1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
65 #include <linux/tracehook.h>
71 #include <asm/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
76 EXPORT_SYMBOL(memory_cgrp_subsys
);
78 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
80 #define MEM_CGROUP_RECLAIM_RETRIES 5
82 /* Socket memory accounting disabled? */
83 static bool cgroup_memory_nosocket
;
85 /* Kernel memory accounting disabled? */
86 static bool cgroup_memory_nokmem
;
88 /* Whether the swap controller is active */
89 #ifdef CONFIG_MEMCG_SWAP
90 int do_swap_account __read_mostly
;
92 #define do_swap_account 0
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys
) && do_swap_account
;
101 static const char * const mem_cgroup_stat_names
[] = {
111 static const char * const mem_cgroup_events_names
[] = {
118 static const char * const mem_cgroup_lru_names
[] = {
126 #define THRESHOLDS_EVENTS_TARGET 128
127 #define SOFTLIMIT_EVENTS_TARGET 1024
128 #define NUMAINFO_EVENTS_TARGET 1024
131 * Cgroups above their limits are maintained in a RB-Tree, independent of
132 * their hierarchy representation
135 struct mem_cgroup_tree_per_node
{
136 struct rb_root rb_root
;
140 struct mem_cgroup_tree
{
141 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
144 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
147 struct mem_cgroup_eventfd_list
{
148 struct list_head list
;
149 struct eventfd_ctx
*eventfd
;
153 * cgroup_event represents events which userspace want to receive.
155 struct mem_cgroup_event
{
157 * memcg which the event belongs to.
159 struct mem_cgroup
*memcg
;
161 * eventfd to signal userspace about the event.
163 struct eventfd_ctx
*eventfd
;
165 * Each of these stored in a list by the cgroup.
167 struct list_head list
;
169 * register_event() callback will be used to add new userspace
170 * waiter for changes related to this event. Use eventfd_signal()
171 * on eventfd to send notification to userspace.
173 int (*register_event
)(struct mem_cgroup
*memcg
,
174 struct eventfd_ctx
*eventfd
, const char *args
);
176 * unregister_event() callback will be called when userspace closes
177 * the eventfd or on cgroup removing. This callback must be set,
178 * if you want provide notification functionality.
180 void (*unregister_event
)(struct mem_cgroup
*memcg
,
181 struct eventfd_ctx
*eventfd
);
183 * All fields below needed to unregister event when
184 * userspace closes eventfd.
187 wait_queue_head_t
*wqh
;
189 struct work_struct remove
;
192 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
193 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
195 /* Stuffs for move charges at task migration. */
197 * Types of charges to be moved.
199 #define MOVE_ANON 0x1U
200 #define MOVE_FILE 0x2U
201 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
203 /* "mc" and its members are protected by cgroup_mutex */
204 static struct move_charge_struct
{
205 spinlock_t lock
; /* for from, to */
206 struct mm_struct
*mm
;
207 struct mem_cgroup
*from
;
208 struct mem_cgroup
*to
;
210 unsigned long precharge
;
211 unsigned long moved_charge
;
212 unsigned long moved_swap
;
213 struct task_struct
*moving_task
; /* a task moving charges */
214 wait_queue_head_t waitq
; /* a waitq for other context */
216 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
217 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
221 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
222 * limit reclaim to prevent infinite loops, if they ever occur.
224 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
225 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
228 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
229 MEM_CGROUP_CHARGE_TYPE_ANON
,
230 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
231 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
235 /* for encoding cft->private value on file */
244 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
245 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
246 #define MEMFILE_ATTR(val) ((val) & 0xffff)
247 /* Used for OOM nofiier */
248 #define OOM_CONTROL (0)
250 /* Some nice accessors for the vmpressure. */
251 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
254 memcg
= root_mem_cgroup
;
255 return &memcg
->vmpressure
;
258 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
260 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
263 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
265 return (memcg
== root_mem_cgroup
);
270 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
271 * The main reason for not using cgroup id for this:
272 * this works better in sparse environments, where we have a lot of memcgs,
273 * but only a few kmem-limited. Or also, if we have, for instance, 200
274 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
275 * 200 entry array for that.
277 * The current size of the caches array is stored in memcg_nr_cache_ids. It
278 * will double each time we have to increase it.
280 static DEFINE_IDA(memcg_cache_ida
);
281 int memcg_nr_cache_ids
;
283 /* Protects memcg_nr_cache_ids */
284 static DECLARE_RWSEM(memcg_cache_ids_sem
);
286 void memcg_get_cache_ids(void)
288 down_read(&memcg_cache_ids_sem
);
291 void memcg_put_cache_ids(void)
293 up_read(&memcg_cache_ids_sem
);
297 * MIN_SIZE is different than 1, because we would like to avoid going through
298 * the alloc/free process all the time. In a small machine, 4 kmem-limited
299 * cgroups is a reasonable guess. In the future, it could be a parameter or
300 * tunable, but that is strictly not necessary.
302 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
303 * this constant directly from cgroup, but it is understandable that this is
304 * better kept as an internal representation in cgroup.c. In any case, the
305 * cgrp_id space is not getting any smaller, and we don't have to necessarily
306 * increase ours as well if it increases.
308 #define MEMCG_CACHES_MIN_SIZE 4
309 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
312 * A lot of the calls to the cache allocation functions are expected to be
313 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
314 * conditional to this static branch, we'll have to allow modules that does
315 * kmem_cache_alloc and the such to see this symbol as well
317 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
318 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
320 #endif /* !CONFIG_SLOB */
323 * mem_cgroup_css_from_page - css of the memcg associated with a page
324 * @page: page of interest
326 * If memcg is bound to the default hierarchy, css of the memcg associated
327 * with @page is returned. The returned css remains associated with @page
328 * until it is released.
330 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
333 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
335 struct mem_cgroup
*memcg
;
337 memcg
= page
->mem_cgroup
;
339 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
340 memcg
= root_mem_cgroup
;
346 * page_cgroup_ino - return inode number of the memcg a page is charged to
349 * Look up the closest online ancestor of the memory cgroup @page is charged to
350 * and return its inode number or 0 if @page is not charged to any cgroup. It
351 * is safe to call this function without holding a reference to @page.
353 * Note, this function is inherently racy, because there is nothing to prevent
354 * the cgroup inode from getting torn down and potentially reallocated a moment
355 * after page_cgroup_ino() returns, so it only should be used by callers that
356 * do not care (such as procfs interfaces).
358 ino_t
page_cgroup_ino(struct page
*page
)
360 struct mem_cgroup
*memcg
;
361 unsigned long ino
= 0;
364 memcg
= READ_ONCE(page
->mem_cgroup
);
365 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
366 memcg
= parent_mem_cgroup(memcg
);
368 ino
= cgroup_ino(memcg
->css
.cgroup
);
373 static struct mem_cgroup_per_node
*
374 mem_cgroup_page_nodeinfo(struct mem_cgroup
*memcg
, struct page
*page
)
376 int nid
= page_to_nid(page
);
378 return memcg
->nodeinfo
[nid
];
381 static struct mem_cgroup_tree_per_node
*
382 soft_limit_tree_node(int nid
)
384 return soft_limit_tree
.rb_tree_per_node
[nid
];
387 static struct mem_cgroup_tree_per_node
*
388 soft_limit_tree_from_page(struct page
*page
)
390 int nid
= page_to_nid(page
);
392 return soft_limit_tree
.rb_tree_per_node
[nid
];
395 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node
*mz
,
396 struct mem_cgroup_tree_per_node
*mctz
,
397 unsigned long new_usage_in_excess
)
399 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
400 struct rb_node
*parent
= NULL
;
401 struct mem_cgroup_per_node
*mz_node
;
406 mz
->usage_in_excess
= new_usage_in_excess
;
407 if (!mz
->usage_in_excess
)
411 mz_node
= rb_entry(parent
, struct mem_cgroup_per_node
,
413 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
416 * We can't avoid mem cgroups that are over their soft
417 * limit by the same amount
419 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
422 rb_link_node(&mz
->tree_node
, parent
, p
);
423 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
427 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
428 struct mem_cgroup_tree_per_node
*mctz
)
432 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
436 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
437 struct mem_cgroup_tree_per_node
*mctz
)
441 spin_lock_irqsave(&mctz
->lock
, flags
);
442 __mem_cgroup_remove_exceeded(mz
, mctz
);
443 spin_unlock_irqrestore(&mctz
->lock
, flags
);
446 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
448 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
449 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
450 unsigned long excess
= 0;
452 if (nr_pages
> soft_limit
)
453 excess
= nr_pages
- soft_limit
;
458 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
460 unsigned long excess
;
461 struct mem_cgroup_per_node
*mz
;
462 struct mem_cgroup_tree_per_node
*mctz
;
464 mctz
= soft_limit_tree_from_page(page
);
466 * Necessary to update all ancestors when hierarchy is used.
467 * because their event counter is not touched.
469 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
470 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
471 excess
= soft_limit_excess(memcg
);
473 * We have to update the tree if mz is on RB-tree or
474 * mem is over its softlimit.
476 if (excess
|| mz
->on_tree
) {
479 spin_lock_irqsave(&mctz
->lock
, flags
);
480 /* if on-tree, remove it */
482 __mem_cgroup_remove_exceeded(mz
, mctz
);
484 * Insert again. mz->usage_in_excess will be updated.
485 * If excess is 0, no tree ops.
487 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
488 spin_unlock_irqrestore(&mctz
->lock
, flags
);
493 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
495 struct mem_cgroup_tree_per_node
*mctz
;
496 struct mem_cgroup_per_node
*mz
;
500 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
501 mctz
= soft_limit_tree_node(nid
);
502 mem_cgroup_remove_exceeded(mz
, mctz
);
506 static struct mem_cgroup_per_node
*
507 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
509 struct rb_node
*rightmost
= NULL
;
510 struct mem_cgroup_per_node
*mz
;
514 rightmost
= rb_last(&mctz
->rb_root
);
516 goto done
; /* Nothing to reclaim from */
518 mz
= rb_entry(rightmost
, struct mem_cgroup_per_node
, tree_node
);
520 * Remove the node now but someone else can add it back,
521 * we will to add it back at the end of reclaim to its correct
522 * position in the tree.
524 __mem_cgroup_remove_exceeded(mz
, mctz
);
525 if (!soft_limit_excess(mz
->memcg
) ||
526 !css_tryget_online(&mz
->memcg
->css
))
532 static struct mem_cgroup_per_node
*
533 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
535 struct mem_cgroup_per_node
*mz
;
537 spin_lock_irq(&mctz
->lock
);
538 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
539 spin_unlock_irq(&mctz
->lock
);
544 * Return page count for single (non recursive) @memcg.
546 * Implementation Note: reading percpu statistics for memcg.
548 * Both of vmstat[] and percpu_counter has threshold and do periodic
549 * synchronization to implement "quick" read. There are trade-off between
550 * reading cost and precision of value. Then, we may have a chance to implement
551 * a periodic synchronization of counter in memcg's counter.
553 * But this _read() function is used for user interface now. The user accounts
554 * memory usage by memory cgroup and he _always_ requires exact value because
555 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
556 * have to visit all online cpus and make sum. So, for now, unnecessary
557 * synchronization is not implemented. (just implemented for cpu hotplug)
559 * If there are kernel internal actions which can make use of some not-exact
560 * value, and reading all cpu value can be performance bottleneck in some
561 * common workload, threshold and synchronization as vmstat[] should be
565 mem_cgroup_read_stat(struct mem_cgroup
*memcg
, enum mem_cgroup_stat_index idx
)
570 /* Per-cpu values can be negative, use a signed accumulator */
571 for_each_possible_cpu(cpu
)
572 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
574 * Summing races with updates, so val may be negative. Avoid exposing
575 * transient negative values.
582 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
583 enum mem_cgroup_events_index idx
)
585 unsigned long val
= 0;
588 for_each_possible_cpu(cpu
)
589 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
593 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
595 bool compound
, int nr_pages
)
598 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
599 * counted as CACHE even if it's on ANON LRU.
602 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
605 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
609 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
610 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
614 /* pagein of a big page is an event. So, ignore page size */
616 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
618 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
619 nr_pages
= -nr_pages
; /* for event */
622 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
625 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
626 int nid
, unsigned int lru_mask
)
628 unsigned long nr
= 0;
629 struct mem_cgroup_per_node
*mz
;
632 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
635 if (!(BIT(lru
) & lru_mask
))
637 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
638 nr
+= mz
->lru_size
[lru
];
643 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
644 unsigned int lru_mask
)
646 unsigned long nr
= 0;
649 for_each_node_state(nid
, N_MEMORY
)
650 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
654 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
655 enum mem_cgroup_events_target target
)
657 unsigned long val
, next
;
659 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
660 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
661 /* from time_after() in jiffies.h */
662 if ((long)next
- (long)val
< 0) {
664 case MEM_CGROUP_TARGET_THRESH
:
665 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
667 case MEM_CGROUP_TARGET_SOFTLIMIT
:
668 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
670 case MEM_CGROUP_TARGET_NUMAINFO
:
671 next
= val
+ NUMAINFO_EVENTS_TARGET
;
676 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
683 * Check events in order.
686 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
688 /* threshold event is triggered in finer grain than soft limit */
689 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
690 MEM_CGROUP_TARGET_THRESH
))) {
692 bool do_numainfo __maybe_unused
;
694 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
695 MEM_CGROUP_TARGET_SOFTLIMIT
);
697 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
698 MEM_CGROUP_TARGET_NUMAINFO
);
700 mem_cgroup_threshold(memcg
);
701 if (unlikely(do_softlimit
))
702 mem_cgroup_update_tree(memcg
, page
);
704 if (unlikely(do_numainfo
))
705 atomic_inc(&memcg
->numainfo_events
);
710 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
713 * mm_update_next_owner() may clear mm->owner to NULL
714 * if it races with swapoff, page migration, etc.
715 * So this can be called with p == NULL.
720 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
722 EXPORT_SYMBOL(mem_cgroup_from_task
);
724 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
726 struct mem_cgroup
*memcg
= NULL
;
731 * Page cache insertions can happen withou an
732 * actual mm context, e.g. during disk probing
733 * on boot, loopback IO, acct() writes etc.
736 memcg
= root_mem_cgroup
;
738 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
739 if (unlikely(!memcg
))
740 memcg
= root_mem_cgroup
;
742 } while (!css_tryget_online(&memcg
->css
));
748 * mem_cgroup_iter - iterate over memory cgroup hierarchy
749 * @root: hierarchy root
750 * @prev: previously returned memcg, NULL on first invocation
751 * @reclaim: cookie for shared reclaim walks, NULL for full walks
753 * Returns references to children of the hierarchy below @root, or
754 * @root itself, or %NULL after a full round-trip.
756 * Caller must pass the return value in @prev on subsequent
757 * invocations for reference counting, or use mem_cgroup_iter_break()
758 * to cancel a hierarchy walk before the round-trip is complete.
760 * Reclaimers can specify a zone and a priority level in @reclaim to
761 * divide up the memcgs in the hierarchy among all concurrent
762 * reclaimers operating on the same zone and priority.
764 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
765 struct mem_cgroup
*prev
,
766 struct mem_cgroup_reclaim_cookie
*reclaim
)
768 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
769 struct cgroup_subsys_state
*css
= NULL
;
770 struct mem_cgroup
*memcg
= NULL
;
771 struct mem_cgroup
*pos
= NULL
;
773 if (mem_cgroup_disabled())
777 root
= root_mem_cgroup
;
779 if (prev
&& !reclaim
)
782 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
791 struct mem_cgroup_per_node
*mz
;
793 mz
= mem_cgroup_nodeinfo(root
, reclaim
->pgdat
->node_id
);
794 iter
= &mz
->iter
[reclaim
->priority
];
796 if (prev
&& reclaim
->generation
!= iter
->generation
)
800 pos
= READ_ONCE(iter
->position
);
801 if (!pos
|| css_tryget(&pos
->css
))
804 * css reference reached zero, so iter->position will
805 * be cleared by ->css_released. However, we should not
806 * rely on this happening soon, because ->css_released
807 * is called from a work queue, and by busy-waiting we
808 * might block it. So we clear iter->position right
811 (void)cmpxchg(&iter
->position
, pos
, NULL
);
819 css
= css_next_descendant_pre(css
, &root
->css
);
822 * Reclaimers share the hierarchy walk, and a
823 * new one might jump in right at the end of
824 * the hierarchy - make sure they see at least
825 * one group and restart from the beginning.
833 * Verify the css and acquire a reference. The root
834 * is provided by the caller, so we know it's alive
835 * and kicking, and don't take an extra reference.
837 memcg
= mem_cgroup_from_css(css
);
839 if (css
== &root
->css
)
850 * The position could have already been updated by a competing
851 * thread, so check that the value hasn't changed since we read
852 * it to avoid reclaiming from the same cgroup twice.
854 (void)cmpxchg(&iter
->position
, pos
, memcg
);
862 reclaim
->generation
= iter
->generation
;
868 if (prev
&& prev
!= root
)
875 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
876 * @root: hierarchy root
877 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
879 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
880 struct mem_cgroup
*prev
)
883 root
= root_mem_cgroup
;
884 if (prev
&& prev
!= root
)
888 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
890 struct mem_cgroup
*memcg
= dead_memcg
;
891 struct mem_cgroup_reclaim_iter
*iter
;
892 struct mem_cgroup_per_node
*mz
;
896 while ((memcg
= parent_mem_cgroup(memcg
))) {
898 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
899 for (i
= 0; i
<= DEF_PRIORITY
; i
++) {
901 cmpxchg(&iter
->position
,
909 * Iteration constructs for visiting all cgroups (under a tree). If
910 * loops are exited prematurely (break), mem_cgroup_iter_break() must
911 * be used for reference counting.
913 #define for_each_mem_cgroup_tree(iter, root) \
914 for (iter = mem_cgroup_iter(root, NULL, NULL); \
916 iter = mem_cgroup_iter(root, iter, NULL))
918 #define for_each_mem_cgroup(iter) \
919 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
921 iter = mem_cgroup_iter(NULL, iter, NULL))
924 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
926 * @zone: zone of the page
928 * This function is only safe when following the LRU page isolation
929 * and putback protocol: the LRU lock must be held, and the page must
930 * either be PageLRU() or the caller must have isolated/allocated it.
932 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct pglist_data
*pgdat
)
934 struct mem_cgroup_per_node
*mz
;
935 struct mem_cgroup
*memcg
;
936 struct lruvec
*lruvec
;
938 if (mem_cgroup_disabled()) {
939 lruvec
= &pgdat
->lruvec
;
943 memcg
= page
->mem_cgroup
;
945 * Swapcache readahead pages are added to the LRU - and
946 * possibly migrated - before they are charged.
949 memcg
= root_mem_cgroup
;
951 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
952 lruvec
= &mz
->lruvec
;
955 * Since a node can be onlined after the mem_cgroup was created,
956 * we have to be prepared to initialize lruvec->zone here;
957 * and if offlined then reonlined, we need to reinitialize it.
959 if (unlikely(lruvec
->pgdat
!= pgdat
))
960 lruvec
->pgdat
= pgdat
;
965 * mem_cgroup_update_lru_size - account for adding or removing an lru page
966 * @lruvec: mem_cgroup per zone lru vector
967 * @lru: index of lru list the page is sitting on
968 * @nr_pages: positive when adding or negative when removing
970 * This function must be called under lru_lock, just before a page is added
971 * to or just after a page is removed from an lru list (that ordering being
972 * so as to allow it to check that lru_size 0 is consistent with list_empty).
974 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
977 struct mem_cgroup_per_node
*mz
;
978 unsigned long *lru_size
;
982 if (mem_cgroup_disabled())
985 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
986 lru_size
= mz
->lru_size
+ lru
;
987 empty
= list_empty(lruvec
->lists
+ lru
);
990 *lru_size
+= nr_pages
;
993 if (WARN_ONCE(size
< 0 || empty
!= !size
,
994 "%s(%p, %d, %d): lru_size %ld but %sempty\n",
995 __func__
, lruvec
, lru
, nr_pages
, size
, empty
? "" : "not ")) {
1001 *lru_size
+= nr_pages
;
1004 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1006 struct mem_cgroup
*task_memcg
;
1007 struct task_struct
*p
;
1010 p
= find_lock_task_mm(task
);
1012 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1016 * All threads may have already detached their mm's, but the oom
1017 * killer still needs to detect if they have already been oom
1018 * killed to prevent needlessly killing additional tasks.
1021 task_memcg
= mem_cgroup_from_task(task
);
1022 css_get(&task_memcg
->css
);
1025 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1026 css_put(&task_memcg
->css
);
1031 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1032 * @memcg: the memory cgroup
1034 * Returns the maximum amount of memory @mem can be charged with, in
1037 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1039 unsigned long margin
= 0;
1040 unsigned long count
;
1041 unsigned long limit
;
1043 count
= page_counter_read(&memcg
->memory
);
1044 limit
= READ_ONCE(memcg
->memory
.limit
);
1046 margin
= limit
- count
;
1048 if (do_memsw_account()) {
1049 count
= page_counter_read(&memcg
->memsw
);
1050 limit
= READ_ONCE(memcg
->memsw
.limit
);
1052 margin
= min(margin
, limit
- count
);
1061 * A routine for checking "mem" is under move_account() or not.
1063 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1064 * moving cgroups. This is for waiting at high-memory pressure
1067 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1069 struct mem_cgroup
*from
;
1070 struct mem_cgroup
*to
;
1073 * Unlike task_move routines, we access mc.to, mc.from not under
1074 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1076 spin_lock(&mc
.lock
);
1082 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1083 mem_cgroup_is_descendant(to
, memcg
);
1085 spin_unlock(&mc
.lock
);
1089 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1091 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1092 if (mem_cgroup_under_move(memcg
)) {
1094 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1095 /* moving charge context might have finished. */
1098 finish_wait(&mc
.waitq
, &wait
);
1105 #define K(x) ((x) << (PAGE_SHIFT-10))
1107 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1108 * @memcg: The memory cgroup that went over limit
1109 * @p: Task that is going to be killed
1111 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1114 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1116 struct mem_cgroup
*iter
;
1122 pr_info("Task in ");
1123 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1124 pr_cont(" killed as a result of limit of ");
1126 pr_info("Memory limit reached of cgroup ");
1129 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1134 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1135 K((u64
)page_counter_read(&memcg
->memory
)),
1136 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1137 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1138 K((u64
)page_counter_read(&memcg
->memsw
)),
1139 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1140 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1141 K((u64
)page_counter_read(&memcg
->kmem
)),
1142 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1144 for_each_mem_cgroup_tree(iter
, memcg
) {
1145 pr_info("Memory cgroup stats for ");
1146 pr_cont_cgroup_path(iter
->css
.cgroup
);
1149 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1150 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1152 pr_cont(" %s:%luKB", mem_cgroup_stat_names
[i
],
1153 K(mem_cgroup_read_stat(iter
, i
)));
1156 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1157 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1158 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1165 * This function returns the number of memcg under hierarchy tree. Returns
1166 * 1(self count) if no children.
1168 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1171 struct mem_cgroup
*iter
;
1173 for_each_mem_cgroup_tree(iter
, memcg
)
1179 * Return the memory (and swap, if configured) limit for a memcg.
1181 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1183 unsigned long limit
;
1185 limit
= memcg
->memory
.limit
;
1186 if (mem_cgroup_swappiness(memcg
)) {
1187 unsigned long memsw_limit
;
1188 unsigned long swap_limit
;
1190 memsw_limit
= memcg
->memsw
.limit
;
1191 swap_limit
= memcg
->swap
.limit
;
1192 swap_limit
= min(swap_limit
, (unsigned long)total_swap_pages
);
1193 limit
= min(limit
+ swap_limit
, memsw_limit
);
1198 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1201 struct oom_control oc
= {
1205 .gfp_mask
= gfp_mask
,
1208 struct mem_cgroup
*iter
;
1209 unsigned long chosen_points
= 0;
1210 unsigned long totalpages
;
1211 unsigned int points
= 0;
1212 struct task_struct
*chosen
= NULL
;
1214 mutex_lock(&oom_lock
);
1217 * If current has a pending SIGKILL or is exiting, then automatically
1218 * select it. The goal is to allow it to allocate so that it may
1219 * quickly exit and free its memory.
1221 if (task_will_free_mem(current
)) {
1222 mark_oom_victim(current
);
1223 wake_oom_reaper(current
);
1227 check_panic_on_oom(&oc
, CONSTRAINT_MEMCG
);
1228 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1229 for_each_mem_cgroup_tree(iter
, memcg
) {
1230 struct css_task_iter it
;
1231 struct task_struct
*task
;
1233 css_task_iter_start(&iter
->css
, &it
);
1234 while ((task
= css_task_iter_next(&it
))) {
1235 switch (oom_scan_process_thread(&oc
, task
)) {
1236 case OOM_SCAN_SELECT
:
1238 put_task_struct(chosen
);
1240 chosen_points
= ULONG_MAX
;
1241 get_task_struct(chosen
);
1243 case OOM_SCAN_CONTINUE
:
1245 case OOM_SCAN_ABORT
:
1246 css_task_iter_end(&it
);
1247 mem_cgroup_iter_break(memcg
, iter
);
1249 put_task_struct(chosen
);
1250 /* Set a dummy value to return "true". */
1251 chosen
= (void *) 1;
1256 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1257 if (!points
|| points
< chosen_points
)
1259 /* Prefer thread group leaders for display purposes */
1260 if (points
== chosen_points
&&
1261 thread_group_leader(chosen
))
1265 put_task_struct(chosen
);
1267 chosen_points
= points
;
1268 get_task_struct(chosen
);
1270 css_task_iter_end(&it
);
1274 points
= chosen_points
* 1000 / totalpages
;
1275 oom_kill_process(&oc
, chosen
, points
, totalpages
,
1276 "Memory cgroup out of memory");
1279 mutex_unlock(&oom_lock
);
1283 #if MAX_NUMNODES > 1
1286 * test_mem_cgroup_node_reclaimable
1287 * @memcg: the target memcg
1288 * @nid: the node ID to be checked.
1289 * @noswap : specify true here if the user wants flle only information.
1291 * This function returns whether the specified memcg contains any
1292 * reclaimable pages on a node. Returns true if there are any reclaimable
1293 * pages in the node.
1295 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1296 int nid
, bool noswap
)
1298 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1300 if (noswap
|| !total_swap_pages
)
1302 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1309 * Always updating the nodemask is not very good - even if we have an empty
1310 * list or the wrong list here, we can start from some node and traverse all
1311 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1314 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1318 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1319 * pagein/pageout changes since the last update.
1321 if (!atomic_read(&memcg
->numainfo_events
))
1323 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1326 /* make a nodemask where this memcg uses memory from */
1327 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1329 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1331 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1332 node_clear(nid
, memcg
->scan_nodes
);
1335 atomic_set(&memcg
->numainfo_events
, 0);
1336 atomic_set(&memcg
->numainfo_updating
, 0);
1340 * Selecting a node where we start reclaim from. Because what we need is just
1341 * reducing usage counter, start from anywhere is O,K. Considering
1342 * memory reclaim from current node, there are pros. and cons.
1344 * Freeing memory from current node means freeing memory from a node which
1345 * we'll use or we've used. So, it may make LRU bad. And if several threads
1346 * hit limits, it will see a contention on a node. But freeing from remote
1347 * node means more costs for memory reclaim because of memory latency.
1349 * Now, we use round-robin. Better algorithm is welcomed.
1351 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1355 mem_cgroup_may_update_nodemask(memcg
);
1356 node
= memcg
->last_scanned_node
;
1358 node
= next_node_in(node
, memcg
->scan_nodes
);
1360 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1361 * last time it really checked all the LRUs due to rate limiting.
1362 * Fallback to the current node in that case for simplicity.
1364 if (unlikely(node
== MAX_NUMNODES
))
1365 node
= numa_node_id();
1367 memcg
->last_scanned_node
= node
;
1371 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1377 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1380 unsigned long *total_scanned
)
1382 struct mem_cgroup
*victim
= NULL
;
1385 unsigned long excess
;
1386 unsigned long nr_scanned
;
1387 struct mem_cgroup_reclaim_cookie reclaim
= {
1392 excess
= soft_limit_excess(root_memcg
);
1395 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1400 * If we have not been able to reclaim
1401 * anything, it might because there are
1402 * no reclaimable pages under this hierarchy
1407 * We want to do more targeted reclaim.
1408 * excess >> 2 is not to excessive so as to
1409 * reclaim too much, nor too less that we keep
1410 * coming back to reclaim from this cgroup
1412 if (total
>= (excess
>> 2) ||
1413 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1418 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1419 pgdat
, &nr_scanned
);
1420 *total_scanned
+= nr_scanned
;
1421 if (!soft_limit_excess(root_memcg
))
1424 mem_cgroup_iter_break(root_memcg
, victim
);
1428 #ifdef CONFIG_LOCKDEP
1429 static struct lockdep_map memcg_oom_lock_dep_map
= {
1430 .name
= "memcg_oom_lock",
1434 static DEFINE_SPINLOCK(memcg_oom_lock
);
1437 * Check OOM-Killer is already running under our hierarchy.
1438 * If someone is running, return false.
1440 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1442 struct mem_cgroup
*iter
, *failed
= NULL
;
1444 spin_lock(&memcg_oom_lock
);
1446 for_each_mem_cgroup_tree(iter
, memcg
) {
1447 if (iter
->oom_lock
) {
1449 * this subtree of our hierarchy is already locked
1450 * so we cannot give a lock.
1453 mem_cgroup_iter_break(memcg
, iter
);
1456 iter
->oom_lock
= true;
1461 * OK, we failed to lock the whole subtree so we have
1462 * to clean up what we set up to the failing subtree
1464 for_each_mem_cgroup_tree(iter
, memcg
) {
1465 if (iter
== failed
) {
1466 mem_cgroup_iter_break(memcg
, iter
);
1469 iter
->oom_lock
= false;
1472 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1474 spin_unlock(&memcg_oom_lock
);
1479 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1481 struct mem_cgroup
*iter
;
1483 spin_lock(&memcg_oom_lock
);
1484 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1485 for_each_mem_cgroup_tree(iter
, memcg
)
1486 iter
->oom_lock
= false;
1487 spin_unlock(&memcg_oom_lock
);
1490 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1492 struct mem_cgroup
*iter
;
1494 spin_lock(&memcg_oom_lock
);
1495 for_each_mem_cgroup_tree(iter
, memcg
)
1497 spin_unlock(&memcg_oom_lock
);
1500 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1502 struct mem_cgroup
*iter
;
1505 * When a new child is created while the hierarchy is under oom,
1506 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1508 spin_lock(&memcg_oom_lock
);
1509 for_each_mem_cgroup_tree(iter
, memcg
)
1510 if (iter
->under_oom
> 0)
1512 spin_unlock(&memcg_oom_lock
);
1515 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1517 struct oom_wait_info
{
1518 struct mem_cgroup
*memcg
;
1522 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1523 unsigned mode
, int sync
, void *arg
)
1525 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1526 struct mem_cgroup
*oom_wait_memcg
;
1527 struct oom_wait_info
*oom_wait_info
;
1529 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1530 oom_wait_memcg
= oom_wait_info
->memcg
;
1532 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1533 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1535 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1538 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1541 * For the following lockless ->under_oom test, the only required
1542 * guarantee is that it must see the state asserted by an OOM when
1543 * this function is called as a result of userland actions
1544 * triggered by the notification of the OOM. This is trivially
1545 * achieved by invoking mem_cgroup_mark_under_oom() before
1546 * triggering notification.
1548 if (memcg
&& memcg
->under_oom
)
1549 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1552 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1554 if (!current
->memcg_may_oom
)
1557 * We are in the middle of the charge context here, so we
1558 * don't want to block when potentially sitting on a callstack
1559 * that holds all kinds of filesystem and mm locks.
1561 * Also, the caller may handle a failed allocation gracefully
1562 * (like optional page cache readahead) and so an OOM killer
1563 * invocation might not even be necessary.
1565 * That's why we don't do anything here except remember the
1566 * OOM context and then deal with it at the end of the page
1567 * fault when the stack is unwound, the locks are released,
1568 * and when we know whether the fault was overall successful.
1570 css_get(&memcg
->css
);
1571 current
->memcg_in_oom
= memcg
;
1572 current
->memcg_oom_gfp_mask
= mask
;
1573 current
->memcg_oom_order
= order
;
1577 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1578 * @handle: actually kill/wait or just clean up the OOM state
1580 * This has to be called at the end of a page fault if the memcg OOM
1581 * handler was enabled.
1583 * Memcg supports userspace OOM handling where failed allocations must
1584 * sleep on a waitqueue until the userspace task resolves the
1585 * situation. Sleeping directly in the charge context with all kinds
1586 * of locks held is not a good idea, instead we remember an OOM state
1587 * in the task and mem_cgroup_oom_synchronize() has to be called at
1588 * the end of the page fault to complete the OOM handling.
1590 * Returns %true if an ongoing memcg OOM situation was detected and
1591 * completed, %false otherwise.
1593 bool mem_cgroup_oom_synchronize(bool handle
)
1595 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1596 struct oom_wait_info owait
;
1599 /* OOM is global, do not handle */
1603 if (!handle
|| oom_killer_disabled
)
1606 owait
.memcg
= memcg
;
1607 owait
.wait
.flags
= 0;
1608 owait
.wait
.func
= memcg_oom_wake_function
;
1609 owait
.wait
.private = current
;
1610 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1612 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1613 mem_cgroup_mark_under_oom(memcg
);
1615 locked
= mem_cgroup_oom_trylock(memcg
);
1618 mem_cgroup_oom_notify(memcg
);
1620 if (locked
&& !memcg
->oom_kill_disable
) {
1621 mem_cgroup_unmark_under_oom(memcg
);
1622 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1623 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1624 current
->memcg_oom_order
);
1627 mem_cgroup_unmark_under_oom(memcg
);
1628 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1632 mem_cgroup_oom_unlock(memcg
);
1634 * There is no guarantee that an OOM-lock contender
1635 * sees the wakeups triggered by the OOM kill
1636 * uncharges. Wake any sleepers explicitely.
1638 memcg_oom_recover(memcg
);
1641 current
->memcg_in_oom
= NULL
;
1642 css_put(&memcg
->css
);
1647 * lock_page_memcg - lock a page->mem_cgroup binding
1650 * This function protects unlocked LRU pages from being moved to
1651 * another cgroup and stabilizes their page->mem_cgroup binding.
1653 void lock_page_memcg(struct page
*page
)
1655 struct mem_cgroup
*memcg
;
1656 unsigned long flags
;
1659 * The RCU lock is held throughout the transaction. The fast
1660 * path can get away without acquiring the memcg->move_lock
1661 * because page moving starts with an RCU grace period.
1665 if (mem_cgroup_disabled())
1668 memcg
= page
->mem_cgroup
;
1669 if (unlikely(!memcg
))
1672 if (atomic_read(&memcg
->moving_account
) <= 0)
1675 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1676 if (memcg
!= page
->mem_cgroup
) {
1677 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1682 * When charge migration first begins, we can have locked and
1683 * unlocked page stat updates happening concurrently. Track
1684 * the task who has the lock for unlock_page_memcg().
1686 memcg
->move_lock_task
= current
;
1687 memcg
->move_lock_flags
= flags
;
1691 EXPORT_SYMBOL(lock_page_memcg
);
1694 * unlock_page_memcg - unlock a page->mem_cgroup binding
1697 void unlock_page_memcg(struct page
*page
)
1699 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
1701 if (memcg
&& memcg
->move_lock_task
== current
) {
1702 unsigned long flags
= memcg
->move_lock_flags
;
1704 memcg
->move_lock_task
= NULL
;
1705 memcg
->move_lock_flags
= 0;
1707 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1712 EXPORT_SYMBOL(unlock_page_memcg
);
1715 * size of first charge trial. "32" comes from vmscan.c's magic value.
1716 * TODO: maybe necessary to use big numbers in big irons.
1718 #define CHARGE_BATCH 32U
1719 struct memcg_stock_pcp
{
1720 struct mem_cgroup
*cached
; /* this never be root cgroup */
1721 unsigned int nr_pages
;
1722 struct work_struct work
;
1723 unsigned long flags
;
1724 #define FLUSHING_CACHED_CHARGE 0
1726 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1727 static DEFINE_MUTEX(percpu_charge_mutex
);
1730 * consume_stock: Try to consume stocked charge on this cpu.
1731 * @memcg: memcg to consume from.
1732 * @nr_pages: how many pages to charge.
1734 * The charges will only happen if @memcg matches the current cpu's memcg
1735 * stock, and at least @nr_pages are available in that stock. Failure to
1736 * service an allocation will refill the stock.
1738 * returns true if successful, false otherwise.
1740 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1742 struct memcg_stock_pcp
*stock
;
1745 if (nr_pages
> CHARGE_BATCH
)
1748 stock
= &get_cpu_var(memcg_stock
);
1749 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
1750 stock
->nr_pages
-= nr_pages
;
1753 put_cpu_var(memcg_stock
);
1758 * Returns stocks cached in percpu and reset cached information.
1760 static void drain_stock(struct memcg_stock_pcp
*stock
)
1762 struct mem_cgroup
*old
= stock
->cached
;
1764 if (stock
->nr_pages
) {
1765 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
1766 if (do_memsw_account())
1767 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
1768 css_put_many(&old
->css
, stock
->nr_pages
);
1769 stock
->nr_pages
= 0;
1771 stock
->cached
= NULL
;
1775 * This must be called under preempt disabled or must be called by
1776 * a thread which is pinned to local cpu.
1778 static void drain_local_stock(struct work_struct
*dummy
)
1780 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
1782 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
1786 * Cache charges(val) to local per_cpu area.
1787 * This will be consumed by consume_stock() function, later.
1789 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1791 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
1793 if (stock
->cached
!= memcg
) { /* reset if necessary */
1795 stock
->cached
= memcg
;
1797 stock
->nr_pages
+= nr_pages
;
1798 put_cpu_var(memcg_stock
);
1802 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1803 * of the hierarchy under it.
1805 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
1809 /* If someone's already draining, avoid adding running more workers. */
1810 if (!mutex_trylock(&percpu_charge_mutex
))
1812 /* Notify other cpus that system-wide "drain" is running */
1815 for_each_online_cpu(cpu
) {
1816 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1817 struct mem_cgroup
*memcg
;
1819 memcg
= stock
->cached
;
1820 if (!memcg
|| !stock
->nr_pages
)
1822 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
1824 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
1826 drain_local_stock(&stock
->work
);
1828 schedule_work_on(cpu
, &stock
->work
);
1833 mutex_unlock(&percpu_charge_mutex
);
1836 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
1837 unsigned long action
,
1840 int cpu
= (unsigned long)hcpu
;
1841 struct memcg_stock_pcp
*stock
;
1843 if (action
== CPU_ONLINE
)
1846 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
1849 stock
= &per_cpu(memcg_stock
, cpu
);
1854 static void reclaim_high(struct mem_cgroup
*memcg
,
1855 unsigned int nr_pages
,
1859 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
1861 mem_cgroup_events(memcg
, MEMCG_HIGH
, 1);
1862 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
1863 } while ((memcg
= parent_mem_cgroup(memcg
)));
1866 static void high_work_func(struct work_struct
*work
)
1868 struct mem_cgroup
*memcg
;
1870 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
1871 reclaim_high(memcg
, CHARGE_BATCH
, GFP_KERNEL
);
1875 * Scheduled by try_charge() to be executed from the userland return path
1876 * and reclaims memory over the high limit.
1878 void mem_cgroup_handle_over_high(void)
1880 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
1881 struct mem_cgroup
*memcg
;
1883 if (likely(!nr_pages
))
1886 memcg
= get_mem_cgroup_from_mm(current
->mm
);
1887 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
1888 css_put(&memcg
->css
);
1889 current
->memcg_nr_pages_over_high
= 0;
1892 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1893 unsigned int nr_pages
)
1895 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
1896 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1897 struct mem_cgroup
*mem_over_limit
;
1898 struct page_counter
*counter
;
1899 unsigned long nr_reclaimed
;
1900 bool may_swap
= true;
1901 bool drained
= false;
1903 if (mem_cgroup_is_root(memcg
))
1906 if (consume_stock(memcg
, nr_pages
))
1909 if (!do_memsw_account() ||
1910 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
1911 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
1913 if (do_memsw_account())
1914 page_counter_uncharge(&memcg
->memsw
, batch
);
1915 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
1917 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
1921 if (batch
> nr_pages
) {
1927 * Unlike in global OOM situations, memcg is not in a physical
1928 * memory shortage. Allow dying and OOM-killed tasks to
1929 * bypass the last charges so that they can exit quickly and
1930 * free their memory.
1932 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
1933 fatal_signal_pending(current
) ||
1934 current
->flags
& PF_EXITING
))
1937 if (unlikely(task_in_memcg_oom(current
)))
1940 if (!gfpflags_allow_blocking(gfp_mask
))
1943 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
1945 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
1946 gfp_mask
, may_swap
);
1948 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
1952 drain_all_stock(mem_over_limit
);
1957 if (gfp_mask
& __GFP_NORETRY
)
1960 * Even though the limit is exceeded at this point, reclaim
1961 * may have been able to free some pages. Retry the charge
1962 * before killing the task.
1964 * Only for regular pages, though: huge pages are rather
1965 * unlikely to succeed so close to the limit, and we fall back
1966 * to regular pages anyway in case of failure.
1968 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
1971 * At task move, charge accounts can be doubly counted. So, it's
1972 * better to wait until the end of task_move if something is going on.
1974 if (mem_cgroup_wait_acct_move(mem_over_limit
))
1980 if (gfp_mask
& __GFP_NOFAIL
)
1983 if (fatal_signal_pending(current
))
1986 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
1988 mem_cgroup_oom(mem_over_limit
, gfp_mask
,
1989 get_order(nr_pages
* PAGE_SIZE
));
1991 if (!(gfp_mask
& __GFP_NOFAIL
))
1995 * The allocation either can't fail or will lead to more memory
1996 * being freed very soon. Allow memory usage go over the limit
1997 * temporarily by force charging it.
1999 page_counter_charge(&memcg
->memory
, nr_pages
);
2000 if (do_memsw_account())
2001 page_counter_charge(&memcg
->memsw
, nr_pages
);
2002 css_get_many(&memcg
->css
, nr_pages
);
2007 css_get_many(&memcg
->css
, batch
);
2008 if (batch
> nr_pages
)
2009 refill_stock(memcg
, batch
- nr_pages
);
2012 * If the hierarchy is above the normal consumption range, schedule
2013 * reclaim on returning to userland. We can perform reclaim here
2014 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2015 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2016 * not recorded as it most likely matches current's and won't
2017 * change in the meantime. As high limit is checked again before
2018 * reclaim, the cost of mismatch is negligible.
2021 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2022 /* Don't bother a random interrupted task */
2023 if (in_interrupt()) {
2024 schedule_work(&memcg
->high_work
);
2027 current
->memcg_nr_pages_over_high
+= batch
;
2028 set_notify_resume(current
);
2031 } while ((memcg
= parent_mem_cgroup(memcg
)));
2036 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2038 if (mem_cgroup_is_root(memcg
))
2041 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2042 if (do_memsw_account())
2043 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2045 css_put_many(&memcg
->css
, nr_pages
);
2048 static void lock_page_lru(struct page
*page
, int *isolated
)
2050 struct zone
*zone
= page_zone(page
);
2052 spin_lock_irq(zone_lru_lock(zone
));
2053 if (PageLRU(page
)) {
2054 struct lruvec
*lruvec
;
2056 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
2058 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2064 static void unlock_page_lru(struct page
*page
, int isolated
)
2066 struct zone
*zone
= page_zone(page
);
2069 struct lruvec
*lruvec
;
2071 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
2072 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2074 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2076 spin_unlock_irq(zone_lru_lock(zone
));
2079 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2084 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2087 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2088 * may already be on some other mem_cgroup's LRU. Take care of it.
2091 lock_page_lru(page
, &isolated
);
2094 * Nobody should be changing or seriously looking at
2095 * page->mem_cgroup at this point:
2097 * - the page is uncharged
2099 * - the page is off-LRU
2101 * - an anonymous fault has exclusive page access, except for
2102 * a locked page table
2104 * - a page cache insertion, a swapin fault, or a migration
2105 * have the page locked
2107 page
->mem_cgroup
= memcg
;
2110 unlock_page_lru(page
, isolated
);
2114 static int memcg_alloc_cache_id(void)
2119 id
= ida_simple_get(&memcg_cache_ida
,
2120 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2124 if (id
< memcg_nr_cache_ids
)
2128 * There's no space for the new id in memcg_caches arrays,
2129 * so we have to grow them.
2131 down_write(&memcg_cache_ids_sem
);
2133 size
= 2 * (id
+ 1);
2134 if (size
< MEMCG_CACHES_MIN_SIZE
)
2135 size
= MEMCG_CACHES_MIN_SIZE
;
2136 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2137 size
= MEMCG_CACHES_MAX_SIZE
;
2139 err
= memcg_update_all_caches(size
);
2141 err
= memcg_update_all_list_lrus(size
);
2143 memcg_nr_cache_ids
= size
;
2145 up_write(&memcg_cache_ids_sem
);
2148 ida_simple_remove(&memcg_cache_ida
, id
);
2154 static void memcg_free_cache_id(int id
)
2156 ida_simple_remove(&memcg_cache_ida
, id
);
2159 struct memcg_kmem_cache_create_work
{
2160 struct mem_cgroup
*memcg
;
2161 struct kmem_cache
*cachep
;
2162 struct work_struct work
;
2165 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2167 struct memcg_kmem_cache_create_work
*cw
=
2168 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2169 struct mem_cgroup
*memcg
= cw
->memcg
;
2170 struct kmem_cache
*cachep
= cw
->cachep
;
2172 memcg_create_kmem_cache(memcg
, cachep
);
2174 css_put(&memcg
->css
);
2179 * Enqueue the creation of a per-memcg kmem_cache.
2181 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2182 struct kmem_cache
*cachep
)
2184 struct memcg_kmem_cache_create_work
*cw
;
2186 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2190 css_get(&memcg
->css
);
2193 cw
->cachep
= cachep
;
2194 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2196 schedule_work(&cw
->work
);
2199 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2200 struct kmem_cache
*cachep
)
2203 * We need to stop accounting when we kmalloc, because if the
2204 * corresponding kmalloc cache is not yet created, the first allocation
2205 * in __memcg_schedule_kmem_cache_create will recurse.
2207 * However, it is better to enclose the whole function. Depending on
2208 * the debugging options enabled, INIT_WORK(), for instance, can
2209 * trigger an allocation. This too, will make us recurse. Because at
2210 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2211 * the safest choice is to do it like this, wrapping the whole function.
2213 current
->memcg_kmem_skip_account
= 1;
2214 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2215 current
->memcg_kmem_skip_account
= 0;
2218 static inline bool memcg_kmem_bypass(void)
2220 if (in_interrupt() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
2226 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2227 * @cachep: the original global kmem cache
2229 * Return the kmem_cache we're supposed to use for a slab allocation.
2230 * We try to use the current memcg's version of the cache.
2232 * If the cache does not exist yet, if we are the first user of it, we
2233 * create it asynchronously in a workqueue and let the current allocation
2234 * go through with the original cache.
2236 * This function takes a reference to the cache it returns to assure it
2237 * won't get destroyed while we are working with it. Once the caller is
2238 * done with it, memcg_kmem_put_cache() must be called to release the
2241 struct kmem_cache
*memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2243 struct mem_cgroup
*memcg
;
2244 struct kmem_cache
*memcg_cachep
;
2247 VM_BUG_ON(!is_root_cache(cachep
));
2249 if (memcg_kmem_bypass())
2252 if (current
->memcg_kmem_skip_account
)
2255 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2256 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2260 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2261 if (likely(memcg_cachep
))
2262 return memcg_cachep
;
2265 * If we are in a safe context (can wait, and not in interrupt
2266 * context), we could be be predictable and return right away.
2267 * This would guarantee that the allocation being performed
2268 * already belongs in the new cache.
2270 * However, there are some clashes that can arrive from locking.
2271 * For instance, because we acquire the slab_mutex while doing
2272 * memcg_create_kmem_cache, this means no further allocation
2273 * could happen with the slab_mutex held. So it's better to
2276 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2278 css_put(&memcg
->css
);
2283 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2284 * @cachep: the cache returned by memcg_kmem_get_cache
2286 void memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2288 if (!is_root_cache(cachep
))
2289 css_put(&cachep
->memcg_params
.memcg
->css
);
2293 * memcg_kmem_charge: charge a kmem page
2294 * @page: page to charge
2295 * @gfp: reclaim mode
2296 * @order: allocation order
2297 * @memcg: memory cgroup to charge
2299 * Returns 0 on success, an error code on failure.
2301 int memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2302 struct mem_cgroup
*memcg
)
2304 unsigned int nr_pages
= 1 << order
;
2305 struct page_counter
*counter
;
2308 ret
= try_charge(memcg
, gfp
, nr_pages
);
2312 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2313 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2314 cancel_charge(memcg
, nr_pages
);
2318 page
->mem_cgroup
= memcg
;
2324 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2325 * @page: page to charge
2326 * @gfp: reclaim mode
2327 * @order: allocation order
2329 * Returns 0 on success, an error code on failure.
2331 int memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2333 struct mem_cgroup
*memcg
;
2336 if (memcg_kmem_bypass())
2339 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2340 if (!mem_cgroup_is_root(memcg
)) {
2341 ret
= memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2343 __SetPageKmemcg(page
);
2345 css_put(&memcg
->css
);
2349 * memcg_kmem_uncharge: uncharge a kmem page
2350 * @page: page to uncharge
2351 * @order: allocation order
2353 void memcg_kmem_uncharge(struct page
*page
, int order
)
2355 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2356 unsigned int nr_pages
= 1 << order
;
2361 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2363 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2364 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2366 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2367 if (do_memsw_account())
2368 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2370 page
->mem_cgroup
= NULL
;
2372 /* slab pages do not have PageKmemcg flag set */
2373 if (PageKmemcg(page
))
2374 __ClearPageKmemcg(page
);
2376 css_put_many(&memcg
->css
, nr_pages
);
2378 #endif /* !CONFIG_SLOB */
2380 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2383 * Because tail pages are not marked as "used", set it. We're under
2384 * zone_lru_lock and migration entries setup in all page mappings.
2386 void mem_cgroup_split_huge_fixup(struct page
*head
)
2390 if (mem_cgroup_disabled())
2393 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2394 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2396 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2399 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2401 #ifdef CONFIG_MEMCG_SWAP
2402 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2405 int val
= (charge
) ? 1 : -1;
2406 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2410 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2411 * @entry: swap entry to be moved
2412 * @from: mem_cgroup which the entry is moved from
2413 * @to: mem_cgroup which the entry is moved to
2415 * It succeeds only when the swap_cgroup's record for this entry is the same
2416 * as the mem_cgroup's id of @from.
2418 * Returns 0 on success, -EINVAL on failure.
2420 * The caller must have charged to @to, IOW, called page_counter_charge() about
2421 * both res and memsw, and called css_get().
2423 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2424 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2426 unsigned short old_id
, new_id
;
2428 old_id
= mem_cgroup_id(from
);
2429 new_id
= mem_cgroup_id(to
);
2431 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2432 mem_cgroup_swap_statistics(from
, false);
2433 mem_cgroup_swap_statistics(to
, true);
2439 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2440 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2446 static DEFINE_MUTEX(memcg_limit_mutex
);
2448 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2449 unsigned long limit
)
2451 unsigned long curusage
;
2452 unsigned long oldusage
;
2453 bool enlarge
= false;
2458 * For keeping hierarchical_reclaim simple, how long we should retry
2459 * is depends on callers. We set our retry-count to be function
2460 * of # of children which we should visit in this loop.
2462 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2463 mem_cgroup_count_children(memcg
);
2465 oldusage
= page_counter_read(&memcg
->memory
);
2468 if (signal_pending(current
)) {
2473 mutex_lock(&memcg_limit_mutex
);
2474 if (limit
> memcg
->memsw
.limit
) {
2475 mutex_unlock(&memcg_limit_mutex
);
2479 if (limit
> memcg
->memory
.limit
)
2481 ret
= page_counter_limit(&memcg
->memory
, limit
);
2482 mutex_unlock(&memcg_limit_mutex
);
2487 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2489 curusage
= page_counter_read(&memcg
->memory
);
2490 /* Usage is reduced ? */
2491 if (curusage
>= oldusage
)
2494 oldusage
= curusage
;
2495 } while (retry_count
);
2497 if (!ret
&& enlarge
)
2498 memcg_oom_recover(memcg
);
2503 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2504 unsigned long limit
)
2506 unsigned long curusage
;
2507 unsigned long oldusage
;
2508 bool enlarge
= false;
2512 /* see mem_cgroup_resize_res_limit */
2513 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2514 mem_cgroup_count_children(memcg
);
2516 oldusage
= page_counter_read(&memcg
->memsw
);
2519 if (signal_pending(current
)) {
2524 mutex_lock(&memcg_limit_mutex
);
2525 if (limit
< memcg
->memory
.limit
) {
2526 mutex_unlock(&memcg_limit_mutex
);
2530 if (limit
> memcg
->memsw
.limit
)
2532 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2533 mutex_unlock(&memcg_limit_mutex
);
2538 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2540 curusage
= page_counter_read(&memcg
->memsw
);
2541 /* Usage is reduced ? */
2542 if (curusage
>= oldusage
)
2545 oldusage
= curusage
;
2546 } while (retry_count
);
2548 if (!ret
&& enlarge
)
2549 memcg_oom_recover(memcg
);
2554 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
2556 unsigned long *total_scanned
)
2558 unsigned long nr_reclaimed
= 0;
2559 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
2560 unsigned long reclaimed
;
2562 struct mem_cgroup_tree_per_node
*mctz
;
2563 unsigned long excess
;
2564 unsigned long nr_scanned
;
2569 mctz
= soft_limit_tree_node(pgdat
->node_id
);
2572 * Do not even bother to check the largest node if the root
2573 * is empty. Do it lockless to prevent lock bouncing. Races
2574 * are acceptable as soft limit is best effort anyway.
2576 if (RB_EMPTY_ROOT(&mctz
->rb_root
))
2580 * This loop can run a while, specially if mem_cgroup's continuously
2581 * keep exceeding their soft limit and putting the system under
2588 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2593 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
2594 gfp_mask
, &nr_scanned
);
2595 nr_reclaimed
+= reclaimed
;
2596 *total_scanned
+= nr_scanned
;
2597 spin_lock_irq(&mctz
->lock
);
2598 __mem_cgroup_remove_exceeded(mz
, mctz
);
2601 * If we failed to reclaim anything from this memory cgroup
2602 * it is time to move on to the next cgroup
2606 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2608 excess
= soft_limit_excess(mz
->memcg
);
2610 * One school of thought says that we should not add
2611 * back the node to the tree if reclaim returns 0.
2612 * But our reclaim could return 0, simply because due
2613 * to priority we are exposing a smaller subset of
2614 * memory to reclaim from. Consider this as a longer
2617 /* If excess == 0, no tree ops */
2618 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2619 spin_unlock_irq(&mctz
->lock
);
2620 css_put(&mz
->memcg
->css
);
2623 * Could not reclaim anything and there are no more
2624 * mem cgroups to try or we seem to be looping without
2625 * reclaiming anything.
2627 if (!nr_reclaimed
&&
2629 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2631 } while (!nr_reclaimed
);
2633 css_put(&next_mz
->memcg
->css
);
2634 return nr_reclaimed
;
2638 * Test whether @memcg has children, dead or alive. Note that this
2639 * function doesn't care whether @memcg has use_hierarchy enabled and
2640 * returns %true if there are child csses according to the cgroup
2641 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2643 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2648 ret
= css_next_child(NULL
, &memcg
->css
);
2654 * Reclaims as many pages from the given memcg as possible.
2656 * Caller is responsible for holding css reference for memcg.
2658 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2660 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2662 /* we call try-to-free pages for make this cgroup empty */
2663 lru_add_drain_all();
2664 /* try to free all pages in this cgroup */
2665 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2668 if (signal_pending(current
))
2671 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2675 /* maybe some writeback is necessary */
2676 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2684 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
2685 char *buf
, size_t nbytes
,
2688 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2690 if (mem_cgroup_is_root(memcg
))
2692 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
2695 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
2698 return mem_cgroup_from_css(css
)->use_hierarchy
;
2701 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
2702 struct cftype
*cft
, u64 val
)
2705 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2706 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
2708 if (memcg
->use_hierarchy
== val
)
2712 * If parent's use_hierarchy is set, we can't make any modifications
2713 * in the child subtrees. If it is unset, then the change can
2714 * occur, provided the current cgroup has no children.
2716 * For the root cgroup, parent_mem is NULL, we allow value to be
2717 * set if there are no children.
2719 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
2720 (val
== 1 || val
== 0)) {
2721 if (!memcg_has_children(memcg
))
2722 memcg
->use_hierarchy
= val
;
2731 static void tree_stat(struct mem_cgroup
*memcg
, unsigned long *stat
)
2733 struct mem_cgroup
*iter
;
2736 memset(stat
, 0, sizeof(*stat
) * MEMCG_NR_STAT
);
2738 for_each_mem_cgroup_tree(iter
, memcg
) {
2739 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
2740 stat
[i
] += mem_cgroup_read_stat(iter
, i
);
2744 static void tree_events(struct mem_cgroup
*memcg
, unsigned long *events
)
2746 struct mem_cgroup
*iter
;
2749 memset(events
, 0, sizeof(*events
) * MEMCG_NR_EVENTS
);
2751 for_each_mem_cgroup_tree(iter
, memcg
) {
2752 for (i
= 0; i
< MEMCG_NR_EVENTS
; i
++)
2753 events
[i
] += mem_cgroup_read_events(iter
, i
);
2757 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
2759 unsigned long val
= 0;
2761 if (mem_cgroup_is_root(memcg
)) {
2762 struct mem_cgroup
*iter
;
2764 for_each_mem_cgroup_tree(iter
, memcg
) {
2765 val
+= mem_cgroup_read_stat(iter
,
2766 MEM_CGROUP_STAT_CACHE
);
2767 val
+= mem_cgroup_read_stat(iter
,
2768 MEM_CGROUP_STAT_RSS
);
2770 val
+= mem_cgroup_read_stat(iter
,
2771 MEM_CGROUP_STAT_SWAP
);
2775 val
= page_counter_read(&memcg
->memory
);
2777 val
= page_counter_read(&memcg
->memsw
);
2790 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
2793 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2794 struct page_counter
*counter
;
2796 switch (MEMFILE_TYPE(cft
->private)) {
2798 counter
= &memcg
->memory
;
2801 counter
= &memcg
->memsw
;
2804 counter
= &memcg
->kmem
;
2807 counter
= &memcg
->tcpmem
;
2813 switch (MEMFILE_ATTR(cft
->private)) {
2815 if (counter
== &memcg
->memory
)
2816 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
2817 if (counter
== &memcg
->memsw
)
2818 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
2819 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
2821 return (u64
)counter
->limit
* PAGE_SIZE
;
2823 return (u64
)counter
->watermark
* PAGE_SIZE
;
2825 return counter
->failcnt
;
2826 case RES_SOFT_LIMIT
:
2827 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
2834 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2838 if (cgroup_memory_nokmem
)
2841 BUG_ON(memcg
->kmemcg_id
>= 0);
2842 BUG_ON(memcg
->kmem_state
);
2844 memcg_id
= memcg_alloc_cache_id();
2848 static_branch_inc(&memcg_kmem_enabled_key
);
2850 * A memory cgroup is considered kmem-online as soon as it gets
2851 * kmemcg_id. Setting the id after enabling static branching will
2852 * guarantee no one starts accounting before all call sites are
2855 memcg
->kmemcg_id
= memcg_id
;
2856 memcg
->kmem_state
= KMEM_ONLINE
;
2861 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2863 struct cgroup_subsys_state
*css
;
2864 struct mem_cgroup
*parent
, *child
;
2867 if (memcg
->kmem_state
!= KMEM_ONLINE
)
2870 * Clear the online state before clearing memcg_caches array
2871 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2872 * guarantees that no cache will be created for this cgroup
2873 * after we are done (see memcg_create_kmem_cache()).
2875 memcg
->kmem_state
= KMEM_ALLOCATED
;
2877 memcg_deactivate_kmem_caches(memcg
);
2879 kmemcg_id
= memcg
->kmemcg_id
;
2880 BUG_ON(kmemcg_id
< 0);
2882 parent
= parent_mem_cgroup(memcg
);
2884 parent
= root_mem_cgroup
;
2887 * Change kmemcg_id of this cgroup and all its descendants to the
2888 * parent's id, and then move all entries from this cgroup's list_lrus
2889 * to ones of the parent. After we have finished, all list_lrus
2890 * corresponding to this cgroup are guaranteed to remain empty. The
2891 * ordering is imposed by list_lru_node->lock taken by
2892 * memcg_drain_all_list_lrus().
2894 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2895 css_for_each_descendant_pre(css
, &memcg
->css
) {
2896 child
= mem_cgroup_from_css(css
);
2897 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
2898 child
->kmemcg_id
= parent
->kmemcg_id
;
2899 if (!memcg
->use_hierarchy
)
2904 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
2906 memcg_free_cache_id(kmemcg_id
);
2909 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2911 /* css_alloc() failed, offlining didn't happen */
2912 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
2913 memcg_offline_kmem(memcg
);
2915 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
2916 memcg_destroy_kmem_caches(memcg
);
2917 static_branch_dec(&memcg_kmem_enabled_key
);
2918 WARN_ON(page_counter_read(&memcg
->kmem
));
2922 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2926 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2929 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2932 #endif /* !CONFIG_SLOB */
2934 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2935 unsigned long limit
)
2939 mutex_lock(&memcg_limit_mutex
);
2940 ret
= page_counter_limit(&memcg
->kmem
, limit
);
2941 mutex_unlock(&memcg_limit_mutex
);
2945 static int memcg_update_tcp_limit(struct mem_cgroup
*memcg
, unsigned long limit
)
2949 mutex_lock(&memcg_limit_mutex
);
2951 ret
= page_counter_limit(&memcg
->tcpmem
, limit
);
2955 if (!memcg
->tcpmem_active
) {
2957 * The active flag needs to be written after the static_key
2958 * update. This is what guarantees that the socket activation
2959 * function is the last one to run. See sock_update_memcg() for
2960 * details, and note that we don't mark any socket as belonging
2961 * to this memcg until that flag is up.
2963 * We need to do this, because static_keys will span multiple
2964 * sites, but we can't control their order. If we mark a socket
2965 * as accounted, but the accounting functions are not patched in
2966 * yet, we'll lose accounting.
2968 * We never race with the readers in sock_update_memcg(),
2969 * because when this value change, the code to process it is not
2972 static_branch_inc(&memcg_sockets_enabled_key
);
2973 memcg
->tcpmem_active
= true;
2976 mutex_unlock(&memcg_limit_mutex
);
2981 * The user of this function is...
2984 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
2985 char *buf
, size_t nbytes
, loff_t off
)
2987 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2988 unsigned long nr_pages
;
2991 buf
= strstrip(buf
);
2992 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
2996 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
2998 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3002 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3004 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
3007 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
3010 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
3013 ret
= memcg_update_tcp_limit(memcg
, nr_pages
);
3017 case RES_SOFT_LIMIT
:
3018 memcg
->soft_limit
= nr_pages
;
3022 return ret
?: nbytes
;
3025 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3026 size_t nbytes
, loff_t off
)
3028 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3029 struct page_counter
*counter
;
3031 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3033 counter
= &memcg
->memory
;
3036 counter
= &memcg
->memsw
;
3039 counter
= &memcg
->kmem
;
3042 counter
= &memcg
->tcpmem
;
3048 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3050 page_counter_reset_watermark(counter
);
3053 counter
->failcnt
= 0;
3062 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3065 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3069 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3070 struct cftype
*cft
, u64 val
)
3072 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3074 if (val
& ~MOVE_MASK
)
3078 * No kind of locking is needed in here, because ->can_attach() will
3079 * check this value once in the beginning of the process, and then carry
3080 * on with stale data. This means that changes to this value will only
3081 * affect task migrations starting after the change.
3083 memcg
->move_charge_at_immigrate
= val
;
3087 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3088 struct cftype
*cft
, u64 val
)
3095 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3099 unsigned int lru_mask
;
3102 static const struct numa_stat stats
[] = {
3103 { "total", LRU_ALL
},
3104 { "file", LRU_ALL_FILE
},
3105 { "anon", LRU_ALL_ANON
},
3106 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3108 const struct numa_stat
*stat
;
3111 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3113 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3114 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3115 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3116 for_each_node_state(nid
, N_MEMORY
) {
3117 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3119 seq_printf(m
, " N%d=%lu", nid
, nr
);
3124 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3125 struct mem_cgroup
*iter
;
3128 for_each_mem_cgroup_tree(iter
, memcg
)
3129 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3130 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3131 for_each_node_state(nid
, N_MEMORY
) {
3133 for_each_mem_cgroup_tree(iter
, memcg
)
3134 nr
+= mem_cgroup_node_nr_lru_pages(
3135 iter
, nid
, stat
->lru_mask
);
3136 seq_printf(m
, " N%d=%lu", nid
, nr
);
3143 #endif /* CONFIG_NUMA */
3145 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3147 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3148 unsigned long memory
, memsw
;
3149 struct mem_cgroup
*mi
;
3152 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3153 MEM_CGROUP_STAT_NSTATS
);
3154 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3155 MEM_CGROUP_EVENTS_NSTATS
);
3156 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3158 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3159 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3161 seq_printf(m
, "%s %lu\n", mem_cgroup_stat_names
[i
],
3162 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3165 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3166 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3167 mem_cgroup_read_events(memcg
, i
));
3169 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3170 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3171 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3173 /* Hierarchical information */
3174 memory
= memsw
= PAGE_COUNTER_MAX
;
3175 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3176 memory
= min(memory
, mi
->memory
.limit
);
3177 memsw
= min(memsw
, mi
->memsw
.limit
);
3179 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3180 (u64
)memory
* PAGE_SIZE
);
3181 if (do_memsw_account())
3182 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3183 (u64
)memsw
* PAGE_SIZE
);
3185 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3186 unsigned long long val
= 0;
3188 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3190 for_each_mem_cgroup_tree(mi
, memcg
)
3191 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3192 seq_printf(m
, "total_%s %llu\n", mem_cgroup_stat_names
[i
], val
);
3195 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3196 unsigned long long val
= 0;
3198 for_each_mem_cgroup_tree(mi
, memcg
)
3199 val
+= mem_cgroup_read_events(mi
, i
);
3200 seq_printf(m
, "total_%s %llu\n",
3201 mem_cgroup_events_names
[i
], val
);
3204 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3205 unsigned long long val
= 0;
3207 for_each_mem_cgroup_tree(mi
, memcg
)
3208 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3209 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3212 #ifdef CONFIG_DEBUG_VM
3215 struct mem_cgroup_per_node
*mz
;
3216 struct zone_reclaim_stat
*rstat
;
3217 unsigned long recent_rotated
[2] = {0, 0};
3218 unsigned long recent_scanned
[2] = {0, 0};
3220 for_each_online_pgdat(pgdat
) {
3221 mz
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
3222 rstat
= &mz
->lruvec
.reclaim_stat
;
3224 recent_rotated
[0] += rstat
->recent_rotated
[0];
3225 recent_rotated
[1] += rstat
->recent_rotated
[1];
3226 recent_scanned
[0] += rstat
->recent_scanned
[0];
3227 recent_scanned
[1] += rstat
->recent_scanned
[1];
3229 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3230 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3231 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3232 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3239 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3242 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3244 return mem_cgroup_swappiness(memcg
);
3247 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3248 struct cftype
*cft
, u64 val
)
3250 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3256 memcg
->swappiness
= val
;
3258 vm_swappiness
= val
;
3263 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3265 struct mem_cgroup_threshold_ary
*t
;
3266 unsigned long usage
;
3271 t
= rcu_dereference(memcg
->thresholds
.primary
);
3273 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3278 usage
= mem_cgroup_usage(memcg
, swap
);
3281 * current_threshold points to threshold just below or equal to usage.
3282 * If it's not true, a threshold was crossed after last
3283 * call of __mem_cgroup_threshold().
3285 i
= t
->current_threshold
;
3288 * Iterate backward over array of thresholds starting from
3289 * current_threshold and check if a threshold is crossed.
3290 * If none of thresholds below usage is crossed, we read
3291 * only one element of the array here.
3293 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3294 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3296 /* i = current_threshold + 1 */
3300 * Iterate forward over array of thresholds starting from
3301 * current_threshold+1 and check if a threshold is crossed.
3302 * If none of thresholds above usage is crossed, we read
3303 * only one element of the array here.
3305 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3306 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3308 /* Update current_threshold */
3309 t
->current_threshold
= i
- 1;
3314 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3317 __mem_cgroup_threshold(memcg
, false);
3318 if (do_memsw_account())
3319 __mem_cgroup_threshold(memcg
, true);
3321 memcg
= parent_mem_cgroup(memcg
);
3325 static int compare_thresholds(const void *a
, const void *b
)
3327 const struct mem_cgroup_threshold
*_a
= a
;
3328 const struct mem_cgroup_threshold
*_b
= b
;
3330 if (_a
->threshold
> _b
->threshold
)
3333 if (_a
->threshold
< _b
->threshold
)
3339 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3341 struct mem_cgroup_eventfd_list
*ev
;
3343 spin_lock(&memcg_oom_lock
);
3345 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3346 eventfd_signal(ev
->eventfd
, 1);
3348 spin_unlock(&memcg_oom_lock
);
3352 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3354 struct mem_cgroup
*iter
;
3356 for_each_mem_cgroup_tree(iter
, memcg
)
3357 mem_cgroup_oom_notify_cb(iter
);
3360 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3361 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3363 struct mem_cgroup_thresholds
*thresholds
;
3364 struct mem_cgroup_threshold_ary
*new;
3365 unsigned long threshold
;
3366 unsigned long usage
;
3369 ret
= page_counter_memparse(args
, "-1", &threshold
);
3373 mutex_lock(&memcg
->thresholds_lock
);
3376 thresholds
= &memcg
->thresholds
;
3377 usage
= mem_cgroup_usage(memcg
, false);
3378 } else if (type
== _MEMSWAP
) {
3379 thresholds
= &memcg
->memsw_thresholds
;
3380 usage
= mem_cgroup_usage(memcg
, true);
3384 /* Check if a threshold crossed before adding a new one */
3385 if (thresholds
->primary
)
3386 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3388 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3390 /* Allocate memory for new array of thresholds */
3391 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3399 /* Copy thresholds (if any) to new array */
3400 if (thresholds
->primary
) {
3401 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3402 sizeof(struct mem_cgroup_threshold
));
3405 /* Add new threshold */
3406 new->entries
[size
- 1].eventfd
= eventfd
;
3407 new->entries
[size
- 1].threshold
= threshold
;
3409 /* Sort thresholds. Registering of new threshold isn't time-critical */
3410 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3411 compare_thresholds
, NULL
);
3413 /* Find current threshold */
3414 new->current_threshold
= -1;
3415 for (i
= 0; i
< size
; i
++) {
3416 if (new->entries
[i
].threshold
<= usage
) {
3418 * new->current_threshold will not be used until
3419 * rcu_assign_pointer(), so it's safe to increment
3422 ++new->current_threshold
;
3427 /* Free old spare buffer and save old primary buffer as spare */
3428 kfree(thresholds
->spare
);
3429 thresholds
->spare
= thresholds
->primary
;
3431 rcu_assign_pointer(thresholds
->primary
, new);
3433 /* To be sure that nobody uses thresholds */
3437 mutex_unlock(&memcg
->thresholds_lock
);
3442 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3443 struct eventfd_ctx
*eventfd
, const char *args
)
3445 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3448 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3449 struct eventfd_ctx
*eventfd
, const char *args
)
3451 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3454 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3455 struct eventfd_ctx
*eventfd
, enum res_type type
)
3457 struct mem_cgroup_thresholds
*thresholds
;
3458 struct mem_cgroup_threshold_ary
*new;
3459 unsigned long usage
;
3462 mutex_lock(&memcg
->thresholds_lock
);
3465 thresholds
= &memcg
->thresholds
;
3466 usage
= mem_cgroup_usage(memcg
, false);
3467 } else if (type
== _MEMSWAP
) {
3468 thresholds
= &memcg
->memsw_thresholds
;
3469 usage
= mem_cgroup_usage(memcg
, true);
3473 if (!thresholds
->primary
)
3476 /* Check if a threshold crossed before removing */
3477 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3479 /* Calculate new number of threshold */
3481 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3482 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3486 new = thresholds
->spare
;
3488 /* Set thresholds array to NULL if we don't have thresholds */
3497 /* Copy thresholds and find current threshold */
3498 new->current_threshold
= -1;
3499 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3500 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3503 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3504 if (new->entries
[j
].threshold
<= usage
) {
3506 * new->current_threshold will not be used
3507 * until rcu_assign_pointer(), so it's safe to increment
3510 ++new->current_threshold
;
3516 /* Swap primary and spare array */
3517 thresholds
->spare
= thresholds
->primary
;
3519 rcu_assign_pointer(thresholds
->primary
, new);
3521 /* To be sure that nobody uses thresholds */
3524 /* If all events are unregistered, free the spare array */
3526 kfree(thresholds
->spare
);
3527 thresholds
->spare
= NULL
;
3530 mutex_unlock(&memcg
->thresholds_lock
);
3533 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3534 struct eventfd_ctx
*eventfd
)
3536 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3539 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3540 struct eventfd_ctx
*eventfd
)
3542 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3545 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3546 struct eventfd_ctx
*eventfd
, const char *args
)
3548 struct mem_cgroup_eventfd_list
*event
;
3550 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3554 spin_lock(&memcg_oom_lock
);
3556 event
->eventfd
= eventfd
;
3557 list_add(&event
->list
, &memcg
->oom_notify
);
3559 /* already in OOM ? */
3560 if (memcg
->under_oom
)
3561 eventfd_signal(eventfd
, 1);
3562 spin_unlock(&memcg_oom_lock
);
3567 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3568 struct eventfd_ctx
*eventfd
)
3570 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3572 spin_lock(&memcg_oom_lock
);
3574 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3575 if (ev
->eventfd
== eventfd
) {
3576 list_del(&ev
->list
);
3581 spin_unlock(&memcg_oom_lock
);
3584 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3586 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3588 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3589 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3593 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3594 struct cftype
*cft
, u64 val
)
3596 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3598 /* cannot set to root cgroup and only 0 and 1 are allowed */
3599 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3602 memcg
->oom_kill_disable
= val
;
3604 memcg_oom_recover(memcg
);
3609 #ifdef CONFIG_CGROUP_WRITEBACK
3611 struct list_head
*mem_cgroup_cgwb_list(struct mem_cgroup
*memcg
)
3613 return &memcg
->cgwb_list
;
3616 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3618 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3621 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3623 wb_domain_exit(&memcg
->cgwb_domain
);
3626 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3628 wb_domain_size_changed(&memcg
->cgwb_domain
);
3631 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3633 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3635 if (!memcg
->css
.parent
)
3638 return &memcg
->cgwb_domain
;
3642 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3643 * @wb: bdi_writeback in question
3644 * @pfilepages: out parameter for number of file pages
3645 * @pheadroom: out parameter for number of allocatable pages according to memcg
3646 * @pdirty: out parameter for number of dirty pages
3647 * @pwriteback: out parameter for number of pages under writeback
3649 * Determine the numbers of file, headroom, dirty, and writeback pages in
3650 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3651 * is a bit more involved.
3653 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3654 * headroom is calculated as the lowest headroom of itself and the
3655 * ancestors. Note that this doesn't consider the actual amount of
3656 * available memory in the system. The caller should further cap
3657 * *@pheadroom accordingly.
3659 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
3660 unsigned long *pheadroom
, unsigned long *pdirty
,
3661 unsigned long *pwriteback
)
3663 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3664 struct mem_cgroup
*parent
;
3666 *pdirty
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
3668 /* this should eventually include NR_UNSTABLE_NFS */
3669 *pwriteback
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
3670 *pfilepages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
3671 (1 << LRU_ACTIVE_FILE
));
3672 *pheadroom
= PAGE_COUNTER_MAX
;
3674 while ((parent
= parent_mem_cgroup(memcg
))) {
3675 unsigned long ceiling
= min(memcg
->memory
.limit
, memcg
->high
);
3676 unsigned long used
= page_counter_read(&memcg
->memory
);
3678 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
3683 #else /* CONFIG_CGROUP_WRITEBACK */
3685 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3690 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3694 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3698 #endif /* CONFIG_CGROUP_WRITEBACK */
3701 * DO NOT USE IN NEW FILES.
3703 * "cgroup.event_control" implementation.
3705 * This is way over-engineered. It tries to support fully configurable
3706 * events for each user. Such level of flexibility is completely
3707 * unnecessary especially in the light of the planned unified hierarchy.
3709 * Please deprecate this and replace with something simpler if at all
3714 * Unregister event and free resources.
3716 * Gets called from workqueue.
3718 static void memcg_event_remove(struct work_struct
*work
)
3720 struct mem_cgroup_event
*event
=
3721 container_of(work
, struct mem_cgroup_event
, remove
);
3722 struct mem_cgroup
*memcg
= event
->memcg
;
3724 remove_wait_queue(event
->wqh
, &event
->wait
);
3726 event
->unregister_event(memcg
, event
->eventfd
);
3728 /* Notify userspace the event is going away. */
3729 eventfd_signal(event
->eventfd
, 1);
3731 eventfd_ctx_put(event
->eventfd
);
3733 css_put(&memcg
->css
);
3737 * Gets called on POLLHUP on eventfd when user closes it.
3739 * Called with wqh->lock held and interrupts disabled.
3741 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
3742 int sync
, void *key
)
3744 struct mem_cgroup_event
*event
=
3745 container_of(wait
, struct mem_cgroup_event
, wait
);
3746 struct mem_cgroup
*memcg
= event
->memcg
;
3747 unsigned long flags
= (unsigned long)key
;
3749 if (flags
& POLLHUP
) {
3751 * If the event has been detached at cgroup removal, we
3752 * can simply return knowing the other side will cleanup
3755 * We can't race against event freeing since the other
3756 * side will require wqh->lock via remove_wait_queue(),
3759 spin_lock(&memcg
->event_list_lock
);
3760 if (!list_empty(&event
->list
)) {
3761 list_del_init(&event
->list
);
3763 * We are in atomic context, but cgroup_event_remove()
3764 * may sleep, so we have to call it in workqueue.
3766 schedule_work(&event
->remove
);
3768 spin_unlock(&memcg
->event_list_lock
);
3774 static void memcg_event_ptable_queue_proc(struct file
*file
,
3775 wait_queue_head_t
*wqh
, poll_table
*pt
)
3777 struct mem_cgroup_event
*event
=
3778 container_of(pt
, struct mem_cgroup_event
, pt
);
3781 add_wait_queue(wqh
, &event
->wait
);
3785 * DO NOT USE IN NEW FILES.
3787 * Parse input and register new cgroup event handler.
3789 * Input must be in format '<event_fd> <control_fd> <args>'.
3790 * Interpretation of args is defined by control file implementation.
3792 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
3793 char *buf
, size_t nbytes
, loff_t off
)
3795 struct cgroup_subsys_state
*css
= of_css(of
);
3796 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3797 struct mem_cgroup_event
*event
;
3798 struct cgroup_subsys_state
*cfile_css
;
3799 unsigned int efd
, cfd
;
3806 buf
= strstrip(buf
);
3808 efd
= simple_strtoul(buf
, &endp
, 10);
3813 cfd
= simple_strtoul(buf
, &endp
, 10);
3814 if ((*endp
!= ' ') && (*endp
!= '\0'))
3818 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
3822 event
->memcg
= memcg
;
3823 INIT_LIST_HEAD(&event
->list
);
3824 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
3825 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
3826 INIT_WORK(&event
->remove
, memcg_event_remove
);
3834 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
3835 if (IS_ERR(event
->eventfd
)) {
3836 ret
= PTR_ERR(event
->eventfd
);
3843 goto out_put_eventfd
;
3846 /* the process need read permission on control file */
3847 /* AV: shouldn't we check that it's been opened for read instead? */
3848 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
3853 * Determine the event callbacks and set them in @event. This used
3854 * to be done via struct cftype but cgroup core no longer knows
3855 * about these events. The following is crude but the whole thing
3856 * is for compatibility anyway.
3858 * DO NOT ADD NEW FILES.
3860 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
3862 if (!strcmp(name
, "memory.usage_in_bytes")) {
3863 event
->register_event
= mem_cgroup_usage_register_event
;
3864 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
3865 } else if (!strcmp(name
, "memory.oom_control")) {
3866 event
->register_event
= mem_cgroup_oom_register_event
;
3867 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
3868 } else if (!strcmp(name
, "memory.pressure_level")) {
3869 event
->register_event
= vmpressure_register_event
;
3870 event
->unregister_event
= vmpressure_unregister_event
;
3871 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
3872 event
->register_event
= memsw_cgroup_usage_register_event
;
3873 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
3880 * Verify @cfile should belong to @css. Also, remaining events are
3881 * automatically removed on cgroup destruction but the removal is
3882 * asynchronous, so take an extra ref on @css.
3884 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
3885 &memory_cgrp_subsys
);
3887 if (IS_ERR(cfile_css
))
3889 if (cfile_css
!= css
) {
3894 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
3898 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
3900 spin_lock(&memcg
->event_list_lock
);
3901 list_add(&event
->list
, &memcg
->event_list
);
3902 spin_unlock(&memcg
->event_list_lock
);
3914 eventfd_ctx_put(event
->eventfd
);
3923 static struct cftype mem_cgroup_legacy_files
[] = {
3925 .name
= "usage_in_bytes",
3926 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
3927 .read_u64
= mem_cgroup_read_u64
,
3930 .name
= "max_usage_in_bytes",
3931 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
3932 .write
= mem_cgroup_reset
,
3933 .read_u64
= mem_cgroup_read_u64
,
3936 .name
= "limit_in_bytes",
3937 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
3938 .write
= mem_cgroup_write
,
3939 .read_u64
= mem_cgroup_read_u64
,
3942 .name
= "soft_limit_in_bytes",
3943 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
3944 .write
= mem_cgroup_write
,
3945 .read_u64
= mem_cgroup_read_u64
,
3949 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
3950 .write
= mem_cgroup_reset
,
3951 .read_u64
= mem_cgroup_read_u64
,
3955 .seq_show
= memcg_stat_show
,
3958 .name
= "force_empty",
3959 .write
= mem_cgroup_force_empty_write
,
3962 .name
= "use_hierarchy",
3963 .write_u64
= mem_cgroup_hierarchy_write
,
3964 .read_u64
= mem_cgroup_hierarchy_read
,
3967 .name
= "cgroup.event_control", /* XXX: for compat */
3968 .write
= memcg_write_event_control
,
3969 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
3972 .name
= "swappiness",
3973 .read_u64
= mem_cgroup_swappiness_read
,
3974 .write_u64
= mem_cgroup_swappiness_write
,
3977 .name
= "move_charge_at_immigrate",
3978 .read_u64
= mem_cgroup_move_charge_read
,
3979 .write_u64
= mem_cgroup_move_charge_write
,
3982 .name
= "oom_control",
3983 .seq_show
= mem_cgroup_oom_control_read
,
3984 .write_u64
= mem_cgroup_oom_control_write
,
3985 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
3988 .name
= "pressure_level",
3992 .name
= "numa_stat",
3993 .seq_show
= memcg_numa_stat_show
,
3997 .name
= "kmem.limit_in_bytes",
3998 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
3999 .write
= mem_cgroup_write
,
4000 .read_u64
= mem_cgroup_read_u64
,
4003 .name
= "kmem.usage_in_bytes",
4004 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4005 .read_u64
= mem_cgroup_read_u64
,
4008 .name
= "kmem.failcnt",
4009 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4010 .write
= mem_cgroup_reset
,
4011 .read_u64
= mem_cgroup_read_u64
,
4014 .name
= "kmem.max_usage_in_bytes",
4015 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4016 .write
= mem_cgroup_reset
,
4017 .read_u64
= mem_cgroup_read_u64
,
4019 #ifdef CONFIG_SLABINFO
4021 .name
= "kmem.slabinfo",
4022 .seq_start
= slab_start
,
4023 .seq_next
= slab_next
,
4024 .seq_stop
= slab_stop
,
4025 .seq_show
= memcg_slab_show
,
4029 .name
= "kmem.tcp.limit_in_bytes",
4030 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4031 .write
= mem_cgroup_write
,
4032 .read_u64
= mem_cgroup_read_u64
,
4035 .name
= "kmem.tcp.usage_in_bytes",
4036 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4037 .read_u64
= mem_cgroup_read_u64
,
4040 .name
= "kmem.tcp.failcnt",
4041 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4042 .write
= mem_cgroup_reset
,
4043 .read_u64
= mem_cgroup_read_u64
,
4046 .name
= "kmem.tcp.max_usage_in_bytes",
4047 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4048 .write
= mem_cgroup_reset
,
4049 .read_u64
= mem_cgroup_read_u64
,
4051 { }, /* terminate */
4055 * Private memory cgroup IDR
4057 * Swap-out records and page cache shadow entries need to store memcg
4058 * references in constrained space, so we maintain an ID space that is
4059 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4060 * memory-controlled cgroups to 64k.
4062 * However, there usually are many references to the oflline CSS after
4063 * the cgroup has been destroyed, such as page cache or reclaimable
4064 * slab objects, that don't need to hang on to the ID. We want to keep
4065 * those dead CSS from occupying IDs, or we might quickly exhaust the
4066 * relatively small ID space and prevent the creation of new cgroups
4067 * even when there are much fewer than 64k cgroups - possibly none.
4069 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4070 * be freed and recycled when it's no longer needed, which is usually
4071 * when the CSS is offlined.
4073 * The only exception to that are records of swapped out tmpfs/shmem
4074 * pages that need to be attributed to live ancestors on swapin. But
4075 * those references are manageable from userspace.
4078 static DEFINE_IDR(mem_cgroup_idr
);
4080 static void mem_cgroup_id_get_many(struct mem_cgroup
*memcg
, unsigned int n
)
4082 atomic_add(n
, &memcg
->id
.ref
);
4085 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
4087 while (!atomic_inc_not_zero(&memcg
->id
.ref
)) {
4089 * The root cgroup cannot be destroyed, so it's refcount must
4092 if (WARN_ON_ONCE(memcg
== root_mem_cgroup
)) {
4096 memcg
= parent_mem_cgroup(memcg
);
4098 memcg
= root_mem_cgroup
;
4103 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
4105 if (atomic_sub_and_test(n
, &memcg
->id
.ref
)) {
4106 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4109 /* Memcg ID pins CSS */
4110 css_put(&memcg
->css
);
4114 static inline void mem_cgroup_id_get(struct mem_cgroup
*memcg
)
4116 mem_cgroup_id_get_many(memcg
, 1);
4119 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
4121 mem_cgroup_id_put_many(memcg
, 1);
4125 * mem_cgroup_from_id - look up a memcg from a memcg id
4126 * @id: the memcg id to look up
4128 * Caller must hold rcu_read_lock().
4130 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
4132 WARN_ON_ONCE(!rcu_read_lock_held());
4133 return idr_find(&mem_cgroup_idr
, id
);
4136 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4138 struct mem_cgroup_per_node
*pn
;
4141 * This routine is called against possible nodes.
4142 * But it's BUG to call kmalloc() against offline node.
4144 * TODO: this routine can waste much memory for nodes which will
4145 * never be onlined. It's better to use memory hotplug callback
4148 if (!node_state(node
, N_NORMAL_MEMORY
))
4150 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4154 lruvec_init(&pn
->lruvec
);
4155 pn
->usage_in_excess
= 0;
4156 pn
->on_tree
= false;
4159 memcg
->nodeinfo
[node
] = pn
;
4163 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4165 kfree(memcg
->nodeinfo
[node
]);
4168 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4172 memcg_wb_domain_exit(memcg
);
4174 free_mem_cgroup_per_node_info(memcg
, node
);
4175 free_percpu(memcg
->stat
);
4179 static struct mem_cgroup
*mem_cgroup_alloc(void)
4181 struct mem_cgroup
*memcg
;
4185 size
= sizeof(struct mem_cgroup
);
4186 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4188 memcg
= kzalloc(size
, GFP_KERNEL
);
4192 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
4193 1, MEM_CGROUP_ID_MAX
,
4195 if (memcg
->id
.id
< 0)
4198 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4203 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
4206 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4209 INIT_WORK(&memcg
->high_work
, high_work_func
);
4210 memcg
->last_scanned_node
= MAX_NUMNODES
;
4211 INIT_LIST_HEAD(&memcg
->oom_notify
);
4212 mutex_init(&memcg
->thresholds_lock
);
4213 spin_lock_init(&memcg
->move_lock
);
4214 vmpressure_init(&memcg
->vmpressure
);
4215 INIT_LIST_HEAD(&memcg
->event_list
);
4216 spin_lock_init(&memcg
->event_list_lock
);
4217 memcg
->socket_pressure
= jiffies
;
4219 memcg
->kmemcg_id
= -1;
4221 #ifdef CONFIG_CGROUP_WRITEBACK
4222 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4224 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
4227 if (memcg
->id
.id
> 0)
4228 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4229 mem_cgroup_free(memcg
);
4233 static struct cgroup_subsys_state
* __ref
4234 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4236 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
4237 struct mem_cgroup
*memcg
;
4238 long error
= -ENOMEM
;
4240 memcg
= mem_cgroup_alloc();
4242 return ERR_PTR(error
);
4244 memcg
->high
= PAGE_COUNTER_MAX
;
4245 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4247 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4248 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4250 if (parent
&& parent
->use_hierarchy
) {
4251 memcg
->use_hierarchy
= true;
4252 page_counter_init(&memcg
->memory
, &parent
->memory
);
4253 page_counter_init(&memcg
->swap
, &parent
->swap
);
4254 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4255 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4256 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
4258 page_counter_init(&memcg
->memory
, NULL
);
4259 page_counter_init(&memcg
->swap
, NULL
);
4260 page_counter_init(&memcg
->memsw
, NULL
);
4261 page_counter_init(&memcg
->kmem
, NULL
);
4262 page_counter_init(&memcg
->tcpmem
, NULL
);
4264 * Deeper hierachy with use_hierarchy == false doesn't make
4265 * much sense so let cgroup subsystem know about this
4266 * unfortunate state in our controller.
4268 if (parent
!= root_mem_cgroup
)
4269 memory_cgrp_subsys
.broken_hierarchy
= true;
4272 /* The following stuff does not apply to the root */
4274 root_mem_cgroup
= memcg
;
4278 error
= memcg_online_kmem(memcg
);
4282 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4283 static_branch_inc(&memcg_sockets_enabled_key
);
4287 mem_cgroup_free(memcg
);
4288 return ERR_PTR(-ENOMEM
);
4291 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4293 /* Online state pins memcg ID, memcg ID pins CSS */
4294 mem_cgroup_id_get(mem_cgroup_from_css(css
));
4299 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4301 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4302 struct mem_cgroup_event
*event
, *tmp
;
4305 * Unregister events and notify userspace.
4306 * Notify userspace about cgroup removing only after rmdir of cgroup
4307 * directory to avoid race between userspace and kernelspace.
4309 spin_lock(&memcg
->event_list_lock
);
4310 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4311 list_del_init(&event
->list
);
4312 schedule_work(&event
->remove
);
4314 spin_unlock(&memcg
->event_list_lock
);
4316 memcg_offline_kmem(memcg
);
4317 wb_memcg_offline(memcg
);
4319 mem_cgroup_id_put(memcg
);
4322 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4324 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4326 invalidate_reclaim_iterators(memcg
);
4329 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4331 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4333 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4334 static_branch_dec(&memcg_sockets_enabled_key
);
4336 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
4337 static_branch_dec(&memcg_sockets_enabled_key
);
4339 vmpressure_cleanup(&memcg
->vmpressure
);
4340 cancel_work_sync(&memcg
->high_work
);
4341 mem_cgroup_remove_from_trees(memcg
);
4342 memcg_free_kmem(memcg
);
4343 mem_cgroup_free(memcg
);
4347 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4348 * @css: the target css
4350 * Reset the states of the mem_cgroup associated with @css. This is
4351 * invoked when the userland requests disabling on the default hierarchy
4352 * but the memcg is pinned through dependency. The memcg should stop
4353 * applying policies and should revert to the vanilla state as it may be
4354 * made visible again.
4356 * The current implementation only resets the essential configurations.
4357 * This needs to be expanded to cover all the visible parts.
4359 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4361 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4363 page_counter_limit(&memcg
->memory
, PAGE_COUNTER_MAX
);
4364 page_counter_limit(&memcg
->swap
, PAGE_COUNTER_MAX
);
4365 page_counter_limit(&memcg
->memsw
, PAGE_COUNTER_MAX
);
4366 page_counter_limit(&memcg
->kmem
, PAGE_COUNTER_MAX
);
4367 page_counter_limit(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
4369 memcg
->high
= PAGE_COUNTER_MAX
;
4370 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4371 memcg_wb_domain_size_changed(memcg
);
4375 /* Handlers for move charge at task migration. */
4376 static int mem_cgroup_do_precharge(unsigned long count
)
4380 /* Try a single bulk charge without reclaim first, kswapd may wake */
4381 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4383 mc
.precharge
+= count
;
4387 /* Try charges one by one with reclaim */
4389 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4403 enum mc_target_type
{
4409 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4410 unsigned long addr
, pte_t ptent
)
4412 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4414 if (!page
|| !page_mapped(page
))
4416 if (PageAnon(page
)) {
4417 if (!(mc
.flags
& MOVE_ANON
))
4420 if (!(mc
.flags
& MOVE_FILE
))
4423 if (!get_page_unless_zero(page
))
4430 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4431 pte_t ptent
, swp_entry_t
*entry
)
4433 struct page
*page
= NULL
;
4434 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4436 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4439 * Because lookup_swap_cache() updates some statistics counter,
4440 * we call find_get_page() with swapper_space directly.
4442 page
= find_get_page(swap_address_space(ent
), ent
.val
);
4443 if (do_memsw_account())
4444 entry
->val
= ent
.val
;
4449 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4450 pte_t ptent
, swp_entry_t
*entry
)
4456 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4457 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4459 struct page
*page
= NULL
;
4460 struct address_space
*mapping
;
4463 if (!vma
->vm_file
) /* anonymous vma */
4465 if (!(mc
.flags
& MOVE_FILE
))
4468 mapping
= vma
->vm_file
->f_mapping
;
4469 pgoff
= linear_page_index(vma
, addr
);
4471 /* page is moved even if it's not RSS of this task(page-faulted). */
4473 /* shmem/tmpfs may report page out on swap: account for that too. */
4474 if (shmem_mapping(mapping
)) {
4475 page
= find_get_entry(mapping
, pgoff
);
4476 if (radix_tree_exceptional_entry(page
)) {
4477 swp_entry_t swp
= radix_to_swp_entry(page
);
4478 if (do_memsw_account())
4480 page
= find_get_page(swap_address_space(swp
), swp
.val
);
4483 page
= find_get_page(mapping
, pgoff
);
4485 page
= find_get_page(mapping
, pgoff
);
4491 * mem_cgroup_move_account - move account of the page
4493 * @compound: charge the page as compound or small page
4494 * @from: mem_cgroup which the page is moved from.
4495 * @to: mem_cgroup which the page is moved to. @from != @to.
4497 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4499 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4502 static int mem_cgroup_move_account(struct page
*page
,
4504 struct mem_cgroup
*from
,
4505 struct mem_cgroup
*to
)
4507 unsigned long flags
;
4508 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
4512 VM_BUG_ON(from
== to
);
4513 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4514 VM_BUG_ON(compound
&& !PageTransHuge(page
));
4517 * Prevent mem_cgroup_migrate() from looking at
4518 * page->mem_cgroup of its source page while we change it.
4521 if (!trylock_page(page
))
4525 if (page
->mem_cgroup
!= from
)
4528 anon
= PageAnon(page
);
4530 spin_lock_irqsave(&from
->move_lock
, flags
);
4532 if (!anon
&& page_mapped(page
)) {
4533 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4535 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4540 * move_lock grabbed above and caller set from->moving_account, so
4541 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4542 * So mapping should be stable for dirty pages.
4544 if (!anon
&& PageDirty(page
)) {
4545 struct address_space
*mapping
= page_mapping(page
);
4547 if (mapping_cap_account_dirty(mapping
)) {
4548 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4550 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4555 if (PageWriteback(page
)) {
4556 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4558 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4563 * It is safe to change page->mem_cgroup here because the page
4564 * is referenced, charged, and isolated - we can't race with
4565 * uncharging, charging, migration, or LRU putback.
4568 /* caller should have done css_get */
4569 page
->mem_cgroup
= to
;
4570 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4574 local_irq_disable();
4575 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
4576 memcg_check_events(to
, page
);
4577 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
4578 memcg_check_events(from
, page
);
4587 * get_mctgt_type - get target type of moving charge
4588 * @vma: the vma the pte to be checked belongs
4589 * @addr: the address corresponding to the pte to be checked
4590 * @ptent: the pte to be checked
4591 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4594 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4595 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4596 * move charge. if @target is not NULL, the page is stored in target->page
4597 * with extra refcnt got(Callers should handle it).
4598 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4599 * target for charge migration. if @target is not NULL, the entry is stored
4602 * Called with pte lock held.
4605 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4606 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4608 struct page
*page
= NULL
;
4609 enum mc_target_type ret
= MC_TARGET_NONE
;
4610 swp_entry_t ent
= { .val
= 0 };
4612 if (pte_present(ptent
))
4613 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4614 else if (is_swap_pte(ptent
))
4615 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
4616 else if (pte_none(ptent
))
4617 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4619 if (!page
&& !ent
.val
)
4623 * Do only loose check w/o serialization.
4624 * mem_cgroup_move_account() checks the page is valid or
4625 * not under LRU exclusion.
4627 if (page
->mem_cgroup
== mc
.from
) {
4628 ret
= MC_TARGET_PAGE
;
4630 target
->page
= page
;
4632 if (!ret
|| !target
)
4635 /* There is a swap entry and a page doesn't exist or isn't charged */
4636 if (ent
.val
&& !ret
&&
4637 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4638 ret
= MC_TARGET_SWAP
;
4645 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4647 * We don't consider swapping or file mapped pages because THP does not
4648 * support them for now.
4649 * Caller should make sure that pmd_trans_huge(pmd) is true.
4651 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4652 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4654 struct page
*page
= NULL
;
4655 enum mc_target_type ret
= MC_TARGET_NONE
;
4657 page
= pmd_page(pmd
);
4658 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4659 if (!(mc
.flags
& MOVE_ANON
))
4661 if (page
->mem_cgroup
== mc
.from
) {
4662 ret
= MC_TARGET_PAGE
;
4665 target
->page
= page
;
4671 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4672 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4674 return MC_TARGET_NONE
;
4678 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4679 unsigned long addr
, unsigned long end
,
4680 struct mm_walk
*walk
)
4682 struct vm_area_struct
*vma
= walk
->vma
;
4686 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4688 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4689 mc
.precharge
+= HPAGE_PMD_NR
;
4694 if (pmd_trans_unstable(pmd
))
4696 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4697 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4698 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4699 mc
.precharge
++; /* increment precharge temporarily */
4700 pte_unmap_unlock(pte
- 1, ptl
);
4706 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4708 unsigned long precharge
;
4710 struct mm_walk mem_cgroup_count_precharge_walk
= {
4711 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4714 down_read(&mm
->mmap_sem
);
4715 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk
);
4716 up_read(&mm
->mmap_sem
);
4718 precharge
= mc
.precharge
;
4724 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4726 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4728 VM_BUG_ON(mc
.moving_task
);
4729 mc
.moving_task
= current
;
4730 return mem_cgroup_do_precharge(precharge
);
4733 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4734 static void __mem_cgroup_clear_mc(void)
4736 struct mem_cgroup
*from
= mc
.from
;
4737 struct mem_cgroup
*to
= mc
.to
;
4739 /* we must uncharge all the leftover precharges from mc.to */
4741 cancel_charge(mc
.to
, mc
.precharge
);
4745 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4746 * we must uncharge here.
4748 if (mc
.moved_charge
) {
4749 cancel_charge(mc
.from
, mc
.moved_charge
);
4750 mc
.moved_charge
= 0;
4752 /* we must fixup refcnts and charges */
4753 if (mc
.moved_swap
) {
4754 /* uncharge swap account from the old cgroup */
4755 if (!mem_cgroup_is_root(mc
.from
))
4756 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4758 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
4761 * we charged both to->memory and to->memsw, so we
4762 * should uncharge to->memory.
4764 if (!mem_cgroup_is_root(mc
.to
))
4765 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4767 mem_cgroup_id_get_many(mc
.to
, mc
.moved_swap
);
4768 css_put_many(&mc
.to
->css
, mc
.moved_swap
);
4772 memcg_oom_recover(from
);
4773 memcg_oom_recover(to
);
4774 wake_up_all(&mc
.waitq
);
4777 static void mem_cgroup_clear_mc(void)
4779 struct mm_struct
*mm
= mc
.mm
;
4782 * we must clear moving_task before waking up waiters at the end of
4785 mc
.moving_task
= NULL
;
4786 __mem_cgroup_clear_mc();
4787 spin_lock(&mc
.lock
);
4791 spin_unlock(&mc
.lock
);
4796 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4798 struct cgroup_subsys_state
*css
;
4799 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
4800 struct mem_cgroup
*from
;
4801 struct task_struct
*leader
, *p
;
4802 struct mm_struct
*mm
;
4803 unsigned long move_flags
;
4806 /* charge immigration isn't supported on the default hierarchy */
4807 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4811 * Multi-process migrations only happen on the default hierarchy
4812 * where charge immigration is not used. Perform charge
4813 * immigration if @tset contains a leader and whine if there are
4817 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
4820 memcg
= mem_cgroup_from_css(css
);
4826 * We are now commited to this value whatever it is. Changes in this
4827 * tunable will only affect upcoming migrations, not the current one.
4828 * So we need to save it, and keep it going.
4830 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
4834 from
= mem_cgroup_from_task(p
);
4836 VM_BUG_ON(from
== memcg
);
4838 mm
= get_task_mm(p
);
4841 /* We move charges only when we move a owner of the mm */
4842 if (mm
->owner
== p
) {
4845 VM_BUG_ON(mc
.precharge
);
4846 VM_BUG_ON(mc
.moved_charge
);
4847 VM_BUG_ON(mc
.moved_swap
);
4849 spin_lock(&mc
.lock
);
4853 mc
.flags
= move_flags
;
4854 spin_unlock(&mc
.lock
);
4855 /* We set mc.moving_task later */
4857 ret
= mem_cgroup_precharge_mc(mm
);
4859 mem_cgroup_clear_mc();
4866 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4869 mem_cgroup_clear_mc();
4872 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4873 unsigned long addr
, unsigned long end
,
4874 struct mm_walk
*walk
)
4877 struct vm_area_struct
*vma
= walk
->vma
;
4880 enum mc_target_type target_type
;
4881 union mc_target target
;
4884 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4886 if (mc
.precharge
< HPAGE_PMD_NR
) {
4890 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
4891 if (target_type
== MC_TARGET_PAGE
) {
4893 if (!isolate_lru_page(page
)) {
4894 if (!mem_cgroup_move_account(page
, true,
4896 mc
.precharge
-= HPAGE_PMD_NR
;
4897 mc
.moved_charge
+= HPAGE_PMD_NR
;
4899 putback_lru_page(page
);
4907 if (pmd_trans_unstable(pmd
))
4910 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4911 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4912 pte_t ptent
= *(pte
++);
4918 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
4919 case MC_TARGET_PAGE
:
4922 * We can have a part of the split pmd here. Moving it
4923 * can be done but it would be too convoluted so simply
4924 * ignore such a partial THP and keep it in original
4925 * memcg. There should be somebody mapping the head.
4927 if (PageTransCompound(page
))
4929 if (isolate_lru_page(page
))
4931 if (!mem_cgroup_move_account(page
, false,
4934 /* we uncharge from mc.from later. */
4937 putback_lru_page(page
);
4938 put
: /* get_mctgt_type() gets the page */
4941 case MC_TARGET_SWAP
:
4943 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
4945 /* we fixup refcnts and charges later. */
4953 pte_unmap_unlock(pte
- 1, ptl
);
4958 * We have consumed all precharges we got in can_attach().
4959 * We try charge one by one, but don't do any additional
4960 * charges to mc.to if we have failed in charge once in attach()
4963 ret
= mem_cgroup_do_precharge(1);
4971 static void mem_cgroup_move_charge(void)
4973 struct mm_walk mem_cgroup_move_charge_walk
= {
4974 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4978 lru_add_drain_all();
4980 * Signal lock_page_memcg() to take the memcg's move_lock
4981 * while we're moving its pages to another memcg. Then wait
4982 * for already started RCU-only updates to finish.
4984 atomic_inc(&mc
.from
->moving_account
);
4987 if (unlikely(!down_read_trylock(&mc
.mm
->mmap_sem
))) {
4989 * Someone who are holding the mmap_sem might be waiting in
4990 * waitq. So we cancel all extra charges, wake up all waiters,
4991 * and retry. Because we cancel precharges, we might not be able
4992 * to move enough charges, but moving charge is a best-effort
4993 * feature anyway, so it wouldn't be a big problem.
4995 __mem_cgroup_clear_mc();
5000 * When we have consumed all precharges and failed in doing
5001 * additional charge, the page walk just aborts.
5003 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk
);
5004 up_read(&mc
.mm
->mmap_sem
);
5005 atomic_dec(&mc
.from
->moving_account
);
5008 static void mem_cgroup_move_task(void)
5011 mem_cgroup_move_charge();
5012 mem_cgroup_clear_mc();
5015 #else /* !CONFIG_MMU */
5016 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5020 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5023 static void mem_cgroup_move_task(void)
5029 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5030 * to verify whether we're attached to the default hierarchy on each mount
5033 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5036 * use_hierarchy is forced on the default hierarchy. cgroup core
5037 * guarantees that @root doesn't have any children, so turning it
5038 * on for the root memcg is enough.
5040 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5041 root_mem_cgroup
->use_hierarchy
= true;
5043 root_mem_cgroup
->use_hierarchy
= false;
5046 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5049 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5051 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
5054 static int memory_low_show(struct seq_file
*m
, void *v
)
5056 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5057 unsigned long low
= READ_ONCE(memcg
->low
);
5059 if (low
== PAGE_COUNTER_MAX
)
5060 seq_puts(m
, "max\n");
5062 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
5067 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5068 char *buf
, size_t nbytes
, loff_t off
)
5070 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5074 buf
= strstrip(buf
);
5075 err
= page_counter_memparse(buf
, "max", &low
);
5084 static int memory_high_show(struct seq_file
*m
, void *v
)
5086 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5087 unsigned long high
= READ_ONCE(memcg
->high
);
5089 if (high
== PAGE_COUNTER_MAX
)
5090 seq_puts(m
, "max\n");
5092 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
5097 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5098 char *buf
, size_t nbytes
, loff_t off
)
5100 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5101 unsigned long nr_pages
;
5105 buf
= strstrip(buf
);
5106 err
= page_counter_memparse(buf
, "max", &high
);
5112 nr_pages
= page_counter_read(&memcg
->memory
);
5113 if (nr_pages
> high
)
5114 try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
5117 memcg_wb_domain_size_changed(memcg
);
5121 static int memory_max_show(struct seq_file
*m
, void *v
)
5123 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5124 unsigned long max
= READ_ONCE(memcg
->memory
.limit
);
5126 if (max
== PAGE_COUNTER_MAX
)
5127 seq_puts(m
, "max\n");
5129 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5134 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5135 char *buf
, size_t nbytes
, loff_t off
)
5137 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5138 unsigned int nr_reclaims
= MEM_CGROUP_RECLAIM_RETRIES
;
5139 bool drained
= false;
5143 buf
= strstrip(buf
);
5144 err
= page_counter_memparse(buf
, "max", &max
);
5148 xchg(&memcg
->memory
.limit
, max
);
5151 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
5153 if (nr_pages
<= max
)
5156 if (signal_pending(current
)) {
5162 drain_all_stock(memcg
);
5168 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
5174 mem_cgroup_events(memcg
, MEMCG_OOM
, 1);
5175 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
5179 memcg_wb_domain_size_changed(memcg
);
5183 static int memory_events_show(struct seq_file
*m
, void *v
)
5185 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5187 seq_printf(m
, "low %lu\n", mem_cgroup_read_events(memcg
, MEMCG_LOW
));
5188 seq_printf(m
, "high %lu\n", mem_cgroup_read_events(memcg
, MEMCG_HIGH
));
5189 seq_printf(m
, "max %lu\n", mem_cgroup_read_events(memcg
, MEMCG_MAX
));
5190 seq_printf(m
, "oom %lu\n", mem_cgroup_read_events(memcg
, MEMCG_OOM
));
5195 static int memory_stat_show(struct seq_file
*m
, void *v
)
5197 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5198 unsigned long stat
[MEMCG_NR_STAT
];
5199 unsigned long events
[MEMCG_NR_EVENTS
];
5203 * Provide statistics on the state of the memory subsystem as
5204 * well as cumulative event counters that show past behavior.
5206 * This list is ordered following a combination of these gradients:
5207 * 1) generic big picture -> specifics and details
5208 * 2) reflecting userspace activity -> reflecting kernel heuristics
5210 * Current memory state:
5213 tree_stat(memcg
, stat
);
5214 tree_events(memcg
, events
);
5216 seq_printf(m
, "anon %llu\n",
5217 (u64
)stat
[MEM_CGROUP_STAT_RSS
] * PAGE_SIZE
);
5218 seq_printf(m
, "file %llu\n",
5219 (u64
)stat
[MEM_CGROUP_STAT_CACHE
] * PAGE_SIZE
);
5220 seq_printf(m
, "kernel_stack %llu\n",
5221 (u64
)stat
[MEMCG_KERNEL_STACK_KB
] * 1024);
5222 seq_printf(m
, "slab %llu\n",
5223 (u64
)(stat
[MEMCG_SLAB_RECLAIMABLE
] +
5224 stat
[MEMCG_SLAB_UNRECLAIMABLE
]) * PAGE_SIZE
);
5225 seq_printf(m
, "sock %llu\n",
5226 (u64
)stat
[MEMCG_SOCK
] * PAGE_SIZE
);
5228 seq_printf(m
, "file_mapped %llu\n",
5229 (u64
)stat
[MEM_CGROUP_STAT_FILE_MAPPED
] * PAGE_SIZE
);
5230 seq_printf(m
, "file_dirty %llu\n",
5231 (u64
)stat
[MEM_CGROUP_STAT_DIRTY
] * PAGE_SIZE
);
5232 seq_printf(m
, "file_writeback %llu\n",
5233 (u64
)stat
[MEM_CGROUP_STAT_WRITEBACK
] * PAGE_SIZE
);
5235 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5236 struct mem_cgroup
*mi
;
5237 unsigned long val
= 0;
5239 for_each_mem_cgroup_tree(mi
, memcg
)
5240 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
));
5241 seq_printf(m
, "%s %llu\n",
5242 mem_cgroup_lru_names
[i
], (u64
)val
* PAGE_SIZE
);
5245 seq_printf(m
, "slab_reclaimable %llu\n",
5246 (u64
)stat
[MEMCG_SLAB_RECLAIMABLE
] * PAGE_SIZE
);
5247 seq_printf(m
, "slab_unreclaimable %llu\n",
5248 (u64
)stat
[MEMCG_SLAB_UNRECLAIMABLE
] * PAGE_SIZE
);
5250 /* Accumulated memory events */
5252 seq_printf(m
, "pgfault %lu\n",
5253 events
[MEM_CGROUP_EVENTS_PGFAULT
]);
5254 seq_printf(m
, "pgmajfault %lu\n",
5255 events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
5260 static struct cftype memory_files
[] = {
5263 .flags
= CFTYPE_NOT_ON_ROOT
,
5264 .read_u64
= memory_current_read
,
5268 .flags
= CFTYPE_NOT_ON_ROOT
,
5269 .seq_show
= memory_low_show
,
5270 .write
= memory_low_write
,
5274 .flags
= CFTYPE_NOT_ON_ROOT
,
5275 .seq_show
= memory_high_show
,
5276 .write
= memory_high_write
,
5280 .flags
= CFTYPE_NOT_ON_ROOT
,
5281 .seq_show
= memory_max_show
,
5282 .write
= memory_max_write
,
5286 .flags
= CFTYPE_NOT_ON_ROOT
,
5287 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
5288 .seq_show
= memory_events_show
,
5292 .flags
= CFTYPE_NOT_ON_ROOT
,
5293 .seq_show
= memory_stat_show
,
5298 struct cgroup_subsys memory_cgrp_subsys
= {
5299 .css_alloc
= mem_cgroup_css_alloc
,
5300 .css_online
= mem_cgroup_css_online
,
5301 .css_offline
= mem_cgroup_css_offline
,
5302 .css_released
= mem_cgroup_css_released
,
5303 .css_free
= mem_cgroup_css_free
,
5304 .css_reset
= mem_cgroup_css_reset
,
5305 .can_attach
= mem_cgroup_can_attach
,
5306 .cancel_attach
= mem_cgroup_cancel_attach
,
5307 .post_attach
= mem_cgroup_move_task
,
5308 .bind
= mem_cgroup_bind
,
5309 .dfl_cftypes
= memory_files
,
5310 .legacy_cftypes
= mem_cgroup_legacy_files
,
5315 * mem_cgroup_low - check if memory consumption is below the normal range
5316 * @root: the highest ancestor to consider
5317 * @memcg: the memory cgroup to check
5319 * Returns %true if memory consumption of @memcg, and that of all
5320 * configurable ancestors up to @root, is below the normal range.
5322 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5324 if (mem_cgroup_disabled())
5328 * The toplevel group doesn't have a configurable range, so
5329 * it's never low when looked at directly, and it is not
5330 * considered an ancestor when assessing the hierarchy.
5333 if (memcg
== root_mem_cgroup
)
5336 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5339 while (memcg
!= root
) {
5340 memcg
= parent_mem_cgroup(memcg
);
5342 if (memcg
== root_mem_cgroup
)
5345 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5352 * mem_cgroup_try_charge - try charging a page
5353 * @page: page to charge
5354 * @mm: mm context of the victim
5355 * @gfp_mask: reclaim mode
5356 * @memcgp: charged memcg return
5357 * @compound: charge the page as compound or small page
5359 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5360 * pages according to @gfp_mask if necessary.
5362 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5363 * Otherwise, an error code is returned.
5365 * After page->mapping has been set up, the caller must finalize the
5366 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5367 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5369 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5370 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
5373 struct mem_cgroup
*memcg
= NULL
;
5374 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5377 if (mem_cgroup_disabled())
5380 if (PageSwapCache(page
)) {
5382 * Every swap fault against a single page tries to charge the
5383 * page, bail as early as possible. shmem_unuse() encounters
5384 * already charged pages, too. The USED bit is protected by
5385 * the page lock, which serializes swap cache removal, which
5386 * in turn serializes uncharging.
5388 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5389 if (page
->mem_cgroup
)
5392 if (do_swap_account
) {
5393 swp_entry_t ent
= { .val
= page_private(page
), };
5394 unsigned short id
= lookup_swap_cgroup_id(ent
);
5397 memcg
= mem_cgroup_from_id(id
);
5398 if (memcg
&& !css_tryget_online(&memcg
->css
))
5405 memcg
= get_mem_cgroup_from_mm(mm
);
5407 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5409 css_put(&memcg
->css
);
5416 * mem_cgroup_commit_charge - commit a page charge
5417 * @page: page to charge
5418 * @memcg: memcg to charge the page to
5419 * @lrucare: page might be on LRU already
5420 * @compound: charge the page as compound or small page
5422 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5423 * after page->mapping has been set up. This must happen atomically
5424 * as part of the page instantiation, i.e. under the page table lock
5425 * for anonymous pages, under the page lock for page and swap cache.
5427 * In addition, the page must not be on the LRU during the commit, to
5428 * prevent racing with task migration. If it might be, use @lrucare.
5430 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5432 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5433 bool lrucare
, bool compound
)
5435 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5437 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5438 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5440 if (mem_cgroup_disabled())
5443 * Swap faults will attempt to charge the same page multiple
5444 * times. But reuse_swap_page() might have removed the page
5445 * from swapcache already, so we can't check PageSwapCache().
5450 commit_charge(page
, memcg
, lrucare
);
5452 local_irq_disable();
5453 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
5454 memcg_check_events(memcg
, page
);
5457 if (do_memsw_account() && PageSwapCache(page
)) {
5458 swp_entry_t entry
= { .val
= page_private(page
) };
5460 * The swap entry might not get freed for a long time,
5461 * let's not wait for it. The page already received a
5462 * memory+swap charge, drop the swap entry duplicate.
5464 mem_cgroup_uncharge_swap(entry
);
5469 * mem_cgroup_cancel_charge - cancel a page charge
5470 * @page: page to charge
5471 * @memcg: memcg to charge the page to
5472 * @compound: charge the page as compound or small page
5474 * Cancel a charge transaction started by mem_cgroup_try_charge().
5476 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5479 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5481 if (mem_cgroup_disabled())
5484 * Swap faults will attempt to charge the same page multiple
5485 * times. But reuse_swap_page() might have removed the page
5486 * from swapcache already, so we can't check PageSwapCache().
5491 cancel_charge(memcg
, nr_pages
);
5494 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5495 unsigned long nr_anon
, unsigned long nr_file
,
5496 unsigned long nr_huge
, unsigned long nr_kmem
,
5497 struct page
*dummy_page
)
5499 unsigned long nr_pages
= nr_anon
+ nr_file
+ nr_kmem
;
5500 unsigned long flags
;
5502 if (!mem_cgroup_is_root(memcg
)) {
5503 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5504 if (do_memsw_account())
5505 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5506 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && nr_kmem
)
5507 page_counter_uncharge(&memcg
->kmem
, nr_kmem
);
5508 memcg_oom_recover(memcg
);
5511 local_irq_save(flags
);
5512 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5513 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5514 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5515 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5516 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5517 memcg_check_events(memcg
, dummy_page
);
5518 local_irq_restore(flags
);
5520 if (!mem_cgroup_is_root(memcg
))
5521 css_put_many(&memcg
->css
, nr_pages
);
5524 static void uncharge_list(struct list_head
*page_list
)
5526 struct mem_cgroup
*memcg
= NULL
;
5527 unsigned long nr_anon
= 0;
5528 unsigned long nr_file
= 0;
5529 unsigned long nr_huge
= 0;
5530 unsigned long nr_kmem
= 0;
5531 unsigned long pgpgout
= 0;
5532 struct list_head
*next
;
5536 * Note that the list can be a single page->lru; hence the
5537 * do-while loop instead of a simple list_for_each_entry().
5539 next
= page_list
->next
;
5541 page
= list_entry(next
, struct page
, lru
);
5542 next
= page
->lru
.next
;
5544 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5545 VM_BUG_ON_PAGE(page_count(page
), page
);
5547 if (!page
->mem_cgroup
)
5551 * Nobody should be changing or seriously looking at
5552 * page->mem_cgroup at this point, we have fully
5553 * exclusive access to the page.
5556 if (memcg
!= page
->mem_cgroup
) {
5558 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5559 nr_huge
, nr_kmem
, page
);
5560 pgpgout
= nr_anon
= nr_file
=
5561 nr_huge
= nr_kmem
= 0;
5563 memcg
= page
->mem_cgroup
;
5566 if (!PageKmemcg(page
)) {
5567 unsigned int nr_pages
= 1;
5569 if (PageTransHuge(page
)) {
5570 nr_pages
<<= compound_order(page
);
5571 nr_huge
+= nr_pages
;
5574 nr_anon
+= nr_pages
;
5576 nr_file
+= nr_pages
;
5579 nr_kmem
+= 1 << compound_order(page
);
5580 __ClearPageKmemcg(page
);
5583 page
->mem_cgroup
= NULL
;
5584 } while (next
!= page_list
);
5587 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5588 nr_huge
, nr_kmem
, page
);
5592 * mem_cgroup_uncharge - uncharge a page
5593 * @page: page to uncharge
5595 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5596 * mem_cgroup_commit_charge().
5598 void mem_cgroup_uncharge(struct page
*page
)
5600 if (mem_cgroup_disabled())
5603 /* Don't touch page->lru of any random page, pre-check: */
5604 if (!page
->mem_cgroup
)
5607 INIT_LIST_HEAD(&page
->lru
);
5608 uncharge_list(&page
->lru
);
5612 * mem_cgroup_uncharge_list - uncharge a list of page
5613 * @page_list: list of pages to uncharge
5615 * Uncharge a list of pages previously charged with
5616 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5618 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5620 if (mem_cgroup_disabled())
5623 if (!list_empty(page_list
))
5624 uncharge_list(page_list
);
5628 * mem_cgroup_migrate - charge a page's replacement
5629 * @oldpage: currently circulating page
5630 * @newpage: replacement page
5632 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5633 * be uncharged upon free.
5635 * Both pages must be locked, @newpage->mapping must be set up.
5637 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
5639 struct mem_cgroup
*memcg
;
5640 unsigned int nr_pages
;
5642 unsigned long flags
;
5644 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5645 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5646 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5647 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5650 if (mem_cgroup_disabled())
5653 /* Page cache replacement: new page already charged? */
5654 if (newpage
->mem_cgroup
)
5657 /* Swapcache readahead pages can get replaced before being charged */
5658 memcg
= oldpage
->mem_cgroup
;
5662 /* Force-charge the new page. The old one will be freed soon */
5663 compound
= PageTransHuge(newpage
);
5664 nr_pages
= compound
? hpage_nr_pages(newpage
) : 1;
5666 page_counter_charge(&memcg
->memory
, nr_pages
);
5667 if (do_memsw_account())
5668 page_counter_charge(&memcg
->memsw
, nr_pages
);
5669 css_get_many(&memcg
->css
, nr_pages
);
5671 commit_charge(newpage
, memcg
, false);
5673 local_irq_save(flags
);
5674 mem_cgroup_charge_statistics(memcg
, newpage
, compound
, nr_pages
);
5675 memcg_check_events(memcg
, newpage
);
5676 local_irq_restore(flags
);
5679 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
5680 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
5682 void sock_update_memcg(struct sock
*sk
)
5684 struct mem_cgroup
*memcg
;
5686 /* Socket cloning can throw us here with sk_cgrp already
5687 * filled. It won't however, necessarily happen from
5688 * process context. So the test for root memcg given
5689 * the current task's memcg won't help us in this case.
5691 * Respecting the original socket's memcg is a better
5692 * decision in this case.
5695 BUG_ON(mem_cgroup_is_root(sk
->sk_memcg
));
5696 css_get(&sk
->sk_memcg
->css
);
5701 memcg
= mem_cgroup_from_task(current
);
5702 if (memcg
== root_mem_cgroup
)
5704 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
5706 if (css_tryget_online(&memcg
->css
))
5707 sk
->sk_memcg
= memcg
;
5711 EXPORT_SYMBOL(sock_update_memcg
);
5713 void sock_release_memcg(struct sock
*sk
)
5715 WARN_ON(!sk
->sk_memcg
);
5716 css_put(&sk
->sk_memcg
->css
);
5720 * mem_cgroup_charge_skmem - charge socket memory
5721 * @memcg: memcg to charge
5722 * @nr_pages: number of pages to charge
5724 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5725 * @memcg's configured limit, %false if the charge had to be forced.
5727 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5729 gfp_t gfp_mask
= GFP_KERNEL
;
5731 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5732 struct page_counter
*fail
;
5734 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
5735 memcg
->tcpmem_pressure
= 0;
5738 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
5739 memcg
->tcpmem_pressure
= 1;
5743 /* Don't block in the packet receive path */
5745 gfp_mask
= GFP_NOWAIT
;
5747 this_cpu_add(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5749 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
5752 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
5757 * mem_cgroup_uncharge_skmem - uncharge socket memory
5758 * @memcg - memcg to uncharge
5759 * @nr_pages - number of pages to uncharge
5761 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5763 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5764 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
5768 this_cpu_sub(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5770 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5771 css_put_many(&memcg
->css
, nr_pages
);
5774 static int __init
cgroup_memory(char *s
)
5778 while ((token
= strsep(&s
, ",")) != NULL
) {
5781 if (!strcmp(token
, "nosocket"))
5782 cgroup_memory_nosocket
= true;
5783 if (!strcmp(token
, "nokmem"))
5784 cgroup_memory_nokmem
= true;
5788 __setup("cgroup.memory=", cgroup_memory
);
5791 * subsys_initcall() for memory controller.
5793 * Some parts like hotcpu_notifier() have to be initialized from this context
5794 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5795 * everything that doesn't depend on a specific mem_cgroup structure should
5796 * be initialized from here.
5798 static int __init
mem_cgroup_init(void)
5802 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5804 for_each_possible_cpu(cpu
)
5805 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5808 for_each_node(node
) {
5809 struct mem_cgroup_tree_per_node
*rtpn
;
5811 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5812 node_online(node
) ? node
: NUMA_NO_NODE
);
5814 rtpn
->rb_root
= RB_ROOT
;
5815 spin_lock_init(&rtpn
->lock
);
5816 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5821 subsys_initcall(mem_cgroup_init
);
5823 #ifdef CONFIG_MEMCG_SWAP
5825 * mem_cgroup_swapout - transfer a memsw charge to swap
5826 * @page: page whose memsw charge to transfer
5827 * @entry: swap entry to move the charge to
5829 * Transfer the memsw charge of @page to @entry.
5831 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5833 struct mem_cgroup
*memcg
, *swap_memcg
;
5834 unsigned short oldid
;
5836 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5837 VM_BUG_ON_PAGE(page_count(page
), page
);
5839 if (!do_memsw_account())
5842 memcg
= page
->mem_cgroup
;
5844 /* Readahead page, never charged */
5849 * In case the memcg owning these pages has been offlined and doesn't
5850 * have an ID allocated to it anymore, charge the closest online
5851 * ancestor for the swap instead and transfer the memory+swap charge.
5853 swap_memcg
= mem_cgroup_id_get_online(memcg
);
5854 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
));
5855 VM_BUG_ON_PAGE(oldid
, page
);
5856 mem_cgroup_swap_statistics(swap_memcg
, true);
5858 page
->mem_cgroup
= NULL
;
5860 if (!mem_cgroup_is_root(memcg
))
5861 page_counter_uncharge(&memcg
->memory
, 1);
5863 if (memcg
!= swap_memcg
) {
5864 if (!mem_cgroup_is_root(swap_memcg
))
5865 page_counter_charge(&swap_memcg
->memsw
, 1);
5866 page_counter_uncharge(&memcg
->memsw
, 1);
5870 * Interrupts should be disabled here because the caller holds the
5871 * mapping->tree_lock lock which is taken with interrupts-off. It is
5872 * important here to have the interrupts disabled because it is the
5873 * only synchronisation we have for udpating the per-CPU variables.
5875 VM_BUG_ON(!irqs_disabled());
5876 mem_cgroup_charge_statistics(memcg
, page
, false, -1);
5877 memcg_check_events(memcg
, page
);
5879 if (!mem_cgroup_is_root(memcg
))
5880 css_put(&memcg
->css
);
5884 * mem_cgroup_try_charge_swap - try charging a swap entry
5885 * @page: page being added to swap
5886 * @entry: swap entry to charge
5888 * Try to charge @entry to the memcg that @page belongs to.
5890 * Returns 0 on success, -ENOMEM on failure.
5892 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
5894 struct mem_cgroup
*memcg
;
5895 struct page_counter
*counter
;
5896 unsigned short oldid
;
5898 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
5901 memcg
= page
->mem_cgroup
;
5903 /* Readahead page, never charged */
5907 memcg
= mem_cgroup_id_get_online(memcg
);
5909 if (!mem_cgroup_is_root(memcg
) &&
5910 !page_counter_try_charge(&memcg
->swap
, 1, &counter
)) {
5911 mem_cgroup_id_put(memcg
);
5915 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5916 VM_BUG_ON_PAGE(oldid
, page
);
5917 mem_cgroup_swap_statistics(memcg
, true);
5923 * mem_cgroup_uncharge_swap - uncharge a swap entry
5924 * @entry: swap entry to uncharge
5926 * Drop the swap charge associated with @entry.
5928 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5930 struct mem_cgroup
*memcg
;
5933 if (!do_swap_account
)
5936 id
= swap_cgroup_record(entry
, 0);
5938 memcg
= mem_cgroup_from_id(id
);
5940 if (!mem_cgroup_is_root(memcg
)) {
5941 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5942 page_counter_uncharge(&memcg
->swap
, 1);
5944 page_counter_uncharge(&memcg
->memsw
, 1);
5946 mem_cgroup_swap_statistics(memcg
, false);
5947 mem_cgroup_id_put(memcg
);
5952 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
5954 long nr_swap_pages
= get_nr_swap_pages();
5956 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5957 return nr_swap_pages
;
5958 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5959 nr_swap_pages
= min_t(long, nr_swap_pages
,
5960 READ_ONCE(memcg
->swap
.limit
) -
5961 page_counter_read(&memcg
->swap
));
5962 return nr_swap_pages
;
5965 bool mem_cgroup_swap_full(struct page
*page
)
5967 struct mem_cgroup
*memcg
;
5969 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5973 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5976 memcg
= page
->mem_cgroup
;
5980 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5981 if (page_counter_read(&memcg
->swap
) * 2 >= memcg
->swap
.limit
)
5987 /* for remember boot option*/
5988 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5989 static int really_do_swap_account __initdata
= 1;
5991 static int really_do_swap_account __initdata
;
5994 static int __init
enable_swap_account(char *s
)
5996 if (!strcmp(s
, "1"))
5997 really_do_swap_account
= 1;
5998 else if (!strcmp(s
, "0"))
5999 really_do_swap_account
= 0;
6002 __setup("swapaccount=", enable_swap_account
);
6004 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
6007 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6009 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
6012 static int swap_max_show(struct seq_file
*m
, void *v
)
6014 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
6015 unsigned long max
= READ_ONCE(memcg
->swap
.limit
);
6017 if (max
== PAGE_COUNTER_MAX
)
6018 seq_puts(m
, "max\n");
6020 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
6025 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
6026 char *buf
, size_t nbytes
, loff_t off
)
6028 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6032 buf
= strstrip(buf
);
6033 err
= page_counter_memparse(buf
, "max", &max
);
6037 mutex_lock(&memcg_limit_mutex
);
6038 err
= page_counter_limit(&memcg
->swap
, max
);
6039 mutex_unlock(&memcg_limit_mutex
);
6046 static struct cftype swap_files
[] = {
6048 .name
= "swap.current",
6049 .flags
= CFTYPE_NOT_ON_ROOT
,
6050 .read_u64
= swap_current_read
,
6054 .flags
= CFTYPE_NOT_ON_ROOT
,
6055 .seq_show
= swap_max_show
,
6056 .write
= swap_max_write
,
6061 static struct cftype memsw_cgroup_files
[] = {
6063 .name
= "memsw.usage_in_bytes",
6064 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6065 .read_u64
= mem_cgroup_read_u64
,
6068 .name
= "memsw.max_usage_in_bytes",
6069 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6070 .write
= mem_cgroup_reset
,
6071 .read_u64
= mem_cgroup_read_u64
,
6074 .name
= "memsw.limit_in_bytes",
6075 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6076 .write
= mem_cgroup_write
,
6077 .read_u64
= mem_cgroup_read_u64
,
6080 .name
= "memsw.failcnt",
6081 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6082 .write
= mem_cgroup_reset
,
6083 .read_u64
= mem_cgroup_read_u64
,
6085 { }, /* terminate */
6088 static int __init
mem_cgroup_swap_init(void)
6090 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6091 do_swap_account
= 1;
6092 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
6094 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
6095 memsw_cgroup_files
));
6099 subsys_initcall(mem_cgroup_swap_init
);
6101 #endif /* CONFIG_MEMCG_SWAP */