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 * @zid: Zone ID of the zone pages have been added to
969 * @nr_pages: positive when adding or negative when removing
971 * This function must be called under lru_lock, just before a page is added
972 * to or just after a page is removed from an lru list (that ordering being
973 * so as to allow it to check that lru_size 0 is consistent with list_empty).
975 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
976 enum zone_type zid
, int nr_pages
)
978 struct mem_cgroup_per_node
*mz
;
979 unsigned long *lru_size
;
983 __update_lru_size(lruvec
, lru
, zid
, nr_pages
);
985 if (mem_cgroup_disabled())
988 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
989 lru_size
= mz
->lru_size
+ lru
;
990 empty
= list_empty(lruvec
->lists
+ lru
);
993 *lru_size
+= nr_pages
;
996 if (WARN_ONCE(size
< 0 || empty
!= !size
,
997 "%s(%p, %d, %d): lru_size %ld but %sempty\n",
998 __func__
, lruvec
, lru
, nr_pages
, size
, empty
? "" : "not ")) {
1004 *lru_size
+= nr_pages
;
1007 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1009 struct mem_cgroup
*task_memcg
;
1010 struct task_struct
*p
;
1013 p
= find_lock_task_mm(task
);
1015 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1019 * All threads may have already detached their mm's, but the oom
1020 * killer still needs to detect if they have already been oom
1021 * killed to prevent needlessly killing additional tasks.
1024 task_memcg
= mem_cgroup_from_task(task
);
1025 css_get(&task_memcg
->css
);
1028 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1029 css_put(&task_memcg
->css
);
1034 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1035 * @memcg: the memory cgroup
1037 * Returns the maximum amount of memory @mem can be charged with, in
1040 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1042 unsigned long margin
= 0;
1043 unsigned long count
;
1044 unsigned long limit
;
1046 count
= page_counter_read(&memcg
->memory
);
1047 limit
= READ_ONCE(memcg
->memory
.limit
);
1049 margin
= limit
- count
;
1051 if (do_memsw_account()) {
1052 count
= page_counter_read(&memcg
->memsw
);
1053 limit
= READ_ONCE(memcg
->memsw
.limit
);
1055 margin
= min(margin
, limit
- count
);
1064 * A routine for checking "mem" is under move_account() or not.
1066 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1067 * moving cgroups. This is for waiting at high-memory pressure
1070 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1072 struct mem_cgroup
*from
;
1073 struct mem_cgroup
*to
;
1076 * Unlike task_move routines, we access mc.to, mc.from not under
1077 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1079 spin_lock(&mc
.lock
);
1085 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1086 mem_cgroup_is_descendant(to
, memcg
);
1088 spin_unlock(&mc
.lock
);
1092 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1094 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1095 if (mem_cgroup_under_move(memcg
)) {
1097 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1098 /* moving charge context might have finished. */
1101 finish_wait(&mc
.waitq
, &wait
);
1108 #define K(x) ((x) << (PAGE_SHIFT-10))
1110 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1111 * @memcg: The memory cgroup that went over limit
1112 * @p: Task that is going to be killed
1114 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1117 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1119 struct mem_cgroup
*iter
;
1125 pr_info("Task in ");
1126 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1127 pr_cont(" killed as a result of limit of ");
1129 pr_info("Memory limit reached of cgroup ");
1132 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1137 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1138 K((u64
)page_counter_read(&memcg
->memory
)),
1139 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1140 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1141 K((u64
)page_counter_read(&memcg
->memsw
)),
1142 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1143 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1144 K((u64
)page_counter_read(&memcg
->kmem
)),
1145 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1147 for_each_mem_cgroup_tree(iter
, memcg
) {
1148 pr_info("Memory cgroup stats for ");
1149 pr_cont_cgroup_path(iter
->css
.cgroup
);
1152 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1153 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1155 pr_cont(" %s:%luKB", mem_cgroup_stat_names
[i
],
1156 K(mem_cgroup_read_stat(iter
, i
)));
1159 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1160 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1161 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1168 * This function returns the number of memcg under hierarchy tree. Returns
1169 * 1(self count) if no children.
1171 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1174 struct mem_cgroup
*iter
;
1176 for_each_mem_cgroup_tree(iter
, memcg
)
1182 * Return the memory (and swap, if configured) limit for a memcg.
1184 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1186 unsigned long limit
;
1188 limit
= memcg
->memory
.limit
;
1189 if (mem_cgroup_swappiness(memcg
)) {
1190 unsigned long memsw_limit
;
1191 unsigned long swap_limit
;
1193 memsw_limit
= memcg
->memsw
.limit
;
1194 swap_limit
= memcg
->swap
.limit
;
1195 swap_limit
= min(swap_limit
, (unsigned long)total_swap_pages
);
1196 limit
= min(limit
+ swap_limit
, memsw_limit
);
1201 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1204 struct oom_control oc
= {
1208 .gfp_mask
= gfp_mask
,
1211 struct mem_cgroup
*iter
;
1212 unsigned long chosen_points
= 0;
1213 unsigned long totalpages
;
1214 unsigned int points
= 0;
1215 struct task_struct
*chosen
= NULL
;
1217 mutex_lock(&oom_lock
);
1220 * If current has a pending SIGKILL or is exiting, then automatically
1221 * select it. The goal is to allow it to allocate so that it may
1222 * quickly exit and free its memory.
1224 if (task_will_free_mem(current
)) {
1225 mark_oom_victim(current
);
1226 wake_oom_reaper(current
);
1230 check_panic_on_oom(&oc
, CONSTRAINT_MEMCG
);
1231 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1232 for_each_mem_cgroup_tree(iter
, memcg
) {
1233 struct css_task_iter it
;
1234 struct task_struct
*task
;
1236 css_task_iter_start(&iter
->css
, &it
);
1237 while ((task
= css_task_iter_next(&it
))) {
1238 switch (oom_scan_process_thread(&oc
, task
)) {
1239 case OOM_SCAN_SELECT
:
1241 put_task_struct(chosen
);
1243 chosen_points
= ULONG_MAX
;
1244 get_task_struct(chosen
);
1246 case OOM_SCAN_CONTINUE
:
1248 case OOM_SCAN_ABORT
:
1249 css_task_iter_end(&it
);
1250 mem_cgroup_iter_break(memcg
, iter
);
1252 put_task_struct(chosen
);
1253 /* Set a dummy value to return "true". */
1254 chosen
= (void *) 1;
1259 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1260 if (!points
|| points
< chosen_points
)
1262 /* Prefer thread group leaders for display purposes */
1263 if (points
== chosen_points
&&
1264 thread_group_leader(chosen
))
1268 put_task_struct(chosen
);
1270 chosen_points
= points
;
1271 get_task_struct(chosen
);
1273 css_task_iter_end(&it
);
1277 points
= chosen_points
* 1000 / totalpages
;
1278 oom_kill_process(&oc
, chosen
, points
, totalpages
,
1279 "Memory cgroup out of memory");
1282 mutex_unlock(&oom_lock
);
1286 #if MAX_NUMNODES > 1
1289 * test_mem_cgroup_node_reclaimable
1290 * @memcg: the target memcg
1291 * @nid: the node ID to be checked.
1292 * @noswap : specify true here if the user wants flle only information.
1294 * This function returns whether the specified memcg contains any
1295 * reclaimable pages on a node. Returns true if there are any reclaimable
1296 * pages in the node.
1298 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1299 int nid
, bool noswap
)
1301 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1303 if (noswap
|| !total_swap_pages
)
1305 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1312 * Always updating the nodemask is not very good - even if we have an empty
1313 * list or the wrong list here, we can start from some node and traverse all
1314 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1317 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1321 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1322 * pagein/pageout changes since the last update.
1324 if (!atomic_read(&memcg
->numainfo_events
))
1326 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1329 /* make a nodemask where this memcg uses memory from */
1330 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1332 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1334 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1335 node_clear(nid
, memcg
->scan_nodes
);
1338 atomic_set(&memcg
->numainfo_events
, 0);
1339 atomic_set(&memcg
->numainfo_updating
, 0);
1343 * Selecting a node where we start reclaim from. Because what we need is just
1344 * reducing usage counter, start from anywhere is O,K. Considering
1345 * memory reclaim from current node, there are pros. and cons.
1347 * Freeing memory from current node means freeing memory from a node which
1348 * we'll use or we've used. So, it may make LRU bad. And if several threads
1349 * hit limits, it will see a contention on a node. But freeing from remote
1350 * node means more costs for memory reclaim because of memory latency.
1352 * Now, we use round-robin. Better algorithm is welcomed.
1354 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1358 mem_cgroup_may_update_nodemask(memcg
);
1359 node
= memcg
->last_scanned_node
;
1361 node
= next_node_in(node
, memcg
->scan_nodes
);
1363 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1364 * last time it really checked all the LRUs due to rate limiting.
1365 * Fallback to the current node in that case for simplicity.
1367 if (unlikely(node
== MAX_NUMNODES
))
1368 node
= numa_node_id();
1370 memcg
->last_scanned_node
= node
;
1374 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1380 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1383 unsigned long *total_scanned
)
1385 struct mem_cgroup
*victim
= NULL
;
1388 unsigned long excess
;
1389 unsigned long nr_scanned
;
1390 struct mem_cgroup_reclaim_cookie reclaim
= {
1395 excess
= soft_limit_excess(root_memcg
);
1398 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1403 * If we have not been able to reclaim
1404 * anything, it might because there are
1405 * no reclaimable pages under this hierarchy
1410 * We want to do more targeted reclaim.
1411 * excess >> 2 is not to excessive so as to
1412 * reclaim too much, nor too less that we keep
1413 * coming back to reclaim from this cgroup
1415 if (total
>= (excess
>> 2) ||
1416 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1421 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1422 pgdat
, &nr_scanned
);
1423 *total_scanned
+= nr_scanned
;
1424 if (!soft_limit_excess(root_memcg
))
1427 mem_cgroup_iter_break(root_memcg
, victim
);
1431 #ifdef CONFIG_LOCKDEP
1432 static struct lockdep_map memcg_oom_lock_dep_map
= {
1433 .name
= "memcg_oom_lock",
1437 static DEFINE_SPINLOCK(memcg_oom_lock
);
1440 * Check OOM-Killer is already running under our hierarchy.
1441 * If someone is running, return false.
1443 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1445 struct mem_cgroup
*iter
, *failed
= NULL
;
1447 spin_lock(&memcg_oom_lock
);
1449 for_each_mem_cgroup_tree(iter
, memcg
) {
1450 if (iter
->oom_lock
) {
1452 * this subtree of our hierarchy is already locked
1453 * so we cannot give a lock.
1456 mem_cgroup_iter_break(memcg
, iter
);
1459 iter
->oom_lock
= true;
1464 * OK, we failed to lock the whole subtree so we have
1465 * to clean up what we set up to the failing subtree
1467 for_each_mem_cgroup_tree(iter
, memcg
) {
1468 if (iter
== failed
) {
1469 mem_cgroup_iter_break(memcg
, iter
);
1472 iter
->oom_lock
= false;
1475 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1477 spin_unlock(&memcg_oom_lock
);
1482 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1484 struct mem_cgroup
*iter
;
1486 spin_lock(&memcg_oom_lock
);
1487 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1488 for_each_mem_cgroup_tree(iter
, memcg
)
1489 iter
->oom_lock
= false;
1490 spin_unlock(&memcg_oom_lock
);
1493 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1495 struct mem_cgroup
*iter
;
1497 spin_lock(&memcg_oom_lock
);
1498 for_each_mem_cgroup_tree(iter
, memcg
)
1500 spin_unlock(&memcg_oom_lock
);
1503 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1505 struct mem_cgroup
*iter
;
1508 * When a new child is created while the hierarchy is under oom,
1509 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1511 spin_lock(&memcg_oom_lock
);
1512 for_each_mem_cgroup_tree(iter
, memcg
)
1513 if (iter
->under_oom
> 0)
1515 spin_unlock(&memcg_oom_lock
);
1518 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1520 struct oom_wait_info
{
1521 struct mem_cgroup
*memcg
;
1525 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1526 unsigned mode
, int sync
, void *arg
)
1528 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1529 struct mem_cgroup
*oom_wait_memcg
;
1530 struct oom_wait_info
*oom_wait_info
;
1532 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1533 oom_wait_memcg
= oom_wait_info
->memcg
;
1535 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1536 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1538 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1541 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1544 * For the following lockless ->under_oom test, the only required
1545 * guarantee is that it must see the state asserted by an OOM when
1546 * this function is called as a result of userland actions
1547 * triggered by the notification of the OOM. This is trivially
1548 * achieved by invoking mem_cgroup_mark_under_oom() before
1549 * triggering notification.
1551 if (memcg
&& memcg
->under_oom
)
1552 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1555 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1557 if (!current
->memcg_may_oom
)
1560 * We are in the middle of the charge context here, so we
1561 * don't want to block when potentially sitting on a callstack
1562 * that holds all kinds of filesystem and mm locks.
1564 * Also, the caller may handle a failed allocation gracefully
1565 * (like optional page cache readahead) and so an OOM killer
1566 * invocation might not even be necessary.
1568 * That's why we don't do anything here except remember the
1569 * OOM context and then deal with it at the end of the page
1570 * fault when the stack is unwound, the locks are released,
1571 * and when we know whether the fault was overall successful.
1573 css_get(&memcg
->css
);
1574 current
->memcg_in_oom
= memcg
;
1575 current
->memcg_oom_gfp_mask
= mask
;
1576 current
->memcg_oom_order
= order
;
1580 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1581 * @handle: actually kill/wait or just clean up the OOM state
1583 * This has to be called at the end of a page fault if the memcg OOM
1584 * handler was enabled.
1586 * Memcg supports userspace OOM handling where failed allocations must
1587 * sleep on a waitqueue until the userspace task resolves the
1588 * situation. Sleeping directly in the charge context with all kinds
1589 * of locks held is not a good idea, instead we remember an OOM state
1590 * in the task and mem_cgroup_oom_synchronize() has to be called at
1591 * the end of the page fault to complete the OOM handling.
1593 * Returns %true if an ongoing memcg OOM situation was detected and
1594 * completed, %false otherwise.
1596 bool mem_cgroup_oom_synchronize(bool handle
)
1598 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1599 struct oom_wait_info owait
;
1602 /* OOM is global, do not handle */
1606 if (!handle
|| oom_killer_disabled
)
1609 owait
.memcg
= memcg
;
1610 owait
.wait
.flags
= 0;
1611 owait
.wait
.func
= memcg_oom_wake_function
;
1612 owait
.wait
.private = current
;
1613 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1615 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1616 mem_cgroup_mark_under_oom(memcg
);
1618 locked
= mem_cgroup_oom_trylock(memcg
);
1621 mem_cgroup_oom_notify(memcg
);
1623 if (locked
&& !memcg
->oom_kill_disable
) {
1624 mem_cgroup_unmark_under_oom(memcg
);
1625 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1626 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1627 current
->memcg_oom_order
);
1630 mem_cgroup_unmark_under_oom(memcg
);
1631 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1635 mem_cgroup_oom_unlock(memcg
);
1637 * There is no guarantee that an OOM-lock contender
1638 * sees the wakeups triggered by the OOM kill
1639 * uncharges. Wake any sleepers explicitely.
1641 memcg_oom_recover(memcg
);
1644 current
->memcg_in_oom
= NULL
;
1645 css_put(&memcg
->css
);
1650 * lock_page_memcg - lock a page->mem_cgroup binding
1653 * This function protects unlocked LRU pages from being moved to
1654 * another cgroup and stabilizes their page->mem_cgroup binding.
1656 void lock_page_memcg(struct page
*page
)
1658 struct mem_cgroup
*memcg
;
1659 unsigned long flags
;
1662 * The RCU lock is held throughout the transaction. The fast
1663 * path can get away without acquiring the memcg->move_lock
1664 * because page moving starts with an RCU grace period.
1668 if (mem_cgroup_disabled())
1671 memcg
= page
->mem_cgroup
;
1672 if (unlikely(!memcg
))
1675 if (atomic_read(&memcg
->moving_account
) <= 0)
1678 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1679 if (memcg
!= page
->mem_cgroup
) {
1680 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1685 * When charge migration first begins, we can have locked and
1686 * unlocked page stat updates happening concurrently. Track
1687 * the task who has the lock for unlock_page_memcg().
1689 memcg
->move_lock_task
= current
;
1690 memcg
->move_lock_flags
= flags
;
1694 EXPORT_SYMBOL(lock_page_memcg
);
1697 * unlock_page_memcg - unlock a page->mem_cgroup binding
1700 void unlock_page_memcg(struct page
*page
)
1702 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
1704 if (memcg
&& memcg
->move_lock_task
== current
) {
1705 unsigned long flags
= memcg
->move_lock_flags
;
1707 memcg
->move_lock_task
= NULL
;
1708 memcg
->move_lock_flags
= 0;
1710 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1715 EXPORT_SYMBOL(unlock_page_memcg
);
1718 * size of first charge trial. "32" comes from vmscan.c's magic value.
1719 * TODO: maybe necessary to use big numbers in big irons.
1721 #define CHARGE_BATCH 32U
1722 struct memcg_stock_pcp
{
1723 struct mem_cgroup
*cached
; /* this never be root cgroup */
1724 unsigned int nr_pages
;
1725 struct work_struct work
;
1726 unsigned long flags
;
1727 #define FLUSHING_CACHED_CHARGE 0
1729 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1730 static DEFINE_MUTEX(percpu_charge_mutex
);
1733 * consume_stock: Try to consume stocked charge on this cpu.
1734 * @memcg: memcg to consume from.
1735 * @nr_pages: how many pages to charge.
1737 * The charges will only happen if @memcg matches the current cpu's memcg
1738 * stock, and at least @nr_pages are available in that stock. Failure to
1739 * service an allocation will refill the stock.
1741 * returns true if successful, false otherwise.
1743 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1745 struct memcg_stock_pcp
*stock
;
1748 if (nr_pages
> CHARGE_BATCH
)
1751 stock
= &get_cpu_var(memcg_stock
);
1752 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
1753 stock
->nr_pages
-= nr_pages
;
1756 put_cpu_var(memcg_stock
);
1761 * Returns stocks cached in percpu and reset cached information.
1763 static void drain_stock(struct memcg_stock_pcp
*stock
)
1765 struct mem_cgroup
*old
= stock
->cached
;
1767 if (stock
->nr_pages
) {
1768 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
1769 if (do_memsw_account())
1770 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
1771 css_put_many(&old
->css
, stock
->nr_pages
);
1772 stock
->nr_pages
= 0;
1774 stock
->cached
= NULL
;
1778 * This must be called under preempt disabled or must be called by
1779 * a thread which is pinned to local cpu.
1781 static void drain_local_stock(struct work_struct
*dummy
)
1783 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
1785 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
1789 * Cache charges(val) to local per_cpu area.
1790 * This will be consumed by consume_stock() function, later.
1792 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1794 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
1796 if (stock
->cached
!= memcg
) { /* reset if necessary */
1798 stock
->cached
= memcg
;
1800 stock
->nr_pages
+= nr_pages
;
1801 put_cpu_var(memcg_stock
);
1805 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1806 * of the hierarchy under it.
1808 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
1812 /* If someone's already draining, avoid adding running more workers. */
1813 if (!mutex_trylock(&percpu_charge_mutex
))
1815 /* Notify other cpus that system-wide "drain" is running */
1818 for_each_online_cpu(cpu
) {
1819 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1820 struct mem_cgroup
*memcg
;
1822 memcg
= stock
->cached
;
1823 if (!memcg
|| !stock
->nr_pages
)
1825 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
1827 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
1829 drain_local_stock(&stock
->work
);
1831 schedule_work_on(cpu
, &stock
->work
);
1836 mutex_unlock(&percpu_charge_mutex
);
1839 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
1840 unsigned long action
,
1843 int cpu
= (unsigned long)hcpu
;
1844 struct memcg_stock_pcp
*stock
;
1846 if (action
== CPU_ONLINE
)
1849 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
1852 stock
= &per_cpu(memcg_stock
, cpu
);
1857 static void reclaim_high(struct mem_cgroup
*memcg
,
1858 unsigned int nr_pages
,
1862 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
1864 mem_cgroup_events(memcg
, MEMCG_HIGH
, 1);
1865 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
1866 } while ((memcg
= parent_mem_cgroup(memcg
)));
1869 static void high_work_func(struct work_struct
*work
)
1871 struct mem_cgroup
*memcg
;
1873 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
1874 reclaim_high(memcg
, CHARGE_BATCH
, GFP_KERNEL
);
1878 * Scheduled by try_charge() to be executed from the userland return path
1879 * and reclaims memory over the high limit.
1881 void mem_cgroup_handle_over_high(void)
1883 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
1884 struct mem_cgroup
*memcg
;
1886 if (likely(!nr_pages
))
1889 memcg
= get_mem_cgroup_from_mm(current
->mm
);
1890 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
1891 css_put(&memcg
->css
);
1892 current
->memcg_nr_pages_over_high
= 0;
1895 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1896 unsigned int nr_pages
)
1898 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
1899 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1900 struct mem_cgroup
*mem_over_limit
;
1901 struct page_counter
*counter
;
1902 unsigned long nr_reclaimed
;
1903 bool may_swap
= true;
1904 bool drained
= false;
1906 if (mem_cgroup_is_root(memcg
))
1909 if (consume_stock(memcg
, nr_pages
))
1912 if (!do_memsw_account() ||
1913 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
1914 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
1916 if (do_memsw_account())
1917 page_counter_uncharge(&memcg
->memsw
, batch
);
1918 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
1920 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
1924 if (batch
> nr_pages
) {
1930 * Unlike in global OOM situations, memcg is not in a physical
1931 * memory shortage. Allow dying and OOM-killed tasks to
1932 * bypass the last charges so that they can exit quickly and
1933 * free their memory.
1935 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
1936 fatal_signal_pending(current
) ||
1937 current
->flags
& PF_EXITING
))
1940 if (unlikely(task_in_memcg_oom(current
)))
1943 if (!gfpflags_allow_blocking(gfp_mask
))
1946 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
1948 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
1949 gfp_mask
, may_swap
);
1951 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
1955 drain_all_stock(mem_over_limit
);
1960 if (gfp_mask
& __GFP_NORETRY
)
1963 * Even though the limit is exceeded at this point, reclaim
1964 * may have been able to free some pages. Retry the charge
1965 * before killing the task.
1967 * Only for regular pages, though: huge pages are rather
1968 * unlikely to succeed so close to the limit, and we fall back
1969 * to regular pages anyway in case of failure.
1971 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
1974 * At task move, charge accounts can be doubly counted. So, it's
1975 * better to wait until the end of task_move if something is going on.
1977 if (mem_cgroup_wait_acct_move(mem_over_limit
))
1983 if (gfp_mask
& __GFP_NOFAIL
)
1986 if (fatal_signal_pending(current
))
1989 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
1991 mem_cgroup_oom(mem_over_limit
, gfp_mask
,
1992 get_order(nr_pages
* PAGE_SIZE
));
1994 if (!(gfp_mask
& __GFP_NOFAIL
))
1998 * The allocation either can't fail or will lead to more memory
1999 * being freed very soon. Allow memory usage go over the limit
2000 * temporarily by force charging it.
2002 page_counter_charge(&memcg
->memory
, nr_pages
);
2003 if (do_memsw_account())
2004 page_counter_charge(&memcg
->memsw
, nr_pages
);
2005 css_get_many(&memcg
->css
, nr_pages
);
2010 css_get_many(&memcg
->css
, batch
);
2011 if (batch
> nr_pages
)
2012 refill_stock(memcg
, batch
- nr_pages
);
2015 * If the hierarchy is above the normal consumption range, schedule
2016 * reclaim on returning to userland. We can perform reclaim here
2017 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2018 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2019 * not recorded as it most likely matches current's and won't
2020 * change in the meantime. As high limit is checked again before
2021 * reclaim, the cost of mismatch is negligible.
2024 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2025 /* Don't bother a random interrupted task */
2026 if (in_interrupt()) {
2027 schedule_work(&memcg
->high_work
);
2030 current
->memcg_nr_pages_over_high
+= batch
;
2031 set_notify_resume(current
);
2034 } while ((memcg
= parent_mem_cgroup(memcg
)));
2039 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2041 if (mem_cgroup_is_root(memcg
))
2044 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2045 if (do_memsw_account())
2046 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2048 css_put_many(&memcg
->css
, nr_pages
);
2051 static void lock_page_lru(struct page
*page
, int *isolated
)
2053 struct zone
*zone
= page_zone(page
);
2055 spin_lock_irq(zone_lru_lock(zone
));
2056 if (PageLRU(page
)) {
2057 struct lruvec
*lruvec
;
2059 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
2061 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2067 static void unlock_page_lru(struct page
*page
, int isolated
)
2069 struct zone
*zone
= page_zone(page
);
2072 struct lruvec
*lruvec
;
2074 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
2075 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2077 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2079 spin_unlock_irq(zone_lru_lock(zone
));
2082 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2087 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2090 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2091 * may already be on some other mem_cgroup's LRU. Take care of it.
2094 lock_page_lru(page
, &isolated
);
2097 * Nobody should be changing or seriously looking at
2098 * page->mem_cgroup at this point:
2100 * - the page is uncharged
2102 * - the page is off-LRU
2104 * - an anonymous fault has exclusive page access, except for
2105 * a locked page table
2107 * - a page cache insertion, a swapin fault, or a migration
2108 * have the page locked
2110 page
->mem_cgroup
= memcg
;
2113 unlock_page_lru(page
, isolated
);
2117 static int memcg_alloc_cache_id(void)
2122 id
= ida_simple_get(&memcg_cache_ida
,
2123 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2127 if (id
< memcg_nr_cache_ids
)
2131 * There's no space for the new id in memcg_caches arrays,
2132 * so we have to grow them.
2134 down_write(&memcg_cache_ids_sem
);
2136 size
= 2 * (id
+ 1);
2137 if (size
< MEMCG_CACHES_MIN_SIZE
)
2138 size
= MEMCG_CACHES_MIN_SIZE
;
2139 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2140 size
= MEMCG_CACHES_MAX_SIZE
;
2142 err
= memcg_update_all_caches(size
);
2144 err
= memcg_update_all_list_lrus(size
);
2146 memcg_nr_cache_ids
= size
;
2148 up_write(&memcg_cache_ids_sem
);
2151 ida_simple_remove(&memcg_cache_ida
, id
);
2157 static void memcg_free_cache_id(int id
)
2159 ida_simple_remove(&memcg_cache_ida
, id
);
2162 struct memcg_kmem_cache_create_work
{
2163 struct mem_cgroup
*memcg
;
2164 struct kmem_cache
*cachep
;
2165 struct work_struct work
;
2168 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2170 struct memcg_kmem_cache_create_work
*cw
=
2171 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2172 struct mem_cgroup
*memcg
= cw
->memcg
;
2173 struct kmem_cache
*cachep
= cw
->cachep
;
2175 memcg_create_kmem_cache(memcg
, cachep
);
2177 css_put(&memcg
->css
);
2182 * Enqueue the creation of a per-memcg kmem_cache.
2184 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2185 struct kmem_cache
*cachep
)
2187 struct memcg_kmem_cache_create_work
*cw
;
2189 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2193 css_get(&memcg
->css
);
2196 cw
->cachep
= cachep
;
2197 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2199 schedule_work(&cw
->work
);
2202 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2203 struct kmem_cache
*cachep
)
2206 * We need to stop accounting when we kmalloc, because if the
2207 * corresponding kmalloc cache is not yet created, the first allocation
2208 * in __memcg_schedule_kmem_cache_create will recurse.
2210 * However, it is better to enclose the whole function. Depending on
2211 * the debugging options enabled, INIT_WORK(), for instance, can
2212 * trigger an allocation. This too, will make us recurse. Because at
2213 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2214 * the safest choice is to do it like this, wrapping the whole function.
2216 current
->memcg_kmem_skip_account
= 1;
2217 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2218 current
->memcg_kmem_skip_account
= 0;
2221 static inline bool memcg_kmem_bypass(void)
2223 if (in_interrupt() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
2229 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2230 * @cachep: the original global kmem cache
2232 * Return the kmem_cache we're supposed to use for a slab allocation.
2233 * We try to use the current memcg's version of the cache.
2235 * If the cache does not exist yet, if we are the first user of it, we
2236 * create it asynchronously in a workqueue and let the current allocation
2237 * go through with the original cache.
2239 * This function takes a reference to the cache it returns to assure it
2240 * won't get destroyed while we are working with it. Once the caller is
2241 * done with it, memcg_kmem_put_cache() must be called to release the
2244 struct kmem_cache
*memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2246 struct mem_cgroup
*memcg
;
2247 struct kmem_cache
*memcg_cachep
;
2250 VM_BUG_ON(!is_root_cache(cachep
));
2252 if (memcg_kmem_bypass())
2255 if (current
->memcg_kmem_skip_account
)
2258 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2259 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2263 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2264 if (likely(memcg_cachep
))
2265 return memcg_cachep
;
2268 * If we are in a safe context (can wait, and not in interrupt
2269 * context), we could be be predictable and return right away.
2270 * This would guarantee that the allocation being performed
2271 * already belongs in the new cache.
2273 * However, there are some clashes that can arrive from locking.
2274 * For instance, because we acquire the slab_mutex while doing
2275 * memcg_create_kmem_cache, this means no further allocation
2276 * could happen with the slab_mutex held. So it's better to
2279 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2281 css_put(&memcg
->css
);
2286 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2287 * @cachep: the cache returned by memcg_kmem_get_cache
2289 void memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2291 if (!is_root_cache(cachep
))
2292 css_put(&cachep
->memcg_params
.memcg
->css
);
2296 * memcg_kmem_charge: charge a kmem page
2297 * @page: page to charge
2298 * @gfp: reclaim mode
2299 * @order: allocation order
2300 * @memcg: memory cgroup to charge
2302 * Returns 0 on success, an error code on failure.
2304 int memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2305 struct mem_cgroup
*memcg
)
2307 unsigned int nr_pages
= 1 << order
;
2308 struct page_counter
*counter
;
2311 ret
= try_charge(memcg
, gfp
, nr_pages
);
2315 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2316 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2317 cancel_charge(memcg
, nr_pages
);
2321 page
->mem_cgroup
= memcg
;
2327 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2328 * @page: page to charge
2329 * @gfp: reclaim mode
2330 * @order: allocation order
2332 * Returns 0 on success, an error code on failure.
2334 int memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2336 struct mem_cgroup
*memcg
;
2339 if (memcg_kmem_bypass())
2342 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2343 if (!mem_cgroup_is_root(memcg
))
2344 ret
= memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2345 css_put(&memcg
->css
);
2349 * memcg_kmem_uncharge: uncharge a kmem page
2350 * @page: page to uncharge
2351 * @order: allocation order
2353 void memcg_kmem_uncharge(struct page
*page
, int order
)
2355 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2356 unsigned int nr_pages
= 1 << order
;
2361 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2363 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2364 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2366 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2367 if (do_memsw_account())
2368 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2370 page
->mem_cgroup
= NULL
;
2371 css_put_many(&memcg
->css
, nr_pages
);
2373 #endif /* !CONFIG_SLOB */
2375 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2378 * Because tail pages are not marked as "used", set it. We're under
2379 * zone_lru_lock and migration entries setup in all page mappings.
2381 void mem_cgroup_split_huge_fixup(struct page
*head
)
2385 if (mem_cgroup_disabled())
2388 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2389 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2391 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2394 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2396 #ifdef CONFIG_MEMCG_SWAP
2397 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2400 int val
= (charge
) ? 1 : -1;
2401 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2405 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2406 * @entry: swap entry to be moved
2407 * @from: mem_cgroup which the entry is moved from
2408 * @to: mem_cgroup which the entry is moved to
2410 * It succeeds only when the swap_cgroup's record for this entry is the same
2411 * as the mem_cgroup's id of @from.
2413 * Returns 0 on success, -EINVAL on failure.
2415 * The caller must have charged to @to, IOW, called page_counter_charge() about
2416 * both res and memsw, and called css_get().
2418 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2419 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2421 unsigned short old_id
, new_id
;
2423 old_id
= mem_cgroup_id(from
);
2424 new_id
= mem_cgroup_id(to
);
2426 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2427 mem_cgroup_swap_statistics(from
, false);
2428 mem_cgroup_swap_statistics(to
, true);
2434 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2435 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2441 static DEFINE_MUTEX(memcg_limit_mutex
);
2443 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2444 unsigned long limit
)
2446 unsigned long curusage
;
2447 unsigned long oldusage
;
2448 bool enlarge
= false;
2453 * For keeping hierarchical_reclaim simple, how long we should retry
2454 * is depends on callers. We set our retry-count to be function
2455 * of # of children which we should visit in this loop.
2457 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2458 mem_cgroup_count_children(memcg
);
2460 oldusage
= page_counter_read(&memcg
->memory
);
2463 if (signal_pending(current
)) {
2468 mutex_lock(&memcg_limit_mutex
);
2469 if (limit
> memcg
->memsw
.limit
) {
2470 mutex_unlock(&memcg_limit_mutex
);
2474 if (limit
> memcg
->memory
.limit
)
2476 ret
= page_counter_limit(&memcg
->memory
, limit
);
2477 mutex_unlock(&memcg_limit_mutex
);
2482 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2484 curusage
= page_counter_read(&memcg
->memory
);
2485 /* Usage is reduced ? */
2486 if (curusage
>= oldusage
)
2489 oldusage
= curusage
;
2490 } while (retry_count
);
2492 if (!ret
&& enlarge
)
2493 memcg_oom_recover(memcg
);
2498 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2499 unsigned long limit
)
2501 unsigned long curusage
;
2502 unsigned long oldusage
;
2503 bool enlarge
= false;
2507 /* see mem_cgroup_resize_res_limit */
2508 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2509 mem_cgroup_count_children(memcg
);
2511 oldusage
= page_counter_read(&memcg
->memsw
);
2514 if (signal_pending(current
)) {
2519 mutex_lock(&memcg_limit_mutex
);
2520 if (limit
< memcg
->memory
.limit
) {
2521 mutex_unlock(&memcg_limit_mutex
);
2525 if (limit
> memcg
->memsw
.limit
)
2527 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2528 mutex_unlock(&memcg_limit_mutex
);
2533 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2535 curusage
= page_counter_read(&memcg
->memsw
);
2536 /* Usage is reduced ? */
2537 if (curusage
>= oldusage
)
2540 oldusage
= curusage
;
2541 } while (retry_count
);
2543 if (!ret
&& enlarge
)
2544 memcg_oom_recover(memcg
);
2549 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
2551 unsigned long *total_scanned
)
2553 unsigned long nr_reclaimed
= 0;
2554 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
2555 unsigned long reclaimed
;
2557 struct mem_cgroup_tree_per_node
*mctz
;
2558 unsigned long excess
;
2559 unsigned long nr_scanned
;
2564 mctz
= soft_limit_tree_node(pgdat
->node_id
);
2566 * This loop can run a while, specially if mem_cgroup's continuously
2567 * keep exceeding their soft limit and putting the system under
2574 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2579 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
2580 gfp_mask
, &nr_scanned
);
2581 nr_reclaimed
+= reclaimed
;
2582 *total_scanned
+= nr_scanned
;
2583 spin_lock_irq(&mctz
->lock
);
2584 __mem_cgroup_remove_exceeded(mz
, mctz
);
2587 * If we failed to reclaim anything from this memory cgroup
2588 * it is time to move on to the next cgroup
2592 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2594 excess
= soft_limit_excess(mz
->memcg
);
2596 * One school of thought says that we should not add
2597 * back the node to the tree if reclaim returns 0.
2598 * But our reclaim could return 0, simply because due
2599 * to priority we are exposing a smaller subset of
2600 * memory to reclaim from. Consider this as a longer
2603 /* If excess == 0, no tree ops */
2604 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2605 spin_unlock_irq(&mctz
->lock
);
2606 css_put(&mz
->memcg
->css
);
2609 * Could not reclaim anything and there are no more
2610 * mem cgroups to try or we seem to be looping without
2611 * reclaiming anything.
2613 if (!nr_reclaimed
&&
2615 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2617 } while (!nr_reclaimed
);
2619 css_put(&next_mz
->memcg
->css
);
2620 return nr_reclaimed
;
2624 * Test whether @memcg has children, dead or alive. Note that this
2625 * function doesn't care whether @memcg has use_hierarchy enabled and
2626 * returns %true if there are child csses according to the cgroup
2627 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2629 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2634 ret
= css_next_child(NULL
, &memcg
->css
);
2640 * Reclaims as many pages from the given memcg as possible.
2642 * Caller is responsible for holding css reference for memcg.
2644 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2646 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2648 /* we call try-to-free pages for make this cgroup empty */
2649 lru_add_drain_all();
2650 /* try to free all pages in this cgroup */
2651 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2654 if (signal_pending(current
))
2657 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2661 /* maybe some writeback is necessary */
2662 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2670 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
2671 char *buf
, size_t nbytes
,
2674 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2676 if (mem_cgroup_is_root(memcg
))
2678 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
2681 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
2684 return mem_cgroup_from_css(css
)->use_hierarchy
;
2687 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
2688 struct cftype
*cft
, u64 val
)
2691 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2692 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
2694 if (memcg
->use_hierarchy
== val
)
2698 * If parent's use_hierarchy is set, we can't make any modifications
2699 * in the child subtrees. If it is unset, then the change can
2700 * occur, provided the current cgroup has no children.
2702 * For the root cgroup, parent_mem is NULL, we allow value to be
2703 * set if there are no children.
2705 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
2706 (val
== 1 || val
== 0)) {
2707 if (!memcg_has_children(memcg
))
2708 memcg
->use_hierarchy
= val
;
2717 static void tree_stat(struct mem_cgroup
*memcg
, unsigned long *stat
)
2719 struct mem_cgroup
*iter
;
2722 memset(stat
, 0, sizeof(*stat
) * MEMCG_NR_STAT
);
2724 for_each_mem_cgroup_tree(iter
, memcg
) {
2725 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
2726 stat
[i
] += mem_cgroup_read_stat(iter
, i
);
2730 static void tree_events(struct mem_cgroup
*memcg
, unsigned long *events
)
2732 struct mem_cgroup
*iter
;
2735 memset(events
, 0, sizeof(*events
) * MEMCG_NR_EVENTS
);
2737 for_each_mem_cgroup_tree(iter
, memcg
) {
2738 for (i
= 0; i
< MEMCG_NR_EVENTS
; i
++)
2739 events
[i
] += mem_cgroup_read_events(iter
, i
);
2743 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
2745 unsigned long val
= 0;
2747 if (mem_cgroup_is_root(memcg
)) {
2748 struct mem_cgroup
*iter
;
2750 for_each_mem_cgroup_tree(iter
, memcg
) {
2751 val
+= mem_cgroup_read_stat(iter
,
2752 MEM_CGROUP_STAT_CACHE
);
2753 val
+= mem_cgroup_read_stat(iter
,
2754 MEM_CGROUP_STAT_RSS
);
2756 val
+= mem_cgroup_read_stat(iter
,
2757 MEM_CGROUP_STAT_SWAP
);
2761 val
= page_counter_read(&memcg
->memory
);
2763 val
= page_counter_read(&memcg
->memsw
);
2776 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
2779 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2780 struct page_counter
*counter
;
2782 switch (MEMFILE_TYPE(cft
->private)) {
2784 counter
= &memcg
->memory
;
2787 counter
= &memcg
->memsw
;
2790 counter
= &memcg
->kmem
;
2793 counter
= &memcg
->tcpmem
;
2799 switch (MEMFILE_ATTR(cft
->private)) {
2801 if (counter
== &memcg
->memory
)
2802 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
2803 if (counter
== &memcg
->memsw
)
2804 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
2805 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
2807 return (u64
)counter
->limit
* PAGE_SIZE
;
2809 return (u64
)counter
->watermark
* PAGE_SIZE
;
2811 return counter
->failcnt
;
2812 case RES_SOFT_LIMIT
:
2813 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
2820 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2824 if (cgroup_memory_nokmem
)
2827 BUG_ON(memcg
->kmemcg_id
>= 0);
2828 BUG_ON(memcg
->kmem_state
);
2830 memcg_id
= memcg_alloc_cache_id();
2834 static_branch_inc(&memcg_kmem_enabled_key
);
2836 * A memory cgroup is considered kmem-online as soon as it gets
2837 * kmemcg_id. Setting the id after enabling static branching will
2838 * guarantee no one starts accounting before all call sites are
2841 memcg
->kmemcg_id
= memcg_id
;
2842 memcg
->kmem_state
= KMEM_ONLINE
;
2847 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2849 struct cgroup_subsys_state
*css
;
2850 struct mem_cgroup
*parent
, *child
;
2853 if (memcg
->kmem_state
!= KMEM_ONLINE
)
2856 * Clear the online state before clearing memcg_caches array
2857 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2858 * guarantees that no cache will be created for this cgroup
2859 * after we are done (see memcg_create_kmem_cache()).
2861 memcg
->kmem_state
= KMEM_ALLOCATED
;
2863 memcg_deactivate_kmem_caches(memcg
);
2865 kmemcg_id
= memcg
->kmemcg_id
;
2866 BUG_ON(kmemcg_id
< 0);
2868 parent
= parent_mem_cgroup(memcg
);
2870 parent
= root_mem_cgroup
;
2873 * Change kmemcg_id of this cgroup and all its descendants to the
2874 * parent's id, and then move all entries from this cgroup's list_lrus
2875 * to ones of the parent. After we have finished, all list_lrus
2876 * corresponding to this cgroup are guaranteed to remain empty. The
2877 * ordering is imposed by list_lru_node->lock taken by
2878 * memcg_drain_all_list_lrus().
2880 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2881 css_for_each_descendant_pre(css
, &memcg
->css
) {
2882 child
= mem_cgroup_from_css(css
);
2883 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
2884 child
->kmemcg_id
= parent
->kmemcg_id
;
2885 if (!memcg
->use_hierarchy
)
2890 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
2892 memcg_free_cache_id(kmemcg_id
);
2895 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2897 /* css_alloc() failed, offlining didn't happen */
2898 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
2899 memcg_offline_kmem(memcg
);
2901 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
2902 memcg_destroy_kmem_caches(memcg
);
2903 static_branch_dec(&memcg_kmem_enabled_key
);
2904 WARN_ON(page_counter_read(&memcg
->kmem
));
2908 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2912 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2915 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2918 #endif /* !CONFIG_SLOB */
2920 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2921 unsigned long limit
)
2925 mutex_lock(&memcg_limit_mutex
);
2926 ret
= page_counter_limit(&memcg
->kmem
, limit
);
2927 mutex_unlock(&memcg_limit_mutex
);
2931 static int memcg_update_tcp_limit(struct mem_cgroup
*memcg
, unsigned long limit
)
2935 mutex_lock(&memcg_limit_mutex
);
2937 ret
= page_counter_limit(&memcg
->tcpmem
, limit
);
2941 if (!memcg
->tcpmem_active
) {
2943 * The active flag needs to be written after the static_key
2944 * update. This is what guarantees that the socket activation
2945 * function is the last one to run. See sock_update_memcg() for
2946 * details, and note that we don't mark any socket as belonging
2947 * to this memcg until that flag is up.
2949 * We need to do this, because static_keys will span multiple
2950 * sites, but we can't control their order. If we mark a socket
2951 * as accounted, but the accounting functions are not patched in
2952 * yet, we'll lose accounting.
2954 * We never race with the readers in sock_update_memcg(),
2955 * because when this value change, the code to process it is not
2958 static_branch_inc(&memcg_sockets_enabled_key
);
2959 memcg
->tcpmem_active
= true;
2962 mutex_unlock(&memcg_limit_mutex
);
2967 * The user of this function is...
2970 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
2971 char *buf
, size_t nbytes
, loff_t off
)
2973 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2974 unsigned long nr_pages
;
2977 buf
= strstrip(buf
);
2978 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
2982 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
2984 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
2988 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
2990 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
2993 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
2996 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
2999 ret
= memcg_update_tcp_limit(memcg
, nr_pages
);
3003 case RES_SOFT_LIMIT
:
3004 memcg
->soft_limit
= nr_pages
;
3008 return ret
?: nbytes
;
3011 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3012 size_t nbytes
, loff_t off
)
3014 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3015 struct page_counter
*counter
;
3017 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3019 counter
= &memcg
->memory
;
3022 counter
= &memcg
->memsw
;
3025 counter
= &memcg
->kmem
;
3028 counter
= &memcg
->tcpmem
;
3034 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3036 page_counter_reset_watermark(counter
);
3039 counter
->failcnt
= 0;
3048 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3051 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3055 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3056 struct cftype
*cft
, u64 val
)
3058 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3060 if (val
& ~MOVE_MASK
)
3064 * No kind of locking is needed in here, because ->can_attach() will
3065 * check this value once in the beginning of the process, and then carry
3066 * on with stale data. This means that changes to this value will only
3067 * affect task migrations starting after the change.
3069 memcg
->move_charge_at_immigrate
= val
;
3073 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3074 struct cftype
*cft
, u64 val
)
3081 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3085 unsigned int lru_mask
;
3088 static const struct numa_stat stats
[] = {
3089 { "total", LRU_ALL
},
3090 { "file", LRU_ALL_FILE
},
3091 { "anon", LRU_ALL_ANON
},
3092 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3094 const struct numa_stat
*stat
;
3097 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3099 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3100 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3101 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3102 for_each_node_state(nid
, N_MEMORY
) {
3103 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3105 seq_printf(m
, " N%d=%lu", nid
, nr
);
3110 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3111 struct mem_cgroup
*iter
;
3114 for_each_mem_cgroup_tree(iter
, memcg
)
3115 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3116 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3117 for_each_node_state(nid
, N_MEMORY
) {
3119 for_each_mem_cgroup_tree(iter
, memcg
)
3120 nr
+= mem_cgroup_node_nr_lru_pages(
3121 iter
, nid
, stat
->lru_mask
);
3122 seq_printf(m
, " N%d=%lu", nid
, nr
);
3129 #endif /* CONFIG_NUMA */
3131 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3133 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3134 unsigned long memory
, memsw
;
3135 struct mem_cgroup
*mi
;
3138 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3139 MEM_CGROUP_STAT_NSTATS
);
3140 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3141 MEM_CGROUP_EVENTS_NSTATS
);
3142 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3144 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3145 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3147 seq_printf(m
, "%s %lu\n", mem_cgroup_stat_names
[i
],
3148 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3151 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3152 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3153 mem_cgroup_read_events(memcg
, i
));
3155 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3156 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3157 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3159 /* Hierarchical information */
3160 memory
= memsw
= PAGE_COUNTER_MAX
;
3161 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3162 memory
= min(memory
, mi
->memory
.limit
);
3163 memsw
= min(memsw
, mi
->memsw
.limit
);
3165 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3166 (u64
)memory
* PAGE_SIZE
);
3167 if (do_memsw_account())
3168 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3169 (u64
)memsw
* PAGE_SIZE
);
3171 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3172 unsigned long long val
= 0;
3174 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3176 for_each_mem_cgroup_tree(mi
, memcg
)
3177 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3178 seq_printf(m
, "total_%s %llu\n", mem_cgroup_stat_names
[i
], val
);
3181 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3182 unsigned long long val
= 0;
3184 for_each_mem_cgroup_tree(mi
, memcg
)
3185 val
+= mem_cgroup_read_events(mi
, i
);
3186 seq_printf(m
, "total_%s %llu\n",
3187 mem_cgroup_events_names
[i
], val
);
3190 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3191 unsigned long long val
= 0;
3193 for_each_mem_cgroup_tree(mi
, memcg
)
3194 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3195 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3198 #ifdef CONFIG_DEBUG_VM
3201 struct mem_cgroup_per_node
*mz
;
3202 struct zone_reclaim_stat
*rstat
;
3203 unsigned long recent_rotated
[2] = {0, 0};
3204 unsigned long recent_scanned
[2] = {0, 0};
3206 for_each_online_pgdat(pgdat
) {
3207 mz
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
3208 rstat
= &mz
->lruvec
.reclaim_stat
;
3210 recent_rotated
[0] += rstat
->recent_rotated
[0];
3211 recent_rotated
[1] += rstat
->recent_rotated
[1];
3212 recent_scanned
[0] += rstat
->recent_scanned
[0];
3213 recent_scanned
[1] += rstat
->recent_scanned
[1];
3215 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3216 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3217 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3218 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3225 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3228 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3230 return mem_cgroup_swappiness(memcg
);
3233 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3234 struct cftype
*cft
, u64 val
)
3236 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3242 memcg
->swappiness
= val
;
3244 vm_swappiness
= val
;
3249 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3251 struct mem_cgroup_threshold_ary
*t
;
3252 unsigned long usage
;
3257 t
= rcu_dereference(memcg
->thresholds
.primary
);
3259 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3264 usage
= mem_cgroup_usage(memcg
, swap
);
3267 * current_threshold points to threshold just below or equal to usage.
3268 * If it's not true, a threshold was crossed after last
3269 * call of __mem_cgroup_threshold().
3271 i
= t
->current_threshold
;
3274 * Iterate backward over array of thresholds starting from
3275 * current_threshold and check if a threshold is crossed.
3276 * If none of thresholds below usage is crossed, we read
3277 * only one element of the array here.
3279 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3280 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3282 /* i = current_threshold + 1 */
3286 * Iterate forward over array of thresholds starting from
3287 * current_threshold+1 and check if a threshold is crossed.
3288 * If none of thresholds above usage is crossed, we read
3289 * only one element of the array here.
3291 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3292 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3294 /* Update current_threshold */
3295 t
->current_threshold
= i
- 1;
3300 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3303 __mem_cgroup_threshold(memcg
, false);
3304 if (do_memsw_account())
3305 __mem_cgroup_threshold(memcg
, true);
3307 memcg
= parent_mem_cgroup(memcg
);
3311 static int compare_thresholds(const void *a
, const void *b
)
3313 const struct mem_cgroup_threshold
*_a
= a
;
3314 const struct mem_cgroup_threshold
*_b
= b
;
3316 if (_a
->threshold
> _b
->threshold
)
3319 if (_a
->threshold
< _b
->threshold
)
3325 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3327 struct mem_cgroup_eventfd_list
*ev
;
3329 spin_lock(&memcg_oom_lock
);
3331 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3332 eventfd_signal(ev
->eventfd
, 1);
3334 spin_unlock(&memcg_oom_lock
);
3338 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3340 struct mem_cgroup
*iter
;
3342 for_each_mem_cgroup_tree(iter
, memcg
)
3343 mem_cgroup_oom_notify_cb(iter
);
3346 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3347 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3349 struct mem_cgroup_thresholds
*thresholds
;
3350 struct mem_cgroup_threshold_ary
*new;
3351 unsigned long threshold
;
3352 unsigned long usage
;
3355 ret
= page_counter_memparse(args
, "-1", &threshold
);
3359 mutex_lock(&memcg
->thresholds_lock
);
3362 thresholds
= &memcg
->thresholds
;
3363 usage
= mem_cgroup_usage(memcg
, false);
3364 } else if (type
== _MEMSWAP
) {
3365 thresholds
= &memcg
->memsw_thresholds
;
3366 usage
= mem_cgroup_usage(memcg
, true);
3370 /* Check if a threshold crossed before adding a new one */
3371 if (thresholds
->primary
)
3372 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3374 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3376 /* Allocate memory for new array of thresholds */
3377 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3385 /* Copy thresholds (if any) to new array */
3386 if (thresholds
->primary
) {
3387 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3388 sizeof(struct mem_cgroup_threshold
));
3391 /* Add new threshold */
3392 new->entries
[size
- 1].eventfd
= eventfd
;
3393 new->entries
[size
- 1].threshold
= threshold
;
3395 /* Sort thresholds. Registering of new threshold isn't time-critical */
3396 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3397 compare_thresholds
, NULL
);
3399 /* Find current threshold */
3400 new->current_threshold
= -1;
3401 for (i
= 0; i
< size
; i
++) {
3402 if (new->entries
[i
].threshold
<= usage
) {
3404 * new->current_threshold will not be used until
3405 * rcu_assign_pointer(), so it's safe to increment
3408 ++new->current_threshold
;
3413 /* Free old spare buffer and save old primary buffer as spare */
3414 kfree(thresholds
->spare
);
3415 thresholds
->spare
= thresholds
->primary
;
3417 rcu_assign_pointer(thresholds
->primary
, new);
3419 /* To be sure that nobody uses thresholds */
3423 mutex_unlock(&memcg
->thresholds_lock
);
3428 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3429 struct eventfd_ctx
*eventfd
, const char *args
)
3431 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3434 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3435 struct eventfd_ctx
*eventfd
, const char *args
)
3437 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3440 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3441 struct eventfd_ctx
*eventfd
, enum res_type type
)
3443 struct mem_cgroup_thresholds
*thresholds
;
3444 struct mem_cgroup_threshold_ary
*new;
3445 unsigned long usage
;
3448 mutex_lock(&memcg
->thresholds_lock
);
3451 thresholds
= &memcg
->thresholds
;
3452 usage
= mem_cgroup_usage(memcg
, false);
3453 } else if (type
== _MEMSWAP
) {
3454 thresholds
= &memcg
->memsw_thresholds
;
3455 usage
= mem_cgroup_usage(memcg
, true);
3459 if (!thresholds
->primary
)
3462 /* Check if a threshold crossed before removing */
3463 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3465 /* Calculate new number of threshold */
3467 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3468 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3472 new = thresholds
->spare
;
3474 /* Set thresholds array to NULL if we don't have thresholds */
3483 /* Copy thresholds and find current threshold */
3484 new->current_threshold
= -1;
3485 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3486 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3489 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3490 if (new->entries
[j
].threshold
<= usage
) {
3492 * new->current_threshold will not be used
3493 * until rcu_assign_pointer(), so it's safe to increment
3496 ++new->current_threshold
;
3502 /* Swap primary and spare array */
3503 thresholds
->spare
= thresholds
->primary
;
3505 rcu_assign_pointer(thresholds
->primary
, new);
3507 /* To be sure that nobody uses thresholds */
3510 /* If all events are unregistered, free the spare array */
3512 kfree(thresholds
->spare
);
3513 thresholds
->spare
= NULL
;
3516 mutex_unlock(&memcg
->thresholds_lock
);
3519 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3520 struct eventfd_ctx
*eventfd
)
3522 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3525 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3526 struct eventfd_ctx
*eventfd
)
3528 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3531 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3532 struct eventfd_ctx
*eventfd
, const char *args
)
3534 struct mem_cgroup_eventfd_list
*event
;
3536 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3540 spin_lock(&memcg_oom_lock
);
3542 event
->eventfd
= eventfd
;
3543 list_add(&event
->list
, &memcg
->oom_notify
);
3545 /* already in OOM ? */
3546 if (memcg
->under_oom
)
3547 eventfd_signal(eventfd
, 1);
3548 spin_unlock(&memcg_oom_lock
);
3553 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3554 struct eventfd_ctx
*eventfd
)
3556 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3558 spin_lock(&memcg_oom_lock
);
3560 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3561 if (ev
->eventfd
== eventfd
) {
3562 list_del(&ev
->list
);
3567 spin_unlock(&memcg_oom_lock
);
3570 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3572 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3574 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3575 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3579 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3580 struct cftype
*cft
, u64 val
)
3582 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3584 /* cannot set to root cgroup and only 0 and 1 are allowed */
3585 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3588 memcg
->oom_kill_disable
= val
;
3590 memcg_oom_recover(memcg
);
3595 #ifdef CONFIG_CGROUP_WRITEBACK
3597 struct list_head
*mem_cgroup_cgwb_list(struct mem_cgroup
*memcg
)
3599 return &memcg
->cgwb_list
;
3602 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3604 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3607 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3609 wb_domain_exit(&memcg
->cgwb_domain
);
3612 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3614 wb_domain_size_changed(&memcg
->cgwb_domain
);
3617 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3619 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3621 if (!memcg
->css
.parent
)
3624 return &memcg
->cgwb_domain
;
3628 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3629 * @wb: bdi_writeback in question
3630 * @pfilepages: out parameter for number of file pages
3631 * @pheadroom: out parameter for number of allocatable pages according to memcg
3632 * @pdirty: out parameter for number of dirty pages
3633 * @pwriteback: out parameter for number of pages under writeback
3635 * Determine the numbers of file, headroom, dirty, and writeback pages in
3636 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3637 * is a bit more involved.
3639 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3640 * headroom is calculated as the lowest headroom of itself and the
3641 * ancestors. Note that this doesn't consider the actual amount of
3642 * available memory in the system. The caller should further cap
3643 * *@pheadroom accordingly.
3645 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
3646 unsigned long *pheadroom
, unsigned long *pdirty
,
3647 unsigned long *pwriteback
)
3649 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3650 struct mem_cgroup
*parent
;
3652 *pdirty
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
3654 /* this should eventually include NR_UNSTABLE_NFS */
3655 *pwriteback
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
3656 *pfilepages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
3657 (1 << LRU_ACTIVE_FILE
));
3658 *pheadroom
= PAGE_COUNTER_MAX
;
3660 while ((parent
= parent_mem_cgroup(memcg
))) {
3661 unsigned long ceiling
= min(memcg
->memory
.limit
, memcg
->high
);
3662 unsigned long used
= page_counter_read(&memcg
->memory
);
3664 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
3669 #else /* CONFIG_CGROUP_WRITEBACK */
3671 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3676 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3680 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3684 #endif /* CONFIG_CGROUP_WRITEBACK */
3687 * DO NOT USE IN NEW FILES.
3689 * "cgroup.event_control" implementation.
3691 * This is way over-engineered. It tries to support fully configurable
3692 * events for each user. Such level of flexibility is completely
3693 * unnecessary especially in the light of the planned unified hierarchy.
3695 * Please deprecate this and replace with something simpler if at all
3700 * Unregister event and free resources.
3702 * Gets called from workqueue.
3704 static void memcg_event_remove(struct work_struct
*work
)
3706 struct mem_cgroup_event
*event
=
3707 container_of(work
, struct mem_cgroup_event
, remove
);
3708 struct mem_cgroup
*memcg
= event
->memcg
;
3710 remove_wait_queue(event
->wqh
, &event
->wait
);
3712 event
->unregister_event(memcg
, event
->eventfd
);
3714 /* Notify userspace the event is going away. */
3715 eventfd_signal(event
->eventfd
, 1);
3717 eventfd_ctx_put(event
->eventfd
);
3719 css_put(&memcg
->css
);
3723 * Gets called on POLLHUP on eventfd when user closes it.
3725 * Called with wqh->lock held and interrupts disabled.
3727 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
3728 int sync
, void *key
)
3730 struct mem_cgroup_event
*event
=
3731 container_of(wait
, struct mem_cgroup_event
, wait
);
3732 struct mem_cgroup
*memcg
= event
->memcg
;
3733 unsigned long flags
= (unsigned long)key
;
3735 if (flags
& POLLHUP
) {
3737 * If the event has been detached at cgroup removal, we
3738 * can simply return knowing the other side will cleanup
3741 * We can't race against event freeing since the other
3742 * side will require wqh->lock via remove_wait_queue(),
3745 spin_lock(&memcg
->event_list_lock
);
3746 if (!list_empty(&event
->list
)) {
3747 list_del_init(&event
->list
);
3749 * We are in atomic context, but cgroup_event_remove()
3750 * may sleep, so we have to call it in workqueue.
3752 schedule_work(&event
->remove
);
3754 spin_unlock(&memcg
->event_list_lock
);
3760 static void memcg_event_ptable_queue_proc(struct file
*file
,
3761 wait_queue_head_t
*wqh
, poll_table
*pt
)
3763 struct mem_cgroup_event
*event
=
3764 container_of(pt
, struct mem_cgroup_event
, pt
);
3767 add_wait_queue(wqh
, &event
->wait
);
3771 * DO NOT USE IN NEW FILES.
3773 * Parse input and register new cgroup event handler.
3775 * Input must be in format '<event_fd> <control_fd> <args>'.
3776 * Interpretation of args is defined by control file implementation.
3778 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
3779 char *buf
, size_t nbytes
, loff_t off
)
3781 struct cgroup_subsys_state
*css
= of_css(of
);
3782 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3783 struct mem_cgroup_event
*event
;
3784 struct cgroup_subsys_state
*cfile_css
;
3785 unsigned int efd
, cfd
;
3792 buf
= strstrip(buf
);
3794 efd
= simple_strtoul(buf
, &endp
, 10);
3799 cfd
= simple_strtoul(buf
, &endp
, 10);
3800 if ((*endp
!= ' ') && (*endp
!= '\0'))
3804 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
3808 event
->memcg
= memcg
;
3809 INIT_LIST_HEAD(&event
->list
);
3810 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
3811 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
3812 INIT_WORK(&event
->remove
, memcg_event_remove
);
3820 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
3821 if (IS_ERR(event
->eventfd
)) {
3822 ret
= PTR_ERR(event
->eventfd
);
3829 goto out_put_eventfd
;
3832 /* the process need read permission on control file */
3833 /* AV: shouldn't we check that it's been opened for read instead? */
3834 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
3839 * Determine the event callbacks and set them in @event. This used
3840 * to be done via struct cftype but cgroup core no longer knows
3841 * about these events. The following is crude but the whole thing
3842 * is for compatibility anyway.
3844 * DO NOT ADD NEW FILES.
3846 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
3848 if (!strcmp(name
, "memory.usage_in_bytes")) {
3849 event
->register_event
= mem_cgroup_usage_register_event
;
3850 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
3851 } else if (!strcmp(name
, "memory.oom_control")) {
3852 event
->register_event
= mem_cgroup_oom_register_event
;
3853 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
3854 } else if (!strcmp(name
, "memory.pressure_level")) {
3855 event
->register_event
= vmpressure_register_event
;
3856 event
->unregister_event
= vmpressure_unregister_event
;
3857 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
3858 event
->register_event
= memsw_cgroup_usage_register_event
;
3859 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
3866 * Verify @cfile should belong to @css. Also, remaining events are
3867 * automatically removed on cgroup destruction but the removal is
3868 * asynchronous, so take an extra ref on @css.
3870 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
3871 &memory_cgrp_subsys
);
3873 if (IS_ERR(cfile_css
))
3875 if (cfile_css
!= css
) {
3880 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
3884 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
3886 spin_lock(&memcg
->event_list_lock
);
3887 list_add(&event
->list
, &memcg
->event_list
);
3888 spin_unlock(&memcg
->event_list_lock
);
3900 eventfd_ctx_put(event
->eventfd
);
3909 static struct cftype mem_cgroup_legacy_files
[] = {
3911 .name
= "usage_in_bytes",
3912 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
3913 .read_u64
= mem_cgroup_read_u64
,
3916 .name
= "max_usage_in_bytes",
3917 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
3918 .write
= mem_cgroup_reset
,
3919 .read_u64
= mem_cgroup_read_u64
,
3922 .name
= "limit_in_bytes",
3923 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
3924 .write
= mem_cgroup_write
,
3925 .read_u64
= mem_cgroup_read_u64
,
3928 .name
= "soft_limit_in_bytes",
3929 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
3930 .write
= mem_cgroup_write
,
3931 .read_u64
= mem_cgroup_read_u64
,
3935 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
3936 .write
= mem_cgroup_reset
,
3937 .read_u64
= mem_cgroup_read_u64
,
3941 .seq_show
= memcg_stat_show
,
3944 .name
= "force_empty",
3945 .write
= mem_cgroup_force_empty_write
,
3948 .name
= "use_hierarchy",
3949 .write_u64
= mem_cgroup_hierarchy_write
,
3950 .read_u64
= mem_cgroup_hierarchy_read
,
3953 .name
= "cgroup.event_control", /* XXX: for compat */
3954 .write
= memcg_write_event_control
,
3955 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
3958 .name
= "swappiness",
3959 .read_u64
= mem_cgroup_swappiness_read
,
3960 .write_u64
= mem_cgroup_swappiness_write
,
3963 .name
= "move_charge_at_immigrate",
3964 .read_u64
= mem_cgroup_move_charge_read
,
3965 .write_u64
= mem_cgroup_move_charge_write
,
3968 .name
= "oom_control",
3969 .seq_show
= mem_cgroup_oom_control_read
,
3970 .write_u64
= mem_cgroup_oom_control_write
,
3971 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
3974 .name
= "pressure_level",
3978 .name
= "numa_stat",
3979 .seq_show
= memcg_numa_stat_show
,
3983 .name
= "kmem.limit_in_bytes",
3984 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
3985 .write
= mem_cgroup_write
,
3986 .read_u64
= mem_cgroup_read_u64
,
3989 .name
= "kmem.usage_in_bytes",
3990 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
3991 .read_u64
= mem_cgroup_read_u64
,
3994 .name
= "kmem.failcnt",
3995 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
3996 .write
= mem_cgroup_reset
,
3997 .read_u64
= mem_cgroup_read_u64
,
4000 .name
= "kmem.max_usage_in_bytes",
4001 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4002 .write
= mem_cgroup_reset
,
4003 .read_u64
= mem_cgroup_read_u64
,
4005 #ifdef CONFIG_SLABINFO
4007 .name
= "kmem.slabinfo",
4008 .seq_start
= slab_start
,
4009 .seq_next
= slab_next
,
4010 .seq_stop
= slab_stop
,
4011 .seq_show
= memcg_slab_show
,
4015 .name
= "kmem.tcp.limit_in_bytes",
4016 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4017 .write
= mem_cgroup_write
,
4018 .read_u64
= mem_cgroup_read_u64
,
4021 .name
= "kmem.tcp.usage_in_bytes",
4022 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4023 .read_u64
= mem_cgroup_read_u64
,
4026 .name
= "kmem.tcp.failcnt",
4027 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4028 .write
= mem_cgroup_reset
,
4029 .read_u64
= mem_cgroup_read_u64
,
4032 .name
= "kmem.tcp.max_usage_in_bytes",
4033 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4034 .write
= mem_cgroup_reset
,
4035 .read_u64
= mem_cgroup_read_u64
,
4037 { }, /* terminate */
4041 * Private memory cgroup IDR
4043 * Swap-out records and page cache shadow entries need to store memcg
4044 * references in constrained space, so we maintain an ID space that is
4045 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4046 * memory-controlled cgroups to 64k.
4048 * However, there usually are many references to the oflline CSS after
4049 * the cgroup has been destroyed, such as page cache or reclaimable
4050 * slab objects, that don't need to hang on to the ID. We want to keep
4051 * those dead CSS from occupying IDs, or we might quickly exhaust the
4052 * relatively small ID space and prevent the creation of new cgroups
4053 * even when there are much fewer than 64k cgroups - possibly none.
4055 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4056 * be freed and recycled when it's no longer needed, which is usually
4057 * when the CSS is offlined.
4059 * The only exception to that are records of swapped out tmpfs/shmem
4060 * pages that need to be attributed to live ancestors on swapin. But
4061 * those references are manageable from userspace.
4064 static DEFINE_IDR(mem_cgroup_idr
);
4066 static void mem_cgroup_id_get(struct mem_cgroup
*memcg
)
4068 atomic_inc(&memcg
->id
.ref
);
4071 static void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
4073 if (atomic_dec_and_test(&memcg
->id
.ref
)) {
4074 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4077 /* Memcg ID pins CSS */
4078 css_put(&memcg
->css
);
4083 * mem_cgroup_from_id - look up a memcg from a memcg id
4084 * @id: the memcg id to look up
4086 * Caller must hold rcu_read_lock().
4088 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
4090 WARN_ON_ONCE(!rcu_read_lock_held());
4091 return idr_find(&mem_cgroup_idr
, id
);
4094 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4096 struct mem_cgroup_per_node
*pn
;
4099 * This routine is called against possible nodes.
4100 * But it's BUG to call kmalloc() against offline node.
4102 * TODO: this routine can waste much memory for nodes which will
4103 * never be onlined. It's better to use memory hotplug callback
4106 if (!node_state(node
, N_NORMAL_MEMORY
))
4108 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4112 lruvec_init(&pn
->lruvec
);
4113 pn
->usage_in_excess
= 0;
4114 pn
->on_tree
= false;
4117 memcg
->nodeinfo
[node
] = pn
;
4121 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4123 kfree(memcg
->nodeinfo
[node
]);
4126 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4130 memcg_wb_domain_exit(memcg
);
4132 free_mem_cgroup_per_node_info(memcg
, node
);
4133 free_percpu(memcg
->stat
);
4137 static struct mem_cgroup
*mem_cgroup_alloc(void)
4139 struct mem_cgroup
*memcg
;
4143 size
= sizeof(struct mem_cgroup
);
4144 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4146 memcg
= kzalloc(size
, GFP_KERNEL
);
4150 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
4151 1, MEM_CGROUP_ID_MAX
,
4153 if (memcg
->id
.id
< 0)
4156 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4161 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
4164 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4167 INIT_WORK(&memcg
->high_work
, high_work_func
);
4168 memcg
->last_scanned_node
= MAX_NUMNODES
;
4169 INIT_LIST_HEAD(&memcg
->oom_notify
);
4170 mutex_init(&memcg
->thresholds_lock
);
4171 spin_lock_init(&memcg
->move_lock
);
4172 vmpressure_init(&memcg
->vmpressure
);
4173 INIT_LIST_HEAD(&memcg
->event_list
);
4174 spin_lock_init(&memcg
->event_list_lock
);
4175 memcg
->socket_pressure
= jiffies
;
4177 memcg
->kmemcg_id
= -1;
4179 #ifdef CONFIG_CGROUP_WRITEBACK
4180 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4182 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
4185 if (memcg
->id
.id
> 0)
4186 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4187 mem_cgroup_free(memcg
);
4191 static struct cgroup_subsys_state
* __ref
4192 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4194 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
4195 struct mem_cgroup
*memcg
;
4196 long error
= -ENOMEM
;
4198 memcg
= mem_cgroup_alloc();
4200 return ERR_PTR(error
);
4202 memcg
->high
= PAGE_COUNTER_MAX
;
4203 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4205 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4206 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4208 if (parent
&& parent
->use_hierarchy
) {
4209 memcg
->use_hierarchy
= true;
4210 page_counter_init(&memcg
->memory
, &parent
->memory
);
4211 page_counter_init(&memcg
->swap
, &parent
->swap
);
4212 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4213 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4214 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
4216 page_counter_init(&memcg
->memory
, NULL
);
4217 page_counter_init(&memcg
->swap
, NULL
);
4218 page_counter_init(&memcg
->memsw
, NULL
);
4219 page_counter_init(&memcg
->kmem
, NULL
);
4220 page_counter_init(&memcg
->tcpmem
, NULL
);
4222 * Deeper hierachy with use_hierarchy == false doesn't make
4223 * much sense so let cgroup subsystem know about this
4224 * unfortunate state in our controller.
4226 if (parent
!= root_mem_cgroup
)
4227 memory_cgrp_subsys
.broken_hierarchy
= true;
4230 /* The following stuff does not apply to the root */
4232 root_mem_cgroup
= memcg
;
4236 error
= memcg_online_kmem(memcg
);
4240 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4241 static_branch_inc(&memcg_sockets_enabled_key
);
4245 mem_cgroup_free(memcg
);
4246 return ERR_PTR(-ENOMEM
);
4249 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4251 /* Online state pins memcg ID, memcg ID pins CSS */
4252 mem_cgroup_id_get(mem_cgroup_from_css(css
));
4257 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4259 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4260 struct mem_cgroup_event
*event
, *tmp
;
4263 * Unregister events and notify userspace.
4264 * Notify userspace about cgroup removing only after rmdir of cgroup
4265 * directory to avoid race between userspace and kernelspace.
4267 spin_lock(&memcg
->event_list_lock
);
4268 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4269 list_del_init(&event
->list
);
4270 schedule_work(&event
->remove
);
4272 spin_unlock(&memcg
->event_list_lock
);
4274 memcg_offline_kmem(memcg
);
4275 wb_memcg_offline(memcg
);
4277 mem_cgroup_id_put(memcg
);
4280 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4282 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4284 invalidate_reclaim_iterators(memcg
);
4287 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4289 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4291 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4292 static_branch_dec(&memcg_sockets_enabled_key
);
4294 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
4295 static_branch_dec(&memcg_sockets_enabled_key
);
4297 vmpressure_cleanup(&memcg
->vmpressure
);
4298 cancel_work_sync(&memcg
->high_work
);
4299 mem_cgroup_remove_from_trees(memcg
);
4300 memcg_free_kmem(memcg
);
4301 mem_cgroup_free(memcg
);
4305 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4306 * @css: the target css
4308 * Reset the states of the mem_cgroup associated with @css. This is
4309 * invoked when the userland requests disabling on the default hierarchy
4310 * but the memcg is pinned through dependency. The memcg should stop
4311 * applying policies and should revert to the vanilla state as it may be
4312 * made visible again.
4314 * The current implementation only resets the essential configurations.
4315 * This needs to be expanded to cover all the visible parts.
4317 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4319 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4321 page_counter_limit(&memcg
->memory
, PAGE_COUNTER_MAX
);
4322 page_counter_limit(&memcg
->swap
, PAGE_COUNTER_MAX
);
4323 page_counter_limit(&memcg
->memsw
, PAGE_COUNTER_MAX
);
4324 page_counter_limit(&memcg
->kmem
, PAGE_COUNTER_MAX
);
4325 page_counter_limit(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
4327 memcg
->high
= PAGE_COUNTER_MAX
;
4328 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4329 memcg_wb_domain_size_changed(memcg
);
4333 /* Handlers for move charge at task migration. */
4334 static int mem_cgroup_do_precharge(unsigned long count
)
4338 /* Try a single bulk charge without reclaim first, kswapd may wake */
4339 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4341 mc
.precharge
+= count
;
4345 /* Try charges one by one with reclaim */
4347 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4361 enum mc_target_type
{
4367 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4368 unsigned long addr
, pte_t ptent
)
4370 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4372 if (!page
|| !page_mapped(page
))
4374 if (PageAnon(page
)) {
4375 if (!(mc
.flags
& MOVE_ANON
))
4378 if (!(mc
.flags
& MOVE_FILE
))
4381 if (!get_page_unless_zero(page
))
4388 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4389 pte_t ptent
, swp_entry_t
*entry
)
4391 struct page
*page
= NULL
;
4392 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4394 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4397 * Because lookup_swap_cache() updates some statistics counter,
4398 * we call find_get_page() with swapper_space directly.
4400 page
= find_get_page(swap_address_space(ent
), ent
.val
);
4401 if (do_memsw_account())
4402 entry
->val
= ent
.val
;
4407 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4408 pte_t ptent
, swp_entry_t
*entry
)
4414 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4415 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4417 struct page
*page
= NULL
;
4418 struct address_space
*mapping
;
4421 if (!vma
->vm_file
) /* anonymous vma */
4423 if (!(mc
.flags
& MOVE_FILE
))
4426 mapping
= vma
->vm_file
->f_mapping
;
4427 pgoff
= linear_page_index(vma
, addr
);
4429 /* page is moved even if it's not RSS of this task(page-faulted). */
4431 /* shmem/tmpfs may report page out on swap: account for that too. */
4432 if (shmem_mapping(mapping
)) {
4433 page
= find_get_entry(mapping
, pgoff
);
4434 if (radix_tree_exceptional_entry(page
)) {
4435 swp_entry_t swp
= radix_to_swp_entry(page
);
4436 if (do_memsw_account())
4438 page
= find_get_page(swap_address_space(swp
), swp
.val
);
4441 page
= find_get_page(mapping
, pgoff
);
4443 page
= find_get_page(mapping
, pgoff
);
4449 * mem_cgroup_move_account - move account of the page
4451 * @compound: charge the page as compound or small page
4452 * @from: mem_cgroup which the page is moved from.
4453 * @to: mem_cgroup which the page is moved to. @from != @to.
4455 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4457 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4460 static int mem_cgroup_move_account(struct page
*page
,
4462 struct mem_cgroup
*from
,
4463 struct mem_cgroup
*to
)
4465 unsigned long flags
;
4466 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
4470 VM_BUG_ON(from
== to
);
4471 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4472 VM_BUG_ON(compound
&& !PageTransHuge(page
));
4475 * Prevent mem_cgroup_migrate() from looking at
4476 * page->mem_cgroup of its source page while we change it.
4479 if (!trylock_page(page
))
4483 if (page
->mem_cgroup
!= from
)
4486 anon
= PageAnon(page
);
4488 spin_lock_irqsave(&from
->move_lock
, flags
);
4490 if (!anon
&& page_mapped(page
)) {
4491 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4493 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4498 * move_lock grabbed above and caller set from->moving_account, so
4499 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4500 * So mapping should be stable for dirty pages.
4502 if (!anon
&& PageDirty(page
)) {
4503 struct address_space
*mapping
= page_mapping(page
);
4505 if (mapping_cap_account_dirty(mapping
)) {
4506 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4508 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4513 if (PageWriteback(page
)) {
4514 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4516 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4521 * It is safe to change page->mem_cgroup here because the page
4522 * is referenced, charged, and isolated - we can't race with
4523 * uncharging, charging, migration, or LRU putback.
4526 /* caller should have done css_get */
4527 page
->mem_cgroup
= to
;
4528 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4532 local_irq_disable();
4533 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
4534 memcg_check_events(to
, page
);
4535 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
4536 memcg_check_events(from
, page
);
4545 * get_mctgt_type - get target type of moving charge
4546 * @vma: the vma the pte to be checked belongs
4547 * @addr: the address corresponding to the pte to be checked
4548 * @ptent: the pte to be checked
4549 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4552 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4553 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4554 * move charge. if @target is not NULL, the page is stored in target->page
4555 * with extra refcnt got(Callers should handle it).
4556 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4557 * target for charge migration. if @target is not NULL, the entry is stored
4560 * Called with pte lock held.
4563 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4564 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4566 struct page
*page
= NULL
;
4567 enum mc_target_type ret
= MC_TARGET_NONE
;
4568 swp_entry_t ent
= { .val
= 0 };
4570 if (pte_present(ptent
))
4571 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4572 else if (is_swap_pte(ptent
))
4573 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
4574 else if (pte_none(ptent
))
4575 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4577 if (!page
&& !ent
.val
)
4581 * Do only loose check w/o serialization.
4582 * mem_cgroup_move_account() checks the page is valid or
4583 * not under LRU exclusion.
4585 if (page
->mem_cgroup
== mc
.from
) {
4586 ret
= MC_TARGET_PAGE
;
4588 target
->page
= page
;
4590 if (!ret
|| !target
)
4593 /* There is a swap entry and a page doesn't exist or isn't charged */
4594 if (ent
.val
&& !ret
&&
4595 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4596 ret
= MC_TARGET_SWAP
;
4603 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4605 * We don't consider swapping or file mapped pages because THP does not
4606 * support them for now.
4607 * Caller should make sure that pmd_trans_huge(pmd) is true.
4609 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4610 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4612 struct page
*page
= NULL
;
4613 enum mc_target_type ret
= MC_TARGET_NONE
;
4615 page
= pmd_page(pmd
);
4616 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4617 if (!(mc
.flags
& MOVE_ANON
))
4619 if (page
->mem_cgroup
== mc
.from
) {
4620 ret
= MC_TARGET_PAGE
;
4623 target
->page
= page
;
4629 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4630 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4632 return MC_TARGET_NONE
;
4636 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4637 unsigned long addr
, unsigned long end
,
4638 struct mm_walk
*walk
)
4640 struct vm_area_struct
*vma
= walk
->vma
;
4644 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4646 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4647 mc
.precharge
+= HPAGE_PMD_NR
;
4652 if (pmd_trans_unstable(pmd
))
4654 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4655 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4656 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4657 mc
.precharge
++; /* increment precharge temporarily */
4658 pte_unmap_unlock(pte
- 1, ptl
);
4664 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4666 unsigned long precharge
;
4668 struct mm_walk mem_cgroup_count_precharge_walk
= {
4669 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4672 down_read(&mm
->mmap_sem
);
4673 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk
);
4674 up_read(&mm
->mmap_sem
);
4676 precharge
= mc
.precharge
;
4682 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4684 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4686 VM_BUG_ON(mc
.moving_task
);
4687 mc
.moving_task
= current
;
4688 return mem_cgroup_do_precharge(precharge
);
4691 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4692 static void __mem_cgroup_clear_mc(void)
4694 struct mem_cgroup
*from
= mc
.from
;
4695 struct mem_cgroup
*to
= mc
.to
;
4697 /* we must uncharge all the leftover precharges from mc.to */
4699 cancel_charge(mc
.to
, mc
.precharge
);
4703 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4704 * we must uncharge here.
4706 if (mc
.moved_charge
) {
4707 cancel_charge(mc
.from
, mc
.moved_charge
);
4708 mc
.moved_charge
= 0;
4710 /* we must fixup refcnts and charges */
4711 if (mc
.moved_swap
) {
4712 /* uncharge swap account from the old cgroup */
4713 if (!mem_cgroup_is_root(mc
.from
))
4714 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4717 * we charged both to->memory and to->memsw, so we
4718 * should uncharge to->memory.
4720 if (!mem_cgroup_is_root(mc
.to
))
4721 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4723 css_put_many(&mc
.from
->css
, mc
.moved_swap
);
4725 /* we've already done css_get(mc.to) */
4728 memcg_oom_recover(from
);
4729 memcg_oom_recover(to
);
4730 wake_up_all(&mc
.waitq
);
4733 static void mem_cgroup_clear_mc(void)
4735 struct mm_struct
*mm
= mc
.mm
;
4738 * we must clear moving_task before waking up waiters at the end of
4741 mc
.moving_task
= NULL
;
4742 __mem_cgroup_clear_mc();
4743 spin_lock(&mc
.lock
);
4747 spin_unlock(&mc
.lock
);
4752 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4754 struct cgroup_subsys_state
*css
;
4755 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
4756 struct mem_cgroup
*from
;
4757 struct task_struct
*leader
, *p
;
4758 struct mm_struct
*mm
;
4759 unsigned long move_flags
;
4762 /* charge immigration isn't supported on the default hierarchy */
4763 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4767 * Multi-process migrations only happen on the default hierarchy
4768 * where charge immigration is not used. Perform charge
4769 * immigration if @tset contains a leader and whine if there are
4773 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
4776 memcg
= mem_cgroup_from_css(css
);
4782 * We are now commited to this value whatever it is. Changes in this
4783 * tunable will only affect upcoming migrations, not the current one.
4784 * So we need to save it, and keep it going.
4786 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
4790 from
= mem_cgroup_from_task(p
);
4792 VM_BUG_ON(from
== memcg
);
4794 mm
= get_task_mm(p
);
4797 /* We move charges only when we move a owner of the mm */
4798 if (mm
->owner
== p
) {
4801 VM_BUG_ON(mc
.precharge
);
4802 VM_BUG_ON(mc
.moved_charge
);
4803 VM_BUG_ON(mc
.moved_swap
);
4805 spin_lock(&mc
.lock
);
4809 mc
.flags
= move_flags
;
4810 spin_unlock(&mc
.lock
);
4811 /* We set mc.moving_task later */
4813 ret
= mem_cgroup_precharge_mc(mm
);
4815 mem_cgroup_clear_mc();
4822 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4825 mem_cgroup_clear_mc();
4828 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4829 unsigned long addr
, unsigned long end
,
4830 struct mm_walk
*walk
)
4833 struct vm_area_struct
*vma
= walk
->vma
;
4836 enum mc_target_type target_type
;
4837 union mc_target target
;
4840 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4842 if (mc
.precharge
< HPAGE_PMD_NR
) {
4846 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
4847 if (target_type
== MC_TARGET_PAGE
) {
4849 if (!isolate_lru_page(page
)) {
4850 if (!mem_cgroup_move_account(page
, true,
4852 mc
.precharge
-= HPAGE_PMD_NR
;
4853 mc
.moved_charge
+= HPAGE_PMD_NR
;
4855 putback_lru_page(page
);
4863 if (pmd_trans_unstable(pmd
))
4866 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4867 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4868 pte_t ptent
= *(pte
++);
4874 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
4875 case MC_TARGET_PAGE
:
4878 * We can have a part of the split pmd here. Moving it
4879 * can be done but it would be too convoluted so simply
4880 * ignore such a partial THP and keep it in original
4881 * memcg. There should be somebody mapping the head.
4883 if (PageTransCompound(page
))
4885 if (isolate_lru_page(page
))
4887 if (!mem_cgroup_move_account(page
, false,
4890 /* we uncharge from mc.from later. */
4893 putback_lru_page(page
);
4894 put
: /* get_mctgt_type() gets the page */
4897 case MC_TARGET_SWAP
:
4899 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
4901 /* we fixup refcnts and charges later. */
4909 pte_unmap_unlock(pte
- 1, ptl
);
4914 * We have consumed all precharges we got in can_attach().
4915 * We try charge one by one, but don't do any additional
4916 * charges to mc.to if we have failed in charge once in attach()
4919 ret
= mem_cgroup_do_precharge(1);
4927 static void mem_cgroup_move_charge(void)
4929 struct mm_walk mem_cgroup_move_charge_walk
= {
4930 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4934 lru_add_drain_all();
4936 * Signal lock_page_memcg() to take the memcg's move_lock
4937 * while we're moving its pages to another memcg. Then wait
4938 * for already started RCU-only updates to finish.
4940 atomic_inc(&mc
.from
->moving_account
);
4943 if (unlikely(!down_read_trylock(&mc
.mm
->mmap_sem
))) {
4945 * Someone who are holding the mmap_sem might be waiting in
4946 * waitq. So we cancel all extra charges, wake up all waiters,
4947 * and retry. Because we cancel precharges, we might not be able
4948 * to move enough charges, but moving charge is a best-effort
4949 * feature anyway, so it wouldn't be a big problem.
4951 __mem_cgroup_clear_mc();
4956 * When we have consumed all precharges and failed in doing
4957 * additional charge, the page walk just aborts.
4959 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk
);
4960 up_read(&mc
.mm
->mmap_sem
);
4961 atomic_dec(&mc
.from
->moving_account
);
4964 static void mem_cgroup_move_task(void)
4967 mem_cgroup_move_charge();
4968 mem_cgroup_clear_mc();
4971 #else /* !CONFIG_MMU */
4972 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4976 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4979 static void mem_cgroup_move_task(void)
4985 * Cgroup retains root cgroups across [un]mount cycles making it necessary
4986 * to verify whether we're attached to the default hierarchy on each mount
4989 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
4992 * use_hierarchy is forced on the default hierarchy. cgroup core
4993 * guarantees that @root doesn't have any children, so turning it
4994 * on for the root memcg is enough.
4996 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4997 root_mem_cgroup
->use_hierarchy
= true;
4999 root_mem_cgroup
->use_hierarchy
= false;
5002 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5005 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5007 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
5010 static int memory_low_show(struct seq_file
*m
, void *v
)
5012 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5013 unsigned long low
= READ_ONCE(memcg
->low
);
5015 if (low
== PAGE_COUNTER_MAX
)
5016 seq_puts(m
, "max\n");
5018 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
5023 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5024 char *buf
, size_t nbytes
, loff_t off
)
5026 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5030 buf
= strstrip(buf
);
5031 err
= page_counter_memparse(buf
, "max", &low
);
5040 static int memory_high_show(struct seq_file
*m
, void *v
)
5042 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5043 unsigned long high
= READ_ONCE(memcg
->high
);
5045 if (high
== PAGE_COUNTER_MAX
)
5046 seq_puts(m
, "max\n");
5048 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
5053 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5054 char *buf
, size_t nbytes
, loff_t off
)
5056 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5057 unsigned long nr_pages
;
5061 buf
= strstrip(buf
);
5062 err
= page_counter_memparse(buf
, "max", &high
);
5068 nr_pages
= page_counter_read(&memcg
->memory
);
5069 if (nr_pages
> high
)
5070 try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
5073 memcg_wb_domain_size_changed(memcg
);
5077 static int memory_max_show(struct seq_file
*m
, void *v
)
5079 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5080 unsigned long max
= READ_ONCE(memcg
->memory
.limit
);
5082 if (max
== PAGE_COUNTER_MAX
)
5083 seq_puts(m
, "max\n");
5085 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5090 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5091 char *buf
, size_t nbytes
, loff_t off
)
5093 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5094 unsigned int nr_reclaims
= MEM_CGROUP_RECLAIM_RETRIES
;
5095 bool drained
= false;
5099 buf
= strstrip(buf
);
5100 err
= page_counter_memparse(buf
, "max", &max
);
5104 xchg(&memcg
->memory
.limit
, max
);
5107 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
5109 if (nr_pages
<= max
)
5112 if (signal_pending(current
)) {
5118 drain_all_stock(memcg
);
5124 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
5130 mem_cgroup_events(memcg
, MEMCG_OOM
, 1);
5131 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
5135 memcg_wb_domain_size_changed(memcg
);
5139 static int memory_events_show(struct seq_file
*m
, void *v
)
5141 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5143 seq_printf(m
, "low %lu\n", mem_cgroup_read_events(memcg
, MEMCG_LOW
));
5144 seq_printf(m
, "high %lu\n", mem_cgroup_read_events(memcg
, MEMCG_HIGH
));
5145 seq_printf(m
, "max %lu\n", mem_cgroup_read_events(memcg
, MEMCG_MAX
));
5146 seq_printf(m
, "oom %lu\n", mem_cgroup_read_events(memcg
, MEMCG_OOM
));
5151 static int memory_stat_show(struct seq_file
*m
, void *v
)
5153 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5154 unsigned long stat
[MEMCG_NR_STAT
];
5155 unsigned long events
[MEMCG_NR_EVENTS
];
5159 * Provide statistics on the state of the memory subsystem as
5160 * well as cumulative event counters that show past behavior.
5162 * This list is ordered following a combination of these gradients:
5163 * 1) generic big picture -> specifics and details
5164 * 2) reflecting userspace activity -> reflecting kernel heuristics
5166 * Current memory state:
5169 tree_stat(memcg
, stat
);
5170 tree_events(memcg
, events
);
5172 seq_printf(m
, "anon %llu\n",
5173 (u64
)stat
[MEM_CGROUP_STAT_RSS
] * PAGE_SIZE
);
5174 seq_printf(m
, "file %llu\n",
5175 (u64
)stat
[MEM_CGROUP_STAT_CACHE
] * PAGE_SIZE
);
5176 seq_printf(m
, "kernel_stack %llu\n",
5177 (u64
)stat
[MEMCG_KERNEL_STACK
] * PAGE_SIZE
);
5178 seq_printf(m
, "slab %llu\n",
5179 (u64
)(stat
[MEMCG_SLAB_RECLAIMABLE
] +
5180 stat
[MEMCG_SLAB_UNRECLAIMABLE
]) * PAGE_SIZE
);
5181 seq_printf(m
, "sock %llu\n",
5182 (u64
)stat
[MEMCG_SOCK
] * PAGE_SIZE
);
5184 seq_printf(m
, "file_mapped %llu\n",
5185 (u64
)stat
[MEM_CGROUP_STAT_FILE_MAPPED
] * PAGE_SIZE
);
5186 seq_printf(m
, "file_dirty %llu\n",
5187 (u64
)stat
[MEM_CGROUP_STAT_DIRTY
] * PAGE_SIZE
);
5188 seq_printf(m
, "file_writeback %llu\n",
5189 (u64
)stat
[MEM_CGROUP_STAT_WRITEBACK
] * PAGE_SIZE
);
5191 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5192 struct mem_cgroup
*mi
;
5193 unsigned long val
= 0;
5195 for_each_mem_cgroup_tree(mi
, memcg
)
5196 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
));
5197 seq_printf(m
, "%s %llu\n",
5198 mem_cgroup_lru_names
[i
], (u64
)val
* PAGE_SIZE
);
5201 seq_printf(m
, "slab_reclaimable %llu\n",
5202 (u64
)stat
[MEMCG_SLAB_RECLAIMABLE
] * PAGE_SIZE
);
5203 seq_printf(m
, "slab_unreclaimable %llu\n",
5204 (u64
)stat
[MEMCG_SLAB_UNRECLAIMABLE
] * PAGE_SIZE
);
5206 /* Accumulated memory events */
5208 seq_printf(m
, "pgfault %lu\n",
5209 events
[MEM_CGROUP_EVENTS_PGFAULT
]);
5210 seq_printf(m
, "pgmajfault %lu\n",
5211 events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
5216 static struct cftype memory_files
[] = {
5219 .flags
= CFTYPE_NOT_ON_ROOT
,
5220 .read_u64
= memory_current_read
,
5224 .flags
= CFTYPE_NOT_ON_ROOT
,
5225 .seq_show
= memory_low_show
,
5226 .write
= memory_low_write
,
5230 .flags
= CFTYPE_NOT_ON_ROOT
,
5231 .seq_show
= memory_high_show
,
5232 .write
= memory_high_write
,
5236 .flags
= CFTYPE_NOT_ON_ROOT
,
5237 .seq_show
= memory_max_show
,
5238 .write
= memory_max_write
,
5242 .flags
= CFTYPE_NOT_ON_ROOT
,
5243 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
5244 .seq_show
= memory_events_show
,
5248 .flags
= CFTYPE_NOT_ON_ROOT
,
5249 .seq_show
= memory_stat_show
,
5254 struct cgroup_subsys memory_cgrp_subsys
= {
5255 .css_alloc
= mem_cgroup_css_alloc
,
5256 .css_online
= mem_cgroup_css_online
,
5257 .css_offline
= mem_cgroup_css_offline
,
5258 .css_released
= mem_cgroup_css_released
,
5259 .css_free
= mem_cgroup_css_free
,
5260 .css_reset
= mem_cgroup_css_reset
,
5261 .can_attach
= mem_cgroup_can_attach
,
5262 .cancel_attach
= mem_cgroup_cancel_attach
,
5263 .post_attach
= mem_cgroup_move_task
,
5264 .bind
= mem_cgroup_bind
,
5265 .dfl_cftypes
= memory_files
,
5266 .legacy_cftypes
= mem_cgroup_legacy_files
,
5271 * mem_cgroup_low - check if memory consumption is below the normal range
5272 * @root: the highest ancestor to consider
5273 * @memcg: the memory cgroup to check
5275 * Returns %true if memory consumption of @memcg, and that of all
5276 * configurable ancestors up to @root, is below the normal range.
5278 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5280 if (mem_cgroup_disabled())
5284 * The toplevel group doesn't have a configurable range, so
5285 * it's never low when looked at directly, and it is not
5286 * considered an ancestor when assessing the hierarchy.
5289 if (memcg
== root_mem_cgroup
)
5292 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5295 while (memcg
!= root
) {
5296 memcg
= parent_mem_cgroup(memcg
);
5298 if (memcg
== root_mem_cgroup
)
5301 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5308 * mem_cgroup_try_charge - try charging a page
5309 * @page: page to charge
5310 * @mm: mm context of the victim
5311 * @gfp_mask: reclaim mode
5312 * @memcgp: charged memcg return
5313 * @compound: charge the page as compound or small page
5315 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5316 * pages according to @gfp_mask if necessary.
5318 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5319 * Otherwise, an error code is returned.
5321 * After page->mapping has been set up, the caller must finalize the
5322 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5323 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5325 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5326 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
5329 struct mem_cgroup
*memcg
= NULL
;
5330 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5333 if (mem_cgroup_disabled())
5336 if (PageSwapCache(page
)) {
5338 * Every swap fault against a single page tries to charge the
5339 * page, bail as early as possible. shmem_unuse() encounters
5340 * already charged pages, too. The USED bit is protected by
5341 * the page lock, which serializes swap cache removal, which
5342 * in turn serializes uncharging.
5344 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5345 if (page
->mem_cgroup
)
5348 if (do_swap_account
) {
5349 swp_entry_t ent
= { .val
= page_private(page
), };
5350 unsigned short id
= lookup_swap_cgroup_id(ent
);
5353 memcg
= mem_cgroup_from_id(id
);
5354 if (memcg
&& !css_tryget_online(&memcg
->css
))
5361 memcg
= get_mem_cgroup_from_mm(mm
);
5363 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5365 css_put(&memcg
->css
);
5372 * mem_cgroup_commit_charge - commit a page charge
5373 * @page: page to charge
5374 * @memcg: memcg to charge the page to
5375 * @lrucare: page might be on LRU already
5376 * @compound: charge the page as compound or small page
5378 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5379 * after page->mapping has been set up. This must happen atomically
5380 * as part of the page instantiation, i.e. under the page table lock
5381 * for anonymous pages, under the page lock for page and swap cache.
5383 * In addition, the page must not be on the LRU during the commit, to
5384 * prevent racing with task migration. If it might be, use @lrucare.
5386 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5388 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5389 bool lrucare
, bool compound
)
5391 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5393 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5394 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5396 if (mem_cgroup_disabled())
5399 * Swap faults will attempt to charge the same page multiple
5400 * times. But reuse_swap_page() might have removed the page
5401 * from swapcache already, so we can't check PageSwapCache().
5406 commit_charge(page
, memcg
, lrucare
);
5408 local_irq_disable();
5409 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
5410 memcg_check_events(memcg
, page
);
5413 if (do_memsw_account() && PageSwapCache(page
)) {
5414 swp_entry_t entry
= { .val
= page_private(page
) };
5416 * The swap entry might not get freed for a long time,
5417 * let's not wait for it. The page already received a
5418 * memory+swap charge, drop the swap entry duplicate.
5420 mem_cgroup_uncharge_swap(entry
);
5425 * mem_cgroup_cancel_charge - cancel a page charge
5426 * @page: page to charge
5427 * @memcg: memcg to charge the page to
5428 * @compound: charge the page as compound or small page
5430 * Cancel a charge transaction started by mem_cgroup_try_charge().
5432 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5435 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5437 if (mem_cgroup_disabled())
5440 * Swap faults will attempt to charge the same page multiple
5441 * times. But reuse_swap_page() might have removed the page
5442 * from swapcache already, so we can't check PageSwapCache().
5447 cancel_charge(memcg
, nr_pages
);
5450 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5451 unsigned long nr_anon
, unsigned long nr_file
,
5452 unsigned long nr_huge
, unsigned long nr_kmem
,
5453 struct page
*dummy_page
)
5455 unsigned long nr_pages
= nr_anon
+ nr_file
+ nr_kmem
;
5456 unsigned long flags
;
5458 if (!mem_cgroup_is_root(memcg
)) {
5459 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5460 if (do_memsw_account())
5461 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5462 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && nr_kmem
)
5463 page_counter_uncharge(&memcg
->kmem
, nr_kmem
);
5464 memcg_oom_recover(memcg
);
5467 local_irq_save(flags
);
5468 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5469 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5470 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5471 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5472 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5473 memcg_check_events(memcg
, dummy_page
);
5474 local_irq_restore(flags
);
5476 if (!mem_cgroup_is_root(memcg
))
5477 css_put_many(&memcg
->css
, nr_pages
);
5480 static void uncharge_list(struct list_head
*page_list
)
5482 struct mem_cgroup
*memcg
= NULL
;
5483 unsigned long nr_anon
= 0;
5484 unsigned long nr_file
= 0;
5485 unsigned long nr_huge
= 0;
5486 unsigned long nr_kmem
= 0;
5487 unsigned long pgpgout
= 0;
5488 struct list_head
*next
;
5492 * Note that the list can be a single page->lru; hence the
5493 * do-while loop instead of a simple list_for_each_entry().
5495 next
= page_list
->next
;
5497 page
= list_entry(next
, struct page
, lru
);
5498 next
= page
->lru
.next
;
5500 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5501 VM_BUG_ON_PAGE(page_count(page
), page
);
5503 if (!page
->mem_cgroup
)
5507 * Nobody should be changing or seriously looking at
5508 * page->mem_cgroup at this point, we have fully
5509 * exclusive access to the page.
5512 if (memcg
!= page
->mem_cgroup
) {
5514 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5515 nr_huge
, nr_kmem
, page
);
5516 pgpgout
= nr_anon
= nr_file
=
5517 nr_huge
= nr_kmem
= 0;
5519 memcg
= page
->mem_cgroup
;
5522 if (!PageKmemcg(page
)) {
5523 unsigned int nr_pages
= 1;
5525 if (PageTransHuge(page
)) {
5526 nr_pages
<<= compound_order(page
);
5527 nr_huge
+= nr_pages
;
5530 nr_anon
+= nr_pages
;
5532 nr_file
+= nr_pages
;
5535 nr_kmem
+= 1 << compound_order(page
);
5537 page
->mem_cgroup
= NULL
;
5538 } while (next
!= page_list
);
5541 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5542 nr_huge
, nr_kmem
, page
);
5546 * mem_cgroup_uncharge - uncharge a page
5547 * @page: page to uncharge
5549 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5550 * mem_cgroup_commit_charge().
5552 void mem_cgroup_uncharge(struct page
*page
)
5554 if (mem_cgroup_disabled())
5557 /* Don't touch page->lru of any random page, pre-check: */
5558 if (!page
->mem_cgroup
)
5561 INIT_LIST_HEAD(&page
->lru
);
5562 uncharge_list(&page
->lru
);
5566 * mem_cgroup_uncharge_list - uncharge a list of page
5567 * @page_list: list of pages to uncharge
5569 * Uncharge a list of pages previously charged with
5570 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5572 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5574 if (mem_cgroup_disabled())
5577 if (!list_empty(page_list
))
5578 uncharge_list(page_list
);
5582 * mem_cgroup_migrate - charge a page's replacement
5583 * @oldpage: currently circulating page
5584 * @newpage: replacement page
5586 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5587 * be uncharged upon free.
5589 * Both pages must be locked, @newpage->mapping must be set up.
5591 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
5593 struct mem_cgroup
*memcg
;
5594 unsigned int nr_pages
;
5596 unsigned long flags
;
5598 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5599 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5600 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5601 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5604 if (mem_cgroup_disabled())
5607 /* Page cache replacement: new page already charged? */
5608 if (newpage
->mem_cgroup
)
5611 /* Swapcache readahead pages can get replaced before being charged */
5612 memcg
= oldpage
->mem_cgroup
;
5616 /* Force-charge the new page. The old one will be freed soon */
5617 compound
= PageTransHuge(newpage
);
5618 nr_pages
= compound
? hpage_nr_pages(newpage
) : 1;
5620 page_counter_charge(&memcg
->memory
, nr_pages
);
5621 if (do_memsw_account())
5622 page_counter_charge(&memcg
->memsw
, nr_pages
);
5623 css_get_many(&memcg
->css
, nr_pages
);
5625 commit_charge(newpage
, memcg
, false);
5627 local_irq_save(flags
);
5628 mem_cgroup_charge_statistics(memcg
, newpage
, compound
, nr_pages
);
5629 memcg_check_events(memcg
, newpage
);
5630 local_irq_restore(flags
);
5633 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
5634 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
5636 void sock_update_memcg(struct sock
*sk
)
5638 struct mem_cgroup
*memcg
;
5640 /* Socket cloning can throw us here with sk_cgrp already
5641 * filled. It won't however, necessarily happen from
5642 * process context. So the test for root memcg given
5643 * the current task's memcg won't help us in this case.
5645 * Respecting the original socket's memcg is a better
5646 * decision in this case.
5649 BUG_ON(mem_cgroup_is_root(sk
->sk_memcg
));
5650 css_get(&sk
->sk_memcg
->css
);
5655 memcg
= mem_cgroup_from_task(current
);
5656 if (memcg
== root_mem_cgroup
)
5658 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
5660 if (css_tryget_online(&memcg
->css
))
5661 sk
->sk_memcg
= memcg
;
5665 EXPORT_SYMBOL(sock_update_memcg
);
5667 void sock_release_memcg(struct sock
*sk
)
5669 WARN_ON(!sk
->sk_memcg
);
5670 css_put(&sk
->sk_memcg
->css
);
5674 * mem_cgroup_charge_skmem - charge socket memory
5675 * @memcg: memcg to charge
5676 * @nr_pages: number of pages to charge
5678 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5679 * @memcg's configured limit, %false if the charge had to be forced.
5681 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5683 gfp_t gfp_mask
= GFP_KERNEL
;
5685 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5686 struct page_counter
*fail
;
5688 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
5689 memcg
->tcpmem_pressure
= 0;
5692 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
5693 memcg
->tcpmem_pressure
= 1;
5697 /* Don't block in the packet receive path */
5699 gfp_mask
= GFP_NOWAIT
;
5701 this_cpu_add(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5703 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
5706 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
5711 * mem_cgroup_uncharge_skmem - uncharge socket memory
5712 * @memcg - memcg to uncharge
5713 * @nr_pages - number of pages to uncharge
5715 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5717 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5718 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
5722 this_cpu_sub(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5724 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5725 css_put_many(&memcg
->css
, nr_pages
);
5728 static int __init
cgroup_memory(char *s
)
5732 while ((token
= strsep(&s
, ",")) != NULL
) {
5735 if (!strcmp(token
, "nosocket"))
5736 cgroup_memory_nosocket
= true;
5737 if (!strcmp(token
, "nokmem"))
5738 cgroup_memory_nokmem
= true;
5742 __setup("cgroup.memory=", cgroup_memory
);
5745 * subsys_initcall() for memory controller.
5747 * Some parts like hotcpu_notifier() have to be initialized from this context
5748 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5749 * everything that doesn't depend on a specific mem_cgroup structure should
5750 * be initialized from here.
5752 static int __init
mem_cgroup_init(void)
5756 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5758 for_each_possible_cpu(cpu
)
5759 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5762 for_each_node(node
) {
5763 struct mem_cgroup_tree_per_node
*rtpn
;
5765 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5766 node_online(node
) ? node
: NUMA_NO_NODE
);
5768 rtpn
->rb_root
= RB_ROOT
;
5769 spin_lock_init(&rtpn
->lock
);
5770 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5775 subsys_initcall(mem_cgroup_init
);
5777 #ifdef CONFIG_MEMCG_SWAP
5779 * mem_cgroup_swapout - transfer a memsw charge to swap
5780 * @page: page whose memsw charge to transfer
5781 * @entry: swap entry to move the charge to
5783 * Transfer the memsw charge of @page to @entry.
5785 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5787 struct mem_cgroup
*memcg
;
5788 unsigned short oldid
;
5790 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5791 VM_BUG_ON_PAGE(page_count(page
), page
);
5793 if (!do_memsw_account())
5796 memcg
= page
->mem_cgroup
;
5798 /* Readahead page, never charged */
5802 mem_cgroup_id_get(memcg
);
5803 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5804 VM_BUG_ON_PAGE(oldid
, page
);
5805 mem_cgroup_swap_statistics(memcg
, true);
5807 page
->mem_cgroup
= NULL
;
5809 if (!mem_cgroup_is_root(memcg
))
5810 page_counter_uncharge(&memcg
->memory
, 1);
5813 * Interrupts should be disabled here because the caller holds the
5814 * mapping->tree_lock lock which is taken with interrupts-off. It is
5815 * important here to have the interrupts disabled because it is the
5816 * only synchronisation we have for udpating the per-CPU variables.
5818 VM_BUG_ON(!irqs_disabled());
5819 mem_cgroup_charge_statistics(memcg
, page
, false, -1);
5820 memcg_check_events(memcg
, page
);
5822 if (!mem_cgroup_is_root(memcg
))
5823 css_put(&memcg
->css
);
5827 * mem_cgroup_try_charge_swap - try charging a swap entry
5828 * @page: page being added to swap
5829 * @entry: swap entry to charge
5831 * Try to charge @entry to the memcg that @page belongs to.
5833 * Returns 0 on success, -ENOMEM on failure.
5835 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
5837 struct mem_cgroup
*memcg
;
5838 struct page_counter
*counter
;
5839 unsigned short oldid
;
5841 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
5844 memcg
= page
->mem_cgroup
;
5846 /* Readahead page, never charged */
5850 if (!mem_cgroup_is_root(memcg
) &&
5851 !page_counter_try_charge(&memcg
->swap
, 1, &counter
))
5854 mem_cgroup_id_get(memcg
);
5855 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5856 VM_BUG_ON_PAGE(oldid
, page
);
5857 mem_cgroup_swap_statistics(memcg
, true);
5863 * mem_cgroup_uncharge_swap - uncharge a swap entry
5864 * @entry: swap entry to uncharge
5866 * Drop the swap charge associated with @entry.
5868 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5870 struct mem_cgroup
*memcg
;
5873 if (!do_swap_account
)
5876 id
= swap_cgroup_record(entry
, 0);
5878 memcg
= mem_cgroup_from_id(id
);
5880 if (!mem_cgroup_is_root(memcg
)) {
5881 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5882 page_counter_uncharge(&memcg
->swap
, 1);
5884 page_counter_uncharge(&memcg
->memsw
, 1);
5886 mem_cgroup_swap_statistics(memcg
, false);
5887 mem_cgroup_id_put(memcg
);
5892 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
5894 long nr_swap_pages
= get_nr_swap_pages();
5896 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5897 return nr_swap_pages
;
5898 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5899 nr_swap_pages
= min_t(long, nr_swap_pages
,
5900 READ_ONCE(memcg
->swap
.limit
) -
5901 page_counter_read(&memcg
->swap
));
5902 return nr_swap_pages
;
5905 bool mem_cgroup_swap_full(struct page
*page
)
5907 struct mem_cgroup
*memcg
;
5909 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5913 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5916 memcg
= page
->mem_cgroup
;
5920 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5921 if (page_counter_read(&memcg
->swap
) * 2 >= memcg
->swap
.limit
)
5927 /* for remember boot option*/
5928 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5929 static int really_do_swap_account __initdata
= 1;
5931 static int really_do_swap_account __initdata
;
5934 static int __init
enable_swap_account(char *s
)
5936 if (!strcmp(s
, "1"))
5937 really_do_swap_account
= 1;
5938 else if (!strcmp(s
, "0"))
5939 really_do_swap_account
= 0;
5942 __setup("swapaccount=", enable_swap_account
);
5944 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
5947 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5949 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
5952 static int swap_max_show(struct seq_file
*m
, void *v
)
5954 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5955 unsigned long max
= READ_ONCE(memcg
->swap
.limit
);
5957 if (max
== PAGE_COUNTER_MAX
)
5958 seq_puts(m
, "max\n");
5960 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5965 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
5966 char *buf
, size_t nbytes
, loff_t off
)
5968 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5972 buf
= strstrip(buf
);
5973 err
= page_counter_memparse(buf
, "max", &max
);
5977 mutex_lock(&memcg_limit_mutex
);
5978 err
= page_counter_limit(&memcg
->swap
, max
);
5979 mutex_unlock(&memcg_limit_mutex
);
5986 static struct cftype swap_files
[] = {
5988 .name
= "swap.current",
5989 .flags
= CFTYPE_NOT_ON_ROOT
,
5990 .read_u64
= swap_current_read
,
5994 .flags
= CFTYPE_NOT_ON_ROOT
,
5995 .seq_show
= swap_max_show
,
5996 .write
= swap_max_write
,
6001 static struct cftype memsw_cgroup_files
[] = {
6003 .name
= "memsw.usage_in_bytes",
6004 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6005 .read_u64
= mem_cgroup_read_u64
,
6008 .name
= "memsw.max_usage_in_bytes",
6009 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6010 .write
= mem_cgroup_reset
,
6011 .read_u64
= mem_cgroup_read_u64
,
6014 .name
= "memsw.limit_in_bytes",
6015 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6016 .write
= mem_cgroup_write
,
6017 .read_u64
= mem_cgroup_read_u64
,
6020 .name
= "memsw.failcnt",
6021 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6022 .write
= mem_cgroup_reset
,
6023 .read_u64
= mem_cgroup_read_u64
,
6025 { }, /* terminate */
6028 static int __init
mem_cgroup_swap_init(void)
6030 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6031 do_swap_account
= 1;
6032 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
6034 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
6035 memsw_cgroup_files
));
6039 subsys_initcall(mem_cgroup_swap_init
);
6041 #endif /* CONFIG_MEMCG_SWAP */