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_scan_tasks - iterate over tasks of a memory cgroup hierarchy
925 * @memcg: hierarchy root
926 * @fn: function to call for each task
927 * @arg: argument passed to @fn
929 * This function iterates over tasks attached to @memcg or to any of its
930 * descendants and calls @fn for each task. If @fn returns a non-zero
931 * value, the function breaks the iteration loop and returns the value.
932 * Otherwise, it will iterate over all tasks and return 0.
934 * This function must not be called for the root memory cgroup.
936 int mem_cgroup_scan_tasks(struct mem_cgroup
*memcg
,
937 int (*fn
)(struct task_struct
*, void *), void *arg
)
939 struct mem_cgroup
*iter
;
942 BUG_ON(memcg
== root_mem_cgroup
);
944 for_each_mem_cgroup_tree(iter
, memcg
) {
945 struct css_task_iter it
;
946 struct task_struct
*task
;
948 css_task_iter_start(&iter
->css
, &it
);
949 while (!ret
&& (task
= css_task_iter_next(&it
)))
951 css_task_iter_end(&it
);
953 mem_cgroup_iter_break(memcg
, iter
);
961 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
963 * @zone: zone of the page
965 * This function is only safe when following the LRU page isolation
966 * and putback protocol: the LRU lock must be held, and the page must
967 * either be PageLRU() or the caller must have isolated/allocated it.
969 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct pglist_data
*pgdat
)
971 struct mem_cgroup_per_node
*mz
;
972 struct mem_cgroup
*memcg
;
973 struct lruvec
*lruvec
;
975 if (mem_cgroup_disabled()) {
976 lruvec
= &pgdat
->lruvec
;
980 memcg
= page
->mem_cgroup
;
982 * Swapcache readahead pages are added to the LRU - and
983 * possibly migrated - before they are charged.
986 memcg
= root_mem_cgroup
;
988 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
989 lruvec
= &mz
->lruvec
;
992 * Since a node can be onlined after the mem_cgroup was created,
993 * we have to be prepared to initialize lruvec->zone here;
994 * and if offlined then reonlined, we need to reinitialize it.
996 if (unlikely(lruvec
->pgdat
!= pgdat
))
997 lruvec
->pgdat
= pgdat
;
1002 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1003 * @lruvec: mem_cgroup per zone lru vector
1004 * @lru: index of lru list the page is sitting on
1005 * @nr_pages: positive when adding or negative when removing
1007 * This function must be called under lru_lock, just before a page is added
1008 * to or just after a page is removed from an lru list (that ordering being
1009 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1011 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1014 struct mem_cgroup_per_node
*mz
;
1015 unsigned long *lru_size
;
1019 if (mem_cgroup_disabled())
1022 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
1023 lru_size
= mz
->lru_size
+ lru
;
1024 empty
= list_empty(lruvec
->lists
+ lru
);
1027 *lru_size
+= nr_pages
;
1030 if (WARN_ONCE(size
< 0 || empty
!= !size
,
1031 "%s(%p, %d, %d): lru_size %ld but %sempty\n",
1032 __func__
, lruvec
, lru
, nr_pages
, size
, empty
? "" : "not ")) {
1038 *lru_size
+= nr_pages
;
1041 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1043 struct mem_cgroup
*task_memcg
;
1044 struct task_struct
*p
;
1047 p
= find_lock_task_mm(task
);
1049 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1053 * All threads may have already detached their mm's, but the oom
1054 * killer still needs to detect if they have already been oom
1055 * killed to prevent needlessly killing additional tasks.
1058 task_memcg
= mem_cgroup_from_task(task
);
1059 css_get(&task_memcg
->css
);
1062 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1063 css_put(&task_memcg
->css
);
1068 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1069 * @memcg: the memory cgroup
1071 * Returns the maximum amount of memory @mem can be charged with, in
1074 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1076 unsigned long margin
= 0;
1077 unsigned long count
;
1078 unsigned long limit
;
1080 count
= page_counter_read(&memcg
->memory
);
1081 limit
= READ_ONCE(memcg
->memory
.limit
);
1083 margin
= limit
- count
;
1085 if (do_memsw_account()) {
1086 count
= page_counter_read(&memcg
->memsw
);
1087 limit
= READ_ONCE(memcg
->memsw
.limit
);
1089 margin
= min(margin
, limit
- count
);
1098 * A routine for checking "mem" is under move_account() or not.
1100 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1101 * moving cgroups. This is for waiting at high-memory pressure
1104 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1106 struct mem_cgroup
*from
;
1107 struct mem_cgroup
*to
;
1110 * Unlike task_move routines, we access mc.to, mc.from not under
1111 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1113 spin_lock(&mc
.lock
);
1119 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1120 mem_cgroup_is_descendant(to
, memcg
);
1122 spin_unlock(&mc
.lock
);
1126 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1128 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1129 if (mem_cgroup_under_move(memcg
)) {
1131 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1132 /* moving charge context might have finished. */
1135 finish_wait(&mc
.waitq
, &wait
);
1142 #define K(x) ((x) << (PAGE_SHIFT-10))
1144 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1145 * @memcg: The memory cgroup that went over limit
1146 * @p: Task that is going to be killed
1148 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1151 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1153 struct mem_cgroup
*iter
;
1159 pr_info("Task in ");
1160 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1161 pr_cont(" killed as a result of limit of ");
1163 pr_info("Memory limit reached of cgroup ");
1166 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1171 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1172 K((u64
)page_counter_read(&memcg
->memory
)),
1173 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1174 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1175 K((u64
)page_counter_read(&memcg
->memsw
)),
1176 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1177 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1178 K((u64
)page_counter_read(&memcg
->kmem
)),
1179 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1181 for_each_mem_cgroup_tree(iter
, memcg
) {
1182 pr_info("Memory cgroup stats for ");
1183 pr_cont_cgroup_path(iter
->css
.cgroup
);
1186 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1187 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1189 pr_cont(" %s:%luKB", mem_cgroup_stat_names
[i
],
1190 K(mem_cgroup_read_stat(iter
, i
)));
1193 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1194 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1195 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1202 * This function returns the number of memcg under hierarchy tree. Returns
1203 * 1(self count) if no children.
1205 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1208 struct mem_cgroup
*iter
;
1210 for_each_mem_cgroup_tree(iter
, memcg
)
1216 * Return the memory (and swap, if configured) limit for a memcg.
1218 unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1220 unsigned long limit
;
1222 limit
= memcg
->memory
.limit
;
1223 if (mem_cgroup_swappiness(memcg
)) {
1224 unsigned long memsw_limit
;
1225 unsigned long swap_limit
;
1227 memsw_limit
= memcg
->memsw
.limit
;
1228 swap_limit
= memcg
->swap
.limit
;
1229 swap_limit
= min(swap_limit
, (unsigned long)total_swap_pages
);
1230 limit
= min(limit
+ swap_limit
, memsw_limit
);
1235 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1238 struct oom_control oc
= {
1242 .gfp_mask
= gfp_mask
,
1247 mutex_lock(&oom_lock
);
1248 ret
= out_of_memory(&oc
);
1249 mutex_unlock(&oom_lock
);
1253 #if MAX_NUMNODES > 1
1256 * test_mem_cgroup_node_reclaimable
1257 * @memcg: the target memcg
1258 * @nid: the node ID to be checked.
1259 * @noswap : specify true here if the user wants flle only information.
1261 * This function returns whether the specified memcg contains any
1262 * reclaimable pages on a node. Returns true if there are any reclaimable
1263 * pages in the node.
1265 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1266 int nid
, bool noswap
)
1268 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1270 if (noswap
|| !total_swap_pages
)
1272 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1279 * Always updating the nodemask is not very good - even if we have an empty
1280 * list or the wrong list here, we can start from some node and traverse all
1281 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1284 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1288 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1289 * pagein/pageout changes since the last update.
1291 if (!atomic_read(&memcg
->numainfo_events
))
1293 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1296 /* make a nodemask where this memcg uses memory from */
1297 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1299 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1301 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1302 node_clear(nid
, memcg
->scan_nodes
);
1305 atomic_set(&memcg
->numainfo_events
, 0);
1306 atomic_set(&memcg
->numainfo_updating
, 0);
1310 * Selecting a node where we start reclaim from. Because what we need is just
1311 * reducing usage counter, start from anywhere is O,K. Considering
1312 * memory reclaim from current node, there are pros. and cons.
1314 * Freeing memory from current node means freeing memory from a node which
1315 * we'll use or we've used. So, it may make LRU bad. And if several threads
1316 * hit limits, it will see a contention on a node. But freeing from remote
1317 * node means more costs for memory reclaim because of memory latency.
1319 * Now, we use round-robin. Better algorithm is welcomed.
1321 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1325 mem_cgroup_may_update_nodemask(memcg
);
1326 node
= memcg
->last_scanned_node
;
1328 node
= next_node_in(node
, memcg
->scan_nodes
);
1330 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1331 * last time it really checked all the LRUs due to rate limiting.
1332 * Fallback to the current node in that case for simplicity.
1334 if (unlikely(node
== MAX_NUMNODES
))
1335 node
= numa_node_id();
1337 memcg
->last_scanned_node
= node
;
1341 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1347 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1350 unsigned long *total_scanned
)
1352 struct mem_cgroup
*victim
= NULL
;
1355 unsigned long excess
;
1356 unsigned long nr_scanned
;
1357 struct mem_cgroup_reclaim_cookie reclaim
= {
1362 excess
= soft_limit_excess(root_memcg
);
1365 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1370 * If we have not been able to reclaim
1371 * anything, it might because there are
1372 * no reclaimable pages under this hierarchy
1377 * We want to do more targeted reclaim.
1378 * excess >> 2 is not to excessive so as to
1379 * reclaim too much, nor too less that we keep
1380 * coming back to reclaim from this cgroup
1382 if (total
>= (excess
>> 2) ||
1383 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1388 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1389 pgdat
, &nr_scanned
);
1390 *total_scanned
+= nr_scanned
;
1391 if (!soft_limit_excess(root_memcg
))
1394 mem_cgroup_iter_break(root_memcg
, victim
);
1398 #ifdef CONFIG_LOCKDEP
1399 static struct lockdep_map memcg_oom_lock_dep_map
= {
1400 .name
= "memcg_oom_lock",
1404 static DEFINE_SPINLOCK(memcg_oom_lock
);
1407 * Check OOM-Killer is already running under our hierarchy.
1408 * If someone is running, return false.
1410 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1412 struct mem_cgroup
*iter
, *failed
= NULL
;
1414 spin_lock(&memcg_oom_lock
);
1416 for_each_mem_cgroup_tree(iter
, memcg
) {
1417 if (iter
->oom_lock
) {
1419 * this subtree of our hierarchy is already locked
1420 * so we cannot give a lock.
1423 mem_cgroup_iter_break(memcg
, iter
);
1426 iter
->oom_lock
= true;
1431 * OK, we failed to lock the whole subtree so we have
1432 * to clean up what we set up to the failing subtree
1434 for_each_mem_cgroup_tree(iter
, memcg
) {
1435 if (iter
== failed
) {
1436 mem_cgroup_iter_break(memcg
, iter
);
1439 iter
->oom_lock
= false;
1442 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1444 spin_unlock(&memcg_oom_lock
);
1449 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1451 struct mem_cgroup
*iter
;
1453 spin_lock(&memcg_oom_lock
);
1454 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1455 for_each_mem_cgroup_tree(iter
, memcg
)
1456 iter
->oom_lock
= false;
1457 spin_unlock(&memcg_oom_lock
);
1460 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1462 struct mem_cgroup
*iter
;
1464 spin_lock(&memcg_oom_lock
);
1465 for_each_mem_cgroup_tree(iter
, memcg
)
1467 spin_unlock(&memcg_oom_lock
);
1470 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1472 struct mem_cgroup
*iter
;
1475 * When a new child is created while the hierarchy is under oom,
1476 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1478 spin_lock(&memcg_oom_lock
);
1479 for_each_mem_cgroup_tree(iter
, memcg
)
1480 if (iter
->under_oom
> 0)
1482 spin_unlock(&memcg_oom_lock
);
1485 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1487 struct oom_wait_info
{
1488 struct mem_cgroup
*memcg
;
1492 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1493 unsigned mode
, int sync
, void *arg
)
1495 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1496 struct mem_cgroup
*oom_wait_memcg
;
1497 struct oom_wait_info
*oom_wait_info
;
1499 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1500 oom_wait_memcg
= oom_wait_info
->memcg
;
1502 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1503 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1505 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1508 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1511 * For the following lockless ->under_oom test, the only required
1512 * guarantee is that it must see the state asserted by an OOM when
1513 * this function is called as a result of userland actions
1514 * triggered by the notification of the OOM. This is trivially
1515 * achieved by invoking mem_cgroup_mark_under_oom() before
1516 * triggering notification.
1518 if (memcg
&& memcg
->under_oom
)
1519 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1522 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1524 if (!current
->memcg_may_oom
)
1527 * We are in the middle of the charge context here, so we
1528 * don't want to block when potentially sitting on a callstack
1529 * that holds all kinds of filesystem and mm locks.
1531 * Also, the caller may handle a failed allocation gracefully
1532 * (like optional page cache readahead) and so an OOM killer
1533 * invocation might not even be necessary.
1535 * That's why we don't do anything here except remember the
1536 * OOM context and then deal with it at the end of the page
1537 * fault when the stack is unwound, the locks are released,
1538 * and when we know whether the fault was overall successful.
1540 css_get(&memcg
->css
);
1541 current
->memcg_in_oom
= memcg
;
1542 current
->memcg_oom_gfp_mask
= mask
;
1543 current
->memcg_oom_order
= order
;
1547 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1548 * @handle: actually kill/wait or just clean up the OOM state
1550 * This has to be called at the end of a page fault if the memcg OOM
1551 * handler was enabled.
1553 * Memcg supports userspace OOM handling where failed allocations must
1554 * sleep on a waitqueue until the userspace task resolves the
1555 * situation. Sleeping directly in the charge context with all kinds
1556 * of locks held is not a good idea, instead we remember an OOM state
1557 * in the task and mem_cgroup_oom_synchronize() has to be called at
1558 * the end of the page fault to complete the OOM handling.
1560 * Returns %true if an ongoing memcg OOM situation was detected and
1561 * completed, %false otherwise.
1563 bool mem_cgroup_oom_synchronize(bool handle
)
1565 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1566 struct oom_wait_info owait
;
1569 /* OOM is global, do not handle */
1576 owait
.memcg
= memcg
;
1577 owait
.wait
.flags
= 0;
1578 owait
.wait
.func
= memcg_oom_wake_function
;
1579 owait
.wait
.private = current
;
1580 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1582 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1583 mem_cgroup_mark_under_oom(memcg
);
1585 locked
= mem_cgroup_oom_trylock(memcg
);
1588 mem_cgroup_oom_notify(memcg
);
1590 if (locked
&& !memcg
->oom_kill_disable
) {
1591 mem_cgroup_unmark_under_oom(memcg
);
1592 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1593 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1594 current
->memcg_oom_order
);
1597 mem_cgroup_unmark_under_oom(memcg
);
1598 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1602 mem_cgroup_oom_unlock(memcg
);
1604 * There is no guarantee that an OOM-lock contender
1605 * sees the wakeups triggered by the OOM kill
1606 * uncharges. Wake any sleepers explicitely.
1608 memcg_oom_recover(memcg
);
1611 current
->memcg_in_oom
= NULL
;
1612 css_put(&memcg
->css
);
1617 * lock_page_memcg - lock a page->mem_cgroup binding
1620 * This function protects unlocked LRU pages from being moved to
1621 * another cgroup and stabilizes their page->mem_cgroup binding.
1623 void lock_page_memcg(struct page
*page
)
1625 struct mem_cgroup
*memcg
;
1626 unsigned long flags
;
1629 * The RCU lock is held throughout the transaction. The fast
1630 * path can get away without acquiring the memcg->move_lock
1631 * because page moving starts with an RCU grace period.
1635 if (mem_cgroup_disabled())
1638 memcg
= page
->mem_cgroup
;
1639 if (unlikely(!memcg
))
1642 if (atomic_read(&memcg
->moving_account
) <= 0)
1645 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1646 if (memcg
!= page
->mem_cgroup
) {
1647 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1652 * When charge migration first begins, we can have locked and
1653 * unlocked page stat updates happening concurrently. Track
1654 * the task who has the lock for unlock_page_memcg().
1656 memcg
->move_lock_task
= current
;
1657 memcg
->move_lock_flags
= flags
;
1661 EXPORT_SYMBOL(lock_page_memcg
);
1664 * unlock_page_memcg - unlock a page->mem_cgroup binding
1667 void unlock_page_memcg(struct page
*page
)
1669 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
1671 if (memcg
&& memcg
->move_lock_task
== current
) {
1672 unsigned long flags
= memcg
->move_lock_flags
;
1674 memcg
->move_lock_task
= NULL
;
1675 memcg
->move_lock_flags
= 0;
1677 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1682 EXPORT_SYMBOL(unlock_page_memcg
);
1685 * size of first charge trial. "32" comes from vmscan.c's magic value.
1686 * TODO: maybe necessary to use big numbers in big irons.
1688 #define CHARGE_BATCH 32U
1689 struct memcg_stock_pcp
{
1690 struct mem_cgroup
*cached
; /* this never be root cgroup */
1691 unsigned int nr_pages
;
1692 struct work_struct work
;
1693 unsigned long flags
;
1694 #define FLUSHING_CACHED_CHARGE 0
1696 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1697 static DEFINE_MUTEX(percpu_charge_mutex
);
1700 * consume_stock: Try to consume stocked charge on this cpu.
1701 * @memcg: memcg to consume from.
1702 * @nr_pages: how many pages to charge.
1704 * The charges will only happen if @memcg matches the current cpu's memcg
1705 * stock, and at least @nr_pages are available in that stock. Failure to
1706 * service an allocation will refill the stock.
1708 * returns true if successful, false otherwise.
1710 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1712 struct memcg_stock_pcp
*stock
;
1715 if (nr_pages
> CHARGE_BATCH
)
1718 stock
= &get_cpu_var(memcg_stock
);
1719 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
1720 stock
->nr_pages
-= nr_pages
;
1723 put_cpu_var(memcg_stock
);
1728 * Returns stocks cached in percpu and reset cached information.
1730 static void drain_stock(struct memcg_stock_pcp
*stock
)
1732 struct mem_cgroup
*old
= stock
->cached
;
1734 if (stock
->nr_pages
) {
1735 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
1736 if (do_memsw_account())
1737 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
1738 css_put_many(&old
->css
, stock
->nr_pages
);
1739 stock
->nr_pages
= 0;
1741 stock
->cached
= NULL
;
1745 * This must be called under preempt disabled or must be called by
1746 * a thread which is pinned to local cpu.
1748 static void drain_local_stock(struct work_struct
*dummy
)
1750 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
1752 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
1756 * Cache charges(val) to local per_cpu area.
1757 * This will be consumed by consume_stock() function, later.
1759 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1761 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
1763 if (stock
->cached
!= memcg
) { /* reset if necessary */
1765 stock
->cached
= memcg
;
1767 stock
->nr_pages
+= nr_pages
;
1768 put_cpu_var(memcg_stock
);
1772 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1773 * of the hierarchy under it.
1775 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
1779 /* If someone's already draining, avoid adding running more workers. */
1780 if (!mutex_trylock(&percpu_charge_mutex
))
1782 /* Notify other cpus that system-wide "drain" is running */
1785 for_each_online_cpu(cpu
) {
1786 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1787 struct mem_cgroup
*memcg
;
1789 memcg
= stock
->cached
;
1790 if (!memcg
|| !stock
->nr_pages
)
1792 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
1794 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
1796 drain_local_stock(&stock
->work
);
1798 schedule_work_on(cpu
, &stock
->work
);
1803 mutex_unlock(&percpu_charge_mutex
);
1806 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
1807 unsigned long action
,
1810 int cpu
= (unsigned long)hcpu
;
1811 struct memcg_stock_pcp
*stock
;
1813 if (action
== CPU_ONLINE
)
1816 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
1819 stock
= &per_cpu(memcg_stock
, cpu
);
1824 static void reclaim_high(struct mem_cgroup
*memcg
,
1825 unsigned int nr_pages
,
1829 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
1831 mem_cgroup_events(memcg
, MEMCG_HIGH
, 1);
1832 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
1833 } while ((memcg
= parent_mem_cgroup(memcg
)));
1836 static void high_work_func(struct work_struct
*work
)
1838 struct mem_cgroup
*memcg
;
1840 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
1841 reclaim_high(memcg
, CHARGE_BATCH
, GFP_KERNEL
);
1845 * Scheduled by try_charge() to be executed from the userland return path
1846 * and reclaims memory over the high limit.
1848 void mem_cgroup_handle_over_high(void)
1850 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
1851 struct mem_cgroup
*memcg
;
1853 if (likely(!nr_pages
))
1856 memcg
= get_mem_cgroup_from_mm(current
->mm
);
1857 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
1858 css_put(&memcg
->css
);
1859 current
->memcg_nr_pages_over_high
= 0;
1862 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1863 unsigned int nr_pages
)
1865 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
1866 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1867 struct mem_cgroup
*mem_over_limit
;
1868 struct page_counter
*counter
;
1869 unsigned long nr_reclaimed
;
1870 bool may_swap
= true;
1871 bool drained
= false;
1873 if (mem_cgroup_is_root(memcg
))
1876 if (consume_stock(memcg
, nr_pages
))
1879 if (!do_memsw_account() ||
1880 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
1881 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
1883 if (do_memsw_account())
1884 page_counter_uncharge(&memcg
->memsw
, batch
);
1885 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
1887 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
1891 if (batch
> nr_pages
) {
1897 * Unlike in global OOM situations, memcg is not in a physical
1898 * memory shortage. Allow dying and OOM-killed tasks to
1899 * bypass the last charges so that they can exit quickly and
1900 * free their memory.
1902 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
1903 fatal_signal_pending(current
) ||
1904 current
->flags
& PF_EXITING
))
1907 if (unlikely(task_in_memcg_oom(current
)))
1910 if (!gfpflags_allow_blocking(gfp_mask
))
1913 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
1915 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
1916 gfp_mask
, may_swap
);
1918 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
1922 drain_all_stock(mem_over_limit
);
1927 if (gfp_mask
& __GFP_NORETRY
)
1930 * Even though the limit is exceeded at this point, reclaim
1931 * may have been able to free some pages. Retry the charge
1932 * before killing the task.
1934 * Only for regular pages, though: huge pages are rather
1935 * unlikely to succeed so close to the limit, and we fall back
1936 * to regular pages anyway in case of failure.
1938 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
1941 * At task move, charge accounts can be doubly counted. So, it's
1942 * better to wait until the end of task_move if something is going on.
1944 if (mem_cgroup_wait_acct_move(mem_over_limit
))
1950 if (gfp_mask
& __GFP_NOFAIL
)
1953 if (fatal_signal_pending(current
))
1956 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
1958 mem_cgroup_oom(mem_over_limit
, gfp_mask
,
1959 get_order(nr_pages
* PAGE_SIZE
));
1961 if (!(gfp_mask
& __GFP_NOFAIL
))
1965 * The allocation either can't fail or will lead to more memory
1966 * being freed very soon. Allow memory usage go over the limit
1967 * temporarily by force charging it.
1969 page_counter_charge(&memcg
->memory
, nr_pages
);
1970 if (do_memsw_account())
1971 page_counter_charge(&memcg
->memsw
, nr_pages
);
1972 css_get_many(&memcg
->css
, nr_pages
);
1977 css_get_many(&memcg
->css
, batch
);
1978 if (batch
> nr_pages
)
1979 refill_stock(memcg
, batch
- nr_pages
);
1982 * If the hierarchy is above the normal consumption range, schedule
1983 * reclaim on returning to userland. We can perform reclaim here
1984 * if __GFP_RECLAIM but let's always punt for simplicity and so that
1985 * GFP_KERNEL can consistently be used during reclaim. @memcg is
1986 * not recorded as it most likely matches current's and won't
1987 * change in the meantime. As high limit is checked again before
1988 * reclaim, the cost of mismatch is negligible.
1991 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
1992 /* Don't bother a random interrupted task */
1993 if (in_interrupt()) {
1994 schedule_work(&memcg
->high_work
);
1997 current
->memcg_nr_pages_over_high
+= batch
;
1998 set_notify_resume(current
);
2001 } while ((memcg
= parent_mem_cgroup(memcg
)));
2006 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2008 if (mem_cgroup_is_root(memcg
))
2011 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2012 if (do_memsw_account())
2013 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2015 css_put_many(&memcg
->css
, nr_pages
);
2018 static void lock_page_lru(struct page
*page
, int *isolated
)
2020 struct zone
*zone
= page_zone(page
);
2022 spin_lock_irq(zone_lru_lock(zone
));
2023 if (PageLRU(page
)) {
2024 struct lruvec
*lruvec
;
2026 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
2028 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2034 static void unlock_page_lru(struct page
*page
, int isolated
)
2036 struct zone
*zone
= page_zone(page
);
2039 struct lruvec
*lruvec
;
2041 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
2042 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2044 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2046 spin_unlock_irq(zone_lru_lock(zone
));
2049 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2054 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2057 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2058 * may already be on some other mem_cgroup's LRU. Take care of it.
2061 lock_page_lru(page
, &isolated
);
2064 * Nobody should be changing or seriously looking at
2065 * page->mem_cgroup at this point:
2067 * - the page is uncharged
2069 * - the page is off-LRU
2071 * - an anonymous fault has exclusive page access, except for
2072 * a locked page table
2074 * - a page cache insertion, a swapin fault, or a migration
2075 * have the page locked
2077 page
->mem_cgroup
= memcg
;
2080 unlock_page_lru(page
, isolated
);
2084 static int memcg_alloc_cache_id(void)
2089 id
= ida_simple_get(&memcg_cache_ida
,
2090 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2094 if (id
< memcg_nr_cache_ids
)
2098 * There's no space for the new id in memcg_caches arrays,
2099 * so we have to grow them.
2101 down_write(&memcg_cache_ids_sem
);
2103 size
= 2 * (id
+ 1);
2104 if (size
< MEMCG_CACHES_MIN_SIZE
)
2105 size
= MEMCG_CACHES_MIN_SIZE
;
2106 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2107 size
= MEMCG_CACHES_MAX_SIZE
;
2109 err
= memcg_update_all_caches(size
);
2111 err
= memcg_update_all_list_lrus(size
);
2113 memcg_nr_cache_ids
= size
;
2115 up_write(&memcg_cache_ids_sem
);
2118 ida_simple_remove(&memcg_cache_ida
, id
);
2124 static void memcg_free_cache_id(int id
)
2126 ida_simple_remove(&memcg_cache_ida
, id
);
2129 struct memcg_kmem_cache_create_work
{
2130 struct mem_cgroup
*memcg
;
2131 struct kmem_cache
*cachep
;
2132 struct work_struct work
;
2135 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2137 struct memcg_kmem_cache_create_work
*cw
=
2138 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2139 struct mem_cgroup
*memcg
= cw
->memcg
;
2140 struct kmem_cache
*cachep
= cw
->cachep
;
2142 memcg_create_kmem_cache(memcg
, cachep
);
2144 css_put(&memcg
->css
);
2149 * Enqueue the creation of a per-memcg kmem_cache.
2151 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2152 struct kmem_cache
*cachep
)
2154 struct memcg_kmem_cache_create_work
*cw
;
2156 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2160 css_get(&memcg
->css
);
2163 cw
->cachep
= cachep
;
2164 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2166 schedule_work(&cw
->work
);
2169 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2170 struct kmem_cache
*cachep
)
2173 * We need to stop accounting when we kmalloc, because if the
2174 * corresponding kmalloc cache is not yet created, the first allocation
2175 * in __memcg_schedule_kmem_cache_create will recurse.
2177 * However, it is better to enclose the whole function. Depending on
2178 * the debugging options enabled, INIT_WORK(), for instance, can
2179 * trigger an allocation. This too, will make us recurse. Because at
2180 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2181 * the safest choice is to do it like this, wrapping the whole function.
2183 current
->memcg_kmem_skip_account
= 1;
2184 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2185 current
->memcg_kmem_skip_account
= 0;
2188 static inline bool memcg_kmem_bypass(void)
2190 if (in_interrupt() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
2196 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2197 * @cachep: the original global kmem cache
2199 * Return the kmem_cache we're supposed to use for a slab allocation.
2200 * We try to use the current memcg's version of the cache.
2202 * If the cache does not exist yet, if we are the first user of it, we
2203 * create it asynchronously in a workqueue and let the current allocation
2204 * go through with the original cache.
2206 * This function takes a reference to the cache it returns to assure it
2207 * won't get destroyed while we are working with it. Once the caller is
2208 * done with it, memcg_kmem_put_cache() must be called to release the
2211 struct kmem_cache
*memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2213 struct mem_cgroup
*memcg
;
2214 struct kmem_cache
*memcg_cachep
;
2217 VM_BUG_ON(!is_root_cache(cachep
));
2219 if (memcg_kmem_bypass())
2222 if (current
->memcg_kmem_skip_account
)
2225 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2226 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2230 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2231 if (likely(memcg_cachep
))
2232 return memcg_cachep
;
2235 * If we are in a safe context (can wait, and not in interrupt
2236 * context), we could be be predictable and return right away.
2237 * This would guarantee that the allocation being performed
2238 * already belongs in the new cache.
2240 * However, there are some clashes that can arrive from locking.
2241 * For instance, because we acquire the slab_mutex while doing
2242 * memcg_create_kmem_cache, this means no further allocation
2243 * could happen with the slab_mutex held. So it's better to
2246 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2248 css_put(&memcg
->css
);
2253 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2254 * @cachep: the cache returned by memcg_kmem_get_cache
2256 void memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2258 if (!is_root_cache(cachep
))
2259 css_put(&cachep
->memcg_params
.memcg
->css
);
2263 * memcg_kmem_charge: charge a kmem page
2264 * @page: page to charge
2265 * @gfp: reclaim mode
2266 * @order: allocation order
2267 * @memcg: memory cgroup to charge
2269 * Returns 0 on success, an error code on failure.
2271 int memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2272 struct mem_cgroup
*memcg
)
2274 unsigned int nr_pages
= 1 << order
;
2275 struct page_counter
*counter
;
2278 ret
= try_charge(memcg
, gfp
, nr_pages
);
2282 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2283 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2284 cancel_charge(memcg
, nr_pages
);
2288 page
->mem_cgroup
= memcg
;
2294 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2295 * @page: page to charge
2296 * @gfp: reclaim mode
2297 * @order: allocation order
2299 * Returns 0 on success, an error code on failure.
2301 int memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2303 struct mem_cgroup
*memcg
;
2306 if (memcg_kmem_bypass())
2309 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2310 if (!mem_cgroup_is_root(memcg
)) {
2311 ret
= memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2313 __SetPageKmemcg(page
);
2315 css_put(&memcg
->css
);
2319 * memcg_kmem_uncharge: uncharge a kmem page
2320 * @page: page to uncharge
2321 * @order: allocation order
2323 void memcg_kmem_uncharge(struct page
*page
, int order
)
2325 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2326 unsigned int nr_pages
= 1 << order
;
2331 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2333 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2334 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2336 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2337 if (do_memsw_account())
2338 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2340 page
->mem_cgroup
= NULL
;
2342 /* slab pages do not have PageKmemcg flag set */
2343 if (PageKmemcg(page
))
2344 __ClearPageKmemcg(page
);
2346 css_put_many(&memcg
->css
, nr_pages
);
2348 #endif /* !CONFIG_SLOB */
2350 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2353 * Because tail pages are not marked as "used", set it. We're under
2354 * zone_lru_lock and migration entries setup in all page mappings.
2356 void mem_cgroup_split_huge_fixup(struct page
*head
)
2360 if (mem_cgroup_disabled())
2363 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2364 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2366 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2369 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2371 #ifdef CONFIG_MEMCG_SWAP
2372 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2375 int val
= (charge
) ? 1 : -1;
2376 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2380 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2381 * @entry: swap entry to be moved
2382 * @from: mem_cgroup which the entry is moved from
2383 * @to: mem_cgroup which the entry is moved to
2385 * It succeeds only when the swap_cgroup's record for this entry is the same
2386 * as the mem_cgroup's id of @from.
2388 * Returns 0 on success, -EINVAL on failure.
2390 * The caller must have charged to @to, IOW, called page_counter_charge() about
2391 * both res and memsw, and called css_get().
2393 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2394 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2396 unsigned short old_id
, new_id
;
2398 old_id
= mem_cgroup_id(from
);
2399 new_id
= mem_cgroup_id(to
);
2401 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2402 mem_cgroup_swap_statistics(from
, false);
2403 mem_cgroup_swap_statistics(to
, true);
2409 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2410 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2416 static DEFINE_MUTEX(memcg_limit_mutex
);
2418 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2419 unsigned long limit
)
2421 unsigned long curusage
;
2422 unsigned long oldusage
;
2423 bool enlarge
= false;
2428 * For keeping hierarchical_reclaim simple, how long we should retry
2429 * is depends on callers. We set our retry-count to be function
2430 * of # of children which we should visit in this loop.
2432 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2433 mem_cgroup_count_children(memcg
);
2435 oldusage
= page_counter_read(&memcg
->memory
);
2438 if (signal_pending(current
)) {
2443 mutex_lock(&memcg_limit_mutex
);
2444 if (limit
> memcg
->memsw
.limit
) {
2445 mutex_unlock(&memcg_limit_mutex
);
2449 if (limit
> memcg
->memory
.limit
)
2451 ret
= page_counter_limit(&memcg
->memory
, limit
);
2452 mutex_unlock(&memcg_limit_mutex
);
2457 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2459 curusage
= page_counter_read(&memcg
->memory
);
2460 /* Usage is reduced ? */
2461 if (curusage
>= oldusage
)
2464 oldusage
= curusage
;
2465 } while (retry_count
);
2467 if (!ret
&& enlarge
)
2468 memcg_oom_recover(memcg
);
2473 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2474 unsigned long limit
)
2476 unsigned long curusage
;
2477 unsigned long oldusage
;
2478 bool enlarge
= false;
2482 /* see mem_cgroup_resize_res_limit */
2483 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2484 mem_cgroup_count_children(memcg
);
2486 oldusage
= page_counter_read(&memcg
->memsw
);
2489 if (signal_pending(current
)) {
2494 mutex_lock(&memcg_limit_mutex
);
2495 if (limit
< memcg
->memory
.limit
) {
2496 mutex_unlock(&memcg_limit_mutex
);
2500 if (limit
> memcg
->memsw
.limit
)
2502 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2503 mutex_unlock(&memcg_limit_mutex
);
2508 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2510 curusage
= page_counter_read(&memcg
->memsw
);
2511 /* Usage is reduced ? */
2512 if (curusage
>= oldusage
)
2515 oldusage
= curusage
;
2516 } while (retry_count
);
2518 if (!ret
&& enlarge
)
2519 memcg_oom_recover(memcg
);
2524 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
2526 unsigned long *total_scanned
)
2528 unsigned long nr_reclaimed
= 0;
2529 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
2530 unsigned long reclaimed
;
2532 struct mem_cgroup_tree_per_node
*mctz
;
2533 unsigned long excess
;
2534 unsigned long nr_scanned
;
2539 mctz
= soft_limit_tree_node(pgdat
->node_id
);
2542 * Do not even bother to check the largest node if the root
2543 * is empty. Do it lockless to prevent lock bouncing. Races
2544 * are acceptable as soft limit is best effort anyway.
2546 if (RB_EMPTY_ROOT(&mctz
->rb_root
))
2550 * This loop can run a while, specially if mem_cgroup's continuously
2551 * keep exceeding their soft limit and putting the system under
2558 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2563 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
2564 gfp_mask
, &nr_scanned
);
2565 nr_reclaimed
+= reclaimed
;
2566 *total_scanned
+= nr_scanned
;
2567 spin_lock_irq(&mctz
->lock
);
2568 __mem_cgroup_remove_exceeded(mz
, mctz
);
2571 * If we failed to reclaim anything from this memory cgroup
2572 * it is time to move on to the next cgroup
2576 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2578 excess
= soft_limit_excess(mz
->memcg
);
2580 * One school of thought says that we should not add
2581 * back the node to the tree if reclaim returns 0.
2582 * But our reclaim could return 0, simply because due
2583 * to priority we are exposing a smaller subset of
2584 * memory to reclaim from. Consider this as a longer
2587 /* If excess == 0, no tree ops */
2588 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2589 spin_unlock_irq(&mctz
->lock
);
2590 css_put(&mz
->memcg
->css
);
2593 * Could not reclaim anything and there are no more
2594 * mem cgroups to try or we seem to be looping without
2595 * reclaiming anything.
2597 if (!nr_reclaimed
&&
2599 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2601 } while (!nr_reclaimed
);
2603 css_put(&next_mz
->memcg
->css
);
2604 return nr_reclaimed
;
2608 * Test whether @memcg has children, dead or alive. Note that this
2609 * function doesn't care whether @memcg has use_hierarchy enabled and
2610 * returns %true if there are child csses according to the cgroup
2611 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2613 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2618 ret
= css_next_child(NULL
, &memcg
->css
);
2624 * Reclaims as many pages from the given memcg as possible.
2626 * Caller is responsible for holding css reference for memcg.
2628 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2630 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2632 /* we call try-to-free pages for make this cgroup empty */
2633 lru_add_drain_all();
2634 /* try to free all pages in this cgroup */
2635 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2638 if (signal_pending(current
))
2641 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2645 /* maybe some writeback is necessary */
2646 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2654 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
2655 char *buf
, size_t nbytes
,
2658 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2660 if (mem_cgroup_is_root(memcg
))
2662 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
2665 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
2668 return mem_cgroup_from_css(css
)->use_hierarchy
;
2671 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
2672 struct cftype
*cft
, u64 val
)
2675 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2676 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
2678 if (memcg
->use_hierarchy
== val
)
2682 * If parent's use_hierarchy is set, we can't make any modifications
2683 * in the child subtrees. If it is unset, then the change can
2684 * occur, provided the current cgroup has no children.
2686 * For the root cgroup, parent_mem is NULL, we allow value to be
2687 * set if there are no children.
2689 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
2690 (val
== 1 || val
== 0)) {
2691 if (!memcg_has_children(memcg
))
2692 memcg
->use_hierarchy
= val
;
2701 static void tree_stat(struct mem_cgroup
*memcg
, unsigned long *stat
)
2703 struct mem_cgroup
*iter
;
2706 memset(stat
, 0, sizeof(*stat
) * MEMCG_NR_STAT
);
2708 for_each_mem_cgroup_tree(iter
, memcg
) {
2709 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
2710 stat
[i
] += mem_cgroup_read_stat(iter
, i
);
2714 static void tree_events(struct mem_cgroup
*memcg
, unsigned long *events
)
2716 struct mem_cgroup
*iter
;
2719 memset(events
, 0, sizeof(*events
) * MEMCG_NR_EVENTS
);
2721 for_each_mem_cgroup_tree(iter
, memcg
) {
2722 for (i
= 0; i
< MEMCG_NR_EVENTS
; i
++)
2723 events
[i
] += mem_cgroup_read_events(iter
, i
);
2727 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
2729 unsigned long val
= 0;
2731 if (mem_cgroup_is_root(memcg
)) {
2732 struct mem_cgroup
*iter
;
2734 for_each_mem_cgroup_tree(iter
, memcg
) {
2735 val
+= mem_cgroup_read_stat(iter
,
2736 MEM_CGROUP_STAT_CACHE
);
2737 val
+= mem_cgroup_read_stat(iter
,
2738 MEM_CGROUP_STAT_RSS
);
2740 val
+= mem_cgroup_read_stat(iter
,
2741 MEM_CGROUP_STAT_SWAP
);
2745 val
= page_counter_read(&memcg
->memory
);
2747 val
= page_counter_read(&memcg
->memsw
);
2760 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
2763 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2764 struct page_counter
*counter
;
2766 switch (MEMFILE_TYPE(cft
->private)) {
2768 counter
= &memcg
->memory
;
2771 counter
= &memcg
->memsw
;
2774 counter
= &memcg
->kmem
;
2777 counter
= &memcg
->tcpmem
;
2783 switch (MEMFILE_ATTR(cft
->private)) {
2785 if (counter
== &memcg
->memory
)
2786 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
2787 if (counter
== &memcg
->memsw
)
2788 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
2789 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
2791 return (u64
)counter
->limit
* PAGE_SIZE
;
2793 return (u64
)counter
->watermark
* PAGE_SIZE
;
2795 return counter
->failcnt
;
2796 case RES_SOFT_LIMIT
:
2797 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
2804 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2808 if (cgroup_memory_nokmem
)
2811 BUG_ON(memcg
->kmemcg_id
>= 0);
2812 BUG_ON(memcg
->kmem_state
);
2814 memcg_id
= memcg_alloc_cache_id();
2818 static_branch_inc(&memcg_kmem_enabled_key
);
2820 * A memory cgroup is considered kmem-online as soon as it gets
2821 * kmemcg_id. Setting the id after enabling static branching will
2822 * guarantee no one starts accounting before all call sites are
2825 memcg
->kmemcg_id
= memcg_id
;
2826 memcg
->kmem_state
= KMEM_ONLINE
;
2831 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2833 struct cgroup_subsys_state
*css
;
2834 struct mem_cgroup
*parent
, *child
;
2837 if (memcg
->kmem_state
!= KMEM_ONLINE
)
2840 * Clear the online state before clearing memcg_caches array
2841 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2842 * guarantees that no cache will be created for this cgroup
2843 * after we are done (see memcg_create_kmem_cache()).
2845 memcg
->kmem_state
= KMEM_ALLOCATED
;
2847 memcg_deactivate_kmem_caches(memcg
);
2849 kmemcg_id
= memcg
->kmemcg_id
;
2850 BUG_ON(kmemcg_id
< 0);
2852 parent
= parent_mem_cgroup(memcg
);
2854 parent
= root_mem_cgroup
;
2857 * Change kmemcg_id of this cgroup and all its descendants to the
2858 * parent's id, and then move all entries from this cgroup's list_lrus
2859 * to ones of the parent. After we have finished, all list_lrus
2860 * corresponding to this cgroup are guaranteed to remain empty. The
2861 * ordering is imposed by list_lru_node->lock taken by
2862 * memcg_drain_all_list_lrus().
2864 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2865 css_for_each_descendant_pre(css
, &memcg
->css
) {
2866 child
= mem_cgroup_from_css(css
);
2867 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
2868 child
->kmemcg_id
= parent
->kmemcg_id
;
2869 if (!memcg
->use_hierarchy
)
2874 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
2876 memcg_free_cache_id(kmemcg_id
);
2879 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2881 /* css_alloc() failed, offlining didn't happen */
2882 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
2883 memcg_offline_kmem(memcg
);
2885 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
2886 memcg_destroy_kmem_caches(memcg
);
2887 static_branch_dec(&memcg_kmem_enabled_key
);
2888 WARN_ON(page_counter_read(&memcg
->kmem
));
2892 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2896 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2899 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2902 #endif /* !CONFIG_SLOB */
2904 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2905 unsigned long limit
)
2909 mutex_lock(&memcg_limit_mutex
);
2910 ret
= page_counter_limit(&memcg
->kmem
, limit
);
2911 mutex_unlock(&memcg_limit_mutex
);
2915 static int memcg_update_tcp_limit(struct mem_cgroup
*memcg
, unsigned long limit
)
2919 mutex_lock(&memcg_limit_mutex
);
2921 ret
= page_counter_limit(&memcg
->tcpmem
, limit
);
2925 if (!memcg
->tcpmem_active
) {
2927 * The active flag needs to be written after the static_key
2928 * update. This is what guarantees that the socket activation
2929 * function is the last one to run. See sock_update_memcg() for
2930 * details, and note that we don't mark any socket as belonging
2931 * to this memcg until that flag is up.
2933 * We need to do this, because static_keys will span multiple
2934 * sites, but we can't control their order. If we mark a socket
2935 * as accounted, but the accounting functions are not patched in
2936 * yet, we'll lose accounting.
2938 * We never race with the readers in sock_update_memcg(),
2939 * because when this value change, the code to process it is not
2942 static_branch_inc(&memcg_sockets_enabled_key
);
2943 memcg
->tcpmem_active
= true;
2946 mutex_unlock(&memcg_limit_mutex
);
2951 * The user of this function is...
2954 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
2955 char *buf
, size_t nbytes
, loff_t off
)
2957 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2958 unsigned long nr_pages
;
2961 buf
= strstrip(buf
);
2962 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
2966 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
2968 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
2972 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
2974 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
2977 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
2980 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
2983 ret
= memcg_update_tcp_limit(memcg
, nr_pages
);
2987 case RES_SOFT_LIMIT
:
2988 memcg
->soft_limit
= nr_pages
;
2992 return ret
?: nbytes
;
2995 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
2996 size_t nbytes
, loff_t off
)
2998 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2999 struct page_counter
*counter
;
3001 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3003 counter
= &memcg
->memory
;
3006 counter
= &memcg
->memsw
;
3009 counter
= &memcg
->kmem
;
3012 counter
= &memcg
->tcpmem
;
3018 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3020 page_counter_reset_watermark(counter
);
3023 counter
->failcnt
= 0;
3032 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3035 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3039 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3040 struct cftype
*cft
, u64 val
)
3042 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3044 if (val
& ~MOVE_MASK
)
3048 * No kind of locking is needed in here, because ->can_attach() will
3049 * check this value once in the beginning of the process, and then carry
3050 * on with stale data. This means that changes to this value will only
3051 * affect task migrations starting after the change.
3053 memcg
->move_charge_at_immigrate
= val
;
3057 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3058 struct cftype
*cft
, u64 val
)
3065 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3069 unsigned int lru_mask
;
3072 static const struct numa_stat stats
[] = {
3073 { "total", LRU_ALL
},
3074 { "file", LRU_ALL_FILE
},
3075 { "anon", LRU_ALL_ANON
},
3076 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3078 const struct numa_stat
*stat
;
3081 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3083 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3084 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3085 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3086 for_each_node_state(nid
, N_MEMORY
) {
3087 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3089 seq_printf(m
, " N%d=%lu", nid
, nr
);
3094 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3095 struct mem_cgroup
*iter
;
3098 for_each_mem_cgroup_tree(iter
, memcg
)
3099 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3100 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3101 for_each_node_state(nid
, N_MEMORY
) {
3103 for_each_mem_cgroup_tree(iter
, memcg
)
3104 nr
+= mem_cgroup_node_nr_lru_pages(
3105 iter
, nid
, stat
->lru_mask
);
3106 seq_printf(m
, " N%d=%lu", nid
, nr
);
3113 #endif /* CONFIG_NUMA */
3115 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3117 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3118 unsigned long memory
, memsw
;
3119 struct mem_cgroup
*mi
;
3122 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3123 MEM_CGROUP_STAT_NSTATS
);
3124 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3125 MEM_CGROUP_EVENTS_NSTATS
);
3126 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3128 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3129 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3131 seq_printf(m
, "%s %lu\n", mem_cgroup_stat_names
[i
],
3132 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3135 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3136 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3137 mem_cgroup_read_events(memcg
, i
));
3139 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3140 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3141 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3143 /* Hierarchical information */
3144 memory
= memsw
= PAGE_COUNTER_MAX
;
3145 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3146 memory
= min(memory
, mi
->memory
.limit
);
3147 memsw
= min(memsw
, mi
->memsw
.limit
);
3149 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3150 (u64
)memory
* PAGE_SIZE
);
3151 if (do_memsw_account())
3152 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3153 (u64
)memsw
* PAGE_SIZE
);
3155 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3156 unsigned long long val
= 0;
3158 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3160 for_each_mem_cgroup_tree(mi
, memcg
)
3161 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3162 seq_printf(m
, "total_%s %llu\n", mem_cgroup_stat_names
[i
], val
);
3165 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3166 unsigned long long val
= 0;
3168 for_each_mem_cgroup_tree(mi
, memcg
)
3169 val
+= mem_cgroup_read_events(mi
, i
);
3170 seq_printf(m
, "total_%s %llu\n",
3171 mem_cgroup_events_names
[i
], val
);
3174 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3175 unsigned long long val
= 0;
3177 for_each_mem_cgroup_tree(mi
, memcg
)
3178 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3179 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3182 #ifdef CONFIG_DEBUG_VM
3185 struct mem_cgroup_per_node
*mz
;
3186 struct zone_reclaim_stat
*rstat
;
3187 unsigned long recent_rotated
[2] = {0, 0};
3188 unsigned long recent_scanned
[2] = {0, 0};
3190 for_each_online_pgdat(pgdat
) {
3191 mz
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
3192 rstat
= &mz
->lruvec
.reclaim_stat
;
3194 recent_rotated
[0] += rstat
->recent_rotated
[0];
3195 recent_rotated
[1] += rstat
->recent_rotated
[1];
3196 recent_scanned
[0] += rstat
->recent_scanned
[0];
3197 recent_scanned
[1] += rstat
->recent_scanned
[1];
3199 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3200 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3201 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3202 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3209 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3212 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3214 return mem_cgroup_swappiness(memcg
);
3217 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3218 struct cftype
*cft
, u64 val
)
3220 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3226 memcg
->swappiness
= val
;
3228 vm_swappiness
= val
;
3233 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3235 struct mem_cgroup_threshold_ary
*t
;
3236 unsigned long usage
;
3241 t
= rcu_dereference(memcg
->thresholds
.primary
);
3243 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3248 usage
= mem_cgroup_usage(memcg
, swap
);
3251 * current_threshold points to threshold just below or equal to usage.
3252 * If it's not true, a threshold was crossed after last
3253 * call of __mem_cgroup_threshold().
3255 i
= t
->current_threshold
;
3258 * Iterate backward over array of thresholds starting from
3259 * current_threshold and check if a threshold is crossed.
3260 * If none of thresholds below usage is crossed, we read
3261 * only one element of the array here.
3263 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3264 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3266 /* i = current_threshold + 1 */
3270 * Iterate forward over array of thresholds starting from
3271 * current_threshold+1 and check if a threshold is crossed.
3272 * If none of thresholds above usage is crossed, we read
3273 * only one element of the array here.
3275 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3276 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3278 /* Update current_threshold */
3279 t
->current_threshold
= i
- 1;
3284 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3287 __mem_cgroup_threshold(memcg
, false);
3288 if (do_memsw_account())
3289 __mem_cgroup_threshold(memcg
, true);
3291 memcg
= parent_mem_cgroup(memcg
);
3295 static int compare_thresholds(const void *a
, const void *b
)
3297 const struct mem_cgroup_threshold
*_a
= a
;
3298 const struct mem_cgroup_threshold
*_b
= b
;
3300 if (_a
->threshold
> _b
->threshold
)
3303 if (_a
->threshold
< _b
->threshold
)
3309 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3311 struct mem_cgroup_eventfd_list
*ev
;
3313 spin_lock(&memcg_oom_lock
);
3315 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3316 eventfd_signal(ev
->eventfd
, 1);
3318 spin_unlock(&memcg_oom_lock
);
3322 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3324 struct mem_cgroup
*iter
;
3326 for_each_mem_cgroup_tree(iter
, memcg
)
3327 mem_cgroup_oom_notify_cb(iter
);
3330 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3331 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3333 struct mem_cgroup_thresholds
*thresholds
;
3334 struct mem_cgroup_threshold_ary
*new;
3335 unsigned long threshold
;
3336 unsigned long usage
;
3339 ret
= page_counter_memparse(args
, "-1", &threshold
);
3343 mutex_lock(&memcg
->thresholds_lock
);
3346 thresholds
= &memcg
->thresholds
;
3347 usage
= mem_cgroup_usage(memcg
, false);
3348 } else if (type
== _MEMSWAP
) {
3349 thresholds
= &memcg
->memsw_thresholds
;
3350 usage
= mem_cgroup_usage(memcg
, true);
3354 /* Check if a threshold crossed before adding a new one */
3355 if (thresholds
->primary
)
3356 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3358 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3360 /* Allocate memory for new array of thresholds */
3361 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3369 /* Copy thresholds (if any) to new array */
3370 if (thresholds
->primary
) {
3371 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3372 sizeof(struct mem_cgroup_threshold
));
3375 /* Add new threshold */
3376 new->entries
[size
- 1].eventfd
= eventfd
;
3377 new->entries
[size
- 1].threshold
= threshold
;
3379 /* Sort thresholds. Registering of new threshold isn't time-critical */
3380 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3381 compare_thresholds
, NULL
);
3383 /* Find current threshold */
3384 new->current_threshold
= -1;
3385 for (i
= 0; i
< size
; i
++) {
3386 if (new->entries
[i
].threshold
<= usage
) {
3388 * new->current_threshold will not be used until
3389 * rcu_assign_pointer(), so it's safe to increment
3392 ++new->current_threshold
;
3397 /* Free old spare buffer and save old primary buffer as spare */
3398 kfree(thresholds
->spare
);
3399 thresholds
->spare
= thresholds
->primary
;
3401 rcu_assign_pointer(thresholds
->primary
, new);
3403 /* To be sure that nobody uses thresholds */
3407 mutex_unlock(&memcg
->thresholds_lock
);
3412 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3413 struct eventfd_ctx
*eventfd
, const char *args
)
3415 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3418 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3419 struct eventfd_ctx
*eventfd
, const char *args
)
3421 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3424 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3425 struct eventfd_ctx
*eventfd
, enum res_type type
)
3427 struct mem_cgroup_thresholds
*thresholds
;
3428 struct mem_cgroup_threshold_ary
*new;
3429 unsigned long usage
;
3432 mutex_lock(&memcg
->thresholds_lock
);
3435 thresholds
= &memcg
->thresholds
;
3436 usage
= mem_cgroup_usage(memcg
, false);
3437 } else if (type
== _MEMSWAP
) {
3438 thresholds
= &memcg
->memsw_thresholds
;
3439 usage
= mem_cgroup_usage(memcg
, true);
3443 if (!thresholds
->primary
)
3446 /* Check if a threshold crossed before removing */
3447 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3449 /* Calculate new number of threshold */
3451 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3452 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3456 new = thresholds
->spare
;
3458 /* Set thresholds array to NULL if we don't have thresholds */
3467 /* Copy thresholds and find current threshold */
3468 new->current_threshold
= -1;
3469 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3470 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3473 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3474 if (new->entries
[j
].threshold
<= usage
) {
3476 * new->current_threshold will not be used
3477 * until rcu_assign_pointer(), so it's safe to increment
3480 ++new->current_threshold
;
3486 /* Swap primary and spare array */
3487 thresholds
->spare
= thresholds
->primary
;
3489 rcu_assign_pointer(thresholds
->primary
, new);
3491 /* To be sure that nobody uses thresholds */
3494 /* If all events are unregistered, free the spare array */
3496 kfree(thresholds
->spare
);
3497 thresholds
->spare
= NULL
;
3500 mutex_unlock(&memcg
->thresholds_lock
);
3503 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3504 struct eventfd_ctx
*eventfd
)
3506 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3509 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3510 struct eventfd_ctx
*eventfd
)
3512 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3515 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3516 struct eventfd_ctx
*eventfd
, const char *args
)
3518 struct mem_cgroup_eventfd_list
*event
;
3520 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3524 spin_lock(&memcg_oom_lock
);
3526 event
->eventfd
= eventfd
;
3527 list_add(&event
->list
, &memcg
->oom_notify
);
3529 /* already in OOM ? */
3530 if (memcg
->under_oom
)
3531 eventfd_signal(eventfd
, 1);
3532 spin_unlock(&memcg_oom_lock
);
3537 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3538 struct eventfd_ctx
*eventfd
)
3540 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3542 spin_lock(&memcg_oom_lock
);
3544 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3545 if (ev
->eventfd
== eventfd
) {
3546 list_del(&ev
->list
);
3551 spin_unlock(&memcg_oom_lock
);
3554 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3556 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3558 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3559 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3563 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3564 struct cftype
*cft
, u64 val
)
3566 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3568 /* cannot set to root cgroup and only 0 and 1 are allowed */
3569 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3572 memcg
->oom_kill_disable
= val
;
3574 memcg_oom_recover(memcg
);
3579 #ifdef CONFIG_CGROUP_WRITEBACK
3581 struct list_head
*mem_cgroup_cgwb_list(struct mem_cgroup
*memcg
)
3583 return &memcg
->cgwb_list
;
3586 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3588 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3591 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3593 wb_domain_exit(&memcg
->cgwb_domain
);
3596 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3598 wb_domain_size_changed(&memcg
->cgwb_domain
);
3601 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3603 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3605 if (!memcg
->css
.parent
)
3608 return &memcg
->cgwb_domain
;
3612 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3613 * @wb: bdi_writeback in question
3614 * @pfilepages: out parameter for number of file pages
3615 * @pheadroom: out parameter for number of allocatable pages according to memcg
3616 * @pdirty: out parameter for number of dirty pages
3617 * @pwriteback: out parameter for number of pages under writeback
3619 * Determine the numbers of file, headroom, dirty, and writeback pages in
3620 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3621 * is a bit more involved.
3623 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3624 * headroom is calculated as the lowest headroom of itself and the
3625 * ancestors. Note that this doesn't consider the actual amount of
3626 * available memory in the system. The caller should further cap
3627 * *@pheadroom accordingly.
3629 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
3630 unsigned long *pheadroom
, unsigned long *pdirty
,
3631 unsigned long *pwriteback
)
3633 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3634 struct mem_cgroup
*parent
;
3636 *pdirty
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
3638 /* this should eventually include NR_UNSTABLE_NFS */
3639 *pwriteback
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
3640 *pfilepages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
3641 (1 << LRU_ACTIVE_FILE
));
3642 *pheadroom
= PAGE_COUNTER_MAX
;
3644 while ((parent
= parent_mem_cgroup(memcg
))) {
3645 unsigned long ceiling
= min(memcg
->memory
.limit
, memcg
->high
);
3646 unsigned long used
= page_counter_read(&memcg
->memory
);
3648 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
3653 #else /* CONFIG_CGROUP_WRITEBACK */
3655 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3660 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3664 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3668 #endif /* CONFIG_CGROUP_WRITEBACK */
3671 * DO NOT USE IN NEW FILES.
3673 * "cgroup.event_control" implementation.
3675 * This is way over-engineered. It tries to support fully configurable
3676 * events for each user. Such level of flexibility is completely
3677 * unnecessary especially in the light of the planned unified hierarchy.
3679 * Please deprecate this and replace with something simpler if at all
3684 * Unregister event and free resources.
3686 * Gets called from workqueue.
3688 static void memcg_event_remove(struct work_struct
*work
)
3690 struct mem_cgroup_event
*event
=
3691 container_of(work
, struct mem_cgroup_event
, remove
);
3692 struct mem_cgroup
*memcg
= event
->memcg
;
3694 remove_wait_queue(event
->wqh
, &event
->wait
);
3696 event
->unregister_event(memcg
, event
->eventfd
);
3698 /* Notify userspace the event is going away. */
3699 eventfd_signal(event
->eventfd
, 1);
3701 eventfd_ctx_put(event
->eventfd
);
3703 css_put(&memcg
->css
);
3707 * Gets called on POLLHUP on eventfd when user closes it.
3709 * Called with wqh->lock held and interrupts disabled.
3711 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
3712 int sync
, void *key
)
3714 struct mem_cgroup_event
*event
=
3715 container_of(wait
, struct mem_cgroup_event
, wait
);
3716 struct mem_cgroup
*memcg
= event
->memcg
;
3717 unsigned long flags
= (unsigned long)key
;
3719 if (flags
& POLLHUP
) {
3721 * If the event has been detached at cgroup removal, we
3722 * can simply return knowing the other side will cleanup
3725 * We can't race against event freeing since the other
3726 * side will require wqh->lock via remove_wait_queue(),
3729 spin_lock(&memcg
->event_list_lock
);
3730 if (!list_empty(&event
->list
)) {
3731 list_del_init(&event
->list
);
3733 * We are in atomic context, but cgroup_event_remove()
3734 * may sleep, so we have to call it in workqueue.
3736 schedule_work(&event
->remove
);
3738 spin_unlock(&memcg
->event_list_lock
);
3744 static void memcg_event_ptable_queue_proc(struct file
*file
,
3745 wait_queue_head_t
*wqh
, poll_table
*pt
)
3747 struct mem_cgroup_event
*event
=
3748 container_of(pt
, struct mem_cgroup_event
, pt
);
3751 add_wait_queue(wqh
, &event
->wait
);
3755 * DO NOT USE IN NEW FILES.
3757 * Parse input and register new cgroup event handler.
3759 * Input must be in format '<event_fd> <control_fd> <args>'.
3760 * Interpretation of args is defined by control file implementation.
3762 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
3763 char *buf
, size_t nbytes
, loff_t off
)
3765 struct cgroup_subsys_state
*css
= of_css(of
);
3766 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3767 struct mem_cgroup_event
*event
;
3768 struct cgroup_subsys_state
*cfile_css
;
3769 unsigned int efd
, cfd
;
3776 buf
= strstrip(buf
);
3778 efd
= simple_strtoul(buf
, &endp
, 10);
3783 cfd
= simple_strtoul(buf
, &endp
, 10);
3784 if ((*endp
!= ' ') && (*endp
!= '\0'))
3788 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
3792 event
->memcg
= memcg
;
3793 INIT_LIST_HEAD(&event
->list
);
3794 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
3795 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
3796 INIT_WORK(&event
->remove
, memcg_event_remove
);
3804 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
3805 if (IS_ERR(event
->eventfd
)) {
3806 ret
= PTR_ERR(event
->eventfd
);
3813 goto out_put_eventfd
;
3816 /* the process need read permission on control file */
3817 /* AV: shouldn't we check that it's been opened for read instead? */
3818 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
3823 * Determine the event callbacks and set them in @event. This used
3824 * to be done via struct cftype but cgroup core no longer knows
3825 * about these events. The following is crude but the whole thing
3826 * is for compatibility anyway.
3828 * DO NOT ADD NEW FILES.
3830 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
3832 if (!strcmp(name
, "memory.usage_in_bytes")) {
3833 event
->register_event
= mem_cgroup_usage_register_event
;
3834 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
3835 } else if (!strcmp(name
, "memory.oom_control")) {
3836 event
->register_event
= mem_cgroup_oom_register_event
;
3837 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
3838 } else if (!strcmp(name
, "memory.pressure_level")) {
3839 event
->register_event
= vmpressure_register_event
;
3840 event
->unregister_event
= vmpressure_unregister_event
;
3841 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
3842 event
->register_event
= memsw_cgroup_usage_register_event
;
3843 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
3850 * Verify @cfile should belong to @css. Also, remaining events are
3851 * automatically removed on cgroup destruction but the removal is
3852 * asynchronous, so take an extra ref on @css.
3854 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
3855 &memory_cgrp_subsys
);
3857 if (IS_ERR(cfile_css
))
3859 if (cfile_css
!= css
) {
3864 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
3868 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
3870 spin_lock(&memcg
->event_list_lock
);
3871 list_add(&event
->list
, &memcg
->event_list
);
3872 spin_unlock(&memcg
->event_list_lock
);
3884 eventfd_ctx_put(event
->eventfd
);
3893 static struct cftype mem_cgroup_legacy_files
[] = {
3895 .name
= "usage_in_bytes",
3896 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
3897 .read_u64
= mem_cgroup_read_u64
,
3900 .name
= "max_usage_in_bytes",
3901 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
3902 .write
= mem_cgroup_reset
,
3903 .read_u64
= mem_cgroup_read_u64
,
3906 .name
= "limit_in_bytes",
3907 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
3908 .write
= mem_cgroup_write
,
3909 .read_u64
= mem_cgroup_read_u64
,
3912 .name
= "soft_limit_in_bytes",
3913 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
3914 .write
= mem_cgroup_write
,
3915 .read_u64
= mem_cgroup_read_u64
,
3919 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
3920 .write
= mem_cgroup_reset
,
3921 .read_u64
= mem_cgroup_read_u64
,
3925 .seq_show
= memcg_stat_show
,
3928 .name
= "force_empty",
3929 .write
= mem_cgroup_force_empty_write
,
3932 .name
= "use_hierarchy",
3933 .write_u64
= mem_cgroup_hierarchy_write
,
3934 .read_u64
= mem_cgroup_hierarchy_read
,
3937 .name
= "cgroup.event_control", /* XXX: for compat */
3938 .write
= memcg_write_event_control
,
3939 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
3942 .name
= "swappiness",
3943 .read_u64
= mem_cgroup_swappiness_read
,
3944 .write_u64
= mem_cgroup_swappiness_write
,
3947 .name
= "move_charge_at_immigrate",
3948 .read_u64
= mem_cgroup_move_charge_read
,
3949 .write_u64
= mem_cgroup_move_charge_write
,
3952 .name
= "oom_control",
3953 .seq_show
= mem_cgroup_oom_control_read
,
3954 .write_u64
= mem_cgroup_oom_control_write
,
3955 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
3958 .name
= "pressure_level",
3962 .name
= "numa_stat",
3963 .seq_show
= memcg_numa_stat_show
,
3967 .name
= "kmem.limit_in_bytes",
3968 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
3969 .write
= mem_cgroup_write
,
3970 .read_u64
= mem_cgroup_read_u64
,
3973 .name
= "kmem.usage_in_bytes",
3974 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
3975 .read_u64
= mem_cgroup_read_u64
,
3978 .name
= "kmem.failcnt",
3979 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
3980 .write
= mem_cgroup_reset
,
3981 .read_u64
= mem_cgroup_read_u64
,
3984 .name
= "kmem.max_usage_in_bytes",
3985 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
3986 .write
= mem_cgroup_reset
,
3987 .read_u64
= mem_cgroup_read_u64
,
3989 #ifdef CONFIG_SLABINFO
3991 .name
= "kmem.slabinfo",
3992 .seq_start
= slab_start
,
3993 .seq_next
= slab_next
,
3994 .seq_stop
= slab_stop
,
3995 .seq_show
= memcg_slab_show
,
3999 .name
= "kmem.tcp.limit_in_bytes",
4000 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4001 .write
= mem_cgroup_write
,
4002 .read_u64
= mem_cgroup_read_u64
,
4005 .name
= "kmem.tcp.usage_in_bytes",
4006 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4007 .read_u64
= mem_cgroup_read_u64
,
4010 .name
= "kmem.tcp.failcnt",
4011 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4012 .write
= mem_cgroup_reset
,
4013 .read_u64
= mem_cgroup_read_u64
,
4016 .name
= "kmem.tcp.max_usage_in_bytes",
4017 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4018 .write
= mem_cgroup_reset
,
4019 .read_u64
= mem_cgroup_read_u64
,
4021 { }, /* terminate */
4025 * Private memory cgroup IDR
4027 * Swap-out records and page cache shadow entries need to store memcg
4028 * references in constrained space, so we maintain an ID space that is
4029 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4030 * memory-controlled cgroups to 64k.
4032 * However, there usually are many references to the oflline CSS after
4033 * the cgroup has been destroyed, such as page cache or reclaimable
4034 * slab objects, that don't need to hang on to the ID. We want to keep
4035 * those dead CSS from occupying IDs, or we might quickly exhaust the
4036 * relatively small ID space and prevent the creation of new cgroups
4037 * even when there are much fewer than 64k cgroups - possibly none.
4039 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4040 * be freed and recycled when it's no longer needed, which is usually
4041 * when the CSS is offlined.
4043 * The only exception to that are records of swapped out tmpfs/shmem
4044 * pages that need to be attributed to live ancestors on swapin. But
4045 * those references are manageable from userspace.
4048 static DEFINE_IDR(mem_cgroup_idr
);
4050 static void mem_cgroup_id_get_many(struct mem_cgroup
*memcg
, unsigned int n
)
4052 VM_BUG_ON(atomic_read(&memcg
->id
.ref
) <= 0);
4053 atomic_add(n
, &memcg
->id
.ref
);
4056 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
4058 VM_BUG_ON(atomic_read(&memcg
->id
.ref
) < n
);
4059 if (atomic_sub_and_test(n
, &memcg
->id
.ref
)) {
4060 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4063 /* Memcg ID pins CSS */
4064 css_put(&memcg
->css
);
4068 static inline void mem_cgroup_id_get(struct mem_cgroup
*memcg
)
4070 mem_cgroup_id_get_many(memcg
, 1);
4073 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
4075 mem_cgroup_id_put_many(memcg
, 1);
4079 * mem_cgroup_from_id - look up a memcg from a memcg id
4080 * @id: the memcg id to look up
4082 * Caller must hold rcu_read_lock().
4084 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
4086 WARN_ON_ONCE(!rcu_read_lock_held());
4087 return idr_find(&mem_cgroup_idr
, id
);
4090 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4092 struct mem_cgroup_per_node
*pn
;
4095 * This routine is called against possible nodes.
4096 * But it's BUG to call kmalloc() against offline node.
4098 * TODO: this routine can waste much memory for nodes which will
4099 * never be onlined. It's better to use memory hotplug callback
4102 if (!node_state(node
, N_NORMAL_MEMORY
))
4104 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4108 lruvec_init(&pn
->lruvec
);
4109 pn
->usage_in_excess
= 0;
4110 pn
->on_tree
= false;
4113 memcg
->nodeinfo
[node
] = pn
;
4117 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4119 kfree(memcg
->nodeinfo
[node
]);
4122 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4126 memcg_wb_domain_exit(memcg
);
4128 free_mem_cgroup_per_node_info(memcg
, node
);
4129 free_percpu(memcg
->stat
);
4133 static struct mem_cgroup
*mem_cgroup_alloc(void)
4135 struct mem_cgroup
*memcg
;
4139 size
= sizeof(struct mem_cgroup
);
4140 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4142 memcg
= kzalloc(size
, GFP_KERNEL
);
4146 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
4147 1, MEM_CGROUP_ID_MAX
,
4149 if (memcg
->id
.id
< 0)
4152 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4157 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
4160 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4163 INIT_WORK(&memcg
->high_work
, high_work_func
);
4164 memcg
->last_scanned_node
= MAX_NUMNODES
;
4165 INIT_LIST_HEAD(&memcg
->oom_notify
);
4166 mutex_init(&memcg
->thresholds_lock
);
4167 spin_lock_init(&memcg
->move_lock
);
4168 vmpressure_init(&memcg
->vmpressure
);
4169 INIT_LIST_HEAD(&memcg
->event_list
);
4170 spin_lock_init(&memcg
->event_list_lock
);
4171 memcg
->socket_pressure
= jiffies
;
4173 memcg
->kmemcg_id
= -1;
4175 #ifdef CONFIG_CGROUP_WRITEBACK
4176 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4178 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
4181 if (memcg
->id
.id
> 0)
4182 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4183 mem_cgroup_free(memcg
);
4187 static struct cgroup_subsys_state
* __ref
4188 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4190 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
4191 struct mem_cgroup
*memcg
;
4192 long error
= -ENOMEM
;
4194 memcg
= mem_cgroup_alloc();
4196 return ERR_PTR(error
);
4198 memcg
->high
= PAGE_COUNTER_MAX
;
4199 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4201 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4202 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4204 if (parent
&& parent
->use_hierarchy
) {
4205 memcg
->use_hierarchy
= true;
4206 page_counter_init(&memcg
->memory
, &parent
->memory
);
4207 page_counter_init(&memcg
->swap
, &parent
->swap
);
4208 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4209 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4210 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
4212 page_counter_init(&memcg
->memory
, NULL
);
4213 page_counter_init(&memcg
->swap
, NULL
);
4214 page_counter_init(&memcg
->memsw
, NULL
);
4215 page_counter_init(&memcg
->kmem
, NULL
);
4216 page_counter_init(&memcg
->tcpmem
, NULL
);
4218 * Deeper hierachy with use_hierarchy == false doesn't make
4219 * much sense so let cgroup subsystem know about this
4220 * unfortunate state in our controller.
4222 if (parent
!= root_mem_cgroup
)
4223 memory_cgrp_subsys
.broken_hierarchy
= true;
4226 /* The following stuff does not apply to the root */
4228 root_mem_cgroup
= memcg
;
4232 error
= memcg_online_kmem(memcg
);
4236 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4237 static_branch_inc(&memcg_sockets_enabled_key
);
4241 mem_cgroup_free(memcg
);
4242 return ERR_PTR(-ENOMEM
);
4245 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4247 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4249 /* Online state pins memcg ID, memcg ID pins CSS */
4250 atomic_set(&memcg
->id
.ref
, 1);
4255 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4257 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4258 struct mem_cgroup_event
*event
, *tmp
;
4261 * Unregister events and notify userspace.
4262 * Notify userspace about cgroup removing only after rmdir of cgroup
4263 * directory to avoid race between userspace and kernelspace.
4265 spin_lock(&memcg
->event_list_lock
);
4266 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4267 list_del_init(&event
->list
);
4268 schedule_work(&event
->remove
);
4270 spin_unlock(&memcg
->event_list_lock
);
4272 memcg_offline_kmem(memcg
);
4273 wb_memcg_offline(memcg
);
4275 mem_cgroup_id_put(memcg
);
4278 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4280 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4282 invalidate_reclaim_iterators(memcg
);
4285 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4287 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4289 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4290 static_branch_dec(&memcg_sockets_enabled_key
);
4292 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
4293 static_branch_dec(&memcg_sockets_enabled_key
);
4295 vmpressure_cleanup(&memcg
->vmpressure
);
4296 cancel_work_sync(&memcg
->high_work
);
4297 mem_cgroup_remove_from_trees(memcg
);
4298 memcg_free_kmem(memcg
);
4299 mem_cgroup_free(memcg
);
4303 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4304 * @css: the target css
4306 * Reset the states of the mem_cgroup associated with @css. This is
4307 * invoked when the userland requests disabling on the default hierarchy
4308 * but the memcg is pinned through dependency. The memcg should stop
4309 * applying policies and should revert to the vanilla state as it may be
4310 * made visible again.
4312 * The current implementation only resets the essential configurations.
4313 * This needs to be expanded to cover all the visible parts.
4315 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4317 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4319 page_counter_limit(&memcg
->memory
, PAGE_COUNTER_MAX
);
4320 page_counter_limit(&memcg
->swap
, PAGE_COUNTER_MAX
);
4321 page_counter_limit(&memcg
->memsw
, PAGE_COUNTER_MAX
);
4322 page_counter_limit(&memcg
->kmem
, PAGE_COUNTER_MAX
);
4323 page_counter_limit(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
4325 memcg
->high
= PAGE_COUNTER_MAX
;
4326 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4327 memcg_wb_domain_size_changed(memcg
);
4331 /* Handlers for move charge at task migration. */
4332 static int mem_cgroup_do_precharge(unsigned long count
)
4336 /* Try a single bulk charge without reclaim first, kswapd may wake */
4337 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4339 mc
.precharge
+= count
;
4343 /* Try charges one by one with reclaim */
4345 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4359 enum mc_target_type
{
4365 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4366 unsigned long addr
, pte_t ptent
)
4368 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4370 if (!page
|| !page_mapped(page
))
4372 if (PageAnon(page
)) {
4373 if (!(mc
.flags
& MOVE_ANON
))
4376 if (!(mc
.flags
& MOVE_FILE
))
4379 if (!get_page_unless_zero(page
))
4386 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4387 pte_t ptent
, swp_entry_t
*entry
)
4389 struct page
*page
= NULL
;
4390 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4392 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4395 * Because lookup_swap_cache() updates some statistics counter,
4396 * we call find_get_page() with swapper_space directly.
4398 page
= find_get_page(swap_address_space(ent
), swp_offset(ent
));
4399 if (do_memsw_account())
4400 entry
->val
= ent
.val
;
4405 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4406 pte_t ptent
, swp_entry_t
*entry
)
4412 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4413 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4415 struct page
*page
= NULL
;
4416 struct address_space
*mapping
;
4419 if (!vma
->vm_file
) /* anonymous vma */
4421 if (!(mc
.flags
& MOVE_FILE
))
4424 mapping
= vma
->vm_file
->f_mapping
;
4425 pgoff
= linear_page_index(vma
, addr
);
4427 /* page is moved even if it's not RSS of this task(page-faulted). */
4429 /* shmem/tmpfs may report page out on swap: account for that too. */
4430 if (shmem_mapping(mapping
)) {
4431 page
= find_get_entry(mapping
, pgoff
);
4432 if (radix_tree_exceptional_entry(page
)) {
4433 swp_entry_t swp
= radix_to_swp_entry(page
);
4434 if (do_memsw_account())
4436 page
= find_get_page(swap_address_space(swp
),
4440 page
= find_get_page(mapping
, pgoff
);
4442 page
= find_get_page(mapping
, pgoff
);
4448 * mem_cgroup_move_account - move account of the page
4450 * @compound: charge the page as compound or small page
4451 * @from: mem_cgroup which the page is moved from.
4452 * @to: mem_cgroup which the page is moved to. @from != @to.
4454 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4456 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4459 static int mem_cgroup_move_account(struct page
*page
,
4461 struct mem_cgroup
*from
,
4462 struct mem_cgroup
*to
)
4464 unsigned long flags
;
4465 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
4469 VM_BUG_ON(from
== to
);
4470 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4471 VM_BUG_ON(compound
&& !PageTransHuge(page
));
4474 * Prevent mem_cgroup_migrate() from looking at
4475 * page->mem_cgroup of its source page while we change it.
4478 if (!trylock_page(page
))
4482 if (page
->mem_cgroup
!= from
)
4485 anon
= PageAnon(page
);
4487 spin_lock_irqsave(&from
->move_lock
, flags
);
4489 if (!anon
&& page_mapped(page
)) {
4490 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4492 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4497 * move_lock grabbed above and caller set from->moving_account, so
4498 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4499 * So mapping should be stable for dirty pages.
4501 if (!anon
&& PageDirty(page
)) {
4502 struct address_space
*mapping
= page_mapping(page
);
4504 if (mapping_cap_account_dirty(mapping
)) {
4505 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4507 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4512 if (PageWriteback(page
)) {
4513 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4515 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4520 * It is safe to change page->mem_cgroup here because the page
4521 * is referenced, charged, and isolated - we can't race with
4522 * uncharging, charging, migration, or LRU putback.
4525 /* caller should have done css_get */
4526 page
->mem_cgroup
= to
;
4527 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4531 local_irq_disable();
4532 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
4533 memcg_check_events(to
, page
);
4534 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
4535 memcg_check_events(from
, page
);
4544 * get_mctgt_type - get target type of moving charge
4545 * @vma: the vma the pte to be checked belongs
4546 * @addr: the address corresponding to the pte to be checked
4547 * @ptent: the pte to be checked
4548 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4551 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4552 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4553 * move charge. if @target is not NULL, the page is stored in target->page
4554 * with extra refcnt got(Callers should handle it).
4555 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4556 * target for charge migration. if @target is not NULL, the entry is stored
4559 * Called with pte lock held.
4562 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4563 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4565 struct page
*page
= NULL
;
4566 enum mc_target_type ret
= MC_TARGET_NONE
;
4567 swp_entry_t ent
= { .val
= 0 };
4569 if (pte_present(ptent
))
4570 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4571 else if (is_swap_pte(ptent
))
4572 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
4573 else if (pte_none(ptent
))
4574 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4576 if (!page
&& !ent
.val
)
4580 * Do only loose check w/o serialization.
4581 * mem_cgroup_move_account() checks the page is valid or
4582 * not under LRU exclusion.
4584 if (page
->mem_cgroup
== mc
.from
) {
4585 ret
= MC_TARGET_PAGE
;
4587 target
->page
= page
;
4589 if (!ret
|| !target
)
4592 /* There is a swap entry and a page doesn't exist or isn't charged */
4593 if (ent
.val
&& !ret
&&
4594 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4595 ret
= MC_TARGET_SWAP
;
4602 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4604 * We don't consider swapping or file mapped pages because THP does not
4605 * support them for now.
4606 * Caller should make sure that pmd_trans_huge(pmd) is true.
4608 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4609 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4611 struct page
*page
= NULL
;
4612 enum mc_target_type ret
= MC_TARGET_NONE
;
4614 page
= pmd_page(pmd
);
4615 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4616 if (!(mc
.flags
& MOVE_ANON
))
4618 if (page
->mem_cgroup
== mc
.from
) {
4619 ret
= MC_TARGET_PAGE
;
4622 target
->page
= page
;
4628 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4629 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4631 return MC_TARGET_NONE
;
4635 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4636 unsigned long addr
, unsigned long end
,
4637 struct mm_walk
*walk
)
4639 struct vm_area_struct
*vma
= walk
->vma
;
4643 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4645 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4646 mc
.precharge
+= HPAGE_PMD_NR
;
4651 if (pmd_trans_unstable(pmd
))
4653 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4654 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4655 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4656 mc
.precharge
++; /* increment precharge temporarily */
4657 pte_unmap_unlock(pte
- 1, ptl
);
4663 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4665 unsigned long precharge
;
4667 struct mm_walk mem_cgroup_count_precharge_walk
= {
4668 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4671 down_read(&mm
->mmap_sem
);
4672 walk_page_range(0, mm
->highest_vm_end
,
4673 &mem_cgroup_count_precharge_walk
);
4674 up_read(&mm
->mmap_sem
);
4676 precharge
= mc
.precharge
;
4682 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4684 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4686 VM_BUG_ON(mc
.moving_task
);
4687 mc
.moving_task
= current
;
4688 return mem_cgroup_do_precharge(precharge
);
4691 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4692 static void __mem_cgroup_clear_mc(void)
4694 struct mem_cgroup
*from
= mc
.from
;
4695 struct mem_cgroup
*to
= mc
.to
;
4697 /* we must uncharge all the leftover precharges from mc.to */
4699 cancel_charge(mc
.to
, mc
.precharge
);
4703 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4704 * we must uncharge here.
4706 if (mc
.moved_charge
) {
4707 cancel_charge(mc
.from
, mc
.moved_charge
);
4708 mc
.moved_charge
= 0;
4710 /* we must fixup refcnts and charges */
4711 if (mc
.moved_swap
) {
4712 /* uncharge swap account from the old cgroup */
4713 if (!mem_cgroup_is_root(mc
.from
))
4714 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4716 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
4719 * we charged both to->memory and to->memsw, so we
4720 * should uncharge to->memory.
4722 if (!mem_cgroup_is_root(mc
.to
))
4723 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4725 mem_cgroup_id_get_many(mc
.to
, mc
.moved_swap
);
4726 css_put_many(&mc
.to
->css
, mc
.moved_swap
);
4730 memcg_oom_recover(from
);
4731 memcg_oom_recover(to
);
4732 wake_up_all(&mc
.waitq
);
4735 static void mem_cgroup_clear_mc(void)
4737 struct mm_struct
*mm
= mc
.mm
;
4740 * we must clear moving_task before waking up waiters at the end of
4743 mc
.moving_task
= NULL
;
4744 __mem_cgroup_clear_mc();
4745 spin_lock(&mc
.lock
);
4749 spin_unlock(&mc
.lock
);
4754 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4756 struct cgroup_subsys_state
*css
;
4757 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
4758 struct mem_cgroup
*from
;
4759 struct task_struct
*leader
, *p
;
4760 struct mm_struct
*mm
;
4761 unsigned long move_flags
;
4764 /* charge immigration isn't supported on the default hierarchy */
4765 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4769 * Multi-process migrations only happen on the default hierarchy
4770 * where charge immigration is not used. Perform charge
4771 * immigration if @tset contains a leader and whine if there are
4775 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
4778 memcg
= mem_cgroup_from_css(css
);
4784 * We are now commited to this value whatever it is. Changes in this
4785 * tunable will only affect upcoming migrations, not the current one.
4786 * So we need to save it, and keep it going.
4788 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
4792 from
= mem_cgroup_from_task(p
);
4794 VM_BUG_ON(from
== memcg
);
4796 mm
= get_task_mm(p
);
4799 /* We move charges only when we move a owner of the mm */
4800 if (mm
->owner
== p
) {
4803 VM_BUG_ON(mc
.precharge
);
4804 VM_BUG_ON(mc
.moved_charge
);
4805 VM_BUG_ON(mc
.moved_swap
);
4807 spin_lock(&mc
.lock
);
4811 mc
.flags
= move_flags
;
4812 spin_unlock(&mc
.lock
);
4813 /* We set mc.moving_task later */
4815 ret
= mem_cgroup_precharge_mc(mm
);
4817 mem_cgroup_clear_mc();
4824 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4827 mem_cgroup_clear_mc();
4830 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4831 unsigned long addr
, unsigned long end
,
4832 struct mm_walk
*walk
)
4835 struct vm_area_struct
*vma
= walk
->vma
;
4838 enum mc_target_type target_type
;
4839 union mc_target target
;
4842 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4844 if (mc
.precharge
< HPAGE_PMD_NR
) {
4848 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
4849 if (target_type
== MC_TARGET_PAGE
) {
4851 if (!isolate_lru_page(page
)) {
4852 if (!mem_cgroup_move_account(page
, true,
4854 mc
.precharge
-= HPAGE_PMD_NR
;
4855 mc
.moved_charge
+= HPAGE_PMD_NR
;
4857 putback_lru_page(page
);
4865 if (pmd_trans_unstable(pmd
))
4868 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4869 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4870 pte_t ptent
= *(pte
++);
4876 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
4877 case MC_TARGET_PAGE
:
4880 * We can have a part of the split pmd here. Moving it
4881 * can be done but it would be too convoluted so simply
4882 * ignore such a partial THP and keep it in original
4883 * memcg. There should be somebody mapping the head.
4885 if (PageTransCompound(page
))
4887 if (isolate_lru_page(page
))
4889 if (!mem_cgroup_move_account(page
, false,
4892 /* we uncharge from mc.from later. */
4895 putback_lru_page(page
);
4896 put
: /* get_mctgt_type() gets the page */
4899 case MC_TARGET_SWAP
:
4901 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
4903 /* we fixup refcnts and charges later. */
4911 pte_unmap_unlock(pte
- 1, ptl
);
4916 * We have consumed all precharges we got in can_attach().
4917 * We try charge one by one, but don't do any additional
4918 * charges to mc.to if we have failed in charge once in attach()
4921 ret
= mem_cgroup_do_precharge(1);
4929 static void mem_cgroup_move_charge(void)
4931 struct mm_walk mem_cgroup_move_charge_walk
= {
4932 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4936 lru_add_drain_all();
4938 * Signal lock_page_memcg() to take the memcg's move_lock
4939 * while we're moving its pages to another memcg. Then wait
4940 * for already started RCU-only updates to finish.
4942 atomic_inc(&mc
.from
->moving_account
);
4945 if (unlikely(!down_read_trylock(&mc
.mm
->mmap_sem
))) {
4947 * Someone who are holding the mmap_sem might be waiting in
4948 * waitq. So we cancel all extra charges, wake up all waiters,
4949 * and retry. Because we cancel precharges, we might not be able
4950 * to move enough charges, but moving charge is a best-effort
4951 * feature anyway, so it wouldn't be a big problem.
4953 __mem_cgroup_clear_mc();
4958 * When we have consumed all precharges and failed in doing
4959 * additional charge, the page walk just aborts.
4961 walk_page_range(0, mc
.mm
->highest_vm_end
, &mem_cgroup_move_charge_walk
);
4963 up_read(&mc
.mm
->mmap_sem
);
4964 atomic_dec(&mc
.from
->moving_account
);
4967 static void mem_cgroup_move_task(void)
4970 mem_cgroup_move_charge();
4971 mem_cgroup_clear_mc();
4974 #else /* !CONFIG_MMU */
4975 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4979 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4982 static void mem_cgroup_move_task(void)
4988 * Cgroup retains root cgroups across [un]mount cycles making it necessary
4989 * to verify whether we're attached to the default hierarchy on each mount
4992 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
4995 * use_hierarchy is forced on the default hierarchy. cgroup core
4996 * guarantees that @root doesn't have any children, so turning it
4997 * on for the root memcg is enough.
4999 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5000 root_mem_cgroup
->use_hierarchy
= true;
5002 root_mem_cgroup
->use_hierarchy
= false;
5005 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5008 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5010 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
5013 static int memory_low_show(struct seq_file
*m
, void *v
)
5015 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5016 unsigned long low
= READ_ONCE(memcg
->low
);
5018 if (low
== PAGE_COUNTER_MAX
)
5019 seq_puts(m
, "max\n");
5021 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
5026 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5027 char *buf
, size_t nbytes
, loff_t off
)
5029 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5033 buf
= strstrip(buf
);
5034 err
= page_counter_memparse(buf
, "max", &low
);
5043 static int memory_high_show(struct seq_file
*m
, void *v
)
5045 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5046 unsigned long high
= READ_ONCE(memcg
->high
);
5048 if (high
== PAGE_COUNTER_MAX
)
5049 seq_puts(m
, "max\n");
5051 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
5056 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5057 char *buf
, size_t nbytes
, loff_t off
)
5059 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5060 unsigned long nr_pages
;
5064 buf
= strstrip(buf
);
5065 err
= page_counter_memparse(buf
, "max", &high
);
5071 nr_pages
= page_counter_read(&memcg
->memory
);
5072 if (nr_pages
> high
)
5073 try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
5076 memcg_wb_domain_size_changed(memcg
);
5080 static int memory_max_show(struct seq_file
*m
, void *v
)
5082 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5083 unsigned long max
= READ_ONCE(memcg
->memory
.limit
);
5085 if (max
== PAGE_COUNTER_MAX
)
5086 seq_puts(m
, "max\n");
5088 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5093 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5094 char *buf
, size_t nbytes
, loff_t off
)
5096 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5097 unsigned int nr_reclaims
= MEM_CGROUP_RECLAIM_RETRIES
;
5098 bool drained
= false;
5102 buf
= strstrip(buf
);
5103 err
= page_counter_memparse(buf
, "max", &max
);
5107 xchg(&memcg
->memory
.limit
, max
);
5110 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
5112 if (nr_pages
<= max
)
5115 if (signal_pending(current
)) {
5121 drain_all_stock(memcg
);
5127 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
5133 mem_cgroup_events(memcg
, MEMCG_OOM
, 1);
5134 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
5138 memcg_wb_domain_size_changed(memcg
);
5142 static int memory_events_show(struct seq_file
*m
, void *v
)
5144 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5146 seq_printf(m
, "low %lu\n", mem_cgroup_read_events(memcg
, MEMCG_LOW
));
5147 seq_printf(m
, "high %lu\n", mem_cgroup_read_events(memcg
, MEMCG_HIGH
));
5148 seq_printf(m
, "max %lu\n", mem_cgroup_read_events(memcg
, MEMCG_MAX
));
5149 seq_printf(m
, "oom %lu\n", mem_cgroup_read_events(memcg
, MEMCG_OOM
));
5154 static int memory_stat_show(struct seq_file
*m
, void *v
)
5156 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5157 unsigned long stat
[MEMCG_NR_STAT
];
5158 unsigned long events
[MEMCG_NR_EVENTS
];
5162 * Provide statistics on the state of the memory subsystem as
5163 * well as cumulative event counters that show past behavior.
5165 * This list is ordered following a combination of these gradients:
5166 * 1) generic big picture -> specifics and details
5167 * 2) reflecting userspace activity -> reflecting kernel heuristics
5169 * Current memory state:
5172 tree_stat(memcg
, stat
);
5173 tree_events(memcg
, events
);
5175 seq_printf(m
, "anon %llu\n",
5176 (u64
)stat
[MEM_CGROUP_STAT_RSS
] * PAGE_SIZE
);
5177 seq_printf(m
, "file %llu\n",
5178 (u64
)stat
[MEM_CGROUP_STAT_CACHE
] * PAGE_SIZE
);
5179 seq_printf(m
, "kernel_stack %llu\n",
5180 (u64
)stat
[MEMCG_KERNEL_STACK_KB
] * 1024);
5181 seq_printf(m
, "slab %llu\n",
5182 (u64
)(stat
[MEMCG_SLAB_RECLAIMABLE
] +
5183 stat
[MEMCG_SLAB_UNRECLAIMABLE
]) * PAGE_SIZE
);
5184 seq_printf(m
, "sock %llu\n",
5185 (u64
)stat
[MEMCG_SOCK
] * PAGE_SIZE
);
5187 seq_printf(m
, "file_mapped %llu\n",
5188 (u64
)stat
[MEM_CGROUP_STAT_FILE_MAPPED
] * PAGE_SIZE
);
5189 seq_printf(m
, "file_dirty %llu\n",
5190 (u64
)stat
[MEM_CGROUP_STAT_DIRTY
] * PAGE_SIZE
);
5191 seq_printf(m
, "file_writeback %llu\n",
5192 (u64
)stat
[MEM_CGROUP_STAT_WRITEBACK
] * PAGE_SIZE
);
5194 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5195 struct mem_cgroup
*mi
;
5196 unsigned long val
= 0;
5198 for_each_mem_cgroup_tree(mi
, memcg
)
5199 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
));
5200 seq_printf(m
, "%s %llu\n",
5201 mem_cgroup_lru_names
[i
], (u64
)val
* PAGE_SIZE
);
5204 seq_printf(m
, "slab_reclaimable %llu\n",
5205 (u64
)stat
[MEMCG_SLAB_RECLAIMABLE
] * PAGE_SIZE
);
5206 seq_printf(m
, "slab_unreclaimable %llu\n",
5207 (u64
)stat
[MEMCG_SLAB_UNRECLAIMABLE
] * PAGE_SIZE
);
5209 /* Accumulated memory events */
5211 seq_printf(m
, "pgfault %lu\n",
5212 events
[MEM_CGROUP_EVENTS_PGFAULT
]);
5213 seq_printf(m
, "pgmajfault %lu\n",
5214 events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
5219 static struct cftype memory_files
[] = {
5222 .flags
= CFTYPE_NOT_ON_ROOT
,
5223 .read_u64
= memory_current_read
,
5227 .flags
= CFTYPE_NOT_ON_ROOT
,
5228 .seq_show
= memory_low_show
,
5229 .write
= memory_low_write
,
5233 .flags
= CFTYPE_NOT_ON_ROOT
,
5234 .seq_show
= memory_high_show
,
5235 .write
= memory_high_write
,
5239 .flags
= CFTYPE_NOT_ON_ROOT
,
5240 .seq_show
= memory_max_show
,
5241 .write
= memory_max_write
,
5245 .flags
= CFTYPE_NOT_ON_ROOT
,
5246 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
5247 .seq_show
= memory_events_show
,
5251 .flags
= CFTYPE_NOT_ON_ROOT
,
5252 .seq_show
= memory_stat_show
,
5257 struct cgroup_subsys memory_cgrp_subsys
= {
5258 .css_alloc
= mem_cgroup_css_alloc
,
5259 .css_online
= mem_cgroup_css_online
,
5260 .css_offline
= mem_cgroup_css_offline
,
5261 .css_released
= mem_cgroup_css_released
,
5262 .css_free
= mem_cgroup_css_free
,
5263 .css_reset
= mem_cgroup_css_reset
,
5264 .can_attach
= mem_cgroup_can_attach
,
5265 .cancel_attach
= mem_cgroup_cancel_attach
,
5266 .post_attach
= mem_cgroup_move_task
,
5267 .bind
= mem_cgroup_bind
,
5268 .dfl_cftypes
= memory_files
,
5269 .legacy_cftypes
= mem_cgroup_legacy_files
,
5274 * mem_cgroup_low - check if memory consumption is below the normal range
5275 * @root: the highest ancestor to consider
5276 * @memcg: the memory cgroup to check
5278 * Returns %true if memory consumption of @memcg, and that of all
5279 * configurable ancestors up to @root, is below the normal range.
5281 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5283 if (mem_cgroup_disabled())
5287 * The toplevel group doesn't have a configurable range, so
5288 * it's never low when looked at directly, and it is not
5289 * considered an ancestor when assessing the hierarchy.
5292 if (memcg
== root_mem_cgroup
)
5295 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5298 while (memcg
!= root
) {
5299 memcg
= parent_mem_cgroup(memcg
);
5301 if (memcg
== root_mem_cgroup
)
5304 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5311 * mem_cgroup_try_charge - try charging a page
5312 * @page: page to charge
5313 * @mm: mm context of the victim
5314 * @gfp_mask: reclaim mode
5315 * @memcgp: charged memcg return
5316 * @compound: charge the page as compound or small page
5318 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5319 * pages according to @gfp_mask if necessary.
5321 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5322 * Otherwise, an error code is returned.
5324 * After page->mapping has been set up, the caller must finalize the
5325 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5326 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5328 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5329 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
5332 struct mem_cgroup
*memcg
= NULL
;
5333 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5336 if (mem_cgroup_disabled())
5339 if (PageSwapCache(page
)) {
5341 * Every swap fault against a single page tries to charge the
5342 * page, bail as early as possible. shmem_unuse() encounters
5343 * already charged pages, too. The USED bit is protected by
5344 * the page lock, which serializes swap cache removal, which
5345 * in turn serializes uncharging.
5347 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5348 if (page
->mem_cgroup
)
5351 if (do_swap_account
) {
5352 swp_entry_t ent
= { .val
= page_private(page
), };
5353 unsigned short id
= lookup_swap_cgroup_id(ent
);
5356 memcg
= mem_cgroup_from_id(id
);
5357 if (memcg
&& !css_tryget_online(&memcg
->css
))
5364 memcg
= get_mem_cgroup_from_mm(mm
);
5366 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5368 css_put(&memcg
->css
);
5375 * mem_cgroup_commit_charge - commit a page charge
5376 * @page: page to charge
5377 * @memcg: memcg to charge the page to
5378 * @lrucare: page might be on LRU already
5379 * @compound: charge the page as compound or small page
5381 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5382 * after page->mapping has been set up. This must happen atomically
5383 * as part of the page instantiation, i.e. under the page table lock
5384 * for anonymous pages, under the page lock for page and swap cache.
5386 * In addition, the page must not be on the LRU during the commit, to
5387 * prevent racing with task migration. If it might be, use @lrucare.
5389 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5391 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5392 bool lrucare
, bool compound
)
5394 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5396 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5397 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5399 if (mem_cgroup_disabled())
5402 * Swap faults will attempt to charge the same page multiple
5403 * times. But reuse_swap_page() might have removed the page
5404 * from swapcache already, so we can't check PageSwapCache().
5409 commit_charge(page
, memcg
, lrucare
);
5411 local_irq_disable();
5412 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
5413 memcg_check_events(memcg
, page
);
5416 if (do_memsw_account() && PageSwapCache(page
)) {
5417 swp_entry_t entry
= { .val
= page_private(page
) };
5419 * The swap entry might not get freed for a long time,
5420 * let's not wait for it. The page already received a
5421 * memory+swap charge, drop the swap entry duplicate.
5423 mem_cgroup_uncharge_swap(entry
);
5428 * mem_cgroup_cancel_charge - cancel a page charge
5429 * @page: page to charge
5430 * @memcg: memcg to charge the page to
5431 * @compound: charge the page as compound or small page
5433 * Cancel a charge transaction started by mem_cgroup_try_charge().
5435 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5438 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
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 cancel_charge(memcg
, nr_pages
);
5453 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5454 unsigned long nr_anon
, unsigned long nr_file
,
5455 unsigned long nr_huge
, unsigned long nr_kmem
,
5456 struct page
*dummy_page
)
5458 unsigned long nr_pages
= nr_anon
+ nr_file
+ nr_kmem
;
5459 unsigned long flags
;
5461 if (!mem_cgroup_is_root(memcg
)) {
5462 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5463 if (do_memsw_account())
5464 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5465 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && nr_kmem
)
5466 page_counter_uncharge(&memcg
->kmem
, nr_kmem
);
5467 memcg_oom_recover(memcg
);
5470 local_irq_save(flags
);
5471 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5472 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5473 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5474 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5475 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5476 memcg_check_events(memcg
, dummy_page
);
5477 local_irq_restore(flags
);
5479 if (!mem_cgroup_is_root(memcg
))
5480 css_put_many(&memcg
->css
, nr_pages
);
5483 static void uncharge_list(struct list_head
*page_list
)
5485 struct mem_cgroup
*memcg
= NULL
;
5486 unsigned long nr_anon
= 0;
5487 unsigned long nr_file
= 0;
5488 unsigned long nr_huge
= 0;
5489 unsigned long nr_kmem
= 0;
5490 unsigned long pgpgout
= 0;
5491 struct list_head
*next
;
5495 * Note that the list can be a single page->lru; hence the
5496 * do-while loop instead of a simple list_for_each_entry().
5498 next
= page_list
->next
;
5500 page
= list_entry(next
, struct page
, lru
);
5501 next
= page
->lru
.next
;
5503 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5504 VM_BUG_ON_PAGE(page_count(page
), page
);
5506 if (!page
->mem_cgroup
)
5510 * Nobody should be changing or seriously looking at
5511 * page->mem_cgroup at this point, we have fully
5512 * exclusive access to the page.
5515 if (memcg
!= page
->mem_cgroup
) {
5517 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5518 nr_huge
, nr_kmem
, page
);
5519 pgpgout
= nr_anon
= nr_file
=
5520 nr_huge
= nr_kmem
= 0;
5522 memcg
= page
->mem_cgroup
;
5525 if (!PageKmemcg(page
)) {
5526 unsigned int nr_pages
= 1;
5528 if (PageTransHuge(page
)) {
5529 nr_pages
<<= compound_order(page
);
5530 nr_huge
+= nr_pages
;
5533 nr_anon
+= nr_pages
;
5535 nr_file
+= nr_pages
;
5538 nr_kmem
+= 1 << compound_order(page
);
5539 __ClearPageKmemcg(page
);
5542 page
->mem_cgroup
= NULL
;
5543 } while (next
!= page_list
);
5546 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5547 nr_huge
, nr_kmem
, page
);
5551 * mem_cgroup_uncharge - uncharge a page
5552 * @page: page to uncharge
5554 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5555 * mem_cgroup_commit_charge().
5557 void mem_cgroup_uncharge(struct page
*page
)
5559 if (mem_cgroup_disabled())
5562 /* Don't touch page->lru of any random page, pre-check: */
5563 if (!page
->mem_cgroup
)
5566 INIT_LIST_HEAD(&page
->lru
);
5567 uncharge_list(&page
->lru
);
5571 * mem_cgroup_uncharge_list - uncharge a list of page
5572 * @page_list: list of pages to uncharge
5574 * Uncharge a list of pages previously charged with
5575 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5577 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5579 if (mem_cgroup_disabled())
5582 if (!list_empty(page_list
))
5583 uncharge_list(page_list
);
5587 * mem_cgroup_migrate - charge a page's replacement
5588 * @oldpage: currently circulating page
5589 * @newpage: replacement page
5591 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5592 * be uncharged upon free.
5594 * Both pages must be locked, @newpage->mapping must be set up.
5596 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
5598 struct mem_cgroup
*memcg
;
5599 unsigned int nr_pages
;
5601 unsigned long flags
;
5603 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5604 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5605 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5606 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5609 if (mem_cgroup_disabled())
5612 /* Page cache replacement: new page already charged? */
5613 if (newpage
->mem_cgroup
)
5616 /* Swapcache readahead pages can get replaced before being charged */
5617 memcg
= oldpage
->mem_cgroup
;
5621 /* Force-charge the new page. The old one will be freed soon */
5622 compound
= PageTransHuge(newpage
);
5623 nr_pages
= compound
? hpage_nr_pages(newpage
) : 1;
5625 page_counter_charge(&memcg
->memory
, nr_pages
);
5626 if (do_memsw_account())
5627 page_counter_charge(&memcg
->memsw
, nr_pages
);
5628 css_get_many(&memcg
->css
, nr_pages
);
5630 commit_charge(newpage
, memcg
, false);
5632 local_irq_save(flags
);
5633 mem_cgroup_charge_statistics(memcg
, newpage
, compound
, nr_pages
);
5634 memcg_check_events(memcg
, newpage
);
5635 local_irq_restore(flags
);
5638 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
5639 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
5641 void sock_update_memcg(struct sock
*sk
)
5643 struct mem_cgroup
*memcg
;
5645 /* Socket cloning can throw us here with sk_cgrp already
5646 * filled. It won't however, necessarily happen from
5647 * process context. So the test for root memcg given
5648 * the current task's memcg won't help us in this case.
5650 * Respecting the original socket's memcg is a better
5651 * decision in this case.
5654 BUG_ON(mem_cgroup_is_root(sk
->sk_memcg
));
5655 css_get(&sk
->sk_memcg
->css
);
5660 memcg
= mem_cgroup_from_task(current
);
5661 if (memcg
== root_mem_cgroup
)
5663 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
5665 if (css_tryget_online(&memcg
->css
))
5666 sk
->sk_memcg
= memcg
;
5670 EXPORT_SYMBOL(sock_update_memcg
);
5672 void sock_release_memcg(struct sock
*sk
)
5674 WARN_ON(!sk
->sk_memcg
);
5675 css_put(&sk
->sk_memcg
->css
);
5679 * mem_cgroup_charge_skmem - charge socket memory
5680 * @memcg: memcg to charge
5681 * @nr_pages: number of pages to charge
5683 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5684 * @memcg's configured limit, %false if the charge had to be forced.
5686 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5688 gfp_t gfp_mask
= GFP_KERNEL
;
5690 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5691 struct page_counter
*fail
;
5693 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
5694 memcg
->tcpmem_pressure
= 0;
5697 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
5698 memcg
->tcpmem_pressure
= 1;
5702 /* Don't block in the packet receive path */
5704 gfp_mask
= GFP_NOWAIT
;
5706 this_cpu_add(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5708 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
5711 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
5716 * mem_cgroup_uncharge_skmem - uncharge socket memory
5717 * @memcg - memcg to uncharge
5718 * @nr_pages - number of pages to uncharge
5720 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5722 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5723 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
5727 this_cpu_sub(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5729 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5730 css_put_many(&memcg
->css
, nr_pages
);
5733 static int __init
cgroup_memory(char *s
)
5737 while ((token
= strsep(&s
, ",")) != NULL
) {
5740 if (!strcmp(token
, "nosocket"))
5741 cgroup_memory_nosocket
= true;
5742 if (!strcmp(token
, "nokmem"))
5743 cgroup_memory_nokmem
= true;
5747 __setup("cgroup.memory=", cgroup_memory
);
5750 * subsys_initcall() for memory controller.
5752 * Some parts like hotcpu_notifier() have to be initialized from this context
5753 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5754 * everything that doesn't depend on a specific mem_cgroup structure should
5755 * be initialized from here.
5757 static int __init
mem_cgroup_init(void)
5761 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5763 for_each_possible_cpu(cpu
)
5764 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5767 for_each_node(node
) {
5768 struct mem_cgroup_tree_per_node
*rtpn
;
5770 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5771 node_online(node
) ? node
: NUMA_NO_NODE
);
5773 rtpn
->rb_root
= RB_ROOT
;
5774 spin_lock_init(&rtpn
->lock
);
5775 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5780 subsys_initcall(mem_cgroup_init
);
5782 #ifdef CONFIG_MEMCG_SWAP
5783 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
5785 while (!atomic_inc_not_zero(&memcg
->id
.ref
)) {
5787 * The root cgroup cannot be destroyed, so it's refcount must
5790 if (WARN_ON_ONCE(memcg
== root_mem_cgroup
)) {
5794 memcg
= parent_mem_cgroup(memcg
);
5796 memcg
= root_mem_cgroup
;
5802 * mem_cgroup_swapout - transfer a memsw charge to swap
5803 * @page: page whose memsw charge to transfer
5804 * @entry: swap entry to move the charge to
5806 * Transfer the memsw charge of @page to @entry.
5808 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5810 struct mem_cgroup
*memcg
, *swap_memcg
;
5811 unsigned short oldid
;
5813 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5814 VM_BUG_ON_PAGE(page_count(page
), page
);
5816 if (!do_memsw_account())
5819 memcg
= page
->mem_cgroup
;
5821 /* Readahead page, never charged */
5826 * In case the memcg owning these pages has been offlined and doesn't
5827 * have an ID allocated to it anymore, charge the closest online
5828 * ancestor for the swap instead and transfer the memory+swap charge.
5830 swap_memcg
= mem_cgroup_id_get_online(memcg
);
5831 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
));
5832 VM_BUG_ON_PAGE(oldid
, page
);
5833 mem_cgroup_swap_statistics(swap_memcg
, true);
5835 page
->mem_cgroup
= NULL
;
5837 if (!mem_cgroup_is_root(memcg
))
5838 page_counter_uncharge(&memcg
->memory
, 1);
5840 if (memcg
!= swap_memcg
) {
5841 if (!mem_cgroup_is_root(swap_memcg
))
5842 page_counter_charge(&swap_memcg
->memsw
, 1);
5843 page_counter_uncharge(&memcg
->memsw
, 1);
5847 * Interrupts should be disabled here because the caller holds the
5848 * mapping->tree_lock lock which is taken with interrupts-off. It is
5849 * important here to have the interrupts disabled because it is the
5850 * only synchronisation we have for udpating the per-CPU variables.
5852 VM_BUG_ON(!irqs_disabled());
5853 mem_cgroup_charge_statistics(memcg
, page
, false, -1);
5854 memcg_check_events(memcg
, page
);
5856 if (!mem_cgroup_is_root(memcg
))
5857 css_put(&memcg
->css
);
5861 * mem_cgroup_try_charge_swap - try charging a swap entry
5862 * @page: page being added to swap
5863 * @entry: swap entry to charge
5865 * Try to charge @entry to the memcg that @page belongs to.
5867 * Returns 0 on success, -ENOMEM on failure.
5869 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
5871 struct mem_cgroup
*memcg
;
5872 struct page_counter
*counter
;
5873 unsigned short oldid
;
5875 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
5878 memcg
= page
->mem_cgroup
;
5880 /* Readahead page, never charged */
5884 memcg
= mem_cgroup_id_get_online(memcg
);
5886 if (!mem_cgroup_is_root(memcg
) &&
5887 !page_counter_try_charge(&memcg
->swap
, 1, &counter
)) {
5888 mem_cgroup_id_put(memcg
);
5892 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5893 VM_BUG_ON_PAGE(oldid
, page
);
5894 mem_cgroup_swap_statistics(memcg
, true);
5900 * mem_cgroup_uncharge_swap - uncharge a swap entry
5901 * @entry: swap entry to uncharge
5903 * Drop the swap charge associated with @entry.
5905 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5907 struct mem_cgroup
*memcg
;
5910 if (!do_swap_account
)
5913 id
= swap_cgroup_record(entry
, 0);
5915 memcg
= mem_cgroup_from_id(id
);
5917 if (!mem_cgroup_is_root(memcg
)) {
5918 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5919 page_counter_uncharge(&memcg
->swap
, 1);
5921 page_counter_uncharge(&memcg
->memsw
, 1);
5923 mem_cgroup_swap_statistics(memcg
, false);
5924 mem_cgroup_id_put(memcg
);
5929 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
5931 long nr_swap_pages
= get_nr_swap_pages();
5933 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5934 return nr_swap_pages
;
5935 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5936 nr_swap_pages
= min_t(long, nr_swap_pages
,
5937 READ_ONCE(memcg
->swap
.limit
) -
5938 page_counter_read(&memcg
->swap
));
5939 return nr_swap_pages
;
5942 bool mem_cgroup_swap_full(struct page
*page
)
5944 struct mem_cgroup
*memcg
;
5946 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5950 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5953 memcg
= page
->mem_cgroup
;
5957 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5958 if (page_counter_read(&memcg
->swap
) * 2 >= memcg
->swap
.limit
)
5964 /* for remember boot option*/
5965 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5966 static int really_do_swap_account __initdata
= 1;
5968 static int really_do_swap_account __initdata
;
5971 static int __init
enable_swap_account(char *s
)
5973 if (!strcmp(s
, "1"))
5974 really_do_swap_account
= 1;
5975 else if (!strcmp(s
, "0"))
5976 really_do_swap_account
= 0;
5979 __setup("swapaccount=", enable_swap_account
);
5981 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
5984 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5986 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
5989 static int swap_max_show(struct seq_file
*m
, void *v
)
5991 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5992 unsigned long max
= READ_ONCE(memcg
->swap
.limit
);
5994 if (max
== PAGE_COUNTER_MAX
)
5995 seq_puts(m
, "max\n");
5997 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
6002 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
6003 char *buf
, size_t nbytes
, loff_t off
)
6005 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6009 buf
= strstrip(buf
);
6010 err
= page_counter_memparse(buf
, "max", &max
);
6014 mutex_lock(&memcg_limit_mutex
);
6015 err
= page_counter_limit(&memcg
->swap
, max
);
6016 mutex_unlock(&memcg_limit_mutex
);
6023 static struct cftype swap_files
[] = {
6025 .name
= "swap.current",
6026 .flags
= CFTYPE_NOT_ON_ROOT
,
6027 .read_u64
= swap_current_read
,
6031 .flags
= CFTYPE_NOT_ON_ROOT
,
6032 .seq_show
= swap_max_show
,
6033 .write
= swap_max_write
,
6038 static struct cftype memsw_cgroup_files
[] = {
6040 .name
= "memsw.usage_in_bytes",
6041 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6042 .read_u64
= mem_cgroup_read_u64
,
6045 .name
= "memsw.max_usage_in_bytes",
6046 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6047 .write
= mem_cgroup_reset
,
6048 .read_u64
= mem_cgroup_read_u64
,
6051 .name
= "memsw.limit_in_bytes",
6052 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6053 .write
= mem_cgroup_write
,
6054 .read_u64
= mem_cgroup_read_u64
,
6057 .name
= "memsw.failcnt",
6058 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6059 .write
= mem_cgroup_reset
,
6060 .read_u64
= mem_cgroup_read_u64
,
6062 { }, /* terminate */
6065 static int __init
mem_cgroup_swap_init(void)
6067 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6068 do_swap_account
= 1;
6069 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
6071 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
6072 memsw_cgroup_files
));
6076 subsys_initcall(mem_cgroup_swap_init
);
6078 #endif /* CONFIG_MEMCG_SWAP */