_file *m, void *v)
7386 return seq_puts_memcg_tunable(m,
7387 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7390 static ssize_t swap_high_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", &high);
7402 page_counter_set_high(&memcg->swap, high);
7407 static int swap_max_show(struct seq_file *m, void *v)
7409 return seq_puts_memcg_tunable(m,
7410 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7413 static ssize_t swap_max_write(struct kernfs_open_file *of,
7414 char *buf, size_t nbytes, loff_t off)
7416 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7420 buf = strstrip(buf);
7421 err = page_counter_memparse(buf, "max", &max);
7425 xchg(&memcg->swap.max, max);
7430 static int swap_events_show(struct seq_file *m, void *v)
7432 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7434 seq_printf(m, "high %lu\n",
7435 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7436 seq_printf(m, "max %lu\n",
7437 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7438 seq_printf(m, "fail %lu\n",
7439 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7444 static struct cftype swap_files[] = {
7446 .name = "swap.current",
7447 .flags = CFTYPE_NOT_ON_ROOT,
7448 .read_u64 = swap_current_read,
7451 .name = "swap.high",
7452 .flags = CFTYPE_NOT_ON_ROOT,
7453 .seq_show = swap_high_show,
7454 .write = swap_high_write,
7458 .flags = CFTYPE_NOT_ON_ROOT,
7459 .seq_show = swap_max_show,
7460 .write = swap_max_write,
7463 .name = "swap.events",
7464 .flags = CFTYPE_NOT_ON_ROOT,
7465 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7466 .seq_show = swap_events_show,
7471 static struct cftype memsw_files[] = {
7473 .name = "memsw.usage_in_bytes",
7474 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7475 .read_u64 = mem_cgroup_read_u64,
7478 .name = "memsw.max_usage_in_bytes",
7479 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7480 .write = mem_cgroup_reset,
7481 .read_u64 = mem_cgroup_read_u64,
7484 .name = "memsw.limit_in_bytes",
7485 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7486 .write = mem_cgroup_write,
7487 .read_u64 = mem_cgroup_read_u64,
7490 .name = "memsw.failcnt",
7491 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7492 .write = mem_cgroup_reset,
7493 .read_u64 = mem_cgroup_read_u64,
7495 { }, /* terminate */
7499 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7500 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7501 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7502 * boot parameter. This may result in premature OOPS inside
7503 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7505 static int __init mem_cgroup_swap_init(void)
7507 /* No memory control -> no swap control */
7508 if (mem_cgroup_disabled())
7509 cgroup_memory_noswap = true;
7511 if (cgroup_memory_noswap)
7514 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7515 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7519 core_initcall(mem_cgroup_swap_init);
7521 #endif /* CONFIG_MEMCG_SWAP */