1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
65 #include <linux/tracehook.h>
71 #include <asm/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
76 EXPORT_SYMBOL(memory_cgrp_subsys
);
78 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
80 #define MEM_CGROUP_RECLAIM_RETRIES 5
82 /* Socket memory accounting disabled? */
83 static bool cgroup_memory_nosocket
;
85 /* Kernel memory accounting disabled? */
86 static bool cgroup_memory_nokmem
;
88 /* Whether the swap controller is active */
89 #ifdef CONFIG_MEMCG_SWAP
90 int do_swap_account __read_mostly
;
92 #define do_swap_account 0
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys
) && do_swap_account
;
101 static const char * const mem_cgroup_stat_names
[] = {
111 static const char * const mem_cgroup_events_names
[] = {
118 static const char * const mem_cgroup_lru_names
[] = {
126 #define THRESHOLDS_EVENTS_TARGET 128
127 #define SOFTLIMIT_EVENTS_TARGET 1024
128 #define NUMAINFO_EVENTS_TARGET 1024
131 * Cgroups above their limits are maintained in a RB-Tree, independent of
132 * their hierarchy representation
135 struct mem_cgroup_tree_per_node
{
136 struct rb_root rb_root
;
140 struct mem_cgroup_tree
{
141 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
144 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
147 struct mem_cgroup_eventfd_list
{
148 struct list_head list
;
149 struct eventfd_ctx
*eventfd
;
153 * cgroup_event represents events which userspace want to receive.
155 struct mem_cgroup_event
{
157 * memcg which the event belongs to.
159 struct mem_cgroup
*memcg
;
161 * eventfd to signal userspace about the event.
163 struct eventfd_ctx
*eventfd
;
165 * Each of these stored in a list by the cgroup.
167 struct list_head list
;
169 * register_event() callback will be used to add new userspace
170 * waiter for changes related to this event. Use eventfd_signal()
171 * on eventfd to send notification to userspace.
173 int (*register_event
)(struct mem_cgroup
*memcg
,
174 struct eventfd_ctx
*eventfd
, const char *args
);
176 * unregister_event() callback will be called when userspace closes
177 * the eventfd or on cgroup removing. This callback must be set,
178 * if you want provide notification functionality.
180 void (*unregister_event
)(struct mem_cgroup
*memcg
,
181 struct eventfd_ctx
*eventfd
);
183 * All fields below needed to unregister event when
184 * userspace closes eventfd.
187 wait_queue_head_t
*wqh
;
189 struct work_struct remove
;
192 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
193 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
195 /* Stuffs for move charges at task migration. */
197 * Types of charges to be moved.
199 #define MOVE_ANON 0x1U
200 #define MOVE_FILE 0x2U
201 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
203 /* "mc" and its members are protected by cgroup_mutex */
204 static struct move_charge_struct
{
205 spinlock_t lock
; /* for from, to */
206 struct mm_struct
*mm
;
207 struct mem_cgroup
*from
;
208 struct mem_cgroup
*to
;
210 unsigned long precharge
;
211 unsigned long moved_charge
;
212 unsigned long moved_swap
;
213 struct task_struct
*moving_task
; /* a task moving charges */
214 wait_queue_head_t waitq
; /* a waitq for other context */
216 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
217 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
221 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
222 * limit reclaim to prevent infinite loops, if they ever occur.
224 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
225 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
228 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
229 MEM_CGROUP_CHARGE_TYPE_ANON
,
230 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
231 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
235 /* for encoding cft->private value on file */
244 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
245 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
246 #define MEMFILE_ATTR(val) ((val) & 0xffff)
247 /* Used for OOM nofiier */
248 #define OOM_CONTROL (0)
250 /* Some nice accessors for the vmpressure. */
251 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
254 memcg
= root_mem_cgroup
;
255 return &memcg
->vmpressure
;
258 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
260 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
263 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
265 return (memcg
== root_mem_cgroup
);
270 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
271 * The main reason for not using cgroup id for this:
272 * this works better in sparse environments, where we have a lot of memcgs,
273 * but only a few kmem-limited. Or also, if we have, for instance, 200
274 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
275 * 200 entry array for that.
277 * The current size of the caches array is stored in memcg_nr_cache_ids. It
278 * will double each time we have to increase it.
280 static DEFINE_IDA(memcg_cache_ida
);
281 int memcg_nr_cache_ids
;
283 /* Protects memcg_nr_cache_ids */
284 static DECLARE_RWSEM(memcg_cache_ids_sem
);
286 void memcg_get_cache_ids(void)
288 down_read(&memcg_cache_ids_sem
);
291 void memcg_put_cache_ids(void)
293 up_read(&memcg_cache_ids_sem
);
297 * MIN_SIZE is different than 1, because we would like to avoid going through
298 * the alloc/free process all the time. In a small machine, 4 kmem-limited
299 * cgroups is a reasonable guess. In the future, it could be a parameter or
300 * tunable, but that is strictly not necessary.
302 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
303 * this constant directly from cgroup, but it is understandable that this is
304 * better kept as an internal representation in cgroup.c. In any case, the
305 * cgrp_id space is not getting any smaller, and we don't have to necessarily
306 * increase ours as well if it increases.
308 #define MEMCG_CACHES_MIN_SIZE 4
309 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
312 * A lot of the calls to the cache allocation functions are expected to be
313 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
314 * conditional to this static branch, we'll have to allow modules that does
315 * kmem_cache_alloc and the such to see this symbol as well
317 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
318 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
320 #endif /* !CONFIG_SLOB */
323 * mem_cgroup_css_from_page - css of the memcg associated with a page
324 * @page: page of interest
326 * If memcg is bound to the default hierarchy, css of the memcg associated
327 * with @page is returned. The returned css remains associated with @page
328 * until it is released.
330 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
333 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
335 struct mem_cgroup
*memcg
;
337 memcg
= page
->mem_cgroup
;
339 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
340 memcg
= root_mem_cgroup
;
346 * page_cgroup_ino - return inode number of the memcg a page is charged to
349 * Look up the closest online ancestor of the memory cgroup @page is charged to
350 * and return its inode number or 0 if @page is not charged to any cgroup. It
351 * is safe to call this function without holding a reference to @page.
353 * Note, this function is inherently racy, because there is nothing to prevent
354 * the cgroup inode from getting torn down and potentially reallocated a moment
355 * after page_cgroup_ino() returns, so it only should be used by callers that
356 * do not care (such as procfs interfaces).
358 ino_t
page_cgroup_ino(struct page
*page
)
360 struct mem_cgroup
*memcg
;
361 unsigned long ino
= 0;
364 memcg
= READ_ONCE(page
->mem_cgroup
);
365 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
366 memcg
= parent_mem_cgroup(memcg
);
368 ino
= cgroup_ino(memcg
->css
.cgroup
);
373 static struct mem_cgroup_per_node
*
374 mem_cgroup_page_nodeinfo(struct mem_cgroup
*memcg
, struct page
*page
)
376 int nid
= page_to_nid(page
);
378 return memcg
->nodeinfo
[nid
];
381 static struct mem_cgroup_tree_per_node
*
382 soft_limit_tree_node(int nid
)
384 return soft_limit_tree
.rb_tree_per_node
[nid
];
387 static struct mem_cgroup_tree_per_node
*
388 soft_limit_tree_from_page(struct page
*page
)
390 int nid
= page_to_nid(page
);
392 return soft_limit_tree
.rb_tree_per_node
[nid
];
395 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node
*mz
,
396 struct mem_cgroup_tree_per_node
*mctz
,
397 unsigned long new_usage_in_excess
)
399 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
400 struct rb_node
*parent
= NULL
;
401 struct mem_cgroup_per_node
*mz_node
;
406 mz
->usage_in_excess
= new_usage_in_excess
;
407 if (!mz
->usage_in_excess
)
411 mz_node
= rb_entry(parent
, struct mem_cgroup_per_node
,
413 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
416 * We can't avoid mem cgroups that are over their soft
417 * limit by the same amount
419 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
422 rb_link_node(&mz
->tree_node
, parent
, p
);
423 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
427 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
428 struct mem_cgroup_tree_per_node
*mctz
)
432 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
436 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
437 struct mem_cgroup_tree_per_node
*mctz
)
441 spin_lock_irqsave(&mctz
->lock
, flags
);
442 __mem_cgroup_remove_exceeded(mz
, mctz
);
443 spin_unlock_irqrestore(&mctz
->lock
, flags
);
446 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
448 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
449 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
450 unsigned long excess
= 0;
452 if (nr_pages
> soft_limit
)
453 excess
= nr_pages
- soft_limit
;
458 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
460 unsigned long excess
;
461 struct mem_cgroup_per_node
*mz
;
462 struct mem_cgroup_tree_per_node
*mctz
;
464 mctz
= soft_limit_tree_from_page(page
);
466 * Necessary to update all ancestors when hierarchy is used.
467 * because their event counter is not touched.
469 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
470 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
471 excess
= soft_limit_excess(memcg
);
473 * We have to update the tree if mz is on RB-tree or
474 * mem is over its softlimit.
476 if (excess
|| mz
->on_tree
) {
479 spin_lock_irqsave(&mctz
->lock
, flags
);
480 /* if on-tree, remove it */
482 __mem_cgroup_remove_exceeded(mz
, mctz
);
484 * Insert again. mz->usage_in_excess will be updated.
485 * If excess is 0, no tree ops.
487 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
488 spin_unlock_irqrestore(&mctz
->lock
, flags
);
493 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
495 struct mem_cgroup_tree_per_node
*mctz
;
496 struct mem_cgroup_per_node
*mz
;
500 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
501 mctz
= soft_limit_tree_node(nid
);
502 mem_cgroup_remove_exceeded(mz
, mctz
);
506 static struct mem_cgroup_per_node
*
507 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
509 struct rb_node
*rightmost
= NULL
;
510 struct mem_cgroup_per_node
*mz
;
514 rightmost
= rb_last(&mctz
->rb_root
);
516 goto done
; /* Nothing to reclaim from */
518 mz
= rb_entry(rightmost
, struct mem_cgroup_per_node
, tree_node
);
520 * Remove the node now but someone else can add it back,
521 * we will to add it back at the end of reclaim to its correct
522 * position in the tree.
524 __mem_cgroup_remove_exceeded(mz
, mctz
);
525 if (!soft_limit_excess(mz
->memcg
) ||
526 !css_tryget_online(&mz
->memcg
->css
))
532 static struct mem_cgroup_per_node
*
533 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
535 struct mem_cgroup_per_node
*mz
;
537 spin_lock_irq(&mctz
->lock
);
538 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
539 spin_unlock_irq(&mctz
->lock
);
544 * Return page count for single (non recursive) @memcg.
546 * Implementation Note: reading percpu statistics for memcg.
548 * Both of vmstat[] and percpu_counter has threshold and do periodic
549 * synchronization to implement "quick" read. There are trade-off between
550 * reading cost and precision of value. Then, we may have a chance to implement
551 * a periodic synchronization of counter in memcg's counter.
553 * But this _read() function is used for user interface now. The user accounts
554 * memory usage by memory cgroup and he _always_ requires exact value because
555 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
556 * have to visit all online cpus and make sum. So, for now, unnecessary
557 * synchronization is not implemented. (just implemented for cpu hotplug)
559 * If there are kernel internal actions which can make use of some not-exact
560 * value, and reading all cpu value can be performance bottleneck in some
561 * common workload, threshold and synchronization as vmstat[] should be
565 mem_cgroup_read_stat(struct mem_cgroup
*memcg
, enum mem_cgroup_stat_index idx
)
570 /* Per-cpu values can be negative, use a signed accumulator */
571 for_each_possible_cpu(cpu
)
572 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
574 * Summing races with updates, so val may be negative. Avoid exposing
575 * transient negative values.
582 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
583 enum mem_cgroup_events_index idx
)
585 unsigned long val
= 0;
588 for_each_possible_cpu(cpu
)
589 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
593 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
595 bool compound
, int nr_pages
)
598 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
599 * counted as CACHE even if it's on ANON LRU.
602 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
605 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
609 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
610 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
614 /* pagein of a big page is an event. So, ignore page size */
616 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
618 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
619 nr_pages
= -nr_pages
; /* for event */
622 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
625 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
626 int nid
, unsigned int lru_mask
)
628 unsigned long nr
= 0;
629 struct mem_cgroup_per_node
*mz
;
632 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
635 if (!(BIT(lru
) & lru_mask
))
637 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
638 nr
+= mz
->lru_size
[lru
];
643 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
644 unsigned int lru_mask
)
646 unsigned long nr
= 0;
649 for_each_node_state(nid
, N_MEMORY
)
650 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
654 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
655 enum mem_cgroup_events_target target
)
657 unsigned long val
, next
;
659 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
660 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
661 /* from time_after() in jiffies.h */
662 if ((long)next
- (long)val
< 0) {
664 case MEM_CGROUP_TARGET_THRESH
:
665 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
667 case MEM_CGROUP_TARGET_SOFTLIMIT
:
668 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
670 case MEM_CGROUP_TARGET_NUMAINFO
:
671 next
= val
+ NUMAINFO_EVENTS_TARGET
;
676 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
683 * Check events in order.
686 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
688 /* threshold event is triggered in finer grain than soft limit */
689 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
690 MEM_CGROUP_TARGET_THRESH
))) {
692 bool do_numainfo __maybe_unused
;
694 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
695 MEM_CGROUP_TARGET_SOFTLIMIT
);
697 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
698 MEM_CGROUP_TARGET_NUMAINFO
);
700 mem_cgroup_threshold(memcg
);
701 if (unlikely(do_softlimit
))
702 mem_cgroup_update_tree(memcg
, page
);
704 if (unlikely(do_numainfo
))
705 atomic_inc(&memcg
->numainfo_events
);
710 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
713 * mm_update_next_owner() may clear mm->owner to NULL
714 * if it races with swapoff, page migration, etc.
715 * So this can be called with p == NULL.
720 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
722 EXPORT_SYMBOL(mem_cgroup_from_task
);
724 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
726 struct mem_cgroup
*memcg
= NULL
;
731 * Page cache insertions can happen withou an
732 * actual mm context, e.g. during disk probing
733 * on boot, loopback IO, acct() writes etc.
736 memcg
= root_mem_cgroup
;
738 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
739 if (unlikely(!memcg
))
740 memcg
= root_mem_cgroup
;
742 } while (!css_tryget_online(&memcg
->css
));
748 * mem_cgroup_iter - iterate over memory cgroup hierarchy
749 * @root: hierarchy root
750 * @prev: previously returned memcg, NULL on first invocation
751 * @reclaim: cookie for shared reclaim walks, NULL for full walks
753 * Returns references to children of the hierarchy below @root, or
754 * @root itself, or %NULL after a full round-trip.
756 * Caller must pass the return value in @prev on subsequent
757 * invocations for reference counting, or use mem_cgroup_iter_break()
758 * to cancel a hierarchy walk before the round-trip is complete.
760 * Reclaimers can specify a zone and a priority level in @reclaim to
761 * divide up the memcgs in the hierarchy among all concurrent
762 * reclaimers operating on the same zone and priority.
764 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
765 struct mem_cgroup
*prev
,
766 struct mem_cgroup_reclaim_cookie
*reclaim
)
768 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
769 struct cgroup_subsys_state
*css
= NULL
;
770 struct mem_cgroup
*memcg
= NULL
;
771 struct mem_cgroup
*pos
= NULL
;
773 if (mem_cgroup_disabled())
777 root
= root_mem_cgroup
;
779 if (prev
&& !reclaim
)
782 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
791 struct mem_cgroup_per_node
*mz
;
793 mz
= mem_cgroup_nodeinfo(root
, reclaim
->pgdat
->node_id
);
794 iter
= &mz
->iter
[reclaim
->priority
];
796 if (prev
&& reclaim
->generation
!= iter
->generation
)
800 pos
= READ_ONCE(iter
->position
);
801 if (!pos
|| css_tryget(&pos
->css
))
804 * css reference reached zero, so iter->position will
805 * be cleared by ->css_released. However, we should not
806 * rely on this happening soon, because ->css_released
807 * is called from a work queue, and by busy-waiting we
808 * might block it. So we clear iter->position right
811 (void)cmpxchg(&iter
->position
, pos
, NULL
);
819 css
= css_next_descendant_pre(css
, &root
->css
);
822 * Reclaimers share the hierarchy walk, and a
823 * new one might jump in right at the end of
824 * the hierarchy - make sure they see at least
825 * one group and restart from the beginning.
833 * Verify the css and acquire a reference. The root
834 * is provided by the caller, so we know it's alive
835 * and kicking, and don't take an extra reference.
837 memcg
= mem_cgroup_from_css(css
);
839 if (css
== &root
->css
)
850 * The position could have already been updated by a competing
851 * thread, so check that the value hasn't changed since we read
852 * it to avoid reclaiming from the same cgroup twice.
854 (void)cmpxchg(&iter
->position
, pos
, memcg
);
862 reclaim
->generation
= iter
->generation
;
868 if (prev
&& prev
!= root
)
875 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
876 * @root: hierarchy root
877 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
879 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
880 struct mem_cgroup
*prev
)
883 root
= root_mem_cgroup
;
884 if (prev
&& prev
!= root
)
888 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
890 struct mem_cgroup
*memcg
= dead_memcg
;
891 struct mem_cgroup_reclaim_iter
*iter
;
892 struct mem_cgroup_per_node
*mz
;
896 while ((memcg
= parent_mem_cgroup(memcg
))) {
898 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
899 for (i
= 0; i
<= DEF_PRIORITY
; i
++) {
901 cmpxchg(&iter
->position
,
909 * Iteration constructs for visiting all cgroups (under a tree). If
910 * loops are exited prematurely (break), mem_cgroup_iter_break() must
911 * be used for reference counting.
913 #define for_each_mem_cgroup_tree(iter, root) \
914 for (iter = mem_cgroup_iter(root, NULL, NULL); \
916 iter = mem_cgroup_iter(root, iter, NULL))
918 #define for_each_mem_cgroup(iter) \
919 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
921 iter = mem_cgroup_iter(NULL, iter, NULL))
924 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
926 * @zone: zone of the page
928 * This function is only safe when following the LRU page isolation
929 * and putback protocol: the LRU lock must be held, and the page must
930 * either be PageLRU() or the caller must have isolated/allocated it.
932 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct pglist_data
*pgdat
)
934 struct mem_cgroup_per_node
*mz
;
935 struct mem_cgroup
*memcg
;
936 struct lruvec
*lruvec
;
938 if (mem_cgroup_disabled()) {
939 lruvec
= &pgdat
->lruvec
;
943 memcg
= page
->mem_cgroup
;
945 * Swapcache readahead pages are added to the LRU - and
946 * possibly migrated - before they are charged.
949 memcg
= root_mem_cgroup
;
951 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
952 lruvec
= &mz
->lruvec
;
955 * Since a node can be onlined after the mem_cgroup was created,
956 * we have to be prepared to initialize lruvec->zone here;
957 * and if offlined then reonlined, we need to reinitialize it.
959 if (unlikely(lruvec
->pgdat
!= pgdat
))
960 lruvec
->pgdat
= pgdat
;
965 * mem_cgroup_update_lru_size - account for adding or removing an lru page
966 * @lruvec: mem_cgroup per zone lru vector
967 * @lru: index of lru list the page is sitting on
968 * @nr_pages: positive when adding or negative when removing
970 * This function must be called under lru_lock, just before a page is added
971 * to or just after a page is removed from an lru list (that ordering being
972 * so as to allow it to check that lru_size 0 is consistent with list_empty).
974 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
977 struct mem_cgroup_per_node
*mz
;
978 unsigned long *lru_size
;
982 if (mem_cgroup_disabled())
985 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
986 lru_size
= mz
->lru_size
+ lru
;
987 empty
= list_empty(lruvec
->lists
+ lru
);
990 *lru_size
+= nr_pages
;
993 if (WARN_ONCE(size
< 0 || empty
!= !size
,
994 "%s(%p, %d, %d): lru_size %ld but %sempty\n",
995 __func__
, lruvec
, lru
, nr_pages
, size
, empty
? "" : "not ")) {
1001 *lru_size
+= nr_pages
;
1004 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1006 struct mem_cgroup
*task_memcg
;
1007 struct task_struct
*p
;
1010 p
= find_lock_task_mm(task
);
1012 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1016 * All threads may have already detached their mm's, but the oom
1017 * killer still needs to detect if they have already been oom
1018 * killed to prevent needlessly killing additional tasks.
1021 task_memcg
= mem_cgroup_from_task(task
);
1022 css_get(&task_memcg
->css
);
1025 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1026 css_put(&task_memcg
->css
);
1031 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1032 * @memcg: the memory cgroup
1034 * Returns the maximum amount of memory @mem can be charged with, in
1037 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1039 unsigned long margin
= 0;
1040 unsigned long count
;
1041 unsigned long limit
;
1043 count
= page_counter_read(&memcg
->memory
);
1044 limit
= READ_ONCE(memcg
->memory
.limit
);
1046 margin
= limit
- count
;
1048 if (do_memsw_account()) {
1049 count
= page_counter_read(&memcg
->memsw
);
1050 limit
= READ_ONCE(memcg
->memsw
.limit
);
1052 margin
= min(margin
, limit
- count
);
1061 * A routine for checking "mem" is under move_account() or not.
1063 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1064 * moving cgroups. This is for waiting at high-memory pressure
1067 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1069 struct mem_cgroup
*from
;
1070 struct mem_cgroup
*to
;
1073 * Unlike task_move routines, we access mc.to, mc.from not under
1074 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1076 spin_lock(&mc
.lock
);
1082 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1083 mem_cgroup_is_descendant(to
, memcg
);
1085 spin_unlock(&mc
.lock
);
1089 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1091 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1092 if (mem_cgroup_under_move(memcg
)) {
1094 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1095 /* moving charge context might have finished. */
1098 finish_wait(&mc
.waitq
, &wait
);
1105 #define K(x) ((x) << (PAGE_SHIFT-10))
1107 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1108 * @memcg: The memory cgroup that went over limit
1109 * @p: Task that is going to be killed
1111 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1114 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1116 struct mem_cgroup
*iter
;
1122 pr_info("Task in ");
1123 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1124 pr_cont(" killed as a result of limit of ");
1126 pr_info("Memory limit reached of cgroup ");
1129 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1134 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1135 K((u64
)page_counter_read(&memcg
->memory
)),
1136 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1137 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1138 K((u64
)page_counter_read(&memcg
->memsw
)),
1139 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1140 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1141 K((u64
)page_counter_read(&memcg
->kmem
)),
1142 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1144 for_each_mem_cgroup_tree(iter
, memcg
) {
1145 pr_info("Memory cgroup stats for ");
1146 pr_cont_cgroup_path(iter
->css
.cgroup
);
1149 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1150 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1152 pr_cont(" %s:%luKB", mem_cgroup_stat_names
[i
],
1153 K(mem_cgroup_read_stat(iter
, i
)));
1156 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1157 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1158 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1165 * This function returns the number of memcg under hierarchy tree. Returns
1166 * 1(self count) if no children.
1168 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1171 struct mem_cgroup
*iter
;
1173 for_each_mem_cgroup_tree(iter
, memcg
)
1179 * Return the memory (and swap, if configured) limit for a memcg.
1181 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1183 unsigned long limit
;
1185 limit
= memcg
->memory
.limit
;
1186 if (mem_cgroup_swappiness(memcg
)) {
1187 unsigned long memsw_limit
;
1188 unsigned long swap_limit
;
1190 memsw_limit
= memcg
->memsw
.limit
;
1191 swap_limit
= memcg
->swap
.limit
;
1192 swap_limit
= min(swap_limit
, (unsigned long)total_swap_pages
);
1193 limit
= min(limit
+ swap_limit
, memsw_limit
);
1198 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1201 struct oom_control oc
= {
1205 .gfp_mask
= gfp_mask
,
1208 struct mem_cgroup
*iter
;
1209 unsigned long chosen_points
= 0;
1210 unsigned long totalpages
;
1211 unsigned int points
= 0;
1212 struct task_struct
*chosen
= NULL
;
1214 mutex_lock(&oom_lock
);
1217 * If current has a pending SIGKILL or is exiting, then automatically
1218 * select it. The goal is to allow it to allocate so that it may
1219 * quickly exit and free its memory.
1221 if (task_will_free_mem(current
)) {
1222 mark_oom_victim(current
);
1223 wake_oom_reaper(current
);
1227 check_panic_on_oom(&oc
, CONSTRAINT_MEMCG
);
1228 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1229 for_each_mem_cgroup_tree(iter
, memcg
) {
1230 struct css_task_iter it
;
1231 struct task_struct
*task
;
1233 css_task_iter_start(&iter
->css
, &it
);
1234 while ((task
= css_task_iter_next(&it
))) {
1235 switch (oom_scan_process_thread(&oc
, task
)) {
1236 case OOM_SCAN_SELECT
:
1238 put_task_struct(chosen
);
1240 chosen_points
= ULONG_MAX
;
1241 get_task_struct(chosen
);
1243 case OOM_SCAN_CONTINUE
:
1245 case OOM_SCAN_ABORT
:
1246 css_task_iter_end(&it
);
1247 mem_cgroup_iter_break(memcg
, iter
);
1249 put_task_struct(chosen
);
1250 /* Set a dummy value to return "true". */
1251 chosen
= (void *) 1;
1256 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1257 if (!points
|| points
< chosen_points
)
1259 /* Prefer thread group leaders for display purposes */
1260 if (points
== chosen_points
&&
1261 thread_group_leader(chosen
))
1265 put_task_struct(chosen
);
1267 chosen_points
= points
;
1268 get_task_struct(chosen
);
1270 css_task_iter_end(&it
);
1274 points
= chosen_points
* 1000 / totalpages
;
1275 oom_kill_process(&oc
, chosen
, points
, totalpages
,
1276 "Memory cgroup out of memory");
1279 mutex_unlock(&oom_lock
);
1283 #if MAX_NUMNODES > 1
1286 * test_mem_cgroup_node_reclaimable
1287 * @memcg: the target memcg
1288 * @nid: the node ID to be checked.
1289 * @noswap : specify true here if the user wants flle only information.
1291 * This function returns whether the specified memcg contains any
1292 * reclaimable pages on a node. Returns true if there are any reclaimable
1293 * pages in the node.
1295 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1296 int nid
, bool noswap
)
1298 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1300 if (noswap
|| !total_swap_pages
)
1302 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1309 * Always updating the nodemask is not very good - even if we have an empty
1310 * list or the wrong list here, we can start from some node and traverse all
1311 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1314 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1318 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1319 * pagein/pageout changes since the last update.
1321 if (!atomic_read(&memcg
->numainfo_events
))
1323 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1326 /* make a nodemask where this memcg uses memory from */
1327 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1329 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1331 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1332 node_clear(nid
, memcg
->scan_nodes
);
1335 atomic_set(&memcg
->numainfo_events
, 0);
1336 atomic_set(&memcg
->numainfo_updating
, 0);
1340 * Selecting a node where we start reclaim from. Because what we need is just
1341 * reducing usage counter, start from anywhere is O,K. Considering
1342 * memory reclaim from current node, there are pros. and cons.
1344 * Freeing memory from current node means freeing memory from a node which
1345 * we'll use or we've used. So, it may make LRU bad. And if several threads
1346 * hit limits, it will see a contention on a node. But freeing from remote
1347 * node means more costs for memory reclaim because of memory latency.
1349 * Now, we use round-robin. Better algorithm is welcomed.
1351 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1355 mem_cgroup_may_update_nodemask(memcg
);
1356 node
= memcg
->last_scanned_node
;
1358 node
= next_node_in(node
, memcg
->scan_nodes
);
1360 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1361 * last time it really checked all the LRUs due to rate limiting.
1362 * Fallback to the current node in that case for simplicity.
1364 if (unlikely(node
== MAX_NUMNODES
))
1365 node
= numa_node_id();
1367 memcg
->last_scanned_node
= node
;
1371 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1377 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1380 unsigned long *total_scanned
)
1382 struct mem_cgroup
*victim
= NULL
;
1385 unsigned long excess
;
1386 unsigned long nr_scanned
;
1387 struct mem_cgroup_reclaim_cookie reclaim
= {
1392 excess
= soft_limit_excess(root_memcg
);
1395 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1400 * If we have not been able to reclaim
1401 * anything, it might because there are
1402 * no reclaimable pages under this hierarchy
1407 * We want to do more targeted reclaim.
1408 * excess >> 2 is not to excessive so as to
1409 * reclaim too much, nor too less that we keep
1410 * coming back to reclaim from this cgroup
1412 if (total
>= (excess
>> 2) ||
1413 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1418 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1419 pgdat
, &nr_scanned
);
1420 *total_scanned
+= nr_scanned
;
1421 if (!soft_limit_excess(root_memcg
))
1424 mem_cgroup_iter_break(root_memcg
, victim
);
1428 #ifdef CONFIG_LOCKDEP
1429 static struct lockdep_map memcg_oom_lock_dep_map
= {
1430 .name
= "memcg_oom_lock",
1434 static DEFINE_SPINLOCK(memcg_oom_lock
);
1437 * Check OOM-Killer is already running under our hierarchy.
1438 * If someone is running, return false.
1440 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1442 struct mem_cgroup
*iter
, *failed
= NULL
;
1444 spin_lock(&memcg_oom_lock
);
1446 for_each_mem_cgroup_tree(iter
, memcg
) {
1447 if (iter
->oom_lock
) {
1449 * this subtree of our hierarchy is already locked
1450 * so we cannot give a lock.
1453 mem_cgroup_iter_break(memcg
, iter
);
1456 iter
->oom_lock
= true;
1461 * OK, we failed to lock the whole subtree so we have
1462 * to clean up what we set up to the failing subtree
1464 for_each_mem_cgroup_tree(iter
, memcg
) {
1465 if (iter
== failed
) {
1466 mem_cgroup_iter_break(memcg
, iter
);
1469 iter
->oom_lock
= false;
1472 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1474 spin_unlock(&memcg_oom_lock
);
1479 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1481 struct mem_cgroup
*iter
;
1483 spin_lock(&memcg_oom_lock
);
1484 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1485 for_each_mem_cgroup_tree(iter
, memcg
)
1486 iter
->oom_lock
= false;
1487 spin_unlock(&memcg_oom_lock
);
1490 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1492 struct mem_cgroup
*iter
;
1494 spin_lock(&memcg_oom_lock
);
1495 for_each_mem_cgroup_tree(iter
, memcg
)
1497 spin_unlock(&memcg_oom_lock
);
1500 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1502 struct mem_cgroup
*iter
;
1505 * When a new child is created while the hierarchy is under oom,
1506 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1508 spin_lock(&memcg_oom_lock
);
1509 for_each_mem_cgroup_tree(iter
, memcg
)
1510 if (iter
->under_oom
> 0)
1512 spin_unlock(&memcg_oom_lock
);
1515 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1517 struct oom_wait_info
{
1518 struct mem_cgroup
*memcg
;
1522 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1523 unsigned mode
, int sync
, void *arg
)
1525 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1526 struct mem_cgroup
*oom_wait_memcg
;
1527 struct oom_wait_info
*oom_wait_info
;
1529 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1530 oom_wait_memcg
= oom_wait_info
->memcg
;
1532 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1533 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1535 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1538 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1541 * For the following lockless ->under_oom test, the only required
1542 * guarantee is that it must see the state asserted by an OOM when
1543 * this function is called as a result of userland actions
1544 * triggered by the notification of the OOM. This is trivially
1545 * achieved by invoking mem_cgroup_mark_under_oom() before
1546 * triggering notification.
1548 if (memcg
&& memcg
->under_oom
)
1549 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1552 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1554 if (!current
->memcg_may_oom
)
1557 * We are in the middle of the charge context here, so we
1558 * don't want to block when potentially sitting on a callstack
1559 * that holds all kinds of filesystem and mm locks.
1561 * Also, the caller may handle a failed allocation gracefully
1562 * (like optional page cache readahead) and so an OOM killer
1563 * invocation might not even be necessary.
1565 * That's why we don't do anything here except remember the
1566 * OOM context and then deal with it at the end of the page
1567 * fault when the stack is unwound, the locks are released,
1568 * and when we know whether the fault was overall successful.
1570 css_get(&memcg
->css
);
1571 current
->memcg_in_oom
= memcg
;
1572 current
->memcg_oom_gfp_mask
= mask
;
1573 current
->memcg_oom_order
= order
;
1577 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1578 * @handle: actually kill/wait or just clean up the OOM state
1580 * This has to be called at the end of a page fault if the memcg OOM
1581 * handler was enabled.
1583 * Memcg supports userspace OOM handling where failed allocations must
1584 * sleep on a waitqueue until the userspace task resolves the
1585 * situation. Sleeping directly in the charge context with all kinds
1586 * of locks held is not a good idea, instead we remember an OOM state
1587 * in the task and mem_cgroup_oom_synchronize() has to be called at
1588 * the end of the page fault to complete the OOM handling.
1590 * Returns %true if an ongoing memcg OOM situation was detected and
1591 * completed, %false otherwise.
1593 bool mem_cgroup_oom_synchronize(bool handle
)
1595 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1596 struct oom_wait_info owait
;
1599 /* OOM is global, do not handle */
1603 if (!handle
|| oom_killer_disabled
)
1606 owait
.memcg
= memcg
;
1607 owait
.wait
.flags
= 0;
1608 owait
.wait
.func
= memcg_oom_wake_function
;
1609 owait
.wait
.private = current
;
1610 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1612 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1613 mem_cgroup_mark_under_oom(memcg
);
1615 locked
= mem_cgroup_oom_trylock(memcg
);
1618 mem_cgroup_oom_notify(memcg
);
1620 if (locked
&& !memcg
->oom_kill_disable
) {
1621 mem_cgroup_unmark_under_oom(memcg
);
1622 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1623 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1624 current
->memcg_oom_order
);
1627 mem_cgroup_unmark_under_oom(memcg
);
1628 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1632 mem_cgroup_oom_unlock(memcg
);
1634 * There is no guarantee that an OOM-lock contender
1635 * sees the wakeups triggered by the OOM kill
1636 * uncharges. Wake any sleepers explicitely.
1638 memcg_oom_recover(memcg
);
1641 current
->memcg_in_oom
= NULL
;
1642 css_put(&memcg
->css
);
1647 * lock_page_memcg - lock a page->mem_cgroup binding
1650 * This function protects unlocked LRU pages from being moved to
1651 * another cgroup and stabilizes their page->mem_cgroup binding.
1653 void lock_page_memcg(struct page
*page
)
1655 struct mem_cgroup
*memcg
;
1656 unsigned long flags
;
1659 * The RCU lock is held throughout the transaction. The fast
1660 * path can get away without acquiring the memcg->move_lock
1661 * because page moving starts with an RCU grace period.
1665 if (mem_cgroup_disabled())
1668 memcg
= page
->mem_cgroup
;
1669 if (unlikely(!memcg
))
1672 if (atomic_read(&memcg
->moving_account
) <= 0)
1675 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1676 if (memcg
!= page
->mem_cgroup
) {
1677 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1682 * When charge migration first begins, we can have locked and
1683 * unlocked page stat updates happening concurrently. Track
1684 * the task who has the lock for unlock_page_memcg().
1686 memcg
->move_lock_task
= current
;
1687 memcg
->move_lock_flags
= flags
;
1691 EXPORT_SYMBOL(lock_page_memcg
);
1694 * unlock_page_memcg - unlock a page->mem_cgroup binding
1697 void unlock_page_memcg(struct page
*page
)
1699 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
1701 if (memcg
&& memcg
->move_lock_task
== current
) {
1702 unsigned long flags
= memcg
->move_lock_flags
;
1704 memcg
->move_lock_task
= NULL
;
1705 memcg
->move_lock_flags
= 0;
1707 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1712 EXPORT_SYMBOL(unlock_page_memcg
);
1715 * size of first charge trial. "32" comes from vmscan.c's magic value.
1716 * TODO: maybe necessary to use big numbers in big irons.
1718 #define CHARGE_BATCH 32U
1719 struct memcg_stock_pcp
{
1720 struct mem_cgroup
*cached
; /* this never be root cgroup */
1721 unsigned int nr_pages
;
1722 struct work_struct work
;
1723 unsigned long flags
;
1724 #define FLUSHING_CACHED_CHARGE 0
1726 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1727 static DEFINE_MUTEX(percpu_charge_mutex
);
1730 * consume_stock: Try to consume stocked charge on this cpu.
1731 * @memcg: memcg to consume from.
1732 * @nr_pages: how many pages to charge.
1734 * The charges will only happen if @memcg matches the current cpu's memcg
1735 * stock, and at least @nr_pages are available in that stock. Failure to
1736 * service an allocation will refill the stock.
1738 * returns true if successful, false otherwise.
1740 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1742 struct memcg_stock_pcp
*stock
;
1745 if (nr_pages
> CHARGE_BATCH
)
1748 stock
= &get_cpu_var(memcg_stock
);
1749 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
1750 stock
->nr_pages
-= nr_pages
;
1753 put_cpu_var(memcg_stock
);
1758 * Returns stocks cached in percpu and reset cached information.
1760 static void drain_stock(struct memcg_stock_pcp
*stock
)
1762 struct mem_cgroup
*old
= stock
->cached
;
1764 if (stock
->nr_pages
) {
1765 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
1766 if (do_memsw_account())
1767 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
1768 css_put_many(&old
->css
, stock
->nr_pages
);
1769 stock
->nr_pages
= 0;
1771 stock
->cached
= NULL
;
1775 * This must be called under preempt disabled or must be called by
1776 * a thread which is pinned to local cpu.
1778 static void drain_local_stock(struct work_struct
*dummy
)
1780 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
1782 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
1786 * Cache charges(val) to local per_cpu area.
1787 * This will be consumed by consume_stock() function, later.
1789 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1791 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
1793 if (stock
->cached
!= memcg
) { /* reset if necessary */
1795 stock
->cached
= memcg
;
1797 stock
->nr_pages
+= nr_pages
;
1798 put_cpu_var(memcg_stock
);
1802 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1803 * of the hierarchy under it.
1805 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
1809 /* If someone's already draining, avoid adding running more workers. */
1810 if (!mutex_trylock(&percpu_charge_mutex
))
1812 /* Notify other cpus that system-wide "drain" is running */
1815 for_each_online_cpu(cpu
) {
1816 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1817 struct mem_cgroup
*memcg
;
1819 memcg
= stock
->cached
;
1820 if (!memcg
|| !stock
->nr_pages
)
1822 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
1824 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
1826 drain_local_stock(&stock
->work
);
1828 schedule_work_on(cpu
, &stock
->work
);
1833 mutex_unlock(&percpu_charge_mutex
);
1836 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
1837 unsigned long action
,
1840 int cpu
= (unsigned long)hcpu
;
1841 struct memcg_stock_pcp
*stock
;
1843 if (action
== CPU_ONLINE
)
1846 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
1849 stock
= &per_cpu(memcg_stock
, cpu
);
1854 static void reclaim_high(struct mem_cgroup
*memcg
,
1855 unsigned int nr_pages
,
1859 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
1861 mem_cgroup_events(memcg
, MEMCG_HIGH
, 1);
1862 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
1863 } while ((memcg
= parent_mem_cgroup(memcg
)));
1866 static void high_work_func(struct work_struct
*work
)
1868 struct mem_cgroup
*memcg
;
1870 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
1871 reclaim_high(memcg
, CHARGE_BATCH
, GFP_KERNEL
);
1875 * Scheduled by try_charge() to be executed from the userland return path
1876 * and reclaims memory over the high limit.
1878 void mem_cgroup_handle_over_high(void)
1880 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
1881 struct mem_cgroup
*memcg
;
1883 if (likely(!nr_pages
))
1886 memcg
= get_mem_cgroup_from_mm(current
->mm
);
1887 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
1888 css_put(&memcg
->css
);
1889 current
->memcg_nr_pages_over_high
= 0;
1892 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1893 unsigned int nr_pages
)
1895 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
1896 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1897 struct mem_cgroup
*mem_over_limit
;
1898 struct page_counter
*counter
;
1899 unsigned long nr_reclaimed
;
1900 bool may_swap
= true;
1901 bool drained
= false;
1903 if (mem_cgroup_is_root(memcg
))
1906 if (consume_stock(memcg
, nr_pages
))
1909 if (!do_memsw_account() ||
1910 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
1911 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
1913 if (do_memsw_account())
1914 page_counter_uncharge(&memcg
->memsw
, batch
);
1915 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
1917 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
1921 if (batch
> nr_pages
) {
1927 * Unlike in global OOM situations, memcg is not in a physical
1928 * memory shortage. Allow dying and OOM-killed tasks to
1929 * bypass the last charges so that they can exit quickly and
1930 * free their memory.
1932 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
1933 fatal_signal_pending(current
) ||
1934 current
->flags
& PF_EXITING
))
1937 if (unlikely(task_in_memcg_oom(current
)))
1940 if (!gfpflags_allow_blocking(gfp_mask
))
1943 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
1945 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
1946 gfp_mask
, may_swap
);
1948 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
1952 drain_all_stock(mem_over_limit
);
1957 if (gfp_mask
& __GFP_NORETRY
)
1960 * Even though the limit is exceeded at this point, reclaim
1961 * may have been able to free some pages. Retry the charge
1962 * before killing the task.
1964 * Only for regular pages, though: huge pages are rather
1965 * unlikely to succeed so close to the limit, and we fall back
1966 * to regular pages anyway in case of failure.
1968 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
1971 * At task move, charge accounts can be doubly counted. So, it's
1972 * better to wait until the end of task_move if something is going on.
1974 if (mem_cgroup_wait_acct_move(mem_over_limit
))
1980 if (gfp_mask
& __GFP_NOFAIL
)
1983 if (fatal_signal_pending(current
))
1986 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
1988 mem_cgroup_oom(mem_over_limit
, gfp_mask
,
1989 get_order(nr_pages
* PAGE_SIZE
));
1991 if (!(gfp_mask
& __GFP_NOFAIL
))
1995 * The allocation either can't fail or will lead to more memory
1996 * being freed very soon. Allow memory usage go over the limit
1997 * temporarily by force charging it.
1999 page_counter_charge(&memcg
->memory
, nr_pages
);
2000 if (do_memsw_account())
2001 page_counter_charge(&memcg
->memsw
, nr_pages
);
2002 css_get_many(&memcg
->css
, nr_pages
);
2007 css_get_many(&memcg
->css
, batch
);
2008 if (batch
> nr_pages
)
2009 refill_stock(memcg
, batch
- nr_pages
);
2012 * If the hierarchy is above the normal consumption range, schedule
2013 * reclaim on returning to userland. We can perform reclaim here
2014 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2015 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2016 * not recorded as it most likely matches current's and won't
2017 * change in the meantime. As high limit is checked again before
2018 * reclaim, the cost of mismatch is negligible.
2021 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2022 /* Don't bother a random interrupted task */
2023 if (in_interrupt()) {
2024 schedule_work(&memcg
->high_work
);
2027 current
->memcg_nr_pages_over_high
+= batch
;
2028 set_notify_resume(current
);
2031 } while ((memcg
= parent_mem_cgroup(memcg
)));
2036 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2038 if (mem_cgroup_is_root(memcg
))
2041 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2042 if (do_memsw_account())
2043 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2045 css_put_many(&memcg
->css
, nr_pages
);
2048 static void lock_page_lru(struct page
*page
, int *isolated
)
2050 struct zone
*zone
= page_zone(page
);
2052 spin_lock_irq(zone_lru_lock(zone
));
2053 if (PageLRU(page
)) {
2054 struct lruvec
*lruvec
;
2056 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
2058 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2064 static void unlock_page_lru(struct page
*page
, int isolated
)
2066 struct zone
*zone
= page_zone(page
);
2069 struct lruvec
*lruvec
;
2071 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
2072 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2074 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2076 spin_unlock_irq(zone_lru_lock(zone
));
2079 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2084 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2087 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2088 * may already be on some other mem_cgroup's LRU. Take care of it.
2091 lock_page_lru(page
, &isolated
);
2094 * Nobody should be changing or seriously looking at
2095 * page->mem_cgroup at this point:
2097 * - the page is uncharged
2099 * - the page is off-LRU
2101 * - an anonymous fault has exclusive page access, except for
2102 * a locked page table
2104 * - a page cache insertion, a swapin fault, or a migration
2105 * have the page locked
2107 page
->mem_cgroup
= memcg
;
2110 unlock_page_lru(page
, isolated
);
2114 static int memcg_alloc_cache_id(void)
2119 id
= ida_simple_get(&memcg_cache_ida
,
2120 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2124 if (id
< memcg_nr_cache_ids
)
2128 * There's no space for the new id in memcg_caches arrays,
2129 * so we have to grow them.
2131 down_write(&memcg_cache_ids_sem
);
2133 size
= 2 * (id
+ 1);
2134 if (size
< MEMCG_CACHES_MIN_SIZE
)
2135 size
= MEMCG_CACHES_MIN_SIZE
;
2136 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2137 size
= MEMCG_CACHES_MAX_SIZE
;
2139 err
= memcg_update_all_caches(size
);
2141 err
= memcg_update_all_list_lrus(size
);
2143 memcg_nr_cache_ids
= size
;
2145 up_write(&memcg_cache_ids_sem
);
2148 ida_simple_remove(&memcg_cache_ida
, id
);
2154 static void memcg_free_cache_id(int id
)
2156 ida_simple_remove(&memcg_cache_ida
, id
);
2159 struct memcg_kmem_cache_create_work
{
2160 struct mem_cgroup
*memcg
;
2161 struct kmem_cache
*cachep
;
2162 struct work_struct work
;
2165 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2167 struct memcg_kmem_cache_create_work
*cw
=
2168 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2169 struct mem_cgroup
*memcg
= cw
->memcg
;
2170 struct kmem_cache
*cachep
= cw
->cachep
;
2172 memcg_create_kmem_cache(memcg
, cachep
);
2174 css_put(&memcg
->css
);
2179 * Enqueue the creation of a per-memcg kmem_cache.
2181 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2182 struct kmem_cache
*cachep
)
2184 struct memcg_kmem_cache_create_work
*cw
;
2186 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2190 css_get(&memcg
->css
);
2193 cw
->cachep
= cachep
;
2194 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2196 schedule_work(&cw
->work
);
2199 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2200 struct kmem_cache
*cachep
)
2203 * We need to stop accounting when we kmalloc, because if the
2204 * corresponding kmalloc cache is not yet created, the first allocation
2205 * in __memcg_schedule_kmem_cache_create will recurse.
2207 * However, it is better to enclose the whole function. Depending on
2208 * the debugging options enabled, INIT_WORK(), for instance, can
2209 * trigger an allocation. This too, will make us recurse. Because at
2210 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2211 * the safest choice is to do it like this, wrapping the whole function.
2213 current
->memcg_kmem_skip_account
= 1;
2214 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2215 current
->memcg_kmem_skip_account
= 0;
2218 static inline bool memcg_kmem_bypass(void)
2220 if (in_interrupt() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
2226 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2227 * @cachep: the original global kmem cache
2229 * Return the kmem_cache we're supposed to use for a slab allocation.
2230 * We try to use the current memcg's version of the cache.
2232 * If the cache does not exist yet, if we are the first user of it, we
2233 * create it asynchronously in a workqueue and let the current allocation
2234 * go through with the original cache.
2236 * This function takes a reference to the cache it returns to assure it
2237 * won't get destroyed while we are working with it. Once the caller is
2238 * done with it, memcg_kmem_put_cache() must be called to release the
2241 struct kmem_cache
*memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2243 struct mem_cgroup
*memcg
;
2244 struct kmem_cache
*memcg_cachep
;
2247 VM_BUG_ON(!is_root_cache(cachep
));
2249 if (memcg_kmem_bypass())
2252 if (current
->memcg_kmem_skip_account
)
2255 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2256 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2260 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2261 if (likely(memcg_cachep
))
2262 return memcg_cachep
;
2265 * If we are in a safe context (can wait, and not in interrupt
2266 * context), we could be be predictable and return right away.
2267 * This would guarantee that the allocation being performed
2268 * already belongs in the new cache.
2270 * However, there are some clashes that can arrive from locking.
2271 * For instance, because we acquire the slab_mutex while doing
2272 * memcg_create_kmem_cache, this means no further allocation
2273 * could happen with the slab_mutex held. So it's better to
2276 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2278 css_put(&memcg
->css
);
2283 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2284 * @cachep: the cache returned by memcg_kmem_get_cache
2286 void memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2288 if (!is_root_cache(cachep
))
2289 css_put(&cachep
->memcg_params
.memcg
->css
);
2293 * memcg_kmem_charge: charge a kmem page
2294 * @page: page to charge
2295 * @gfp: reclaim mode
2296 * @order: allocation order
2297 * @memcg: memory cgroup to charge
2299 * Returns 0 on success, an error code on failure.
2301 int memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2302 struct mem_cgroup
*memcg
)
2304 unsigned int nr_pages
= 1 << order
;
2305 struct page_counter
*counter
;
2308 ret
= try_charge(memcg
, gfp
, nr_pages
);
2312 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2313 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2314 cancel_charge(memcg
, nr_pages
);
2318 page
->mem_cgroup
= memcg
;
2324 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2325 * @page: page to charge
2326 * @gfp: reclaim mode
2327 * @order: allocation order
2329 * Returns 0 on success, an error code on failure.
2331 int memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2333 struct mem_cgroup
*memcg
;
2336 if (memcg_kmem_bypass())
2339 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2340 if (!mem_cgroup_is_root(memcg
))
2341 ret
= memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2342 css_put(&memcg
->css
);
2346 * memcg_kmem_uncharge: uncharge a kmem page
2347 * @page: page to uncharge
2348 * @order: allocation order
2350 void memcg_kmem_uncharge(struct page
*page
, int order
)
2352 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2353 unsigned int nr_pages
= 1 << order
;
2358 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2360 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2361 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2363 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2364 if (do_memsw_account())
2365 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2367 page
->mem_cgroup
= NULL
;
2368 css_put_many(&memcg
->css
, nr_pages
);
2370 #endif /* !CONFIG_SLOB */
2372 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2375 * Because tail pages are not marked as "used", set it. We're under
2376 * zone_lru_lock and migration entries setup in all page mappings.
2378 void mem_cgroup_split_huge_fixup(struct page
*head
)
2382 if (mem_cgroup_disabled())
2385 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2386 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2388 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2391 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2393 #ifdef CONFIG_MEMCG_SWAP
2394 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2397 int val
= (charge
) ? 1 : -1;
2398 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2402 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2403 * @entry: swap entry to be moved
2404 * @from: mem_cgroup which the entry is moved from
2405 * @to: mem_cgroup which the entry is moved to
2407 * It succeeds only when the swap_cgroup's record for this entry is the same
2408 * as the mem_cgroup's id of @from.
2410 * Returns 0 on success, -EINVAL on failure.
2412 * The caller must have charged to @to, IOW, called page_counter_charge() about
2413 * both res and memsw, and called css_get().
2415 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2416 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2418 unsigned short old_id
, new_id
;
2420 old_id
= mem_cgroup_id(from
);
2421 new_id
= mem_cgroup_id(to
);
2423 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2424 mem_cgroup_swap_statistics(from
, false);
2425 mem_cgroup_swap_statistics(to
, true);
2431 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2432 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2438 static DEFINE_MUTEX(memcg_limit_mutex
);
2440 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2441 unsigned long limit
)
2443 unsigned long curusage
;
2444 unsigned long oldusage
;
2445 bool enlarge
= false;
2450 * For keeping hierarchical_reclaim simple, how long we should retry
2451 * is depends on callers. We set our retry-count to be function
2452 * of # of children which we should visit in this loop.
2454 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2455 mem_cgroup_count_children(memcg
);
2457 oldusage
= page_counter_read(&memcg
->memory
);
2460 if (signal_pending(current
)) {
2465 mutex_lock(&memcg_limit_mutex
);
2466 if (limit
> memcg
->memsw
.limit
) {
2467 mutex_unlock(&memcg_limit_mutex
);
2471 if (limit
> memcg
->memory
.limit
)
2473 ret
= page_counter_limit(&memcg
->memory
, limit
);
2474 mutex_unlock(&memcg_limit_mutex
);
2479 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2481 curusage
= page_counter_read(&memcg
->memory
);
2482 /* Usage is reduced ? */
2483 if (curusage
>= oldusage
)
2486 oldusage
= curusage
;
2487 } while (retry_count
);
2489 if (!ret
&& enlarge
)
2490 memcg_oom_recover(memcg
);
2495 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2496 unsigned long limit
)
2498 unsigned long curusage
;
2499 unsigned long oldusage
;
2500 bool enlarge
= false;
2504 /* see mem_cgroup_resize_res_limit */
2505 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2506 mem_cgroup_count_children(memcg
);
2508 oldusage
= page_counter_read(&memcg
->memsw
);
2511 if (signal_pending(current
)) {
2516 mutex_lock(&memcg_limit_mutex
);
2517 if (limit
< memcg
->memory
.limit
) {
2518 mutex_unlock(&memcg_limit_mutex
);
2522 if (limit
> memcg
->memsw
.limit
)
2524 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2525 mutex_unlock(&memcg_limit_mutex
);
2530 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2532 curusage
= page_counter_read(&memcg
->memsw
);
2533 /* Usage is reduced ? */
2534 if (curusage
>= oldusage
)
2537 oldusage
= curusage
;
2538 } while (retry_count
);
2540 if (!ret
&& enlarge
)
2541 memcg_oom_recover(memcg
);
2546 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
2548 unsigned long *total_scanned
)
2550 unsigned long nr_reclaimed
= 0;
2551 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
2552 unsigned long reclaimed
;
2554 struct mem_cgroup_tree_per_node
*mctz
;
2555 unsigned long excess
;
2556 unsigned long nr_scanned
;
2561 mctz
= soft_limit_tree_node(pgdat
->node_id
);
2563 * This loop can run a while, specially if mem_cgroup's continuously
2564 * keep exceeding their soft limit and putting the system under
2571 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2576 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
2577 gfp_mask
, &nr_scanned
);
2578 nr_reclaimed
+= reclaimed
;
2579 *total_scanned
+= nr_scanned
;
2580 spin_lock_irq(&mctz
->lock
);
2581 __mem_cgroup_remove_exceeded(mz
, mctz
);
2584 * If we failed to reclaim anything from this memory cgroup
2585 * it is time to move on to the next cgroup
2589 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2591 excess
= soft_limit_excess(mz
->memcg
);
2593 * One school of thought says that we should not add
2594 * back the node to the tree if reclaim returns 0.
2595 * But our reclaim could return 0, simply because due
2596 * to priority we are exposing a smaller subset of
2597 * memory to reclaim from. Consider this as a longer
2600 /* If excess == 0, no tree ops */
2601 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2602 spin_unlock_irq(&mctz
->lock
);
2603 css_put(&mz
->memcg
->css
);
2606 * Could not reclaim anything and there are no more
2607 * mem cgroups to try or we seem to be looping without
2608 * reclaiming anything.
2610 if (!nr_reclaimed
&&
2612 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2614 } while (!nr_reclaimed
);
2616 css_put(&next_mz
->memcg
->css
);
2617 return nr_reclaimed
;
2621 * Test whether @memcg has children, dead or alive. Note that this
2622 * function doesn't care whether @memcg has use_hierarchy enabled and
2623 * returns %true if there are child csses according to the cgroup
2624 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2626 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2631 ret
= css_next_child(NULL
, &memcg
->css
);
2637 * Reclaims as many pages from the given memcg as possible.
2639 * Caller is responsible for holding css reference for memcg.
2641 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2643 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2645 /* we call try-to-free pages for make this cgroup empty */
2646 lru_add_drain_all();
2647 /* try to free all pages in this cgroup */
2648 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2651 if (signal_pending(current
))
2654 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2658 /* maybe some writeback is necessary */
2659 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2667 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
2668 char *buf
, size_t nbytes
,
2671 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2673 if (mem_cgroup_is_root(memcg
))
2675 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
2678 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
2681 return mem_cgroup_from_css(css
)->use_hierarchy
;
2684 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
2685 struct cftype
*cft
, u64 val
)
2688 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2689 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
2691 if (memcg
->use_hierarchy
== val
)
2695 * If parent's use_hierarchy is set, we can't make any modifications
2696 * in the child subtrees. If it is unset, then the change can
2697 * occur, provided the current cgroup has no children.
2699 * For the root cgroup, parent_mem is NULL, we allow value to be
2700 * set if there are no children.
2702 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
2703 (val
== 1 || val
== 0)) {
2704 if (!memcg_has_children(memcg
))
2705 memcg
->use_hierarchy
= val
;
2714 static void tree_stat(struct mem_cgroup
*memcg
, unsigned long *stat
)
2716 struct mem_cgroup
*iter
;
2719 memset(stat
, 0, sizeof(*stat
) * MEMCG_NR_STAT
);
2721 for_each_mem_cgroup_tree(iter
, memcg
) {
2722 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
2723 stat
[i
] += mem_cgroup_read_stat(iter
, i
);
2727 static void tree_events(struct mem_cgroup
*memcg
, unsigned long *events
)
2729 struct mem_cgroup
*iter
;
2732 memset(events
, 0, sizeof(*events
) * MEMCG_NR_EVENTS
);
2734 for_each_mem_cgroup_tree(iter
, memcg
) {
2735 for (i
= 0; i
< MEMCG_NR_EVENTS
; i
++)
2736 events
[i
] += mem_cgroup_read_events(iter
, i
);
2740 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
2742 unsigned long val
= 0;
2744 if (mem_cgroup_is_root(memcg
)) {
2745 struct mem_cgroup
*iter
;
2747 for_each_mem_cgroup_tree(iter
, memcg
) {
2748 val
+= mem_cgroup_read_stat(iter
,
2749 MEM_CGROUP_STAT_CACHE
);
2750 val
+= mem_cgroup_read_stat(iter
,
2751 MEM_CGROUP_STAT_RSS
);
2753 val
+= mem_cgroup_read_stat(iter
,
2754 MEM_CGROUP_STAT_SWAP
);
2758 val
= page_counter_read(&memcg
->memory
);
2760 val
= page_counter_read(&memcg
->memsw
);
2773 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
2776 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2777 struct page_counter
*counter
;
2779 switch (MEMFILE_TYPE(cft
->private)) {
2781 counter
= &memcg
->memory
;
2784 counter
= &memcg
->memsw
;
2787 counter
= &memcg
->kmem
;
2790 counter
= &memcg
->tcpmem
;
2796 switch (MEMFILE_ATTR(cft
->private)) {
2798 if (counter
== &memcg
->memory
)
2799 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
2800 if (counter
== &memcg
->memsw
)
2801 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
2802 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
2804 return (u64
)counter
->limit
* PAGE_SIZE
;
2806 return (u64
)counter
->watermark
* PAGE_SIZE
;
2808 return counter
->failcnt
;
2809 case RES_SOFT_LIMIT
:
2810 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
2817 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2821 if (cgroup_memory_nokmem
)
2824 BUG_ON(memcg
->kmemcg_id
>= 0);
2825 BUG_ON(memcg
->kmem_state
);
2827 memcg_id
= memcg_alloc_cache_id();
2831 static_branch_inc(&memcg_kmem_enabled_key
);
2833 * A memory cgroup is considered kmem-online as soon as it gets
2834 * kmemcg_id. Setting the id after enabling static branching will
2835 * guarantee no one starts accounting before all call sites are
2838 memcg
->kmemcg_id
= memcg_id
;
2839 memcg
->kmem_state
= KMEM_ONLINE
;
2844 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2846 struct cgroup_subsys_state
*css
;
2847 struct mem_cgroup
*parent
, *child
;
2850 if (memcg
->kmem_state
!= KMEM_ONLINE
)
2853 * Clear the online state before clearing memcg_caches array
2854 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2855 * guarantees that no cache will be created for this cgroup
2856 * after we are done (see memcg_create_kmem_cache()).
2858 memcg
->kmem_state
= KMEM_ALLOCATED
;
2860 memcg_deactivate_kmem_caches(memcg
);
2862 kmemcg_id
= memcg
->kmemcg_id
;
2863 BUG_ON(kmemcg_id
< 0);
2865 parent
= parent_mem_cgroup(memcg
);
2867 parent
= root_mem_cgroup
;
2870 * Change kmemcg_id of this cgroup and all its descendants to the
2871 * parent's id, and then move all entries from this cgroup's list_lrus
2872 * to ones of the parent. After we have finished, all list_lrus
2873 * corresponding to this cgroup are guaranteed to remain empty. The
2874 * ordering is imposed by list_lru_node->lock taken by
2875 * memcg_drain_all_list_lrus().
2877 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2878 css_for_each_descendant_pre(css
, &memcg
->css
) {
2879 child
= mem_cgroup_from_css(css
);
2880 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
2881 child
->kmemcg_id
= parent
->kmemcg_id
;
2882 if (!memcg
->use_hierarchy
)
2887 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
2889 memcg_free_cache_id(kmemcg_id
);
2892 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2894 /* css_alloc() failed, offlining didn't happen */
2895 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
2896 memcg_offline_kmem(memcg
);
2898 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
2899 memcg_destroy_kmem_caches(memcg
);
2900 static_branch_dec(&memcg_kmem_enabled_key
);
2901 WARN_ON(page_counter_read(&memcg
->kmem
));
2905 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2909 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2912 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2915 #endif /* !CONFIG_SLOB */
2917 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2918 unsigned long limit
)
2922 mutex_lock(&memcg_limit_mutex
);
2923 ret
= page_counter_limit(&memcg
->kmem
, limit
);
2924 mutex_unlock(&memcg_limit_mutex
);
2928 static int memcg_update_tcp_limit(struct mem_cgroup
*memcg
, unsigned long limit
)
2932 mutex_lock(&memcg_limit_mutex
);
2934 ret
= page_counter_limit(&memcg
->tcpmem
, limit
);
2938 if (!memcg
->tcpmem_active
) {
2940 * The active flag needs to be written after the static_key
2941 * update. This is what guarantees that the socket activation
2942 * function is the last one to run. See sock_update_memcg() for
2943 * details, and note that we don't mark any socket as belonging
2944 * to this memcg until that flag is up.
2946 * We need to do this, because static_keys will span multiple
2947 * sites, but we can't control their order. If we mark a socket
2948 * as accounted, but the accounting functions are not patched in
2949 * yet, we'll lose accounting.
2951 * We never race with the readers in sock_update_memcg(),
2952 * because when this value change, the code to process it is not
2955 static_branch_inc(&memcg_sockets_enabled_key
);
2956 memcg
->tcpmem_active
= true;
2959 mutex_unlock(&memcg_limit_mutex
);
2964 * The user of this function is...
2967 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
2968 char *buf
, size_t nbytes
, loff_t off
)
2970 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2971 unsigned long nr_pages
;
2974 buf
= strstrip(buf
);
2975 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
2979 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
2981 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
2985 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
2987 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
2990 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
2993 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
2996 ret
= memcg_update_tcp_limit(memcg
, nr_pages
);
3000 case RES_SOFT_LIMIT
:
3001 memcg
->soft_limit
= nr_pages
;
3005 return ret
?: nbytes
;
3008 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3009 size_t nbytes
, loff_t off
)
3011 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3012 struct page_counter
*counter
;
3014 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3016 counter
= &memcg
->memory
;
3019 counter
= &memcg
->memsw
;
3022 counter
= &memcg
->kmem
;
3025 counter
= &memcg
->tcpmem
;
3031 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3033 page_counter_reset_watermark(counter
);
3036 counter
->failcnt
= 0;
3045 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3048 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3052 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3053 struct cftype
*cft
, u64 val
)
3055 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3057 if (val
& ~MOVE_MASK
)
3061 * No kind of locking is needed in here, because ->can_attach() will
3062 * check this value once in the beginning of the process, and then carry
3063 * on with stale data. This means that changes to this value will only
3064 * affect task migrations starting after the change.
3066 memcg
->move_charge_at_immigrate
= val
;
3070 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3071 struct cftype
*cft
, u64 val
)
3078 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3082 unsigned int lru_mask
;
3085 static const struct numa_stat stats
[] = {
3086 { "total", LRU_ALL
},
3087 { "file", LRU_ALL_FILE
},
3088 { "anon", LRU_ALL_ANON
},
3089 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3091 const struct numa_stat
*stat
;
3094 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3096 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3097 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3098 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3099 for_each_node_state(nid
, N_MEMORY
) {
3100 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3102 seq_printf(m
, " N%d=%lu", nid
, nr
);
3107 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3108 struct mem_cgroup
*iter
;
3111 for_each_mem_cgroup_tree(iter
, memcg
)
3112 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3113 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3114 for_each_node_state(nid
, N_MEMORY
) {
3116 for_each_mem_cgroup_tree(iter
, memcg
)
3117 nr
+= mem_cgroup_node_nr_lru_pages(
3118 iter
, nid
, stat
->lru_mask
);
3119 seq_printf(m
, " N%d=%lu", nid
, nr
);
3126 #endif /* CONFIG_NUMA */
3128 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3130 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3131 unsigned long memory
, memsw
;
3132 struct mem_cgroup
*mi
;
3135 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3136 MEM_CGROUP_STAT_NSTATS
);
3137 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3138 MEM_CGROUP_EVENTS_NSTATS
);
3139 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3141 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3142 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3144 seq_printf(m
, "%s %lu\n", mem_cgroup_stat_names
[i
],
3145 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3148 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3149 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3150 mem_cgroup_read_events(memcg
, i
));
3152 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3153 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3154 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3156 /* Hierarchical information */
3157 memory
= memsw
= PAGE_COUNTER_MAX
;
3158 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3159 memory
= min(memory
, mi
->memory
.limit
);
3160 memsw
= min(memsw
, mi
->memsw
.limit
);
3162 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3163 (u64
)memory
* PAGE_SIZE
);
3164 if (do_memsw_account())
3165 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3166 (u64
)memsw
* PAGE_SIZE
);
3168 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3169 unsigned long long val
= 0;
3171 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3173 for_each_mem_cgroup_tree(mi
, memcg
)
3174 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3175 seq_printf(m
, "total_%s %llu\n", mem_cgroup_stat_names
[i
], val
);
3178 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3179 unsigned long long val
= 0;
3181 for_each_mem_cgroup_tree(mi
, memcg
)
3182 val
+= mem_cgroup_read_events(mi
, i
);
3183 seq_printf(m
, "total_%s %llu\n",
3184 mem_cgroup_events_names
[i
], val
);
3187 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3188 unsigned long long val
= 0;
3190 for_each_mem_cgroup_tree(mi
, memcg
)
3191 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3192 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3195 #ifdef CONFIG_DEBUG_VM
3198 struct mem_cgroup_per_node
*mz
;
3199 struct zone_reclaim_stat
*rstat
;
3200 unsigned long recent_rotated
[2] = {0, 0};
3201 unsigned long recent_scanned
[2] = {0, 0};
3203 for_each_online_pgdat(pgdat
) {
3204 mz
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
3205 rstat
= &mz
->lruvec
.reclaim_stat
;
3207 recent_rotated
[0] += rstat
->recent_rotated
[0];
3208 recent_rotated
[1] += rstat
->recent_rotated
[1];
3209 recent_scanned
[0] += rstat
->recent_scanned
[0];
3210 recent_scanned
[1] += rstat
->recent_scanned
[1];
3212 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3213 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3214 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3215 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3222 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3225 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3227 return mem_cgroup_swappiness(memcg
);
3230 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3231 struct cftype
*cft
, u64 val
)
3233 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3239 memcg
->swappiness
= val
;
3241 vm_swappiness
= val
;
3246 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3248 struct mem_cgroup_threshold_ary
*t
;
3249 unsigned long usage
;
3254 t
= rcu_dereference(memcg
->thresholds
.primary
);
3256 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3261 usage
= mem_cgroup_usage(memcg
, swap
);
3264 * current_threshold points to threshold just below or equal to usage.
3265 * If it's not true, a threshold was crossed after last
3266 * call of __mem_cgroup_threshold().
3268 i
= t
->current_threshold
;
3271 * Iterate backward over array of thresholds starting from
3272 * current_threshold and check if a threshold is crossed.
3273 * If none of thresholds below usage is crossed, we read
3274 * only one element of the array here.
3276 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3277 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3279 /* i = current_threshold + 1 */
3283 * Iterate forward over array of thresholds starting from
3284 * current_threshold+1 and check if a threshold is crossed.
3285 * If none of thresholds above usage is crossed, we read
3286 * only one element of the array here.
3288 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3289 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3291 /* Update current_threshold */
3292 t
->current_threshold
= i
- 1;
3297 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3300 __mem_cgroup_threshold(memcg
, false);
3301 if (do_memsw_account())
3302 __mem_cgroup_threshold(memcg
, true);
3304 memcg
= parent_mem_cgroup(memcg
);
3308 static int compare_thresholds(const void *a
, const void *b
)
3310 const struct mem_cgroup_threshold
*_a
= a
;
3311 const struct mem_cgroup_threshold
*_b
= b
;
3313 if (_a
->threshold
> _b
->threshold
)
3316 if (_a
->threshold
< _b
->threshold
)
3322 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3324 struct mem_cgroup_eventfd_list
*ev
;
3326 spin_lock(&memcg_oom_lock
);
3328 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3329 eventfd_signal(ev
->eventfd
, 1);
3331 spin_unlock(&memcg_oom_lock
);
3335 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3337 struct mem_cgroup
*iter
;
3339 for_each_mem_cgroup_tree(iter
, memcg
)
3340 mem_cgroup_oom_notify_cb(iter
);
3343 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3344 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3346 struct mem_cgroup_thresholds
*thresholds
;
3347 struct mem_cgroup_threshold_ary
*new;
3348 unsigned long threshold
;
3349 unsigned long usage
;
3352 ret
= page_counter_memparse(args
, "-1", &threshold
);
3356 mutex_lock(&memcg
->thresholds_lock
);
3359 thresholds
= &memcg
->thresholds
;
3360 usage
= mem_cgroup_usage(memcg
, false);
3361 } else if (type
== _MEMSWAP
) {
3362 thresholds
= &memcg
->memsw_thresholds
;
3363 usage
= mem_cgroup_usage(memcg
, true);
3367 /* Check if a threshold crossed before adding a new one */
3368 if (thresholds
->primary
)
3369 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3371 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3373 /* Allocate memory for new array of thresholds */
3374 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3382 /* Copy thresholds (if any) to new array */
3383 if (thresholds
->primary
) {
3384 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3385 sizeof(struct mem_cgroup_threshold
));
3388 /* Add new threshold */
3389 new->entries
[size
- 1].eventfd
= eventfd
;
3390 new->entries
[size
- 1].threshold
= threshold
;
3392 /* Sort thresholds. Registering of new threshold isn't time-critical */
3393 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3394 compare_thresholds
, NULL
);
3396 /* Find current threshold */
3397 new->current_threshold
= -1;
3398 for (i
= 0; i
< size
; i
++) {
3399 if (new->entries
[i
].threshold
<= usage
) {
3401 * new->current_threshold will not be used until
3402 * rcu_assign_pointer(), so it's safe to increment
3405 ++new->current_threshold
;
3410 /* Free old spare buffer and save old primary buffer as spare */
3411 kfree(thresholds
->spare
);
3412 thresholds
->spare
= thresholds
->primary
;
3414 rcu_assign_pointer(thresholds
->primary
, new);
3416 /* To be sure that nobody uses thresholds */
3420 mutex_unlock(&memcg
->thresholds_lock
);
3425 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3426 struct eventfd_ctx
*eventfd
, const char *args
)
3428 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3431 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3432 struct eventfd_ctx
*eventfd
, const char *args
)
3434 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3437 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3438 struct eventfd_ctx
*eventfd
, enum res_type type
)
3440 struct mem_cgroup_thresholds
*thresholds
;
3441 struct mem_cgroup_threshold_ary
*new;
3442 unsigned long usage
;
3445 mutex_lock(&memcg
->thresholds_lock
);
3448 thresholds
= &memcg
->thresholds
;
3449 usage
= mem_cgroup_usage(memcg
, false);
3450 } else if (type
== _MEMSWAP
) {
3451 thresholds
= &memcg
->memsw_thresholds
;
3452 usage
= mem_cgroup_usage(memcg
, true);
3456 if (!thresholds
->primary
)
3459 /* Check if a threshold crossed before removing */
3460 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3462 /* Calculate new number of threshold */
3464 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3465 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3469 new = thresholds
->spare
;
3471 /* Set thresholds array to NULL if we don't have thresholds */
3480 /* Copy thresholds and find current threshold */
3481 new->current_threshold
= -1;
3482 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3483 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3486 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3487 if (new->entries
[j
].threshold
<= usage
) {
3489 * new->current_threshold will not be used
3490 * until rcu_assign_pointer(), so it's safe to increment
3493 ++new->current_threshold
;
3499 /* Swap primary and spare array */
3500 thresholds
->spare
= thresholds
->primary
;
3502 rcu_assign_pointer(thresholds
->primary
, new);
3504 /* To be sure that nobody uses thresholds */
3507 /* If all events are unregistered, free the spare array */
3509 kfree(thresholds
->spare
);
3510 thresholds
->spare
= NULL
;
3513 mutex_unlock(&memcg
->thresholds_lock
);
3516 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3517 struct eventfd_ctx
*eventfd
)
3519 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3522 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3523 struct eventfd_ctx
*eventfd
)
3525 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3528 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3529 struct eventfd_ctx
*eventfd
, const char *args
)
3531 struct mem_cgroup_eventfd_list
*event
;
3533 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3537 spin_lock(&memcg_oom_lock
);
3539 event
->eventfd
= eventfd
;
3540 list_add(&event
->list
, &memcg
->oom_notify
);
3542 /* already in OOM ? */
3543 if (memcg
->under_oom
)
3544 eventfd_signal(eventfd
, 1);
3545 spin_unlock(&memcg_oom_lock
);
3550 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3551 struct eventfd_ctx
*eventfd
)
3553 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3555 spin_lock(&memcg_oom_lock
);
3557 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3558 if (ev
->eventfd
== eventfd
) {
3559 list_del(&ev
->list
);
3564 spin_unlock(&memcg_oom_lock
);
3567 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3569 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3571 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3572 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3576 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3577 struct cftype
*cft
, u64 val
)
3579 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3581 /* cannot set to root cgroup and only 0 and 1 are allowed */
3582 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3585 memcg
->oom_kill_disable
= val
;
3587 memcg_oom_recover(memcg
);
3592 #ifdef CONFIG_CGROUP_WRITEBACK
3594 struct list_head
*mem_cgroup_cgwb_list(struct mem_cgroup
*memcg
)
3596 return &memcg
->cgwb_list
;
3599 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3601 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3604 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3606 wb_domain_exit(&memcg
->cgwb_domain
);
3609 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3611 wb_domain_size_changed(&memcg
->cgwb_domain
);
3614 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3616 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3618 if (!memcg
->css
.parent
)
3621 return &memcg
->cgwb_domain
;
3625 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3626 * @wb: bdi_writeback in question
3627 * @pfilepages: out parameter for number of file pages
3628 * @pheadroom: out parameter for number of allocatable pages according to memcg
3629 * @pdirty: out parameter for number of dirty pages
3630 * @pwriteback: out parameter for number of pages under writeback
3632 * Determine the numbers of file, headroom, dirty, and writeback pages in
3633 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3634 * is a bit more involved.
3636 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3637 * headroom is calculated as the lowest headroom of itself and the
3638 * ancestors. Note that this doesn't consider the actual amount of
3639 * available memory in the system. The caller should further cap
3640 * *@pheadroom accordingly.
3642 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
3643 unsigned long *pheadroom
, unsigned long *pdirty
,
3644 unsigned long *pwriteback
)
3646 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3647 struct mem_cgroup
*parent
;
3649 *pdirty
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
3651 /* this should eventually include NR_UNSTABLE_NFS */
3652 *pwriteback
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
3653 *pfilepages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
3654 (1 << LRU_ACTIVE_FILE
));
3655 *pheadroom
= PAGE_COUNTER_MAX
;
3657 while ((parent
= parent_mem_cgroup(memcg
))) {
3658 unsigned long ceiling
= min(memcg
->memory
.limit
, memcg
->high
);
3659 unsigned long used
= page_counter_read(&memcg
->memory
);
3661 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
3666 #else /* CONFIG_CGROUP_WRITEBACK */
3668 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3673 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3677 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3681 #endif /* CONFIG_CGROUP_WRITEBACK */
3684 * DO NOT USE IN NEW FILES.
3686 * "cgroup.event_control" implementation.
3688 * This is way over-engineered. It tries to support fully configurable
3689 * events for each user. Such level of flexibility is completely
3690 * unnecessary especially in the light of the planned unified hierarchy.
3692 * Please deprecate this and replace with something simpler if at all
3697 * Unregister event and free resources.
3699 * Gets called from workqueue.
3701 static void memcg_event_remove(struct work_struct
*work
)
3703 struct mem_cgroup_event
*event
=
3704 container_of(work
, struct mem_cgroup_event
, remove
);
3705 struct mem_cgroup
*memcg
= event
->memcg
;
3707 remove_wait_queue(event
->wqh
, &event
->wait
);
3709 event
->unregister_event(memcg
, event
->eventfd
);
3711 /* Notify userspace the event is going away. */
3712 eventfd_signal(event
->eventfd
, 1);
3714 eventfd_ctx_put(event
->eventfd
);
3716 css_put(&memcg
->css
);
3720 * Gets called on POLLHUP on eventfd when user closes it.
3722 * Called with wqh->lock held and interrupts disabled.
3724 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
3725 int sync
, void *key
)
3727 struct mem_cgroup_event
*event
=
3728 container_of(wait
, struct mem_cgroup_event
, wait
);
3729 struct mem_cgroup
*memcg
= event
->memcg
;
3730 unsigned long flags
= (unsigned long)key
;
3732 if (flags
& POLLHUP
) {
3734 * If the event has been detached at cgroup removal, we
3735 * can simply return knowing the other side will cleanup
3738 * We can't race against event freeing since the other
3739 * side will require wqh->lock via remove_wait_queue(),
3742 spin_lock(&memcg
->event_list_lock
);
3743 if (!list_empty(&event
->list
)) {
3744 list_del_init(&event
->list
);
3746 * We are in atomic context, but cgroup_event_remove()
3747 * may sleep, so we have to call it in workqueue.
3749 schedule_work(&event
->remove
);
3751 spin_unlock(&memcg
->event_list_lock
);
3757 static void memcg_event_ptable_queue_proc(struct file
*file
,
3758 wait_queue_head_t
*wqh
, poll_table
*pt
)
3760 struct mem_cgroup_event
*event
=
3761 container_of(pt
, struct mem_cgroup_event
, pt
);
3764 add_wait_queue(wqh
, &event
->wait
);
3768 * DO NOT USE IN NEW FILES.
3770 * Parse input and register new cgroup event handler.
3772 * Input must be in format '<event_fd> <control_fd> <args>'.
3773 * Interpretation of args is defined by control file implementation.
3775 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
3776 char *buf
, size_t nbytes
, loff_t off
)
3778 struct cgroup_subsys_state
*css
= of_css(of
);
3779 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3780 struct mem_cgroup_event
*event
;
3781 struct cgroup_subsys_state
*cfile_css
;
3782 unsigned int efd
, cfd
;
3789 buf
= strstrip(buf
);
3791 efd
= simple_strtoul(buf
, &endp
, 10);
3796 cfd
= simple_strtoul(buf
, &endp
, 10);
3797 if ((*endp
!= ' ') && (*endp
!= '\0'))
3801 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
3805 event
->memcg
= memcg
;
3806 INIT_LIST_HEAD(&event
->list
);
3807 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
3808 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
3809 INIT_WORK(&event
->remove
, memcg_event_remove
);
3817 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
3818 if (IS_ERR(event
->eventfd
)) {
3819 ret
= PTR_ERR(event
->eventfd
);
3826 goto out_put_eventfd
;
3829 /* the process need read permission on control file */
3830 /* AV: shouldn't we check that it's been opened for read instead? */
3831 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
3836 * Determine the event callbacks and set them in @event. This used
3837 * to be done via struct cftype but cgroup core no longer knows
3838 * about these events. The following is crude but the whole thing
3839 * is for compatibility anyway.
3841 * DO NOT ADD NEW FILES.
3843 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
3845 if (!strcmp(name
, "memory.usage_in_bytes")) {
3846 event
->register_event
= mem_cgroup_usage_register_event
;
3847 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
3848 } else if (!strcmp(name
, "memory.oom_control")) {
3849 event
->register_event
= mem_cgroup_oom_register_event
;
3850 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
3851 } else if (!strcmp(name
, "memory.pressure_level")) {
3852 event
->register_event
= vmpressure_register_event
;
3853 event
->unregister_event
= vmpressure_unregister_event
;
3854 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
3855 event
->register_event
= memsw_cgroup_usage_register_event
;
3856 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
3863 * Verify @cfile should belong to @css. Also, remaining events are
3864 * automatically removed on cgroup destruction but the removal is
3865 * asynchronous, so take an extra ref on @css.
3867 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
3868 &memory_cgrp_subsys
);
3870 if (IS_ERR(cfile_css
))
3872 if (cfile_css
!= css
) {
3877 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
3881 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
3883 spin_lock(&memcg
->event_list_lock
);
3884 list_add(&event
->list
, &memcg
->event_list
);
3885 spin_unlock(&memcg
->event_list_lock
);
3897 eventfd_ctx_put(event
->eventfd
);
3906 static struct cftype mem_cgroup_legacy_files
[] = {
3908 .name
= "usage_in_bytes",
3909 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
3910 .read_u64
= mem_cgroup_read_u64
,
3913 .name
= "max_usage_in_bytes",
3914 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
3915 .write
= mem_cgroup_reset
,
3916 .read_u64
= mem_cgroup_read_u64
,
3919 .name
= "limit_in_bytes",
3920 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
3921 .write
= mem_cgroup_write
,
3922 .read_u64
= mem_cgroup_read_u64
,
3925 .name
= "soft_limit_in_bytes",
3926 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
3927 .write
= mem_cgroup_write
,
3928 .read_u64
= mem_cgroup_read_u64
,
3932 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
3933 .write
= mem_cgroup_reset
,
3934 .read_u64
= mem_cgroup_read_u64
,
3938 .seq_show
= memcg_stat_show
,
3941 .name
= "force_empty",
3942 .write
= mem_cgroup_force_empty_write
,
3945 .name
= "use_hierarchy",
3946 .write_u64
= mem_cgroup_hierarchy_write
,
3947 .read_u64
= mem_cgroup_hierarchy_read
,
3950 .name
= "cgroup.event_control", /* XXX: for compat */
3951 .write
= memcg_write_event_control
,
3952 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
3955 .name
= "swappiness",
3956 .read_u64
= mem_cgroup_swappiness_read
,
3957 .write_u64
= mem_cgroup_swappiness_write
,
3960 .name
= "move_charge_at_immigrate",
3961 .read_u64
= mem_cgroup_move_charge_read
,
3962 .write_u64
= mem_cgroup_move_charge_write
,
3965 .name
= "oom_control",
3966 .seq_show
= mem_cgroup_oom_control_read
,
3967 .write_u64
= mem_cgroup_oom_control_write
,
3968 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
3971 .name
= "pressure_level",
3975 .name
= "numa_stat",
3976 .seq_show
= memcg_numa_stat_show
,
3980 .name
= "kmem.limit_in_bytes",
3981 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
3982 .write
= mem_cgroup_write
,
3983 .read_u64
= mem_cgroup_read_u64
,
3986 .name
= "kmem.usage_in_bytes",
3987 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
3988 .read_u64
= mem_cgroup_read_u64
,
3991 .name
= "kmem.failcnt",
3992 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
3993 .write
= mem_cgroup_reset
,
3994 .read_u64
= mem_cgroup_read_u64
,
3997 .name
= "kmem.max_usage_in_bytes",
3998 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
3999 .write
= mem_cgroup_reset
,
4000 .read_u64
= mem_cgroup_read_u64
,
4002 #ifdef CONFIG_SLABINFO
4004 .name
= "kmem.slabinfo",
4005 .seq_start
= slab_start
,
4006 .seq_next
= slab_next
,
4007 .seq_stop
= slab_stop
,
4008 .seq_show
= memcg_slab_show
,
4012 .name
= "kmem.tcp.limit_in_bytes",
4013 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4014 .write
= mem_cgroup_write
,
4015 .read_u64
= mem_cgroup_read_u64
,
4018 .name
= "kmem.tcp.usage_in_bytes",
4019 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4020 .read_u64
= mem_cgroup_read_u64
,
4023 .name
= "kmem.tcp.failcnt",
4024 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4025 .write
= mem_cgroup_reset
,
4026 .read_u64
= mem_cgroup_read_u64
,
4029 .name
= "kmem.tcp.max_usage_in_bytes",
4030 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4031 .write
= mem_cgroup_reset
,
4032 .read_u64
= mem_cgroup_read_u64
,
4034 { }, /* terminate */
4038 * Private memory cgroup IDR
4040 * Swap-out records and page cache shadow entries need to store memcg
4041 * references in constrained space, so we maintain an ID space that is
4042 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4043 * memory-controlled cgroups to 64k.
4045 * However, there usually are many references to the oflline CSS after
4046 * the cgroup has been destroyed, such as page cache or reclaimable
4047 * slab objects, that don't need to hang on to the ID. We want to keep
4048 * those dead CSS from occupying IDs, or we might quickly exhaust the
4049 * relatively small ID space and prevent the creation of new cgroups
4050 * even when there are much fewer than 64k cgroups - possibly none.
4052 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4053 * be freed and recycled when it's no longer needed, which is usually
4054 * when the CSS is offlined.
4056 * The only exception to that are records of swapped out tmpfs/shmem
4057 * pages that need to be attributed to live ancestors on swapin. But
4058 * those references are manageable from userspace.
4061 static DEFINE_IDR(mem_cgroup_idr
);
4063 static void mem_cgroup_id_get(struct mem_cgroup
*memcg
)
4065 atomic_inc(&memcg
->id
.ref
);
4068 static void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
4070 if (atomic_dec_and_test(&memcg
->id
.ref
)) {
4071 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4074 /* Memcg ID pins CSS */
4075 css_put(&memcg
->css
);
4080 * mem_cgroup_from_id - look up a memcg from a memcg id
4081 * @id: the memcg id to look up
4083 * Caller must hold rcu_read_lock().
4085 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
4087 WARN_ON_ONCE(!rcu_read_lock_held());
4088 return idr_find(&mem_cgroup_idr
, id
);
4091 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4093 struct mem_cgroup_per_node
*pn
;
4096 * This routine is called against possible nodes.
4097 * But it's BUG to call kmalloc() against offline node.
4099 * TODO: this routine can waste much memory for nodes which will
4100 * never be onlined. It's better to use memory hotplug callback
4103 if (!node_state(node
, N_NORMAL_MEMORY
))
4105 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4109 lruvec_init(&pn
->lruvec
);
4110 pn
->usage_in_excess
= 0;
4111 pn
->on_tree
= false;
4114 memcg
->nodeinfo
[node
] = pn
;
4118 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4120 kfree(memcg
->nodeinfo
[node
]);
4123 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4127 memcg_wb_domain_exit(memcg
);
4129 free_mem_cgroup_per_node_info(memcg
, node
);
4130 free_percpu(memcg
->stat
);
4134 static struct mem_cgroup
*mem_cgroup_alloc(void)
4136 struct mem_cgroup
*memcg
;
4140 size
= sizeof(struct mem_cgroup
);
4141 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4143 memcg
= kzalloc(size
, GFP_KERNEL
);
4147 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
4148 1, MEM_CGROUP_ID_MAX
,
4150 if (memcg
->id
.id
< 0)
4153 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4158 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
4161 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4164 INIT_WORK(&memcg
->high_work
, high_work_func
);
4165 memcg
->last_scanned_node
= MAX_NUMNODES
;
4166 INIT_LIST_HEAD(&memcg
->oom_notify
);
4167 mutex_init(&memcg
->thresholds_lock
);
4168 spin_lock_init(&memcg
->move_lock
);
4169 vmpressure_init(&memcg
->vmpressure
);
4170 INIT_LIST_HEAD(&memcg
->event_list
);
4171 spin_lock_init(&memcg
->event_list_lock
);
4172 memcg
->socket_pressure
= jiffies
;
4174 memcg
->kmemcg_id
= -1;
4176 #ifdef CONFIG_CGROUP_WRITEBACK
4177 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4179 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
4182 if (memcg
->id
.id
> 0)
4183 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4184 mem_cgroup_free(memcg
);
4188 static struct cgroup_subsys_state
* __ref
4189 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4191 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
4192 struct mem_cgroup
*memcg
;
4193 long error
= -ENOMEM
;
4195 memcg
= mem_cgroup_alloc();
4197 return ERR_PTR(error
);
4199 memcg
->high
= PAGE_COUNTER_MAX
;
4200 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4202 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4203 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4205 if (parent
&& parent
->use_hierarchy
) {
4206 memcg
->use_hierarchy
= true;
4207 page_counter_init(&memcg
->memory
, &parent
->memory
);
4208 page_counter_init(&memcg
->swap
, &parent
->swap
);
4209 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4210 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4211 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
4213 page_counter_init(&memcg
->memory
, NULL
);
4214 page_counter_init(&memcg
->swap
, NULL
);
4215 page_counter_init(&memcg
->memsw
, NULL
);
4216 page_counter_init(&memcg
->kmem
, NULL
);
4217 page_counter_init(&memcg
->tcpmem
, NULL
);
4219 * Deeper hierachy with use_hierarchy == false doesn't make
4220 * much sense so let cgroup subsystem know about this
4221 * unfortunate state in our controller.
4223 if (parent
!= root_mem_cgroup
)
4224 memory_cgrp_subsys
.broken_hierarchy
= true;
4227 /* The following stuff does not apply to the root */
4229 root_mem_cgroup
= memcg
;
4233 error
= memcg_online_kmem(memcg
);
4237 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4238 static_branch_inc(&memcg_sockets_enabled_key
);
4242 mem_cgroup_free(memcg
);
4243 return ERR_PTR(-ENOMEM
);
4246 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4248 /* Online state pins memcg ID, memcg ID pins CSS */
4249 mem_cgroup_id_get(mem_cgroup_from_css(css
));
4254 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4256 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4257 struct mem_cgroup_event
*event
, *tmp
;
4260 * Unregister events and notify userspace.
4261 * Notify userspace about cgroup removing only after rmdir of cgroup
4262 * directory to avoid race between userspace and kernelspace.
4264 spin_lock(&memcg
->event_list_lock
);
4265 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4266 list_del_init(&event
->list
);
4267 schedule_work(&event
->remove
);
4269 spin_unlock(&memcg
->event_list_lock
);
4271 memcg_offline_kmem(memcg
);
4272 wb_memcg_offline(memcg
);
4274 mem_cgroup_id_put(memcg
);
4277 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4279 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4281 invalidate_reclaim_iterators(memcg
);
4284 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4286 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4288 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4289 static_branch_dec(&memcg_sockets_enabled_key
);
4291 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
4292 static_branch_dec(&memcg_sockets_enabled_key
);
4294 vmpressure_cleanup(&memcg
->vmpressure
);
4295 cancel_work_sync(&memcg
->high_work
);
4296 mem_cgroup_remove_from_trees(memcg
);
4297 memcg_free_kmem(memcg
);
4298 mem_cgroup_free(memcg
);
4302 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4303 * @css: the target css
4305 * Reset the states of the mem_cgroup associated with @css. This is
4306 * invoked when the userland requests disabling on the default hierarchy
4307 * but the memcg is pinned through dependency. The memcg should stop
4308 * applying policies and should revert to the vanilla state as it may be
4309 * made visible again.
4311 * The current implementation only resets the essential configurations.
4312 * This needs to be expanded to cover all the visible parts.
4314 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4316 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4318 page_counter_limit(&memcg
->memory
, PAGE_COUNTER_MAX
);
4319 page_counter_limit(&memcg
->swap
, PAGE_COUNTER_MAX
);
4320 page_counter_limit(&memcg
->memsw
, PAGE_COUNTER_MAX
);
4321 page_counter_limit(&memcg
->kmem
, PAGE_COUNTER_MAX
);
4322 page_counter_limit(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
4324 memcg
->high
= PAGE_COUNTER_MAX
;
4325 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4326 memcg_wb_domain_size_changed(memcg
);
4330 /* Handlers for move charge at task migration. */
4331 static int mem_cgroup_do_precharge(unsigned long count
)
4335 /* Try a single bulk charge without reclaim first, kswapd may wake */
4336 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4338 mc
.precharge
+= count
;
4342 /* Try charges one by one with reclaim */
4344 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4358 enum mc_target_type
{
4364 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4365 unsigned long addr
, pte_t ptent
)
4367 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4369 if (!page
|| !page_mapped(page
))
4371 if (PageAnon(page
)) {
4372 if (!(mc
.flags
& MOVE_ANON
))
4375 if (!(mc
.flags
& MOVE_FILE
))
4378 if (!get_page_unless_zero(page
))
4385 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4386 pte_t ptent
, swp_entry_t
*entry
)
4388 struct page
*page
= NULL
;
4389 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4391 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4394 * Because lookup_swap_cache() updates some statistics counter,
4395 * we call find_get_page() with swapper_space directly.
4397 page
= find_get_page(swap_address_space(ent
), ent
.val
);
4398 if (do_memsw_account())
4399 entry
->val
= ent
.val
;
4404 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4405 pte_t ptent
, swp_entry_t
*entry
)
4411 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4412 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4414 struct page
*page
= NULL
;
4415 struct address_space
*mapping
;
4418 if (!vma
->vm_file
) /* anonymous vma */
4420 if (!(mc
.flags
& MOVE_FILE
))
4423 mapping
= vma
->vm_file
->f_mapping
;
4424 pgoff
= linear_page_index(vma
, addr
);
4426 /* page is moved even if it's not RSS of this task(page-faulted). */
4428 /* shmem/tmpfs may report page out on swap: account for that too. */
4429 if (shmem_mapping(mapping
)) {
4430 page
= find_get_entry(mapping
, pgoff
);
4431 if (radix_tree_exceptional_entry(page
)) {
4432 swp_entry_t swp
= radix_to_swp_entry(page
);
4433 if (do_memsw_account())
4435 page
= find_get_page(swap_address_space(swp
), swp
.val
);
4438 page
= find_get_page(mapping
, pgoff
);
4440 page
= find_get_page(mapping
, pgoff
);
4446 * mem_cgroup_move_account - move account of the page
4448 * @compound: charge the page as compound or small page
4449 * @from: mem_cgroup which the page is moved from.
4450 * @to: mem_cgroup which the page is moved to. @from != @to.
4452 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4454 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4457 static int mem_cgroup_move_account(struct page
*page
,
4459 struct mem_cgroup
*from
,
4460 struct mem_cgroup
*to
)
4462 unsigned long flags
;
4463 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
4467 VM_BUG_ON(from
== to
);
4468 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4469 VM_BUG_ON(compound
&& !PageTransHuge(page
));
4472 * Prevent mem_cgroup_migrate() from looking at
4473 * page->mem_cgroup of its source page while we change it.
4476 if (!trylock_page(page
))
4480 if (page
->mem_cgroup
!= from
)
4483 anon
= PageAnon(page
);
4485 spin_lock_irqsave(&from
->move_lock
, flags
);
4487 if (!anon
&& page_mapped(page
)) {
4488 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4490 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4495 * move_lock grabbed above and caller set from->moving_account, so
4496 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4497 * So mapping should be stable for dirty pages.
4499 if (!anon
&& PageDirty(page
)) {
4500 struct address_space
*mapping
= page_mapping(page
);
4502 if (mapping_cap_account_dirty(mapping
)) {
4503 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4505 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4510 if (PageWriteback(page
)) {
4511 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4513 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4518 * It is safe to change page->mem_cgroup here because the page
4519 * is referenced, charged, and isolated - we can't race with
4520 * uncharging, charging, migration, or LRU putback.
4523 /* caller should have done css_get */
4524 page
->mem_cgroup
= to
;
4525 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4529 local_irq_disable();
4530 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
4531 memcg_check_events(to
, page
);
4532 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
4533 memcg_check_events(from
, page
);
4542 * get_mctgt_type - get target type of moving charge
4543 * @vma: the vma the pte to be checked belongs
4544 * @addr: the address corresponding to the pte to be checked
4545 * @ptent: the pte to be checked
4546 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4549 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4550 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4551 * move charge. if @target is not NULL, the page is stored in target->page
4552 * with extra refcnt got(Callers should handle it).
4553 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4554 * target for charge migration. if @target is not NULL, the entry is stored
4557 * Called with pte lock held.
4560 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4561 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4563 struct page
*page
= NULL
;
4564 enum mc_target_type ret
= MC_TARGET_NONE
;
4565 swp_entry_t ent
= { .val
= 0 };
4567 if (pte_present(ptent
))
4568 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4569 else if (is_swap_pte(ptent
))
4570 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
4571 else if (pte_none(ptent
))
4572 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4574 if (!page
&& !ent
.val
)
4578 * Do only loose check w/o serialization.
4579 * mem_cgroup_move_account() checks the page is valid or
4580 * not under LRU exclusion.
4582 if (page
->mem_cgroup
== mc
.from
) {
4583 ret
= MC_TARGET_PAGE
;
4585 target
->page
= page
;
4587 if (!ret
|| !target
)
4590 /* There is a swap entry and a page doesn't exist or isn't charged */
4591 if (ent
.val
&& !ret
&&
4592 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4593 ret
= MC_TARGET_SWAP
;
4600 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4602 * We don't consider swapping or file mapped pages because THP does not
4603 * support them for now.
4604 * Caller should make sure that pmd_trans_huge(pmd) is true.
4606 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4607 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4609 struct page
*page
= NULL
;
4610 enum mc_target_type ret
= MC_TARGET_NONE
;
4612 page
= pmd_page(pmd
);
4613 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4614 if (!(mc
.flags
& MOVE_ANON
))
4616 if (page
->mem_cgroup
== mc
.from
) {
4617 ret
= MC_TARGET_PAGE
;
4620 target
->page
= page
;
4626 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4627 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4629 return MC_TARGET_NONE
;
4633 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4634 unsigned long addr
, unsigned long end
,
4635 struct mm_walk
*walk
)
4637 struct vm_area_struct
*vma
= walk
->vma
;
4641 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4643 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4644 mc
.precharge
+= HPAGE_PMD_NR
;
4649 if (pmd_trans_unstable(pmd
))
4651 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4652 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4653 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4654 mc
.precharge
++; /* increment precharge temporarily */
4655 pte_unmap_unlock(pte
- 1, ptl
);
4661 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4663 unsigned long precharge
;
4665 struct mm_walk mem_cgroup_count_precharge_walk
= {
4666 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4669 down_read(&mm
->mmap_sem
);
4670 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk
);
4671 up_read(&mm
->mmap_sem
);
4673 precharge
= mc
.precharge
;
4679 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4681 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4683 VM_BUG_ON(mc
.moving_task
);
4684 mc
.moving_task
= current
;
4685 return mem_cgroup_do_precharge(precharge
);
4688 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4689 static void __mem_cgroup_clear_mc(void)
4691 struct mem_cgroup
*from
= mc
.from
;
4692 struct mem_cgroup
*to
= mc
.to
;
4694 /* we must uncharge all the leftover precharges from mc.to */
4696 cancel_charge(mc
.to
, mc
.precharge
);
4700 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4701 * we must uncharge here.
4703 if (mc
.moved_charge
) {
4704 cancel_charge(mc
.from
, mc
.moved_charge
);
4705 mc
.moved_charge
= 0;
4707 /* we must fixup refcnts and charges */
4708 if (mc
.moved_swap
) {
4709 /* uncharge swap account from the old cgroup */
4710 if (!mem_cgroup_is_root(mc
.from
))
4711 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4714 * we charged both to->memory and to->memsw, so we
4715 * should uncharge to->memory.
4717 if (!mem_cgroup_is_root(mc
.to
))
4718 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4720 css_put_many(&mc
.from
->css
, mc
.moved_swap
);
4722 /* we've already done css_get(mc.to) */
4725 memcg_oom_recover(from
);
4726 memcg_oom_recover(to
);
4727 wake_up_all(&mc
.waitq
);
4730 static void mem_cgroup_clear_mc(void)
4732 struct mm_struct
*mm
= mc
.mm
;
4735 * we must clear moving_task before waking up waiters at the end of
4738 mc
.moving_task
= NULL
;
4739 __mem_cgroup_clear_mc();
4740 spin_lock(&mc
.lock
);
4744 spin_unlock(&mc
.lock
);
4749 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4751 struct cgroup_subsys_state
*css
;
4752 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
4753 struct mem_cgroup
*from
;
4754 struct task_struct
*leader
, *p
;
4755 struct mm_struct
*mm
;
4756 unsigned long move_flags
;
4759 /* charge immigration isn't supported on the default hierarchy */
4760 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4764 * Multi-process migrations only happen on the default hierarchy
4765 * where charge immigration is not used. Perform charge
4766 * immigration if @tset contains a leader and whine if there are
4770 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
4773 memcg
= mem_cgroup_from_css(css
);
4779 * We are now commited to this value whatever it is. Changes in this
4780 * tunable will only affect upcoming migrations, not the current one.
4781 * So we need to save it, and keep it going.
4783 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
4787 from
= mem_cgroup_from_task(p
);
4789 VM_BUG_ON(from
== memcg
);
4791 mm
= get_task_mm(p
);
4794 /* We move charges only when we move a owner of the mm */
4795 if (mm
->owner
== p
) {
4798 VM_BUG_ON(mc
.precharge
);
4799 VM_BUG_ON(mc
.moved_charge
);
4800 VM_BUG_ON(mc
.moved_swap
);
4802 spin_lock(&mc
.lock
);
4806 mc
.flags
= move_flags
;
4807 spin_unlock(&mc
.lock
);
4808 /* We set mc.moving_task later */
4810 ret
= mem_cgroup_precharge_mc(mm
);
4812 mem_cgroup_clear_mc();
4819 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4822 mem_cgroup_clear_mc();
4825 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4826 unsigned long addr
, unsigned long end
,
4827 struct mm_walk
*walk
)
4830 struct vm_area_struct
*vma
= walk
->vma
;
4833 enum mc_target_type target_type
;
4834 union mc_target target
;
4837 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4839 if (mc
.precharge
< HPAGE_PMD_NR
) {
4843 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
4844 if (target_type
== MC_TARGET_PAGE
) {
4846 if (!isolate_lru_page(page
)) {
4847 if (!mem_cgroup_move_account(page
, true,
4849 mc
.precharge
-= HPAGE_PMD_NR
;
4850 mc
.moved_charge
+= HPAGE_PMD_NR
;
4852 putback_lru_page(page
);
4860 if (pmd_trans_unstable(pmd
))
4863 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4864 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4865 pte_t ptent
= *(pte
++);
4871 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
4872 case MC_TARGET_PAGE
:
4875 * We can have a part of the split pmd here. Moving it
4876 * can be done but it would be too convoluted so simply
4877 * ignore such a partial THP and keep it in original
4878 * memcg. There should be somebody mapping the head.
4880 if (PageTransCompound(page
))
4882 if (isolate_lru_page(page
))
4884 if (!mem_cgroup_move_account(page
, false,
4887 /* we uncharge from mc.from later. */
4890 putback_lru_page(page
);
4891 put
: /* get_mctgt_type() gets the page */
4894 case MC_TARGET_SWAP
:
4896 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
4898 /* we fixup refcnts and charges later. */
4906 pte_unmap_unlock(pte
- 1, ptl
);
4911 * We have consumed all precharges we got in can_attach().
4912 * We try charge one by one, but don't do any additional
4913 * charges to mc.to if we have failed in charge once in attach()
4916 ret
= mem_cgroup_do_precharge(1);
4924 static void mem_cgroup_move_charge(void)
4926 struct mm_walk mem_cgroup_move_charge_walk
= {
4927 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4931 lru_add_drain_all();
4933 * Signal lock_page_memcg() to take the memcg's move_lock
4934 * while we're moving its pages to another memcg. Then wait
4935 * for already started RCU-only updates to finish.
4937 atomic_inc(&mc
.from
->moving_account
);
4940 if (unlikely(!down_read_trylock(&mc
.mm
->mmap_sem
))) {
4942 * Someone who are holding the mmap_sem might be waiting in
4943 * waitq. So we cancel all extra charges, wake up all waiters,
4944 * and retry. Because we cancel precharges, we might not be able
4945 * to move enough charges, but moving charge is a best-effort
4946 * feature anyway, so it wouldn't be a big problem.
4948 __mem_cgroup_clear_mc();
4953 * When we have consumed all precharges and failed in doing
4954 * additional charge, the page walk just aborts.
4956 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk
);
4957 up_read(&mc
.mm
->mmap_sem
);
4958 atomic_dec(&mc
.from
->moving_account
);
4961 static void mem_cgroup_move_task(void)
4964 mem_cgroup_move_charge();
4965 mem_cgroup_clear_mc();
4968 #else /* !CONFIG_MMU */
4969 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4973 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4976 static void mem_cgroup_move_task(void)
4982 * Cgroup retains root cgroups across [un]mount cycles making it necessary
4983 * to verify whether we're attached to the default hierarchy on each mount
4986 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
4989 * use_hierarchy is forced on the default hierarchy. cgroup core
4990 * guarantees that @root doesn't have any children, so turning it
4991 * on for the root memcg is enough.
4993 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4994 root_mem_cgroup
->use_hierarchy
= true;
4996 root_mem_cgroup
->use_hierarchy
= false;
4999 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5002 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5004 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
5007 static int memory_low_show(struct seq_file
*m
, void *v
)
5009 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5010 unsigned long low
= READ_ONCE(memcg
->low
);
5012 if (low
== PAGE_COUNTER_MAX
)
5013 seq_puts(m
, "max\n");
5015 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
5020 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5021 char *buf
, size_t nbytes
, loff_t off
)
5023 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5027 buf
= strstrip(buf
);
5028 err
= page_counter_memparse(buf
, "max", &low
);
5037 static int memory_high_show(struct seq_file
*m
, void *v
)
5039 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5040 unsigned long high
= READ_ONCE(memcg
->high
);
5042 if (high
== PAGE_COUNTER_MAX
)
5043 seq_puts(m
, "max\n");
5045 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
5050 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5051 char *buf
, size_t nbytes
, loff_t off
)
5053 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5054 unsigned long nr_pages
;
5058 buf
= strstrip(buf
);
5059 err
= page_counter_memparse(buf
, "max", &high
);
5065 nr_pages
= page_counter_read(&memcg
->memory
);
5066 if (nr_pages
> high
)
5067 try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
5070 memcg_wb_domain_size_changed(memcg
);
5074 static int memory_max_show(struct seq_file
*m
, void *v
)
5076 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5077 unsigned long max
= READ_ONCE(memcg
->memory
.limit
);
5079 if (max
== PAGE_COUNTER_MAX
)
5080 seq_puts(m
, "max\n");
5082 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5087 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5088 char *buf
, size_t nbytes
, loff_t off
)
5090 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5091 unsigned int nr_reclaims
= MEM_CGROUP_RECLAIM_RETRIES
;
5092 bool drained
= false;
5096 buf
= strstrip(buf
);
5097 err
= page_counter_memparse(buf
, "max", &max
);
5101 xchg(&memcg
->memory
.limit
, max
);
5104 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
5106 if (nr_pages
<= max
)
5109 if (signal_pending(current
)) {
5115 drain_all_stock(memcg
);
5121 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
5127 mem_cgroup_events(memcg
, MEMCG_OOM
, 1);
5128 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
5132 memcg_wb_domain_size_changed(memcg
);
5136 static int memory_events_show(struct seq_file
*m
, void *v
)
5138 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5140 seq_printf(m
, "low %lu\n", mem_cgroup_read_events(memcg
, MEMCG_LOW
));
5141 seq_printf(m
, "high %lu\n", mem_cgroup_read_events(memcg
, MEMCG_HIGH
));
5142 seq_printf(m
, "max %lu\n", mem_cgroup_read_events(memcg
, MEMCG_MAX
));
5143 seq_printf(m
, "oom %lu\n", mem_cgroup_read_events(memcg
, MEMCG_OOM
));
5148 static int memory_stat_show(struct seq_file
*m
, void *v
)
5150 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5151 unsigned long stat
[MEMCG_NR_STAT
];
5152 unsigned long events
[MEMCG_NR_EVENTS
];
5156 * Provide statistics on the state of the memory subsystem as
5157 * well as cumulative event counters that show past behavior.
5159 * This list is ordered following a combination of these gradients:
5160 * 1) generic big picture -> specifics and details
5161 * 2) reflecting userspace activity -> reflecting kernel heuristics
5163 * Current memory state:
5166 tree_stat(memcg
, stat
);
5167 tree_events(memcg
, events
);
5169 seq_printf(m
, "anon %llu\n",
5170 (u64
)stat
[MEM_CGROUP_STAT_RSS
] * PAGE_SIZE
);
5171 seq_printf(m
, "file %llu\n",
5172 (u64
)stat
[MEM_CGROUP_STAT_CACHE
] * PAGE_SIZE
);
5173 seq_printf(m
, "kernel_stack %llu\n",
5174 (u64
)stat
[MEMCG_KERNEL_STACK
] * PAGE_SIZE
);
5175 seq_printf(m
, "slab %llu\n",
5176 (u64
)(stat
[MEMCG_SLAB_RECLAIMABLE
] +
5177 stat
[MEMCG_SLAB_UNRECLAIMABLE
]) * PAGE_SIZE
);
5178 seq_printf(m
, "sock %llu\n",
5179 (u64
)stat
[MEMCG_SOCK
] * PAGE_SIZE
);
5181 seq_printf(m
, "file_mapped %llu\n",
5182 (u64
)stat
[MEM_CGROUP_STAT_FILE_MAPPED
] * PAGE_SIZE
);
5183 seq_printf(m
, "file_dirty %llu\n",
5184 (u64
)stat
[MEM_CGROUP_STAT_DIRTY
] * PAGE_SIZE
);
5185 seq_printf(m
, "file_writeback %llu\n",
5186 (u64
)stat
[MEM_CGROUP_STAT_WRITEBACK
] * PAGE_SIZE
);
5188 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5189 struct mem_cgroup
*mi
;
5190 unsigned long val
= 0;
5192 for_each_mem_cgroup_tree(mi
, memcg
)
5193 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
));
5194 seq_printf(m
, "%s %llu\n",
5195 mem_cgroup_lru_names
[i
], (u64
)val
* PAGE_SIZE
);
5198 seq_printf(m
, "slab_reclaimable %llu\n",
5199 (u64
)stat
[MEMCG_SLAB_RECLAIMABLE
] * PAGE_SIZE
);
5200 seq_printf(m
, "slab_unreclaimable %llu\n",
5201 (u64
)stat
[MEMCG_SLAB_UNRECLAIMABLE
] * PAGE_SIZE
);
5203 /* Accumulated memory events */
5205 seq_printf(m
, "pgfault %lu\n",
5206 events
[MEM_CGROUP_EVENTS_PGFAULT
]);
5207 seq_printf(m
, "pgmajfault %lu\n",
5208 events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
5213 static struct cftype memory_files
[] = {
5216 .flags
= CFTYPE_NOT_ON_ROOT
,
5217 .read_u64
= memory_current_read
,
5221 .flags
= CFTYPE_NOT_ON_ROOT
,
5222 .seq_show
= memory_low_show
,
5223 .write
= memory_low_write
,
5227 .flags
= CFTYPE_NOT_ON_ROOT
,
5228 .seq_show
= memory_high_show
,
5229 .write
= memory_high_write
,
5233 .flags
= CFTYPE_NOT_ON_ROOT
,
5234 .seq_show
= memory_max_show
,
5235 .write
= memory_max_write
,
5239 .flags
= CFTYPE_NOT_ON_ROOT
,
5240 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
5241 .seq_show
= memory_events_show
,
5245 .flags
= CFTYPE_NOT_ON_ROOT
,
5246 .seq_show
= memory_stat_show
,
5251 struct cgroup_subsys memory_cgrp_subsys
= {
5252 .css_alloc
= mem_cgroup_css_alloc
,
5253 .css_online
= mem_cgroup_css_online
,
5254 .css_offline
= mem_cgroup_css_offline
,
5255 .css_released
= mem_cgroup_css_released
,
5256 .css_free
= mem_cgroup_css_free
,
5257 .css_reset
= mem_cgroup_css_reset
,
5258 .can_attach
= mem_cgroup_can_attach
,
5259 .cancel_attach
= mem_cgroup_cancel_attach
,
5260 .post_attach
= mem_cgroup_move_task
,
5261 .bind
= mem_cgroup_bind
,
5262 .dfl_cftypes
= memory_files
,
5263 .legacy_cftypes
= mem_cgroup_legacy_files
,
5268 * mem_cgroup_low - check if memory consumption is below the normal range
5269 * @root: the highest ancestor to consider
5270 * @memcg: the memory cgroup to check
5272 * Returns %true if memory consumption of @memcg, and that of all
5273 * configurable ancestors up to @root, is below the normal range.
5275 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5277 if (mem_cgroup_disabled())
5281 * The toplevel group doesn't have a configurable range, so
5282 * it's never low when looked at directly, and it is not
5283 * considered an ancestor when assessing the hierarchy.
5286 if (memcg
== root_mem_cgroup
)
5289 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5292 while (memcg
!= root
) {
5293 memcg
= parent_mem_cgroup(memcg
);
5295 if (memcg
== root_mem_cgroup
)
5298 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5305 * mem_cgroup_try_charge - try charging a page
5306 * @page: page to charge
5307 * @mm: mm context of the victim
5308 * @gfp_mask: reclaim mode
5309 * @memcgp: charged memcg return
5310 * @compound: charge the page as compound or small page
5312 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5313 * pages according to @gfp_mask if necessary.
5315 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5316 * Otherwise, an error code is returned.
5318 * After page->mapping has been set up, the caller must finalize the
5319 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5320 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5322 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5323 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
5326 struct mem_cgroup
*memcg
= NULL
;
5327 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5330 if (mem_cgroup_disabled())
5333 if (PageSwapCache(page
)) {
5335 * Every swap fault against a single page tries to charge the
5336 * page, bail as early as possible. shmem_unuse() encounters
5337 * already charged pages, too. The USED bit is protected by
5338 * the page lock, which serializes swap cache removal, which
5339 * in turn serializes uncharging.
5341 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5342 if (page
->mem_cgroup
)
5345 if (do_swap_account
) {
5346 swp_entry_t ent
= { .val
= page_private(page
), };
5347 unsigned short id
= lookup_swap_cgroup_id(ent
);
5350 memcg
= mem_cgroup_from_id(id
);
5351 if (memcg
&& !css_tryget_online(&memcg
->css
))
5358 memcg
= get_mem_cgroup_from_mm(mm
);
5360 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5362 css_put(&memcg
->css
);
5369 * mem_cgroup_commit_charge - commit a page charge
5370 * @page: page to charge
5371 * @memcg: memcg to charge the page to
5372 * @lrucare: page might be on LRU already
5373 * @compound: charge the page as compound or small page
5375 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5376 * after page->mapping has been set up. This must happen atomically
5377 * as part of the page instantiation, i.e. under the page table lock
5378 * for anonymous pages, under the page lock for page and swap cache.
5380 * In addition, the page must not be on the LRU during the commit, to
5381 * prevent racing with task migration. If it might be, use @lrucare.
5383 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5385 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5386 bool lrucare
, bool compound
)
5388 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5390 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5391 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5393 if (mem_cgroup_disabled())
5396 * Swap faults will attempt to charge the same page multiple
5397 * times. But reuse_swap_page() might have removed the page
5398 * from swapcache already, so we can't check PageSwapCache().
5403 commit_charge(page
, memcg
, lrucare
);
5405 local_irq_disable();
5406 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
5407 memcg_check_events(memcg
, page
);
5410 if (do_memsw_account() && PageSwapCache(page
)) {
5411 swp_entry_t entry
= { .val
= page_private(page
) };
5413 * The swap entry might not get freed for a long time,
5414 * let's not wait for it. The page already received a
5415 * memory+swap charge, drop the swap entry duplicate.
5417 mem_cgroup_uncharge_swap(entry
);
5422 * mem_cgroup_cancel_charge - cancel a page charge
5423 * @page: page to charge
5424 * @memcg: memcg to charge the page to
5425 * @compound: charge the page as compound or small page
5427 * Cancel a charge transaction started by mem_cgroup_try_charge().
5429 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5432 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5434 if (mem_cgroup_disabled())
5437 * Swap faults will attempt to charge the same page multiple
5438 * times. But reuse_swap_page() might have removed the page
5439 * from swapcache already, so we can't check PageSwapCache().
5444 cancel_charge(memcg
, nr_pages
);
5447 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5448 unsigned long nr_anon
, unsigned long nr_file
,
5449 unsigned long nr_huge
, unsigned long nr_kmem
,
5450 struct page
*dummy_page
)
5452 unsigned long nr_pages
= nr_anon
+ nr_file
+ nr_kmem
;
5453 unsigned long flags
;
5455 if (!mem_cgroup_is_root(memcg
)) {
5456 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5457 if (do_memsw_account())
5458 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5459 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && nr_kmem
)
5460 page_counter_uncharge(&memcg
->kmem
, nr_kmem
);
5461 memcg_oom_recover(memcg
);
5464 local_irq_save(flags
);
5465 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5466 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5467 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5468 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5469 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5470 memcg_check_events(memcg
, dummy_page
);
5471 local_irq_restore(flags
);
5473 if (!mem_cgroup_is_root(memcg
))
5474 css_put_many(&memcg
->css
, nr_pages
);
5477 static void uncharge_list(struct list_head
*page_list
)
5479 struct mem_cgroup
*memcg
= NULL
;
5480 unsigned long nr_anon
= 0;
5481 unsigned long nr_file
= 0;
5482 unsigned long nr_huge
= 0;
5483 unsigned long nr_kmem
= 0;
5484 unsigned long pgpgout
= 0;
5485 struct list_head
*next
;
5489 * Note that the list can be a single page->lru; hence the
5490 * do-while loop instead of a simple list_for_each_entry().
5492 next
= page_list
->next
;
5494 page
= list_entry(next
, struct page
, lru
);
5495 next
= page
->lru
.next
;
5497 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5498 VM_BUG_ON_PAGE(page_count(page
), page
);
5500 if (!page
->mem_cgroup
)
5504 * Nobody should be changing or seriously looking at
5505 * page->mem_cgroup at this point, we have fully
5506 * exclusive access to the page.
5509 if (memcg
!= page
->mem_cgroup
) {
5511 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5512 nr_huge
, nr_kmem
, page
);
5513 pgpgout
= nr_anon
= nr_file
=
5514 nr_huge
= nr_kmem
= 0;
5516 memcg
= page
->mem_cgroup
;
5519 if (!PageKmemcg(page
)) {
5520 unsigned int nr_pages
= 1;
5522 if (PageTransHuge(page
)) {
5523 nr_pages
<<= compound_order(page
);
5524 nr_huge
+= nr_pages
;
5527 nr_anon
+= nr_pages
;
5529 nr_file
+= nr_pages
;
5532 nr_kmem
+= 1 << compound_order(page
);
5534 page
->mem_cgroup
= NULL
;
5535 } while (next
!= page_list
);
5538 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5539 nr_huge
, nr_kmem
, page
);
5543 * mem_cgroup_uncharge - uncharge a page
5544 * @page: page to uncharge
5546 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5547 * mem_cgroup_commit_charge().
5549 void mem_cgroup_uncharge(struct page
*page
)
5551 if (mem_cgroup_disabled())
5554 /* Don't touch page->lru of any random page, pre-check: */
5555 if (!page
->mem_cgroup
)
5558 INIT_LIST_HEAD(&page
->lru
);
5559 uncharge_list(&page
->lru
);
5563 * mem_cgroup_uncharge_list - uncharge a list of page
5564 * @page_list: list of pages to uncharge
5566 * Uncharge a list of pages previously charged with
5567 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5569 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5571 if (mem_cgroup_disabled())
5574 if (!list_empty(page_list
))
5575 uncharge_list(page_list
);
5579 * mem_cgroup_migrate - charge a page's replacement
5580 * @oldpage: currently circulating page
5581 * @newpage: replacement page
5583 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5584 * be uncharged upon free.
5586 * Both pages must be locked, @newpage->mapping must be set up.
5588 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
5590 struct mem_cgroup
*memcg
;
5591 unsigned int nr_pages
;
5593 unsigned long flags
;
5595 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5596 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5597 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5598 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5601 if (mem_cgroup_disabled())
5604 /* Page cache replacement: new page already charged? */
5605 if (newpage
->mem_cgroup
)
5608 /* Swapcache readahead pages can get replaced before being charged */
5609 memcg
= oldpage
->mem_cgroup
;
5613 /* Force-charge the new page. The old one will be freed soon */
5614 compound
= PageTransHuge(newpage
);
5615 nr_pages
= compound
? hpage_nr_pages(newpage
) : 1;
5617 page_counter_charge(&memcg
->memory
, nr_pages
);
5618 if (do_memsw_account())
5619 page_counter_charge(&memcg
->memsw
, nr_pages
);
5620 css_get_many(&memcg
->css
, nr_pages
);
5622 commit_charge(newpage
, memcg
, false);
5624 local_irq_save(flags
);
5625 mem_cgroup_charge_statistics(memcg
, newpage
, compound
, nr_pages
);
5626 memcg_check_events(memcg
, newpage
);
5627 local_irq_restore(flags
);
5630 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
5631 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
5633 void sock_update_memcg(struct sock
*sk
)
5635 struct mem_cgroup
*memcg
;
5637 /* Socket cloning can throw us here with sk_cgrp already
5638 * filled. It won't however, necessarily happen from
5639 * process context. So the test for root memcg given
5640 * the current task's memcg won't help us in this case.
5642 * Respecting the original socket's memcg is a better
5643 * decision in this case.
5646 BUG_ON(mem_cgroup_is_root(sk
->sk_memcg
));
5647 css_get(&sk
->sk_memcg
->css
);
5652 memcg
= mem_cgroup_from_task(current
);
5653 if (memcg
== root_mem_cgroup
)
5655 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
5657 if (css_tryget_online(&memcg
->css
))
5658 sk
->sk_memcg
= memcg
;
5662 EXPORT_SYMBOL(sock_update_memcg
);
5664 void sock_release_memcg(struct sock
*sk
)
5666 WARN_ON(!sk
->sk_memcg
);
5667 css_put(&sk
->sk_memcg
->css
);
5671 * mem_cgroup_charge_skmem - charge socket memory
5672 * @memcg: memcg to charge
5673 * @nr_pages: number of pages to charge
5675 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5676 * @memcg's configured limit, %false if the charge had to be forced.
5678 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5680 gfp_t gfp_mask
= GFP_KERNEL
;
5682 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5683 struct page_counter
*fail
;
5685 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
5686 memcg
->tcpmem_pressure
= 0;
5689 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
5690 memcg
->tcpmem_pressure
= 1;
5694 /* Don't block in the packet receive path */
5696 gfp_mask
= GFP_NOWAIT
;
5698 this_cpu_add(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5700 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
5703 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
5708 * mem_cgroup_uncharge_skmem - uncharge socket memory
5709 * @memcg - memcg to uncharge
5710 * @nr_pages - number of pages to uncharge
5712 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5714 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5715 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
5719 this_cpu_sub(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5721 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5722 css_put_many(&memcg
->css
, nr_pages
);
5725 static int __init
cgroup_memory(char *s
)
5729 while ((token
= strsep(&s
, ",")) != NULL
) {
5732 if (!strcmp(token
, "nosocket"))
5733 cgroup_memory_nosocket
= true;
5734 if (!strcmp(token
, "nokmem"))
5735 cgroup_memory_nokmem
= true;
5739 __setup("cgroup.memory=", cgroup_memory
);
5742 * subsys_initcall() for memory controller.
5744 * Some parts like hotcpu_notifier() have to be initialized from this context
5745 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5746 * everything that doesn't depend on a specific mem_cgroup structure should
5747 * be initialized from here.
5749 static int __init
mem_cgroup_init(void)
5753 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5755 for_each_possible_cpu(cpu
)
5756 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5759 for_each_node(node
) {
5760 struct mem_cgroup_tree_per_node
*rtpn
;
5762 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5763 node_online(node
) ? node
: NUMA_NO_NODE
);
5765 rtpn
->rb_root
= RB_ROOT
;
5766 spin_lock_init(&rtpn
->lock
);
5767 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5772 subsys_initcall(mem_cgroup_init
);
5774 #ifdef CONFIG_MEMCG_SWAP
5776 * mem_cgroup_swapout - transfer a memsw charge to swap
5777 * @page: page whose memsw charge to transfer
5778 * @entry: swap entry to move the charge to
5780 * Transfer the memsw charge of @page to @entry.
5782 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5784 struct mem_cgroup
*memcg
;
5785 unsigned short oldid
;
5787 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5788 VM_BUG_ON_PAGE(page_count(page
), page
);
5790 if (!do_memsw_account())
5793 memcg
= page
->mem_cgroup
;
5795 /* Readahead page, never charged */
5799 mem_cgroup_id_get(memcg
);
5800 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5801 VM_BUG_ON_PAGE(oldid
, page
);
5802 mem_cgroup_swap_statistics(memcg
, true);
5804 page
->mem_cgroup
= NULL
;
5806 if (!mem_cgroup_is_root(memcg
))
5807 page_counter_uncharge(&memcg
->memory
, 1);
5810 * Interrupts should be disabled here because the caller holds the
5811 * mapping->tree_lock lock which is taken with interrupts-off. It is
5812 * important here to have the interrupts disabled because it is the
5813 * only synchronisation we have for udpating the per-CPU variables.
5815 VM_BUG_ON(!irqs_disabled());
5816 mem_cgroup_charge_statistics(memcg
, page
, false, -1);
5817 memcg_check_events(memcg
, page
);
5819 if (!mem_cgroup_is_root(memcg
))
5820 css_put(&memcg
->css
);
5824 * mem_cgroup_try_charge_swap - try charging a swap entry
5825 * @page: page being added to swap
5826 * @entry: swap entry to charge
5828 * Try to charge @entry to the memcg that @page belongs to.
5830 * Returns 0 on success, -ENOMEM on failure.
5832 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
5834 struct mem_cgroup
*memcg
;
5835 struct page_counter
*counter
;
5836 unsigned short oldid
;
5838 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
5841 memcg
= page
->mem_cgroup
;
5843 /* Readahead page, never charged */
5847 if (!mem_cgroup_is_root(memcg
) &&
5848 !page_counter_try_charge(&memcg
->swap
, 1, &counter
))
5851 mem_cgroup_id_get(memcg
);
5852 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5853 VM_BUG_ON_PAGE(oldid
, page
);
5854 mem_cgroup_swap_statistics(memcg
, true);
5860 * mem_cgroup_uncharge_swap - uncharge a swap entry
5861 * @entry: swap entry to uncharge
5863 * Drop the swap charge associated with @entry.
5865 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5867 struct mem_cgroup
*memcg
;
5870 if (!do_swap_account
)
5873 id
= swap_cgroup_record(entry
, 0);
5875 memcg
= mem_cgroup_from_id(id
);
5877 if (!mem_cgroup_is_root(memcg
)) {
5878 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5879 page_counter_uncharge(&memcg
->swap
, 1);
5881 page_counter_uncharge(&memcg
->memsw
, 1);
5883 mem_cgroup_swap_statistics(memcg
, false);
5884 mem_cgroup_id_put(memcg
);
5889 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
5891 long nr_swap_pages
= get_nr_swap_pages();
5893 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5894 return nr_swap_pages
;
5895 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5896 nr_swap_pages
= min_t(long, nr_swap_pages
,
5897 READ_ONCE(memcg
->swap
.limit
) -
5898 page_counter_read(&memcg
->swap
));
5899 return nr_swap_pages
;
5902 bool mem_cgroup_swap_full(struct page
*page
)
5904 struct mem_cgroup
*memcg
;
5906 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5910 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5913 memcg
= page
->mem_cgroup
;
5917 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5918 if (page_counter_read(&memcg
->swap
) * 2 >= memcg
->swap
.limit
)
5924 /* for remember boot option*/
5925 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5926 static int really_do_swap_account __initdata
= 1;
5928 static int really_do_swap_account __initdata
;
5931 static int __init
enable_swap_account(char *s
)
5933 if (!strcmp(s
, "1"))
5934 really_do_swap_account
= 1;
5935 else if (!strcmp(s
, "0"))
5936 really_do_swap_account
= 0;
5939 __setup("swapaccount=", enable_swap_account
);
5941 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
5944 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5946 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
5949 static int swap_max_show(struct seq_file
*m
, void *v
)
5951 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5952 unsigned long max
= READ_ONCE(memcg
->swap
.limit
);
5954 if (max
== PAGE_COUNTER_MAX
)
5955 seq_puts(m
, "max\n");
5957 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5962 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
5963 char *buf
, size_t nbytes
, loff_t off
)
5965 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5969 buf
= strstrip(buf
);
5970 err
= page_counter_memparse(buf
, "max", &max
);
5974 mutex_lock(&memcg_limit_mutex
);
5975 err
= page_counter_limit(&memcg
->swap
, max
);
5976 mutex_unlock(&memcg_limit_mutex
);
5983 static struct cftype swap_files
[] = {
5985 .name
= "swap.current",
5986 .flags
= CFTYPE_NOT_ON_ROOT
,
5987 .read_u64
= swap_current_read
,
5991 .flags
= CFTYPE_NOT_ON_ROOT
,
5992 .seq_show
= swap_max_show
,
5993 .write
= swap_max_write
,
5998 static struct cftype memsw_cgroup_files
[] = {
6000 .name
= "memsw.usage_in_bytes",
6001 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6002 .read_u64
= mem_cgroup_read_u64
,
6005 .name
= "memsw.max_usage_in_bytes",
6006 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6007 .write
= mem_cgroup_reset
,
6008 .read_u64
= mem_cgroup_read_u64
,
6011 .name
= "memsw.limit_in_bytes",
6012 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6013 .write
= mem_cgroup_write
,
6014 .read_u64
= mem_cgroup_read_u64
,
6017 .name
= "memsw.failcnt",
6018 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6019 .write
= mem_cgroup_reset
,
6020 .read_u64
= mem_cgroup_read_u64
,
6022 { }, /* terminate */
6025 static int __init
mem_cgroup_swap_init(void)
6027 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6028 do_swap_account
= 1;
6029 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
6031 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
6032 memsw_cgroup_files
));
6036 subsys_initcall(mem_cgroup_swap_init
);
6038 #endif /* CONFIG_MEMCG_SWAP */