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_task() || !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 if (stock->nr_slab_reclaimable_b) {
3110 mod_objcg_mlstate(objcg, pgdat, NR_SLAB_RECLAIMABLE_B,
3111 stock->nr_slab_reclaimable_b);
3112 stock->nr_slab_reclaimable_b = 0;
3114 if (stock->nr_slab_unreclaimable_b) {
3115 mod_objcg_mlstate(objcg, pgdat, NR_SLAB_UNRECLAIMABLE_B,
3116 stock->nr_slab_unreclaimable_b);
3117 stock->nr_slab_unreclaimable_b = 0;
3119 stock->cached_pgdat = pgdat;
3122 bytes = (idx == NR_SLAB_RECLAIMABLE_B) ? &stock->nr_slab_reclaimable_b
3123 : &stock->nr_slab_unreclaimable_b;
3125 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3126 * cached locally at least once before pushing it out.
3133 if (abs(*bytes) > PAGE_SIZE) {
3141 mod_objcg_mlstate(objcg, pgdat, idx, nr);
3143 put_obj_stock(flags);
3146 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3148 unsigned long flags;
3149 struct obj_stock *stock = get_obj_stock(&flags);
3152 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3153 stock->nr_bytes -= nr_bytes;
3157 put_obj_stock(flags);
3162 static void drain_obj_stock(struct obj_stock *stock)
3164 struct obj_cgroup *old = stock->cached_objcg;
3169 if (stock->nr_bytes) {
3170 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3171 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3174 obj_cgroup_uncharge_pages(old, nr_pages);
3177 * The leftover is flushed to the centralized per-memcg value.
3178 * On the next attempt to refill obj stock it will be moved
3179 * to a per-cpu stock (probably, on an other CPU), see
3180 * refill_obj_stock().
3182 * How often it's flushed is a trade-off between the memory
3183 * limit enforcement accuracy and potential CPU contention,
3184 * so it might be changed in the future.
3186 atomic_add(nr_bytes, &old->nr_charged_bytes);
3187 stock->nr_bytes = 0;
3191 * Flush the vmstat data in current stock
3193 if (stock->nr_slab_reclaimable_b || stock->nr_slab_unreclaimable_b) {
3194 if (stock->nr_slab_reclaimable_b) {
3195 mod_objcg_mlstate(old, stock->cached_pgdat,
3196 NR_SLAB_RECLAIMABLE_B,
3197 stock->nr_slab_reclaimable_b);
3198 stock->nr_slab_reclaimable_b = 0;
3200 if (stock->nr_slab_unreclaimable_b) {
3201 mod_objcg_mlstate(old, stock->cached_pgdat,
3202 NR_SLAB_UNRECLAIMABLE_B,
3203 stock->nr_slab_unreclaimable_b);
3204 stock->nr_slab_unreclaimable_b = 0;
3206 stock->cached_pgdat = NULL;
3209 obj_cgroup_put(old);
3210 stock->cached_objcg = NULL;
3213 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3214 struct mem_cgroup *root_memcg)
3216 struct mem_cgroup *memcg;
3218 if (in_task() && stock->task_obj.cached_objcg) {
3219 memcg = obj_cgroup_memcg(stock->task_obj.cached_objcg);
3220 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3223 if (stock->irq_obj.cached_objcg) {
3224 memcg = obj_cgroup_memcg(stock->irq_obj.cached_objcg);
3225 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3232 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes,
3233 bool allow_uncharge)
3235 unsigned long flags;
3236 struct obj_stock *stock = get_obj_stock(&flags);
3237 unsigned int nr_pages = 0;
3239 if (stock->cached_objcg != objcg) { /* reset if necessary */
3240 drain_obj_stock(stock);
3241 obj_cgroup_get(objcg);
3242 stock->cached_objcg = objcg;
3243 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3244 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3245 allow_uncharge = true; /* Allow uncharge when objcg changes */
3247 stock->nr_bytes += nr_bytes;
3249 if (allow_uncharge && (stock->nr_bytes > PAGE_SIZE)) {
3250 nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3251 stock->nr_bytes &= (PAGE_SIZE - 1);
3254 put_obj_stock(flags);
3257 obj_cgroup_uncharge_pages(objcg, nr_pages);
3260 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3262 unsigned int nr_pages, nr_bytes;
3265 if (consume_obj_stock(objcg, size))
3269 * In theory, objcg->nr_charged_bytes can have enough
3270 * pre-charged bytes to satisfy the allocation. However,
3271 * flushing objcg->nr_charged_bytes requires two atomic
3272 * operations, and objcg->nr_charged_bytes can't be big.
3273 * The shared objcg->nr_charged_bytes can also become a
3274 * performance bottleneck if all tasks of the same memcg are
3275 * trying to update it. So it's better to ignore it and try
3276 * grab some new pages. The stock's nr_bytes will be flushed to
3277 * objcg->nr_charged_bytes later on when objcg changes.
3279 * The stock's nr_bytes may contain enough pre-charged bytes
3280 * to allow one less page from being charged, but we can't rely
3281 * on the pre-charged bytes not being changed outside of
3282 * consume_obj_stock() or refill_obj_stock(). So ignore those
3283 * pre-charged bytes as well when charging pages. To avoid a
3284 * page uncharge right after a page charge, we set the
3285 * allow_uncharge flag to false when calling refill_obj_stock()
3286 * to temporarily allow the pre-charged bytes to exceed the page
3287 * size limit. The maximum reachable value of the pre-charged
3288 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3291 nr_pages = size >> PAGE_SHIFT;
3292 nr_bytes = size & (PAGE_SIZE - 1);
3297 ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3298 if (!ret && nr_bytes)
3299 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes, false);
3304 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3306 refill_obj_stock(objcg, size, true);
3309 #endif /* CONFIG_MEMCG_KMEM */
3312 * Because page_memcg(head) is not set on tails, set it now.
3314 void split_page_memcg(struct page *head, unsigned int nr)
3316 struct mem_cgroup *memcg = page_memcg(head);
3319 if (mem_cgroup_disabled() || !memcg)
3322 for (i = 1; i < nr; i++)
3323 head[i].memcg_data = head->memcg_data;
3325 if (PageMemcgKmem(head))
3326 obj_cgroup_get_many(__page_objcg(head), nr - 1);
3328 css_get_many(&memcg->css, nr - 1);
3331 #ifdef CONFIG_MEMCG_SWAP
3333 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3334 * @entry: swap entry to be moved
3335 * @from: mem_cgroup which the entry is moved from
3336 * @to: mem_cgroup which the entry is moved to
3338 * It succeeds only when the swap_cgroup's record for this entry is the same
3339 * as the mem_cgroup's id of @from.
3341 * Returns 0 on success, -EINVAL on failure.
3343 * The caller must have charged to @to, IOW, called page_counter_charge() about
3344 * both res and memsw, and called css_get().
3346 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3347 struct mem_cgroup *from, struct mem_cgroup *to)
3349 unsigned short old_id, new_id;
3351 old_id = mem_cgroup_id(from);
3352 new_id = mem_cgroup_id(to);
3354 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3355 mod_memcg_state(from, MEMCG_SWAP, -1);
3356 mod_memcg_state(to, MEMCG_SWAP, 1);
3362 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3363 struct mem_cgroup *from, struct mem_cgroup *to)
3369 static DEFINE_MUTEX(memcg_max_mutex);
3371 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3372 unsigned long max, bool memsw)
3374 bool enlarge = false;
3375 bool drained = false;
3377 bool limits_invariant;
3378 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3381 if (signal_pending(current)) {
3386 mutex_lock(&memcg_max_mutex);
3388 * Make sure that the new limit (memsw or memory limit) doesn't
3389 * break our basic invariant rule memory.max <= memsw.max.
3391 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3392 max <= memcg->memsw.max;
3393 if (!limits_invariant) {
3394 mutex_unlock(&memcg_max_mutex);
3398 if (max > counter->max)
3400 ret = page_counter_set_max(counter, max);
3401 mutex_unlock(&memcg_max_mutex);
3407 drain_all_stock(memcg);
3412 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3413 GFP_KERNEL, !memsw)) {
3419 if (!ret && enlarge)
3420 memcg_oom_recover(memcg);
3425 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3427 unsigned long *total_scanned)
3429 unsigned long nr_reclaimed = 0;
3430 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3431 unsigned long reclaimed;
3433 struct mem_cgroup_tree_per_node *mctz;
3434 unsigned long excess;
3435 unsigned long nr_scanned;
3440 mctz = soft_limit_tree_node(pgdat->node_id);
3443 * Do not even bother to check the largest node if the root
3444 * is empty. Do it lockless to prevent lock bouncing. Races
3445 * are acceptable as soft limit is best effort anyway.
3447 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3451 * This loop can run a while, specially if mem_cgroup's continuously
3452 * keep exceeding their soft limit and putting the system under
3459 mz = mem_cgroup_largest_soft_limit_node(mctz);
3464 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3465 gfp_mask, &nr_scanned);
3466 nr_reclaimed += reclaimed;
3467 *total_scanned += nr_scanned;
3468 spin_lock_irq(&mctz->lock);
3469 __mem_cgroup_remove_exceeded(mz, mctz);
3472 * If we failed to reclaim anything from this memory cgroup
3473 * it is time to move on to the next cgroup
3477 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3479 excess = soft_limit_excess(mz->memcg);
3481 * One school of thought says that we should not add
3482 * back the node to the tree if reclaim returns 0.
3483 * But our reclaim could return 0, simply because due
3484 * to priority we are exposing a smaller subset of
3485 * memory to reclaim from. Consider this as a longer
3488 /* If excess == 0, no tree ops */
3489 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3490 spin_unlock_irq(&mctz->lock);
3491 css_put(&mz->memcg->css);
3494 * Could not reclaim anything and there are no more
3495 * mem cgroups to try or we seem to be looping without
3496 * reclaiming anything.
3498 if (!nr_reclaimed &&
3500 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3502 } while (!nr_reclaimed);
3504 css_put(&next_mz->memcg->css);
3505 return nr_reclaimed;
3509 * Reclaims as many pages from the given memcg as possible.
3511 * Caller is responsible for holding css reference for memcg.
3513 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3515 int nr_retries = MAX_RECLAIM_RETRIES;
3517 /* we call try-to-free pages for make this cgroup empty */
3518 lru_add_drain_all();
3520 drain_all_stock(memcg);
3522 /* try to free all pages in this cgroup */
3523 while (nr_retries && page_counter_read(&memcg->memory)) {
3526 if (signal_pending(current))
3529 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3533 /* maybe some writeback is necessary */
3534 congestion_wait(BLK_RW_ASYNC, HZ/10);
3542 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3543 char *buf, size_t nbytes,
3546 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3548 if (mem_cgroup_is_root(memcg))
3550 return mem_cgroup_force_empty(memcg) ?: nbytes;
3553 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3559 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3560 struct cftype *cft, u64 val)
3565 pr_warn_once("Non-hierarchical mode is deprecated. "
3566 "Please report your usecase to linux-mm@kvack.org if you "
3567 "depend on this functionality.\n");
3572 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3576 if (mem_cgroup_is_root(memcg)) {
3577 /* mem_cgroup_threshold() calls here from irqsafe context */
3578 cgroup_rstat_flush_irqsafe(memcg->css.cgroup);
3579 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3580 memcg_page_state(memcg, NR_ANON_MAPPED);
3582 val += memcg_page_state(memcg, MEMCG_SWAP);
3585 val = page_counter_read(&memcg->memory);
3587 val = page_counter_read(&memcg->memsw);
3600 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3603 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3604 struct page_counter *counter;
3606 switch (MEMFILE_TYPE(cft->private)) {
3608 counter = &memcg->memory;
3611 counter = &memcg->memsw;
3614 counter = &memcg->kmem;
3617 counter = &memcg->tcpmem;
3623 switch (MEMFILE_ATTR(cft->private)) {
3625 if (counter == &memcg->memory)
3626 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3627 if (counter == &memcg->memsw)
3628 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3629 return (u64)page_counter_read(counter) * PAGE_SIZE;
3631 return (u64)counter->max * PAGE_SIZE;
3633 return (u64)counter->watermark * PAGE_SIZE;
3635 return counter->failcnt;
3636 case RES_SOFT_LIMIT:
3637 return (u64)memcg->soft_limit * PAGE_SIZE;
3643 #ifdef CONFIG_MEMCG_KMEM
3644 static int memcg_online_kmem(struct mem_cgroup *memcg)
3646 struct obj_cgroup *objcg;
3649 if (cgroup_memory_nokmem)
3652 BUG_ON(memcg->kmemcg_id >= 0);
3653 BUG_ON(memcg->kmem_state);
3655 memcg_id = memcg_alloc_cache_id();
3659 objcg = obj_cgroup_alloc();
3661 memcg_free_cache_id(memcg_id);
3664 objcg->memcg = memcg;
3665 rcu_assign_pointer(memcg->objcg, objcg);
3667 static_branch_enable(&memcg_kmem_enabled_key);
3669 memcg->kmemcg_id = memcg_id;
3670 memcg->kmem_state = KMEM_ONLINE;
3675 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3677 struct cgroup_subsys_state *css;
3678 struct mem_cgroup *parent, *child;
3681 if (memcg->kmem_state != KMEM_ONLINE)
3684 memcg->kmem_state = KMEM_ALLOCATED;
3686 parent = parent_mem_cgroup(memcg);
3688 parent = root_mem_cgroup;
3690 memcg_reparent_objcgs(memcg, parent);
3692 kmemcg_id = memcg->kmemcg_id;
3693 BUG_ON(kmemcg_id < 0);
3696 * Change kmemcg_id of this cgroup and all its descendants to the
3697 * parent's id, and then move all entries from this cgroup's list_lrus
3698 * to ones of the parent. After we have finished, all list_lrus
3699 * corresponding to this cgroup are guaranteed to remain empty. The
3700 * ordering is imposed by list_lru_node->lock taken by
3701 * memcg_drain_all_list_lrus().
3703 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3704 css_for_each_descendant_pre(css, &memcg->css) {
3705 child = mem_cgroup_from_css(css);
3706 BUG_ON(child->kmemcg_id != kmemcg_id);
3707 child->kmemcg_id = parent->kmemcg_id;
3711 memcg_drain_all_list_lrus(kmemcg_id, parent);
3713 memcg_free_cache_id(kmemcg_id);
3716 static void memcg_free_kmem(struct mem_cgroup *memcg)
3718 /* css_alloc() failed, offlining didn't happen */
3719 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3720 memcg_offline_kmem(memcg);
3723 static int memcg_online_kmem(struct mem_cgroup *memcg)
3727 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3730 static void memcg_free_kmem(struct mem_cgroup *memcg)
3733 #endif /* CONFIG_MEMCG_KMEM */
3735 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3740 mutex_lock(&memcg_max_mutex);
3741 ret = page_counter_set_max(&memcg->kmem, max);
3742 mutex_unlock(&memcg_max_mutex);
3746 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3750 mutex_lock(&memcg_max_mutex);
3752 ret = page_counter_set_max(&memcg->tcpmem, max);
3756 if (!memcg->tcpmem_active) {
3758 * The active flag needs to be written after the static_key
3759 * update. This is what guarantees that the socket activation
3760 * function is the last one to run. See mem_cgroup_sk_alloc()
3761 * for details, and note that we don't mark any socket as
3762 * belonging to this memcg until that flag is up.
3764 * We need to do this, because static_keys will span multiple
3765 * sites, but we can't control their order. If we mark a socket
3766 * as accounted, but the accounting functions are not patched in
3767 * yet, we'll lose accounting.
3769 * We never race with the readers in mem_cgroup_sk_alloc(),
3770 * because when this value change, the code to process it is not
3773 static_branch_inc(&memcg_sockets_enabled_key);
3774 memcg->tcpmem_active = true;
3777 mutex_unlock(&memcg_max_mutex);
3782 * The user of this function is...
3785 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3786 char *buf, size_t nbytes, loff_t off)
3788 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3789 unsigned long nr_pages;
3792 buf = strstrip(buf);
3793 ret = page_counter_memparse(buf, "-1", &nr_pages);
3797 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3799 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3803 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3805 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3808 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3811 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3812 "Please report your usecase to linux-mm@kvack.org if you "
3813 "depend on this functionality.\n");
3814 ret = memcg_update_kmem_max(memcg, nr_pages);
3817 ret = memcg_update_tcp_max(memcg, nr_pages);
3821 case RES_SOFT_LIMIT:
3822 memcg->soft_limit = nr_pages;
3826 return ret ?: nbytes;
3829 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3830 size_t nbytes, loff_t off)
3832 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3833 struct page_counter *counter;
3835 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3837 counter = &memcg->memory;
3840 counter = &memcg->memsw;
3843 counter = &memcg->kmem;
3846 counter = &memcg->tcpmem;
3852 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3854 page_counter_reset_watermark(counter);
3857 counter->failcnt = 0;
3866 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3869 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3873 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3874 struct cftype *cft, u64 val)
3876 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3878 if (val & ~MOVE_MASK)
3882 * No kind of locking is needed in here, because ->can_attach() will
3883 * check this value once in the beginning of the process, and then carry
3884 * on with stale data. This means that changes to this value will only
3885 * affect task migrations starting after the change.
3887 memcg->move_charge_at_immigrate = val;
3891 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3892 struct cftype *cft, u64 val)
3900 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3901 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3902 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3904 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3905 int nid, unsigned int lru_mask, bool tree)
3907 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3908 unsigned long nr = 0;
3911 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3914 if (!(BIT(lru) & lru_mask))
3917 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3919 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3924 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3925 unsigned int lru_mask,
3928 unsigned long nr = 0;
3932 if (!(BIT(lru) & lru_mask))
3935 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3937 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3942 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3946 unsigned int lru_mask;
3949 static const struct numa_stat stats[] = {
3950 { "total", LRU_ALL },
3951 { "file", LRU_ALL_FILE },
3952 { "anon", LRU_ALL_ANON },
3953 { "unevictable", BIT(LRU_UNEVICTABLE) },
3955 const struct numa_stat *stat;
3957 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3959 cgroup_rstat_flush(memcg->css.cgroup);
3961 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3962 seq_printf(m, "%s=%lu", stat->name,
3963 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3965 for_each_node_state(nid, N_MEMORY)
3966 seq_printf(m, " N%d=%lu", nid,
3967 mem_cgroup_node_nr_lru_pages(memcg, nid,
3968 stat->lru_mask, false));
3972 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3974 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3975 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3977 for_each_node_state(nid, N_MEMORY)
3978 seq_printf(m, " N%d=%lu", nid,
3979 mem_cgroup_node_nr_lru_pages(memcg, nid,
3980 stat->lru_mask, true));
3986 #endif /* CONFIG_NUMA */
3988 static const unsigned int memcg1_stats[] = {
3991 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4001 static const char *const memcg1_stat_names[] = {
4004 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4014 /* Universal VM events cgroup1 shows, original sort order */
4015 static const unsigned int memcg1_events[] = {
4022 static int memcg_stat_show(struct seq_file *m, void *v)
4024 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4025 unsigned long memory, memsw;
4026 struct mem_cgroup *mi;
4029 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4031 cgroup_rstat_flush(memcg->css.cgroup);
4033 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4036 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4038 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4039 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4042 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4043 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4044 memcg_events_local(memcg, memcg1_events[i]));
4046 for (i = 0; i < NR_LRU_LISTS; i++)
4047 seq_printf(m, "%s %lu\n", lru_list_name(i),
4048 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4051 /* Hierarchical information */
4052 memory = memsw = PAGE_COUNTER_MAX;
4053 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4054 memory = min(memory, READ_ONCE(mi->memory.max));
4055 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4057 seq_printf(m, "hierarchical_memory_limit %llu\n",
4058 (u64)memory * PAGE_SIZE);
4059 if (do_memsw_account())
4060 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4061 (u64)memsw * PAGE_SIZE);
4063 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4066 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4068 nr = memcg_page_state(memcg, memcg1_stats[i]);
4069 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4070 (u64)nr * PAGE_SIZE);
4073 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4074 seq_printf(m, "total_%s %llu\n",
4075 vm_event_name(memcg1_events[i]),
4076 (u64)memcg_events(memcg, memcg1_events[i]));
4078 for (i = 0; i < NR_LRU_LISTS; i++)
4079 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4080 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4083 #ifdef CONFIG_DEBUG_VM
4086 struct mem_cgroup_per_node *mz;
4087 unsigned long anon_cost = 0;
4088 unsigned long file_cost = 0;
4090 for_each_online_pgdat(pgdat) {
4091 mz = memcg->nodeinfo[pgdat->node_id];
4093 anon_cost += mz->lruvec.anon_cost;
4094 file_cost += mz->lruvec.file_cost;
4096 seq_printf(m, "anon_cost %lu\n", anon_cost);
4097 seq_printf(m, "file_cost %lu\n", file_cost);
4104 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4107 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4109 return mem_cgroup_swappiness(memcg);
4112 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4113 struct cftype *cft, u64 val)
4115 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4120 if (!mem_cgroup_is_root(memcg))
4121 memcg->swappiness = val;
4123 vm_swappiness = val;
4128 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4130 struct mem_cgroup_threshold_ary *t;
4131 unsigned long usage;
4136 t = rcu_dereference(memcg->thresholds.primary);
4138 t = rcu_dereference(memcg->memsw_thresholds.primary);
4143 usage = mem_cgroup_usage(memcg, swap);
4146 * current_threshold points to threshold just below or equal to usage.
4147 * If it's not true, a threshold was crossed after last
4148 * call of __mem_cgroup_threshold().
4150 i = t->current_threshold;
4153 * Iterate backward over array of thresholds starting from
4154 * current_threshold and check if a threshold is crossed.
4155 * If none of thresholds below usage is crossed, we read
4156 * only one element of the array here.
4158 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4159 eventfd_signal(t->entries[i].eventfd, 1);
4161 /* i = current_threshold + 1 */
4165 * Iterate forward over array of thresholds starting from
4166 * current_threshold+1 and check if a threshold is crossed.
4167 * If none of thresholds above usage is crossed, we read
4168 * only one element of the array here.
4170 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4171 eventfd_signal(t->entries[i].eventfd, 1);
4173 /* Update current_threshold */
4174 t->current_threshold = i - 1;
4179 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4182 __mem_cgroup_threshold(memcg, false);
4183 if (do_memsw_account())
4184 __mem_cgroup_threshold(memcg, true);
4186 memcg = parent_mem_cgroup(memcg);
4190 static int compare_thresholds(const void *a, const void *b)
4192 const struct mem_cgroup_threshold *_a = a;
4193 const struct mem_cgroup_threshold *_b = b;
4195 if (_a->threshold > _b->threshold)
4198 if (_a->threshold < _b->threshold)
4204 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4206 struct mem_cgroup_eventfd_list *ev;
4208 spin_lock(&memcg_oom_lock);
4210 list_for_each_entry(ev, &memcg->oom_notify, list)
4211 eventfd_signal(ev->eventfd, 1);
4213 spin_unlock(&memcg_oom_lock);
4217 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4219 struct mem_cgroup *iter;
4221 for_each_mem_cgroup_tree(iter, memcg)
4222 mem_cgroup_oom_notify_cb(iter);
4225 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4226 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4228 struct mem_cgroup_thresholds *thresholds;
4229 struct mem_cgroup_threshold_ary *new;
4230 unsigned long threshold;
4231 unsigned long usage;
4234 ret = page_counter_memparse(args, "-1", &threshold);
4238 mutex_lock(&memcg->thresholds_lock);
4241 thresholds = &memcg->thresholds;
4242 usage = mem_cgroup_usage(memcg, false);
4243 } else if (type == _MEMSWAP) {
4244 thresholds = &memcg->memsw_thresholds;
4245 usage = mem_cgroup_usage(memcg, true);
4249 /* Check if a threshold crossed before adding a new one */
4250 if (thresholds->primary)
4251 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4253 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4255 /* Allocate memory for new array of thresholds */
4256 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4263 /* Copy thresholds (if any) to new array */
4264 if (thresholds->primary)
4265 memcpy(new->entries, thresholds->primary->entries,
4266 flex_array_size(new, entries, size - 1));
4268 /* Add new threshold */
4269 new->entries[size - 1].eventfd = eventfd;
4270 new->entries[size - 1].threshold = threshold;
4272 /* Sort thresholds. Registering of new threshold isn't time-critical */
4273 sort(new->entries, size, sizeof(*new->entries),
4274 compare_thresholds, NULL);
4276 /* Find current threshold */
4277 new->current_threshold = -1;
4278 for (i = 0; i < size; i++) {
4279 if (new->entries[i].threshold <= usage) {
4281 * new->current_threshold will not be used until
4282 * rcu_assign_pointer(), so it's safe to increment
4285 ++new->current_threshold;
4290 /* Free old spare buffer and save old primary buffer as spare */
4291 kfree(thresholds->spare);
4292 thresholds->spare = thresholds->primary;
4294 rcu_assign_pointer(thresholds->primary, new);
4296 /* To be sure that nobody uses thresholds */
4300 mutex_unlock(&memcg->thresholds_lock);
4305 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4306 struct eventfd_ctx *eventfd, const char *args)
4308 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4311 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4312 struct eventfd_ctx *eventfd, const char *args)
4314 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4317 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4318 struct eventfd_ctx *eventfd, enum res_type type)
4320 struct mem_cgroup_thresholds *thresholds;
4321 struct mem_cgroup_threshold_ary *new;
4322 unsigned long usage;
4323 int i, j, size, entries;
4325 mutex_lock(&memcg->thresholds_lock);
4328 thresholds = &memcg->thresholds;
4329 usage = mem_cgroup_usage(memcg, false);
4330 } else if (type == _MEMSWAP) {
4331 thresholds = &memcg->memsw_thresholds;
4332 usage = mem_cgroup_usage(memcg, true);
4336 if (!thresholds->primary)
4339 /* Check if a threshold crossed before removing */
4340 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4342 /* Calculate new number of threshold */
4344 for (i = 0; i < thresholds->primary->size; i++) {
4345 if (thresholds->primary->entries[i].eventfd != eventfd)
4351 new = thresholds->spare;
4353 /* If no items related to eventfd have been cleared, nothing to do */
4357 /* Set thresholds array to NULL if we don't have thresholds */
4366 /* Copy thresholds and find current threshold */
4367 new->current_threshold = -1;
4368 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4369 if (thresholds->primary->entries[i].eventfd == eventfd)
4372 new->entries[j] = thresholds->primary->entries[i];
4373 if (new->entries[j].threshold <= usage) {
4375 * new->current_threshold will not be used
4376 * until rcu_assign_pointer(), so it's safe to increment
4379 ++new->current_threshold;
4385 /* Swap primary and spare array */
4386 thresholds->spare = thresholds->primary;
4388 rcu_assign_pointer(thresholds->primary, new);
4390 /* To be sure that nobody uses thresholds */
4393 /* If all events are unregistered, free the spare array */
4395 kfree(thresholds->spare);
4396 thresholds->spare = NULL;
4399 mutex_unlock(&memcg->thresholds_lock);
4402 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4403 struct eventfd_ctx *eventfd)
4405 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4408 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4409 struct eventfd_ctx *eventfd)
4411 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4414 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4415 struct eventfd_ctx *eventfd, const char *args)
4417 struct mem_cgroup_eventfd_list *event;
4419 event = kmalloc(sizeof(*event), GFP_KERNEL);
4423 spin_lock(&memcg_oom_lock);
4425 event->eventfd = eventfd;
4426 list_add(&event->list, &memcg->oom_notify);
4428 /* already in OOM ? */
4429 if (memcg->under_oom)
4430 eventfd_signal(eventfd, 1);
4431 spin_unlock(&memcg_oom_lock);
4436 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4437 struct eventfd_ctx *eventfd)
4439 struct mem_cgroup_eventfd_list *ev, *tmp;
4441 spin_lock(&memcg_oom_lock);
4443 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4444 if (ev->eventfd == eventfd) {
4445 list_del(&ev->list);
4450 spin_unlock(&memcg_oom_lock);
4453 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4455 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4457 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4458 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4459 seq_printf(sf, "oom_kill %lu\n",
4460 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4464 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4465 struct cftype *cft, u64 val)
4467 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4469 /* cannot set to root cgroup and only 0 and 1 are allowed */
4470 if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4473 memcg->oom_kill_disable = val;
4475 memcg_oom_recover(memcg);
4480 #ifdef CONFIG_CGROUP_WRITEBACK
4482 #include <trace/events/writeback.h>
4484 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4486 return wb_domain_init(&memcg->cgwb_domain, gfp);
4489 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4491 wb_domain_exit(&memcg->cgwb_domain);
4494 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4496 wb_domain_size_changed(&memcg->cgwb_domain);
4499 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4501 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4503 if (!memcg->css.parent)
4506 return &memcg->cgwb_domain;
4510 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4511 * @wb: bdi_writeback in question
4512 * @pfilepages: out parameter for number of file pages
4513 * @pheadroom: out parameter for number of allocatable pages according to memcg
4514 * @pdirty: out parameter for number of dirty pages
4515 * @pwriteback: out parameter for number of pages under writeback
4517 * Determine the numbers of file, headroom, dirty, and writeback pages in
4518 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4519 * is a bit more involved.
4521 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4522 * headroom is calculated as the lowest headroom of itself and the
4523 * ancestors. Note that this doesn't consider the actual amount of
4524 * available memory in the system. The caller should further cap
4525 * *@pheadroom accordingly.
4527 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4528 unsigned long *pheadroom, unsigned long *pdirty,
4529 unsigned long *pwriteback)
4531 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4532 struct mem_cgroup *parent;
4534 cgroup_rstat_flush_irqsafe(memcg->css.cgroup);
4536 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4537 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4538 *pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4539 memcg_page_state(memcg, NR_ACTIVE_FILE);
4541 *pheadroom = PAGE_COUNTER_MAX;
4542 while ((parent = parent_mem_cgroup(memcg))) {
4543 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4544 READ_ONCE(memcg->memory.high));
4545 unsigned long used = page_counter_read(&memcg->memory);
4547 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4553 * Foreign dirty flushing
4555 * There's an inherent mismatch between memcg and writeback. The former
4556 * tracks ownership per-page while the latter per-inode. This was a
4557 * deliberate design decision because honoring per-page ownership in the
4558 * writeback path is complicated, may lead to higher CPU and IO overheads
4559 * and deemed unnecessary given that write-sharing an inode across
4560 * different cgroups isn't a common use-case.
4562 * Combined with inode majority-writer ownership switching, this works well
4563 * enough in most cases but there are some pathological cases. For
4564 * example, let's say there are two cgroups A and B which keep writing to
4565 * different but confined parts of the same inode. B owns the inode and
4566 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4567 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4568 * triggering background writeback. A will be slowed down without a way to
4569 * make writeback of the dirty pages happen.
4571 * Conditions like the above can lead to a cgroup getting repeatedly and
4572 * severely throttled after making some progress after each
4573 * dirty_expire_interval while the underlying IO device is almost
4576 * Solving this problem completely requires matching the ownership tracking
4577 * granularities between memcg and writeback in either direction. However,
4578 * the more egregious behaviors can be avoided by simply remembering the
4579 * most recent foreign dirtying events and initiating remote flushes on
4580 * them when local writeback isn't enough to keep the memory clean enough.
4582 * The following two functions implement such mechanism. When a foreign
4583 * page - a page whose memcg and writeback ownerships don't match - is
4584 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4585 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4586 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4587 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4588 * foreign bdi_writebacks which haven't expired. Both the numbers of
4589 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4590 * limited to MEMCG_CGWB_FRN_CNT.
4592 * The mechanism only remembers IDs and doesn't hold any object references.
4593 * As being wrong occasionally doesn't matter, updates and accesses to the
4594 * records are lockless and racy.
4596 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4597 struct bdi_writeback *wb)
4599 struct mem_cgroup *memcg = page_memcg(page);
4600 struct memcg_cgwb_frn *frn;
4601 u64 now = get_jiffies_64();
4602 u64 oldest_at = now;
4606 trace_track_foreign_dirty(page, wb);
4609 * Pick the slot to use. If there is already a slot for @wb, keep
4610 * using it. If not replace the oldest one which isn't being
4613 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4614 frn = &memcg->cgwb_frn[i];
4615 if (frn->bdi_id == wb->bdi->id &&
4616 frn->memcg_id == wb->memcg_css->id)
4618 if (time_before64(frn->at, oldest_at) &&
4619 atomic_read(&frn->done.cnt) == 1) {
4621 oldest_at = frn->at;
4625 if (i < MEMCG_CGWB_FRN_CNT) {
4627 * Re-using an existing one. Update timestamp lazily to
4628 * avoid making the cacheline hot. We want them to be
4629 * reasonably up-to-date and significantly shorter than
4630 * dirty_expire_interval as that's what expires the record.
4631 * Use the shorter of 1s and dirty_expire_interval / 8.
4633 unsigned long update_intv =
4634 min_t(unsigned long, HZ,
4635 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4637 if (time_before64(frn->at, now - update_intv))
4639 } else if (oldest >= 0) {
4640 /* replace the oldest free one */
4641 frn = &memcg->cgwb_frn[oldest];
4642 frn->bdi_id = wb->bdi->id;
4643 frn->memcg_id = wb->memcg_css->id;
4648 /* issue foreign writeback flushes for recorded foreign dirtying events */
4649 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4651 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4652 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4653 u64 now = jiffies_64;
4656 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4657 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4660 * If the record is older than dirty_expire_interval,
4661 * writeback on it has already started. No need to kick it
4662 * off again. Also, don't start a new one if there's
4663 * already one in flight.
4665 if (time_after64(frn->at, now - intv) &&
4666 atomic_read(&frn->done.cnt) == 1) {
4668 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4669 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4670 WB_REASON_FOREIGN_FLUSH,
4676 #else /* CONFIG_CGROUP_WRITEBACK */
4678 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4683 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4687 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4691 #endif /* CONFIG_CGROUP_WRITEBACK */
4694 * DO NOT USE IN NEW FILES.
4696 * "cgroup.event_control" implementation.
4698 * This is way over-engineered. It tries to support fully configurable
4699 * events for each user. Such level of flexibility is completely
4700 * unnecessary especially in the light of the planned unified hierarchy.
4702 * Please deprecate this and replace with something simpler if at all
4707 * Unregister event and free resources.
4709 * Gets called from workqueue.
4711 static void memcg_event_remove(struct work_struct *work)
4713 struct mem_cgroup_event *event =
4714 container_of(work, struct mem_cgroup_event, remove);
4715 struct mem_cgroup *memcg = event->memcg;
4717 remove_wait_queue(event->wqh, &event->wait);
4719 event->unregister_event(memcg, event->eventfd);
4721 /* Notify userspace the event is going away. */
4722 eventfd_signal(event->eventfd, 1);
4724 eventfd_ctx_put(event->eventfd);
4726 css_put(&memcg->css);
4730 * Gets called on EPOLLHUP on eventfd when user closes it.
4732 * Called with wqh->lock held and interrupts disabled.
4734 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4735 int sync, void *key)
4737 struct mem_cgroup_event *event =
4738 container_of(wait, struct mem_cgroup_event, wait);
4739 struct mem_cgroup *memcg = event->memcg;
4740 __poll_t flags = key_to_poll(key);
4742 if (flags & EPOLLHUP) {
4744 * If the event has been detached at cgroup removal, we
4745 * can simply return knowing the other side will cleanup
4748 * We can't race against event freeing since the other
4749 * side will require wqh->lock via remove_wait_queue(),
4752 spin_lock(&memcg->event_list_lock);
4753 if (!list_empty(&event->list)) {
4754 list_del_init(&event->list);
4756 * We are in atomic context, but cgroup_event_remove()
4757 * may sleep, so we have to call it in workqueue.
4759 schedule_work(&event->remove);
4761 spin_unlock(&memcg->event_list_lock);
4767 static void memcg_event_ptable_queue_proc(struct file *file,
4768 wait_queue_head_t *wqh, poll_table *pt)
4770 struct mem_cgroup_event *event =
4771 container_of(pt, struct mem_cgroup_event, pt);
4774 add_wait_queue(wqh, &event->wait);
4778 * DO NOT USE IN NEW FILES.
4780 * Parse input and register new cgroup event handler.
4782 * Input must be in format '<event_fd> <control_fd> <args>'.
4783 * Interpretation of args is defined by control file implementation.
4785 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4786 char *buf, size_t nbytes, loff_t off)
4788 struct cgroup_subsys_state *css = of_css(of);
4789 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4790 struct mem_cgroup_event *event;
4791 struct cgroup_subsys_state *cfile_css;
4792 unsigned int efd, cfd;
4799 buf = strstrip(buf);
4801 efd = simple_strtoul(buf, &endp, 10);
4806 cfd = simple_strtoul(buf, &endp, 10);
4807 if ((*endp != ' ') && (*endp != '\0'))
4811 event = kzalloc(sizeof(*event), GFP_KERNEL);
4815 event->memcg = memcg;
4816 INIT_LIST_HEAD(&event->list);
4817 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4818 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4819 INIT_WORK(&event->remove, memcg_event_remove);
4827 event->eventfd = eventfd_ctx_fileget(efile.file);
4828 if (IS_ERR(event->eventfd)) {
4829 ret = PTR_ERR(event->eventfd);
4836 goto out_put_eventfd;
4839 /* the process need read permission on control file */
4840 /* AV: shouldn't we check that it's been opened for read instead? */
4841 ret = file_permission(cfile.file, MAY_READ);
4846 * Determine the event callbacks and set them in @event. This used
4847 * to be done via struct cftype but cgroup core no longer knows
4848 * about these events. The following is crude but the whole thing
4849 * is for compatibility anyway.
4851 * DO NOT ADD NEW FILES.
4853 name = cfile.file->f_path.dentry->d_name.name;
4855 if (!strcmp(name, "memory.usage_in_bytes")) {
4856 event->register_event = mem_cgroup_usage_register_event;
4857 event->unregister_event = mem_cgroup_usage_unregister_event;
4858 } else if (!strcmp(name, "memory.oom_control")) {
4859 event->register_event = mem_cgroup_oom_register_event;
4860 event->unregister_event = mem_cgroup_oom_unregister_event;
4861 } else if (!strcmp(name, "memory.pressure_level")) {
4862 event->register_event = vmpressure_register_event;
4863 event->unregister_event = vmpressure_unregister_event;
4864 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4865 event->register_event = memsw_cgroup_usage_register_event;
4866 event->unregister_event = memsw_cgroup_usage_unregister_event;
4873 * Verify @cfile should belong to @css. Also, remaining events are
4874 * automatically removed on cgroup destruction but the removal is
4875 * asynchronous, so take an extra ref on @css.
4877 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4878 &memory_cgrp_subsys);
4880 if (IS_ERR(cfile_css))
4882 if (cfile_css != css) {
4887 ret = event->register_event(memcg, event->eventfd, buf);
4891 vfs_poll(efile.file, &event->pt);
4893 spin_lock(&memcg->event_list_lock);
4894 list_add(&event->list, &memcg->event_list);
4895 spin_unlock(&memcg->event_list_lock);
4907 eventfd_ctx_put(event->eventfd);
4916 static struct cftype mem_cgroup_legacy_files[] = {
4918 .name = "usage_in_bytes",
4919 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4920 .read_u64 = mem_cgroup_read_u64,
4923 .name = "max_usage_in_bytes",
4924 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4925 .write = mem_cgroup_reset,
4926 .read_u64 = mem_cgroup_read_u64,
4929 .name = "limit_in_bytes",
4930 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4931 .write = mem_cgroup_write,
4932 .read_u64 = mem_cgroup_read_u64,
4935 .name = "soft_limit_in_bytes",
4936 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4937 .write = mem_cgroup_write,
4938 .read_u64 = mem_cgroup_read_u64,
4942 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4943 .write = mem_cgroup_reset,
4944 .read_u64 = mem_cgroup_read_u64,
4948 .seq_show = memcg_stat_show,
4951 .name = "force_empty",
4952 .write = mem_cgroup_force_empty_write,
4955 .name = "use_hierarchy",
4956 .write_u64 = mem_cgroup_hierarchy_write,
4957 .read_u64 = mem_cgroup_hierarchy_read,
4960 .name = "cgroup.event_control", /* XXX: for compat */
4961 .write = memcg_write_event_control,
4962 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4965 .name = "swappiness",
4966 .read_u64 = mem_cgroup_swappiness_read,
4967 .write_u64 = mem_cgroup_swappiness_write,
4970 .name = "move_charge_at_immigrate",
4971 .read_u64 = mem_cgroup_move_charge_read,
4972 .write_u64 = mem_cgroup_move_charge_write,
4975 .name = "oom_control",
4976 .seq_show = mem_cgroup_oom_control_read,
4977 .write_u64 = mem_cgroup_oom_control_write,
4978 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4981 .name = "pressure_level",
4985 .name = "numa_stat",
4986 .seq_show = memcg_numa_stat_show,
4990 .name = "kmem.limit_in_bytes",
4991 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4992 .write = mem_cgroup_write,
4993 .read_u64 = mem_cgroup_read_u64,
4996 .name = "kmem.usage_in_bytes",
4997 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4998 .read_u64 = mem_cgroup_read_u64,
5001 .name = "kmem.failcnt",
5002 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5003 .write = mem_cgroup_reset,
5004 .read_u64 = mem_cgroup_read_u64,
5007 .name = "kmem.max_usage_in_bytes",
5008 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5009 .write = mem_cgroup_reset,
5010 .read_u64 = mem_cgroup_read_u64,
5012 #if defined(CONFIG_MEMCG_KMEM) && \
5013 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5015 .name = "kmem.slabinfo",
5016 .seq_show = memcg_slab_show,
5020 .name = "kmem.tcp.limit_in_bytes",
5021 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5022 .write = mem_cgroup_write,
5023 .read_u64 = mem_cgroup_read_u64,
5026 .name = "kmem.tcp.usage_in_bytes",
5027 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5028 .read_u64 = mem_cgroup_read_u64,
5031 .name = "kmem.tcp.failcnt",
5032 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5033 .write = mem_cgroup_reset,
5034 .read_u64 = mem_cgroup_read_u64,
5037 .name = "kmem.tcp.max_usage_in_bytes",
5038 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5039 .write = mem_cgroup_reset,
5040 .read_u64 = mem_cgroup_read_u64,
5042 { }, /* terminate */
5046 * Private memory cgroup IDR
5048 * Swap-out records and page cache shadow entries need to store memcg
5049 * references in constrained space, so we maintain an ID space that is
5050 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5051 * memory-controlled cgroups to 64k.
5053 * However, there usually are many references to the offline CSS after
5054 * the cgroup has been destroyed, such as page cache or reclaimable
5055 * slab objects, that don't need to hang on to the ID. We want to keep
5056 * those dead CSS from occupying IDs, or we might quickly exhaust the
5057 * relatively small ID space and prevent the creation of new cgroups
5058 * even when there are much fewer than 64k cgroups - possibly none.
5060 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5061 * be freed and recycled when it's no longer needed, which is usually
5062 * when the CSS is offlined.
5064 * The only exception to that are records of swapped out tmpfs/shmem
5065 * pages that need to be attributed to live ancestors on swapin. But
5066 * those references are manageable from userspace.
5069 static DEFINE_IDR(mem_cgroup_idr);
5071 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5073 if (memcg->id.id > 0) {
5074 idr_remove(&mem_cgroup_idr, memcg->id.id);
5079 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5082 refcount_add(n, &memcg->id.ref);
5085 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5087 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5088 mem_cgroup_id_remove(memcg);
5090 /* Memcg ID pins CSS */
5091 css_put(&memcg->css);
5095 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5097 mem_cgroup_id_put_many(memcg, 1);
5101 * mem_cgroup_from_id - look up a memcg from a memcg id
5102 * @id: the memcg id to look up
5104 * Caller must hold rcu_read_lock().
5106 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5108 WARN_ON_ONCE(!rcu_read_lock_held());
5109 return idr_find(&mem_cgroup_idr, id);
5112 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5114 struct mem_cgroup_per_node *pn;
5117 * This routine is called against possible nodes.
5118 * But it's BUG to call kmalloc() against offline node.
5120 * TODO: this routine can waste much memory for nodes which will
5121 * never be onlined. It's better to use memory hotplug callback
5124 if (!node_state(node, N_NORMAL_MEMORY))
5126 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5130 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5131 GFP_KERNEL_ACCOUNT);
5132 if (!pn->lruvec_stat_local) {
5137 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct batched_lruvec_stat,
5138 GFP_KERNEL_ACCOUNT);
5139 if (!pn->lruvec_stat_cpu) {
5140 free_percpu(pn->lruvec_stat_local);
5145 lruvec_init(&pn->lruvec);
5146 pn->usage_in_excess = 0;
5147 pn->on_tree = false;
5150 memcg->nodeinfo[node] = pn;
5154 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5156 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5161 free_percpu(pn->lruvec_stat_cpu);
5162 free_percpu(pn->lruvec_stat_local);
5166 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5171 free_mem_cgroup_per_node_info(memcg, node);
5172 free_percpu(memcg->vmstats_percpu);
5176 static void mem_cgroup_free(struct mem_cgroup *memcg)
5180 memcg_wb_domain_exit(memcg);
5182 * Flush percpu lruvec stats to guarantee the value
5183 * correctness on parent's and all ancestor levels.
5185 for_each_online_cpu(cpu)
5186 memcg_flush_lruvec_page_state(memcg, cpu);
5187 __mem_cgroup_free(memcg);
5190 static struct mem_cgroup *mem_cgroup_alloc(void)
5192 struct mem_cgroup *memcg;
5195 int __maybe_unused i;
5196 long error = -ENOMEM;
5198 size = sizeof(struct mem_cgroup);
5199 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5201 memcg = kzalloc(size, GFP_KERNEL);
5203 return ERR_PTR(error);
5205 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5206 1, MEM_CGROUP_ID_MAX,
5208 if (memcg->id.id < 0) {
5209 error = memcg->id.id;
5213 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5214 GFP_KERNEL_ACCOUNT);
5215 if (!memcg->vmstats_percpu)
5219 if (alloc_mem_cgroup_per_node_info(memcg, node))
5222 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5225 INIT_WORK(&memcg->high_work, high_work_func);
5226 INIT_LIST_HEAD(&memcg->oom_notify);
5227 mutex_init(&memcg->thresholds_lock);
5228 spin_lock_init(&memcg->move_lock);
5229 vmpressure_init(&memcg->vmpressure);
5230 INIT_LIST_HEAD(&memcg->event_list);
5231 spin_lock_init(&memcg->event_list_lock);
5232 memcg->socket_pressure = jiffies;
5233 #ifdef CONFIG_MEMCG_KMEM
5234 memcg->kmemcg_id = -1;
5235 INIT_LIST_HEAD(&memcg->objcg_list);
5237 #ifdef CONFIG_CGROUP_WRITEBACK
5238 INIT_LIST_HEAD(&memcg->cgwb_list);
5239 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5240 memcg->cgwb_frn[i].done =
5241 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5243 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5244 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5245 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5246 memcg->deferred_split_queue.split_queue_len = 0;
5248 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5251 mem_cgroup_id_remove(memcg);
5252 __mem_cgroup_free(memcg);
5253 return ERR_PTR(error);
5256 static struct cgroup_subsys_state * __ref
5257 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5259 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5260 struct mem_cgroup *memcg, *old_memcg;
5261 long error = -ENOMEM;
5263 old_memcg = set_active_memcg(parent);
5264 memcg = mem_cgroup_alloc();
5265 set_active_memcg(old_memcg);
5267 return ERR_CAST(memcg);
5269 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5270 memcg->soft_limit = PAGE_COUNTER_MAX;
5271 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5273 memcg->swappiness = mem_cgroup_swappiness(parent);
5274 memcg->oom_kill_disable = parent->oom_kill_disable;
5276 page_counter_init(&memcg->memory, &parent->memory);
5277 page_counter_init(&memcg->swap, &parent->swap);
5278 page_counter_init(&memcg->kmem, &parent->kmem);
5279 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5281 page_counter_init(&memcg->memory, NULL);
5282 page_counter_init(&memcg->swap, NULL);
5283 page_counter_init(&memcg->kmem, NULL);
5284 page_counter_init(&memcg->tcpmem, NULL);
5286 root_mem_cgroup = memcg;
5290 /* The following stuff does not apply to the root */
5291 error = memcg_online_kmem(memcg);
5295 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5296 static_branch_inc(&memcg_sockets_enabled_key);
5300 mem_cgroup_id_remove(memcg);
5301 mem_cgroup_free(memcg);
5302 return ERR_PTR(error);
5305 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5307 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5310 * A memcg must be visible for expand_shrinker_info()
5311 * by the time the maps are allocated. So, we allocate maps
5312 * here, when for_each_mem_cgroup() can't skip it.
5314 if (alloc_shrinker_info(memcg)) {
5315 mem_cgroup_id_remove(memcg);
5319 /* Online state pins memcg ID, memcg ID pins CSS */
5320 refcount_set(&memcg->id.ref, 1);
5325 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5327 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5328 struct mem_cgroup_event *event, *tmp;
5331 * Unregister events and notify userspace.
5332 * Notify userspace about cgroup removing only after rmdir of cgroup
5333 * directory to avoid race between userspace and kernelspace.
5335 spin_lock(&memcg->event_list_lock);
5336 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5337 list_del_init(&event->list);
5338 schedule_work(&event->remove);
5340 spin_unlock(&memcg->event_list_lock);
5342 page_counter_set_min(&memcg->memory, 0);
5343 page_counter_set_low(&memcg->memory, 0);
5345 memcg_offline_kmem(memcg);
5346 reparent_shrinker_deferred(memcg);
5347 wb_memcg_offline(memcg);
5349 drain_all_stock(memcg);
5351 mem_cgroup_id_put(memcg);
5354 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5356 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5358 invalidate_reclaim_iterators(memcg);
5361 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5363 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5364 int __maybe_unused i;
5366 #ifdef CONFIG_CGROUP_WRITEBACK
5367 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5368 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5370 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5371 static_branch_dec(&memcg_sockets_enabled_key);
5373 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5374 static_branch_dec(&memcg_sockets_enabled_key);
5376 vmpressure_cleanup(&memcg->vmpressure);
5377 cancel_work_sync(&memcg->high_work);
5378 mem_cgroup_remove_from_trees(memcg);
5379 free_shrinker_info(memcg);
5380 memcg_free_kmem(memcg);
5381 mem_cgroup_free(memcg);
5385 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5386 * @css: the target css
5388 * Reset the states of the mem_cgroup associated with @css. This is
5389 * invoked when the userland requests disabling on the default hierarchy
5390 * but the memcg is pinned through dependency. The memcg should stop
5391 * applying policies and should revert to the vanilla state as it may be
5392 * made visible again.
5394 * The current implementation only resets the essential configurations.
5395 * This needs to be expanded to cover all the visible parts.
5397 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5399 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5401 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5402 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5403 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5404 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5405 page_counter_set_min(&memcg->memory, 0);
5406 page_counter_set_low(&memcg->memory, 0);
5407 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5408 memcg->soft_limit = PAGE_COUNTER_MAX;
5409 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5410 memcg_wb_domain_size_changed(memcg);
5413 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5415 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5416 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5417 struct memcg_vmstats_percpu *statc;
5421 statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5423 for (i = 0; i < MEMCG_NR_STAT; i++) {
5425 * Collect the aggregated propagation counts of groups
5426 * below us. We're in a per-cpu loop here and this is
5427 * a global counter, so the first cycle will get them.
5429 delta = memcg->vmstats.state_pending[i];
5431 memcg->vmstats.state_pending[i] = 0;
5433 /* Add CPU changes on this level since the last flush */
5434 v = READ_ONCE(statc->state[i]);
5435 if (v != statc->state_prev[i]) {
5436 delta += v - statc->state_prev[i];
5437 statc->state_prev[i] = v;
5443 /* Aggregate counts on this level and propagate upwards */
5444 memcg->vmstats.state[i] += delta;
5446 parent->vmstats.state_pending[i] += delta;
5449 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
5450 delta = memcg->vmstats.events_pending[i];
5452 memcg->vmstats.events_pending[i] = 0;
5454 v = READ_ONCE(statc->events[i]);
5455 if (v != statc->events_prev[i]) {
5456 delta += v - statc->events_prev[i];
5457 statc->events_prev[i] = v;
5463 memcg->vmstats.events[i] += delta;
5465 parent->vmstats.events_pending[i] += delta;
5470 /* Handlers for move charge at task migration. */
5471 static int mem_cgroup_do_precharge(unsigned long count)
5475 /* Try a single bulk charge without reclaim first, kswapd may wake */
5476 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5478 mc.precharge += count;
5482 /* Try charges one by one with reclaim, but do not retry */
5484 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5498 enum mc_target_type {
5505 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5506 unsigned long addr, pte_t ptent)
5508 struct page *page = vm_normal_page(vma, addr, ptent);
5510 if (!page || !page_mapped(page))
5512 if (PageAnon(page)) {
5513 if (!(mc.flags & MOVE_ANON))
5516 if (!(mc.flags & MOVE_FILE))
5519 if (!get_page_unless_zero(page))
5525 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5526 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5527 pte_t ptent, swp_entry_t *entry)
5529 struct page *page = NULL;
5530 swp_entry_t ent = pte_to_swp_entry(ptent);
5532 if (!(mc.flags & MOVE_ANON))
5536 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5537 * a device and because they are not accessible by CPU they are store
5538 * as special swap entry in the CPU page table.
5540 if (is_device_private_entry(ent)) {
5541 page = pfn_swap_entry_to_page(ent);
5543 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5544 * a refcount of 1 when free (unlike normal page)
5546 if (!page_ref_add_unless(page, 1, 1))
5551 if (non_swap_entry(ent))
5555 * Because lookup_swap_cache() updates some statistics counter,
5556 * we call find_get_page() with swapper_space directly.
5558 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5559 entry->val = ent.val;
5564 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5565 pte_t ptent, swp_entry_t *entry)
5571 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5572 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5574 if (!vma->vm_file) /* anonymous vma */
5576 if (!(mc.flags & MOVE_FILE))
5579 /* page is moved even if it's not RSS of this task(page-faulted). */
5580 /* shmem/tmpfs may report page out on swap: account for that too. */
5581 return find_get_incore_page(vma->vm_file->f_mapping,
5582 linear_page_index(vma, addr));
5586 * mem_cgroup_move_account - move account of the page
5588 * @compound: charge the page as compound or small page
5589 * @from: mem_cgroup which the page is moved from.
5590 * @to: mem_cgroup which the page is moved to. @from != @to.
5592 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5594 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5597 static int mem_cgroup_move_account(struct page *page,
5599 struct mem_cgroup *from,
5600 struct mem_cgroup *to)
5602 struct lruvec *from_vec, *to_vec;
5603 struct pglist_data *pgdat;
5604 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5607 VM_BUG_ON(from == to);
5608 VM_BUG_ON_PAGE(PageLRU(page), page);
5609 VM_BUG_ON(compound && !PageTransHuge(page));
5612 * Prevent mem_cgroup_migrate() from looking at
5613 * page's memory cgroup of its source page while we change it.
5616 if (!trylock_page(page))
5620 if (page_memcg(page) != from)
5623 pgdat = page_pgdat(page);
5624 from_vec = mem_cgroup_lruvec(from, pgdat);
5625 to_vec = mem_cgroup_lruvec(to, pgdat);
5627 lock_page_memcg(page);
5629 if (PageAnon(page)) {
5630 if (page_mapped(page)) {
5631 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5632 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5633 if (PageTransHuge(page)) {
5634 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5636 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5641 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5642 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5644 if (PageSwapBacked(page)) {
5645 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5646 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5649 if (page_mapped(page)) {
5650 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5651 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5654 if (PageDirty(page)) {
5655 struct address_space *mapping = page_mapping(page);
5657 if (mapping_can_writeback(mapping)) {
5658 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5660 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5666 if (PageWriteback(page)) {
5667 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5668 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5672 * All state has been migrated, let's switch to the new memcg.
5674 * It is safe to change page's memcg here because the page
5675 * is referenced, charged, isolated, and locked: we can't race
5676 * with (un)charging, migration, LRU putback, or anything else
5677 * that would rely on a stable page's memory cgroup.
5679 * Note that lock_page_memcg is a memcg lock, not a page lock,
5680 * to save space. As soon as we switch page's memory cgroup to a
5681 * new memcg that isn't locked, the above state can change
5682 * concurrently again. Make sure we're truly done with it.
5687 css_put(&from->css);
5689 page->memcg_data = (unsigned long)to;
5691 __unlock_page_memcg(from);
5695 local_irq_disable();
5696 mem_cgroup_charge_statistics(to, page, nr_pages);
5697 memcg_check_events(to, page);
5698 mem_cgroup_charge_statistics(from, page, -nr_pages);
5699 memcg_check_events(from, page);
5708 * get_mctgt_type - get target type of moving charge
5709 * @vma: the vma the pte to be checked belongs
5710 * @addr: the address corresponding to the pte to be checked
5711 * @ptent: the pte to be checked
5712 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5715 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5716 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5717 * move charge. if @target is not NULL, the page is stored in target->page
5718 * with extra refcnt got(Callers should handle it).
5719 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5720 * target for charge migration. if @target is not NULL, the entry is stored
5722 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5723 * (so ZONE_DEVICE page and thus not on the lru).
5724 * For now we such page is charge like a regular page would be as for all
5725 * intent and purposes it is just special memory taking the place of a
5728 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5730 * Called with pte lock held.
5733 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5734 unsigned long addr, pte_t ptent, union mc_target *target)
5736 struct page *page = NULL;
5737 enum mc_target_type ret = MC_TARGET_NONE;
5738 swp_entry_t ent = { .val = 0 };
5740 if (pte_present(ptent))
5741 page = mc_handle_present_pte(vma, addr, ptent);
5742 else if (is_swap_pte(ptent))
5743 page = mc_handle_swap_pte(vma, ptent, &ent);
5744 else if (pte_none(ptent))
5745 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5747 if (!page && !ent.val)
5751 * Do only loose check w/o serialization.
5752 * mem_cgroup_move_account() checks the page is valid or
5753 * not under LRU exclusion.
5755 if (page_memcg(page) == mc.from) {
5756 ret = MC_TARGET_PAGE;
5757 if (is_device_private_page(page))
5758 ret = MC_TARGET_DEVICE;
5760 target->page = page;
5762 if (!ret || !target)
5766 * There is a swap entry and a page doesn't exist or isn't charged.
5767 * But we cannot move a tail-page in a THP.
5769 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5770 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5771 ret = MC_TARGET_SWAP;
5778 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5780 * We don't consider PMD mapped swapping or file mapped pages because THP does
5781 * not support them for now.
5782 * Caller should make sure that pmd_trans_huge(pmd) is true.
5784 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5785 unsigned long addr, pmd_t pmd, union mc_target *target)
5787 struct page *page = NULL;
5788 enum mc_target_type ret = MC_TARGET_NONE;
5790 if (unlikely(is_swap_pmd(pmd))) {
5791 VM_BUG_ON(thp_migration_supported() &&
5792 !is_pmd_migration_entry(pmd));
5795 page = pmd_page(pmd);
5796 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5797 if (!(mc.flags & MOVE_ANON))
5799 if (page_memcg(page) == mc.from) {
5800 ret = MC_TARGET_PAGE;
5803 target->page = page;
5809 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5810 unsigned long addr, pmd_t pmd, union mc_target *target)
5812 return MC_TARGET_NONE;
5816 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5817 unsigned long addr, unsigned long end,
5818 struct mm_walk *walk)
5820 struct vm_area_struct *vma = walk->vma;
5824 ptl = pmd_trans_huge_lock(pmd, vma);
5827 * Note their can not be MC_TARGET_DEVICE for now as we do not
5828 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5829 * this might change.
5831 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5832 mc.precharge += HPAGE_PMD_NR;
5837 if (pmd_trans_unstable(pmd))
5839 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5840 for (; addr != end; pte++, addr += PAGE_SIZE)
5841 if (get_mctgt_type(vma, addr, *pte, NULL))
5842 mc.precharge++; /* increment precharge temporarily */
5843 pte_unmap_unlock(pte - 1, ptl);
5849 static const struct mm_walk_ops precharge_walk_ops = {
5850 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5853 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5855 unsigned long precharge;
5858 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5859 mmap_read_unlock(mm);
5861 precharge = mc.precharge;
5867 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5869 unsigned long precharge = mem_cgroup_count_precharge(mm);
5871 VM_BUG_ON(mc.moving_task);
5872 mc.moving_task = current;
5873 return mem_cgroup_do_precharge(precharge);
5876 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5877 static void __mem_cgroup_clear_mc(void)
5879 struct mem_cgroup *from = mc.from;
5880 struct mem_cgroup *to = mc.to;
5882 /* we must uncharge all the leftover precharges from mc.to */
5884 cancel_charge(mc.to, mc.precharge);
5888 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5889 * we must uncharge here.
5891 if (mc.moved_charge) {
5892 cancel_charge(mc.from, mc.moved_charge);
5893 mc.moved_charge = 0;
5895 /* we must fixup refcnts and charges */
5896 if (mc.moved_swap) {
5897 /* uncharge swap account from the old cgroup */
5898 if (!mem_cgroup_is_root(mc.from))
5899 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5901 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5904 * we charged both to->memory and to->memsw, so we
5905 * should uncharge to->memory.
5907 if (!mem_cgroup_is_root(mc.to))
5908 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5912 memcg_oom_recover(from);
5913 memcg_oom_recover(to);
5914 wake_up_all(&mc.waitq);
5917 static void mem_cgroup_clear_mc(void)
5919 struct mm_struct *mm = mc.mm;
5922 * we must clear moving_task before waking up waiters at the end of
5925 mc.moving_task = NULL;
5926 __mem_cgroup_clear_mc();
5927 spin_lock(&mc.lock);
5931 spin_unlock(&mc.lock);
5936 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5938 struct cgroup_subsys_state *css;
5939 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5940 struct mem_cgroup *from;
5941 struct task_struct *leader, *p;
5942 struct mm_struct *mm;
5943 unsigned long move_flags;
5946 /* charge immigration isn't supported on the default hierarchy */
5947 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5951 * Multi-process migrations only happen on the default hierarchy
5952 * where charge immigration is not used. Perform charge
5953 * immigration if @tset contains a leader and whine if there are
5957 cgroup_taskset_for_each_leader(leader, css, tset) {
5960 memcg = mem_cgroup_from_css(css);
5966 * We are now committed to this value whatever it is. Changes in this
5967 * tunable will only affect upcoming migrations, not the current one.
5968 * So we need to save it, and keep it going.
5970 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5974 from = mem_cgroup_from_task(p);
5976 VM_BUG_ON(from == memcg);
5978 mm = get_task_mm(p);
5981 /* We move charges only when we move a owner of the mm */
5982 if (mm->owner == p) {
5985 VM_BUG_ON(mc.precharge);
5986 VM_BUG_ON(mc.moved_charge);
5987 VM_BUG_ON(mc.moved_swap);
5989 spin_lock(&mc.lock);
5993 mc.flags = move_flags;
5994 spin_unlock(&mc.lock);
5995 /* We set mc.moving_task later */
5997 ret = mem_cgroup_precharge_mc(mm);
5999 mem_cgroup_clear_mc();
6006 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6009 mem_cgroup_clear_mc();
6012 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6013 unsigned long addr, unsigned long end,
6014 struct mm_walk *walk)
6017 struct vm_area_struct *vma = walk->vma;
6020 enum mc_target_type target_type;
6021 union mc_target target;
6024 ptl = pmd_trans_huge_lock(pmd, vma);
6026 if (mc.precharge < HPAGE_PMD_NR) {
6030 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6031 if (target_type == MC_TARGET_PAGE) {
6033 if (!isolate_lru_page(page)) {
6034 if (!mem_cgroup_move_account(page, true,
6036 mc.precharge -= HPAGE_PMD_NR;
6037 mc.moved_charge += HPAGE_PMD_NR;
6039 putback_lru_page(page);
6042 } else if (target_type == MC_TARGET_DEVICE) {
6044 if (!mem_cgroup_move_account(page, true,
6046 mc.precharge -= HPAGE_PMD_NR;
6047 mc.moved_charge += HPAGE_PMD_NR;
6055 if (pmd_trans_unstable(pmd))
6058 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6059 for (; addr != end; addr += PAGE_SIZE) {
6060 pte_t ptent = *(pte++);
6061 bool device = false;
6067 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6068 case MC_TARGET_DEVICE:
6071 case MC_TARGET_PAGE:
6074 * We can have a part of the split pmd here. Moving it
6075 * can be done but it would be too convoluted so simply
6076 * ignore such a partial THP and keep it in original
6077 * memcg. There should be somebody mapping the head.
6079 if (PageTransCompound(page))
6081 if (!device && isolate_lru_page(page))
6083 if (!mem_cgroup_move_account(page, false,
6086 /* we uncharge from mc.from later. */
6090 putback_lru_page(page);
6091 put: /* get_mctgt_type() gets the page */
6094 case MC_TARGET_SWAP:
6096 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6098 mem_cgroup_id_get_many(mc.to, 1);
6099 /* we fixup other refcnts and charges later. */
6107 pte_unmap_unlock(pte - 1, ptl);
6112 * We have consumed all precharges we got in can_attach().
6113 * We try charge one by one, but don't do any additional
6114 * charges to mc.to if we have failed in charge once in attach()
6117 ret = mem_cgroup_do_precharge(1);
6125 static const struct mm_walk_ops charge_walk_ops = {
6126 .pmd_entry = mem_cgroup_move_charge_pte_range,
6129 static void mem_cgroup_move_charge(void)
6131 lru_add_drain_all();
6133 * Signal lock_page_memcg() to take the memcg's move_lock
6134 * while we're moving its pages to another memcg. Then wait
6135 * for already started RCU-only updates to finish.
6137 atomic_inc(&mc.from->moving_account);
6140 if (unlikely(!mmap_read_trylock(mc.mm))) {
6142 * Someone who are holding the mmap_lock might be waiting in
6143 * waitq. So we cancel all extra charges, wake up all waiters,
6144 * and retry. Because we cancel precharges, we might not be able
6145 * to move enough charges, but moving charge is a best-effort
6146 * feature anyway, so it wouldn't be a big problem.
6148 __mem_cgroup_clear_mc();
6153 * When we have consumed all precharges and failed in doing
6154 * additional charge, the page walk just aborts.
6156 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6159 mmap_read_unlock(mc.mm);
6160 atomic_dec(&mc.from->moving_account);
6163 static void mem_cgroup_move_task(void)
6166 mem_cgroup_move_charge();
6167 mem_cgroup_clear_mc();
6170 #else /* !CONFIG_MMU */
6171 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6175 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6178 static void mem_cgroup_move_task(void)
6183 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6185 if (value == PAGE_COUNTER_MAX)
6186 seq_puts(m, "max\n");
6188 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6193 static u64 memory_current_read(struct cgroup_subsys_state *css,
6196 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6198 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6201 static int memory_min_show(struct seq_file *m, void *v)
6203 return seq_puts_memcg_tunable(m,
6204 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6207 static ssize_t memory_min_write(struct kernfs_open_file *of,
6208 char *buf, size_t nbytes, loff_t off)
6210 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6214 buf = strstrip(buf);
6215 err = page_counter_memparse(buf, "max", &min);
6219 page_counter_set_min(&memcg->memory, min);
6224 static int memory_low_show(struct seq_file *m, void *v)
6226 return seq_puts_memcg_tunable(m,
6227 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6230 static ssize_t memory_low_write(struct kernfs_open_file *of,
6231 char *buf, size_t nbytes, loff_t off)
6233 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6237 buf = strstrip(buf);
6238 err = page_counter_memparse(buf, "max", &low);
6242 page_counter_set_low(&memcg->memory, low);
6247 static int memory_high_show(struct seq_file *m, void *v)
6249 return seq_puts_memcg_tunable(m,
6250 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6253 static ssize_t memory_high_write(struct kernfs_open_file *of,
6254 char *buf, size_t nbytes, loff_t off)
6256 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6257 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6258 bool drained = false;
6262 buf = strstrip(buf);
6263 err = page_counter_memparse(buf, "max", &high);
6267 page_counter_set_high(&memcg->memory, high);
6270 unsigned long nr_pages = page_counter_read(&memcg->memory);
6271 unsigned long reclaimed;
6273 if (nr_pages <= high)
6276 if (signal_pending(current))
6280 drain_all_stock(memcg);
6285 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6288 if (!reclaimed && !nr_retries--)
6292 memcg_wb_domain_size_changed(memcg);
6296 static int memory_max_show(struct seq_file *m, void *v)
6298 return seq_puts_memcg_tunable(m,
6299 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6302 static ssize_t memory_max_write(struct kernfs_open_file *of,
6303 char *buf, size_t nbytes, loff_t off)
6305 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6306 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6307 bool drained = false;
6311 buf = strstrip(buf);
6312 err = page_counter_memparse(buf, "max", &max);
6316 xchg(&memcg->memory.max, max);
6319 unsigned long nr_pages = page_counter_read(&memcg->memory);
6321 if (nr_pages <= max)
6324 if (signal_pending(current))
6328 drain_all_stock(memcg);
6334 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6340 memcg_memory_event(memcg, MEMCG_OOM);
6341 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6345 memcg_wb_domain_size_changed(memcg);
6349 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6351 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6352 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6353 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6354 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6355 seq_printf(m, "oom_kill %lu\n",
6356 atomic_long_read(&events[MEMCG_OOM_KILL]));
6359 static int memory_events_show(struct seq_file *m, void *v)
6361 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6363 __memory_events_show(m, memcg->memory_events);
6367 static int memory_events_local_show(struct seq_file *m, void *v)
6369 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6371 __memory_events_show(m, memcg->memory_events_local);
6375 static int memory_stat_show(struct seq_file *m, void *v)
6377 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6380 buf = memory_stat_format(memcg);
6389 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6392 return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6395 static int memory_numa_stat_show(struct seq_file *m, void *v)
6398 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6400 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6403 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6406 seq_printf(m, "%s", memory_stats[i].name);
6407 for_each_node_state(nid, N_MEMORY) {
6409 struct lruvec *lruvec;
6411 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6412 size = lruvec_page_state_output(lruvec,
6413 memory_stats[i].idx);
6414 seq_printf(m, " N%d=%llu", nid, size);
6423 static int memory_oom_group_show(struct seq_file *m, void *v)
6425 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6427 seq_printf(m, "%d\n", memcg->oom_group);
6432 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6433 char *buf, size_t nbytes, loff_t off)
6435 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6438 buf = strstrip(buf);
6442 ret = kstrtoint(buf, 0, &oom_group);
6446 if (oom_group != 0 && oom_group != 1)
6449 memcg->oom_group = oom_group;
6454 static struct cftype memory_files[] = {
6457 .flags = CFTYPE_NOT_ON_ROOT,
6458 .read_u64 = memory_current_read,
6462 .flags = CFTYPE_NOT_ON_ROOT,
6463 .seq_show = memory_min_show,
6464 .write = memory_min_write,
6468 .flags = CFTYPE_NOT_ON_ROOT,
6469 .seq_show = memory_low_show,
6470 .write = memory_low_write,
6474 .flags = CFTYPE_NOT_ON_ROOT,
6475 .seq_show = memory_high_show,
6476 .write = memory_high_write,
6480 .flags = CFTYPE_NOT_ON_ROOT,
6481 .seq_show = memory_max_show,
6482 .write = memory_max_write,
6486 .flags = CFTYPE_NOT_ON_ROOT,
6487 .file_offset = offsetof(struct mem_cgroup, events_file),
6488 .seq_show = memory_events_show,
6491 .name = "events.local",
6492 .flags = CFTYPE_NOT_ON_ROOT,
6493 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6494 .seq_show = memory_events_local_show,
6498 .seq_show = memory_stat_show,
6502 .name = "numa_stat",
6503 .seq_show = memory_numa_stat_show,
6507 .name = "oom.group",
6508 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6509 .seq_show = memory_oom_group_show,
6510 .write = memory_oom_group_write,
6515 struct cgroup_subsys memory_cgrp_subsys = {
6516 .css_alloc = mem_cgroup_css_alloc,
6517 .css_online = mem_cgroup_css_online,
6518 .css_offline = mem_cgroup_css_offline,
6519 .css_released = mem_cgroup_css_released,
6520 .css_free = mem_cgroup_css_free,
6521 .css_reset = mem_cgroup_css_reset,
6522 .css_rstat_flush = mem_cgroup_css_rstat_flush,
6523 .can_attach = mem_cgroup_can_attach,
6524 .cancel_attach = mem_cgroup_cancel_attach,
6525 .post_attach = mem_cgroup_move_task,
6526 .dfl_cftypes = memory_files,
6527 .legacy_cftypes = mem_cgroup_legacy_files,
6532 * This function calculates an individual cgroup's effective
6533 * protection which is derived from its own memory.min/low, its
6534 * parent's and siblings' settings, as well as the actual memory
6535 * distribution in the tree.
6537 * The following rules apply to the effective protection values:
6539 * 1. At the first level of reclaim, effective protection is equal to
6540 * the declared protection in memory.min and memory.low.
6542 * 2. To enable safe delegation of the protection configuration, at
6543 * subsequent levels the effective protection is capped to the
6544 * parent's effective protection.
6546 * 3. To make complex and dynamic subtrees easier to configure, the
6547 * user is allowed to overcommit the declared protection at a given
6548 * level. If that is the case, the parent's effective protection is
6549 * distributed to the children in proportion to how much protection
6550 * they have declared and how much of it they are utilizing.
6552 * This makes distribution proportional, but also work-conserving:
6553 * if one cgroup claims much more protection than it uses memory,
6554 * the unused remainder is available to its siblings.
6556 * 4. Conversely, when the declared protection is undercommitted at a
6557 * given level, the distribution of the larger parental protection
6558 * budget is NOT proportional. A cgroup's protection from a sibling
6559 * is capped to its own memory.min/low setting.
6561 * 5. However, to allow protecting recursive subtrees from each other
6562 * without having to declare each individual cgroup's fixed share
6563 * of the ancestor's claim to protection, any unutilized -
6564 * "floating" - protection from up the tree is distributed in
6565 * proportion to each cgroup's *usage*. This makes the protection
6566 * neutral wrt sibling cgroups and lets them compete freely over
6567 * the shared parental protection budget, but it protects the
6568 * subtree as a whole from neighboring subtrees.
6570 * Note that 4. and 5. are not in conflict: 4. is about protecting
6571 * against immediate siblings whereas 5. is about protecting against
6572 * neighboring subtrees.
6574 static unsigned long effective_protection(unsigned long usage,
6575 unsigned long parent_usage,
6576 unsigned long setting,
6577 unsigned long parent_effective,
6578 unsigned long siblings_protected)
6580 unsigned long protected;
6583 protected = min(usage, setting);
6585 * If all cgroups at this level combined claim and use more
6586 * protection then what the parent affords them, distribute
6587 * shares in proportion to utilization.
6589 * We are using actual utilization rather than the statically
6590 * claimed protection in order to be work-conserving: claimed
6591 * but unused protection is available to siblings that would
6592 * otherwise get a smaller chunk than what they claimed.
6594 if (siblings_protected > parent_effective)
6595 return protected * parent_effective / siblings_protected;
6598 * Ok, utilized protection of all children is within what the
6599 * parent affords them, so we know whatever this child claims
6600 * and utilizes is effectively protected.
6602 * If there is unprotected usage beyond this value, reclaim
6603 * will apply pressure in proportion to that amount.
6605 * If there is unutilized protection, the cgroup will be fully
6606 * shielded from reclaim, but we do return a smaller value for
6607 * protection than what the group could enjoy in theory. This
6608 * is okay. With the overcommit distribution above, effective
6609 * protection is always dependent on how memory is actually
6610 * consumed among the siblings anyway.
6615 * If the children aren't claiming (all of) the protection
6616 * afforded to them by the parent, distribute the remainder in
6617 * proportion to the (unprotected) memory of each cgroup. That
6618 * way, cgroups that aren't explicitly prioritized wrt each
6619 * other compete freely over the allowance, but they are
6620 * collectively protected from neighboring trees.
6622 * We're using unprotected memory for the weight so that if
6623 * some cgroups DO claim explicit protection, we don't protect
6624 * the same bytes twice.
6626 * Check both usage and parent_usage against the respective
6627 * protected values. One should imply the other, but they
6628 * aren't read atomically - make sure the division is sane.
6630 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6632 if (parent_effective > siblings_protected &&
6633 parent_usage > siblings_protected &&
6634 usage > protected) {
6635 unsigned long unclaimed;
6637 unclaimed = parent_effective - siblings_protected;
6638 unclaimed *= usage - protected;
6639 unclaimed /= parent_usage - siblings_protected;
6648 * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
6649 * @root: the top ancestor of the sub-tree being checked
6650 * @memcg: the memory cgroup to check
6652 * WARNING: This function is not stateless! It can only be used as part
6653 * of a top-down tree iteration, not for isolated queries.
6655 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6656 struct mem_cgroup *memcg)
6658 unsigned long usage, parent_usage;
6659 struct mem_cgroup *parent;
6661 if (mem_cgroup_disabled())
6665 root = root_mem_cgroup;
6668 * Effective values of the reclaim targets are ignored so they
6669 * can be stale. Have a look at mem_cgroup_protection for more
6671 * TODO: calculation should be more robust so that we do not need
6672 * that special casing.
6677 usage = page_counter_read(&memcg->memory);
6681 parent = parent_mem_cgroup(memcg);
6682 /* No parent means a non-hierarchical mode on v1 memcg */
6686 if (parent == root) {
6687 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6688 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6692 parent_usage = page_counter_read(&parent->memory);
6694 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6695 READ_ONCE(memcg->memory.min),
6696 READ_ONCE(parent->memory.emin),
6697 atomic_long_read(&parent->memory.children_min_usage)));
6699 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6700 READ_ONCE(memcg->memory.low),
6701 READ_ONCE(parent->memory.elow),
6702 atomic_long_read(&parent->memory.children_low_usage)));
6705 static int __mem_cgroup_charge(struct page *page, struct mem_cgroup *memcg,
6708 unsigned int nr_pages = thp_nr_pages(page);
6711 ret = try_charge(memcg, gfp, nr_pages);
6715 css_get(&memcg->css);
6716 commit_charge(page, memcg);
6718 local_irq_disable();
6719 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6720 memcg_check_events(memcg, page);
6727 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6728 * @page: page to charge
6729 * @mm: mm context of the victim
6730 * @gfp_mask: reclaim mode
6732 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6733 * pages according to @gfp_mask if necessary. if @mm is NULL, try to
6734 * charge to the active memcg.
6736 * Do not use this for pages allocated for swapin.
6738 * Returns 0 on success. Otherwise, an error code is returned.
6740 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6742 struct mem_cgroup *memcg;
6745 if (mem_cgroup_disabled())
6748 memcg = get_mem_cgroup_from_mm(mm);
6749 ret = __mem_cgroup_charge(page, memcg, gfp_mask);
6750 css_put(&memcg->css);
6756 * mem_cgroup_swapin_charge_page - charge a newly allocated page for swapin
6757 * @page: page to charge
6758 * @mm: mm context of the victim
6759 * @gfp: reclaim mode
6760 * @entry: swap entry for which the page is allocated
6762 * This function charges a page allocated for swapin. Please call this before
6763 * adding the page to the swapcache.
6765 * Returns 0 on success. Otherwise, an error code is returned.
6767 int mem_cgroup_swapin_charge_page(struct page *page, struct mm_struct *mm,
6768 gfp_t gfp, swp_entry_t entry)
6770 struct mem_cgroup *memcg;
6774 if (mem_cgroup_disabled())
6777 id = lookup_swap_cgroup_id(entry);
6779 memcg = mem_cgroup_from_id(id);
6780 if (!memcg || !css_tryget_online(&memcg->css))
6781 memcg = get_mem_cgroup_from_mm(mm);
6784 ret = __mem_cgroup_charge(page, memcg, gfp);
6786 css_put(&memcg->css);
6791 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6792 * @entry: swap entry for which the page is charged
6794 * Call this function after successfully adding the charged page to swapcache.
6796 * Note: This function assumes the page for which swap slot is being uncharged
6799 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
6802 * Cgroup1's unified memory+swap counter has been charged with the
6803 * new swapcache page, finish the transfer by uncharging the swap
6804 * slot. The swap slot would also get uncharged when it dies, but
6805 * it can stick around indefinitely and we'd count the page twice
6808 * Cgroup2 has separate resource counters for memory and swap,
6809 * so this is a non-issue here. Memory and swap charge lifetimes
6810 * correspond 1:1 to page and swap slot lifetimes: we charge the
6811 * page to memory here, and uncharge swap when the slot is freed.
6813 if (!mem_cgroup_disabled() && do_memsw_account()) {
6815 * The swap entry might not get freed for a long time,
6816 * let's not wait for it. The page already received a
6817 * memory+swap charge, drop the swap entry duplicate.
6819 mem_cgroup_uncharge_swap(entry, 1);
6823 struct uncharge_gather {
6824 struct mem_cgroup *memcg;
6825 unsigned long nr_memory;
6826 unsigned long pgpgout;
6827 unsigned long nr_kmem;
6828 struct page *dummy_page;
6831 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6833 memset(ug, 0, sizeof(*ug));
6836 static void uncharge_batch(const struct uncharge_gather *ug)
6838 unsigned long flags;
6840 if (ug->nr_memory) {
6841 page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
6842 if (do_memsw_account())
6843 page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
6844 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6845 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6846 memcg_oom_recover(ug->memcg);
6849 local_irq_save(flags);
6850 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6851 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
6852 memcg_check_events(ug->memcg, ug->dummy_page);
6853 local_irq_restore(flags);
6855 /* drop reference from uncharge_page */
6856 css_put(&ug->memcg->css);
6859 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6861 unsigned long nr_pages;
6862 struct mem_cgroup *memcg;
6863 struct obj_cgroup *objcg;
6864 bool use_objcg = PageMemcgKmem(page);
6866 VM_BUG_ON_PAGE(PageLRU(page), page);
6869 * Nobody should be changing or seriously looking at
6870 * page memcg or objcg at this point, we have fully
6871 * exclusive access to the page.
6874 objcg = __page_objcg(page);
6876 * This get matches the put at the end of the function and
6877 * kmem pages do not hold memcg references anymore.
6879 memcg = get_mem_cgroup_from_objcg(objcg);
6881 memcg = __page_memcg(page);
6887 if (ug->memcg != memcg) {
6890 uncharge_gather_clear(ug);
6893 ug->dummy_page = page;
6895 /* pairs with css_put in uncharge_batch */
6896 css_get(&memcg->css);
6899 nr_pages = compound_nr(page);
6902 ug->nr_memory += nr_pages;
6903 ug->nr_kmem += nr_pages;
6905 page->memcg_data = 0;
6906 obj_cgroup_put(objcg);
6908 /* LRU pages aren't accounted at the root level */
6909 if (!mem_cgroup_is_root(memcg))
6910 ug->nr_memory += nr_pages;
6913 page->memcg_data = 0;
6916 css_put(&memcg->css);
6920 * mem_cgroup_uncharge - uncharge a page
6921 * @page: page to uncharge
6923 * Uncharge a page previously charged with mem_cgroup_charge().
6925 void mem_cgroup_uncharge(struct page *page)
6927 struct uncharge_gather ug;
6929 if (mem_cgroup_disabled())
6932 /* Don't touch page->lru of any random page, pre-check: */
6933 if (!page_memcg(page))
6936 uncharge_gather_clear(&ug);
6937 uncharge_page(page, &ug);
6938 uncharge_batch(&ug);
6942 * mem_cgroup_uncharge_list - uncharge a list of page
6943 * @page_list: list of pages to uncharge
6945 * Uncharge a list of pages previously charged with
6946 * mem_cgroup_charge().
6948 void mem_cgroup_uncharge_list(struct list_head *page_list)
6950 struct uncharge_gather ug;
6953 if (mem_cgroup_disabled())
6956 uncharge_gather_clear(&ug);
6957 list_for_each_entry(page, page_list, lru)
6958 uncharge_page(page, &ug);
6960 uncharge_batch(&ug);
6964 * mem_cgroup_migrate - charge a page's replacement
6965 * @oldpage: currently circulating page
6966 * @newpage: replacement page
6968 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6969 * be uncharged upon free.
6971 * Both pages must be locked, @newpage->mapping must be set up.
6973 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6975 struct mem_cgroup *memcg;
6976 unsigned int nr_pages;
6977 unsigned long flags;
6979 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6980 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6981 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6982 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6985 if (mem_cgroup_disabled())
6988 /* Page cache replacement: new page already charged? */
6989 if (page_memcg(newpage))
6992 memcg = page_memcg(oldpage);
6993 VM_WARN_ON_ONCE_PAGE(!memcg, oldpage);
6997 /* Force-charge the new page. The old one will be freed soon */
6998 nr_pages = thp_nr_pages(newpage);
7000 if (!mem_cgroup_is_root(memcg)) {
7001 page_counter_charge(&memcg->memory, nr_pages);
7002 if (do_memsw_account())
7003 page_counter_charge(&memcg->memsw, nr_pages);
7006 css_get(&memcg->css);
7007 commit_charge(newpage, memcg);
7009 local_irq_save(flags);
7010 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
7011 memcg_check_events(memcg, newpage);
7012 local_irq_restore(flags);
7015 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7016 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7018 void mem_cgroup_sk_alloc(struct sock *sk)
7020 struct mem_cgroup *memcg;
7022 if (!mem_cgroup_sockets_enabled)
7025 /* Do not associate the sock with unrelated interrupted task's memcg. */
7030 memcg = mem_cgroup_from_task(current);
7031 if (memcg == root_mem_cgroup)
7033 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7035 if (css_tryget(&memcg->css))
7036 sk->sk_memcg = memcg;
7041 void mem_cgroup_sk_free(struct sock *sk)
7044 css_put(&sk->sk_memcg->css);
7048 * mem_cgroup_charge_skmem - charge socket memory
7049 * @memcg: memcg to charge
7050 * @nr_pages: number of pages to charge
7052 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7053 * @memcg's configured limit, %false if the charge had to be forced.
7055 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7057 gfp_t gfp_mask = GFP_KERNEL;
7059 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7060 struct page_counter *fail;
7062 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7063 memcg->tcpmem_pressure = 0;
7066 page_counter_charge(&memcg->tcpmem, nr_pages);
7067 memcg->tcpmem_pressure = 1;
7071 /* Don't block in the packet receive path */
7073 gfp_mask = GFP_NOWAIT;
7075 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7077 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7080 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7085 * mem_cgroup_uncharge_skmem - uncharge socket memory
7086 * @memcg: memcg to uncharge
7087 * @nr_pages: number of pages to uncharge
7089 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7091 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7092 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7096 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7098 refill_stock(memcg, nr_pages);
7101 static int __init cgroup_memory(char *s)
7105 while ((token = strsep(&s, ",")) != NULL) {
7108 if (!strcmp(token, "nosocket"))
7109 cgroup_memory_nosocket = true;
7110 if (!strcmp(token, "nokmem"))
7111 cgroup_memory_nokmem = true;
7115 __setup("cgroup.memory=", cgroup_memory);
7118 * subsys_initcall() for memory controller.
7120 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7121 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7122 * basically everything that doesn't depend on a specific mem_cgroup structure
7123 * should be initialized from here.
7125 static int __init mem_cgroup_init(void)
7130 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7131 * used for per-memcg-per-cpu caching of per-node statistics. In order
7132 * to work fine, we should make sure that the overfill threshold can't
7133 * exceed S32_MAX / PAGE_SIZE.
7135 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7137 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7138 memcg_hotplug_cpu_dead);
7140 for_each_possible_cpu(cpu)
7141 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7144 for_each_node(node) {
7145 struct mem_cgroup_tree_per_node *rtpn;
7147 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7148 node_online(node) ? node : NUMA_NO_NODE);
7150 rtpn->rb_root = RB_ROOT;
7151 rtpn->rb_rightmost = NULL;
7152 spin_lock_init(&rtpn->lock);
7153 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7158 subsys_initcall(mem_cgroup_init);
7160 #ifdef CONFIG_MEMCG_SWAP
7161 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7163 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7165 * The root cgroup cannot be destroyed, so it's refcount must
7168 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7172 memcg = parent_mem_cgroup(memcg);
7174 memcg = root_mem_cgroup;
7180 * mem_cgroup_swapout - transfer a memsw charge to swap
7181 * @page: page whose memsw charge to transfer
7182 * @entry: swap entry to move the charge to
7184 * Transfer the memsw charge of @page to @entry.
7186 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7188 struct mem_cgroup *memcg, *swap_memcg;
7189 unsigned int nr_entries;
7190 unsigned short oldid;
7192 VM_BUG_ON_PAGE(PageLRU(page), page);
7193 VM_BUG_ON_PAGE(page_count(page), page);
7195 if (mem_cgroup_disabled())
7198 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7201 memcg = page_memcg(page);
7203 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7208 * In case the memcg owning these pages has been offlined and doesn't
7209 * have an ID allocated to it anymore, charge the closest online
7210 * ancestor for the swap instead and transfer the memory+swap charge.
7212 swap_memcg = mem_cgroup_id_get_online(memcg);
7213 nr_entries = thp_nr_pages(page);
7214 /* Get references for the tail pages, too */
7216 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7217 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7219 VM_BUG_ON_PAGE(oldid, page);
7220 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7222 page->memcg_data = 0;
7224 if (!mem_cgroup_is_root(memcg))
7225 page_counter_uncharge(&memcg->memory, nr_entries);
7227 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7228 if (!mem_cgroup_is_root(swap_memcg))
7229 page_counter_charge(&swap_memcg->memsw, nr_entries);
7230 page_counter_uncharge(&memcg->memsw, nr_entries);
7234 * Interrupts should be disabled here because the caller holds the
7235 * i_pages lock which is taken with interrupts-off. It is
7236 * important here to have the interrupts disabled because it is the
7237 * only synchronisation we have for updating the per-CPU variables.
7239 VM_BUG_ON(!irqs_disabled());
7240 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7241 memcg_check_events(memcg, page);
7243 css_put(&memcg->css);
7247 * mem_cgroup_try_charge_swap - try charging swap space for a page
7248 * @page: page being added to swap
7249 * @entry: swap entry to charge
7251 * Try to charge @page's memcg for the swap space at @entry.
7253 * Returns 0 on success, -ENOMEM on failure.
7255 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7257 unsigned int nr_pages = thp_nr_pages(page);
7258 struct page_counter *counter;
7259 struct mem_cgroup *memcg;
7260 unsigned short oldid;
7262 if (mem_cgroup_disabled())
7265 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7268 memcg = page_memcg(page);
7270 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7275 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7279 memcg = mem_cgroup_id_get_online(memcg);
7281 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7282 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7283 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7284 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7285 mem_cgroup_id_put(memcg);
7289 /* Get references for the tail pages, too */
7291 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7292 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7293 VM_BUG_ON_PAGE(oldid, page);
7294 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7300 * mem_cgroup_uncharge_swap - uncharge swap space
7301 * @entry: swap entry to uncharge
7302 * @nr_pages: the amount of swap space to uncharge
7304 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7306 struct mem_cgroup *memcg;
7309 id = swap_cgroup_record(entry, 0, nr_pages);
7311 memcg = mem_cgroup_from_id(id);
7313 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7314 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7315 page_counter_uncharge(&memcg->swap, nr_pages);
7317 page_counter_uncharge(&memcg->memsw, nr_pages);
7319 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7320 mem_cgroup_id_put_many(memcg, nr_pages);
7325 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7327 long nr_swap_pages = get_nr_swap_pages();
7329 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7330 return nr_swap_pages;
7331 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7332 nr_swap_pages = min_t(long, nr_swap_pages,
7333 READ_ONCE(memcg->swap.max) -
7334 page_counter_read(&memcg->swap));
7335 return nr_swap_pages;
7338 bool mem_cgroup_swap_full(struct page *page)
7340 struct mem_cgroup *memcg;
7342 VM_BUG_ON_PAGE(!PageLocked(page), page);
7346 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7349 memcg = page_memcg(page);
7353 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7354 unsigned long usage = page_counter_read(&memcg->swap);
7356 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7357 usage * 2 >= READ_ONCE(memcg->swap.max))
7364 static int __init setup_swap_account(char *s)
7366 if (!strcmp(s, "1"))
7367 cgroup_memory_noswap = false;
7368 else if (!strcmp(s, "0"))
7369 cgroup_memory_noswap = true;
7372 __setup("swapaccount=", setup_swap_account);
7374 static u64 swap_current_read(struct cgroup_subsys_state *css,
7377 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7379 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7382 static int swap_high_show(struct seq_file *m, void *v)
7384 return seq_puts_memcg_tunable(m,
7385 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7388 static ssize_t swap_high_write(struct kernfs_open_file *of,
7389 char *buf, size_t nbytes, loff_t off)
7391 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7395 buf = strstrip(buf);
7396 err = page_counter_memparse(buf, "max", &high);
7400 page_counter_set_high(&memcg->swap, high);
7405 static int swap_max_show(struct seq_file *m, void *v)
7407 return seq_puts_memcg_tunable(m,
7408 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7411 static ssize_t swap_max_write(struct kernfs_open_file *of,
7412 char *buf, size_t nbytes, loff_t off)
7414 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7418 buf = strstrip(buf);
7419 err = page_counter_memparse(buf, "max", &max);
7423 xchg(&memcg->swap.max, max);
7428 static int swap_events_show(struct seq_file *m, void *v)
7430 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7432 seq_printf(m, "high %lu\n",
7433 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7434 seq_printf(m, "max %lu\n",
7435 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7436 seq_printf(m, "fail %lu\n",
7437 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7442 static struct cftype swap_files[] = {
7444 .name = "swap.current",
7445 .flags = CFTYPE_NOT_ON_ROOT,
7446 .read_u64 = swap_current_read,
7449 .name = "swap.high",
7450 .flags = CFTYPE_NOT_ON_ROOT,
7451 .seq_show = swap_high_show,
7452 .write = swap_high_write,
7456 .flags = CFTYPE_NOT_ON_ROOT,
7457 .seq_show = swap_max_show,
7458 .write = swap_max_write,
7461 .name = "swap.events",
7462 .flags = CFTYPE_NOT_ON_ROOT,
7463 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7464 .seq_show = swap_events_show,
7469 static struct cftype memsw_files[] = {
7471 .name = "memsw.usage_in_bytes",
7472 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7473 .read_u64 = mem_cgroup_read_u64,
7476 .name = "memsw.max_usage_in_bytes",
7477 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7478 .write = mem_cgroup_reset,
7479 .read_u64 = mem_cgroup_read_u64,
7482 .name = "memsw.limit_in_bytes",
7483 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7484 .write = mem_cgroup_write,
7485 .read_u64 = mem_cgroup_read_u64,
7488 .name = "memsw.failcnt",
7489 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7490 .write = mem_cgroup_reset,
7491 .read_u64 = mem_cgroup_read_u64,
7493 { }, /* terminate */
7497 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7498 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7499 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7500 * boot parameter. This may result in premature OOPS inside
7501 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7503 static int __init mem_cgroup_swap_init(void)
7505 /* No memory control -> no swap control */
7506 if (mem_cgroup_disabled())
7507 cgroup_memory_noswap = true;
7509 if (cgroup_memory_noswap)
7512 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7513 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7517 core_initcall(mem_cgroup_swap_init);
7519 #endif /* CONFIG_MEMCG_SWAP */