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 <linux/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
76 EXPORT_SYMBOL(memory_cgrp_subsys
);
78 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
80 #define MEM_CGROUP_RECLAIM_RETRIES 5
82 /* Socket memory accounting disabled? */
83 static bool cgroup_memory_nosocket
;
85 /* Kernel memory accounting disabled? */
86 static bool cgroup_memory_nokmem
;
88 /* Whether the swap controller is active */
89 #ifdef CONFIG_MEMCG_SWAP
90 int do_swap_account __read_mostly
;
92 #define do_swap_account 0
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys
) && do_swap_account
;
101 static const char * const mem_cgroup_stat_names
[] = {
111 static const char * const mem_cgroup_events_names
[] = {
118 static const char * const mem_cgroup_lru_names
[] = {
126 #define THRESHOLDS_EVENTS_TARGET 128
127 #define SOFTLIMIT_EVENTS_TARGET 1024
128 #define NUMAINFO_EVENTS_TARGET 1024
131 * Cgroups above their limits are maintained in a RB-Tree, independent of
132 * their hierarchy representation
135 struct mem_cgroup_tree_per_node
{
136 struct rb_root rb_root
;
140 struct mem_cgroup_tree
{
141 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
144 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
147 struct mem_cgroup_eventfd_list
{
148 struct list_head list
;
149 struct eventfd_ctx
*eventfd
;
153 * cgroup_event represents events which userspace want to receive.
155 struct mem_cgroup_event
{
157 * memcg which the event belongs to.
159 struct mem_cgroup
*memcg
;
161 * eventfd to signal userspace about the event.
163 struct eventfd_ctx
*eventfd
;
165 * Each of these stored in a list by the cgroup.
167 struct list_head list
;
169 * register_event() callback will be used to add new userspace
170 * waiter for changes related to this event. Use eventfd_signal()
171 * on eventfd to send notification to userspace.
173 int (*register_event
)(struct mem_cgroup
*memcg
,
174 struct eventfd_ctx
*eventfd
, const char *args
);
176 * unregister_event() callback will be called when userspace closes
177 * the eventfd or on cgroup removing. This callback must be set,
178 * if you want provide notification functionality.
180 void (*unregister_event
)(struct mem_cgroup
*memcg
,
181 struct eventfd_ctx
*eventfd
);
183 * All fields below needed to unregister event when
184 * userspace closes eventfd.
187 wait_queue_head_t
*wqh
;
189 struct work_struct remove
;
192 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
193 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
195 /* Stuffs for move charges at task migration. */
197 * Types of charges to be moved.
199 #define MOVE_ANON 0x1U
200 #define MOVE_FILE 0x2U
201 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
203 /* "mc" and its members are protected by cgroup_mutex */
204 static struct move_charge_struct
{
205 spinlock_t lock
; /* for from, to */
206 struct mm_struct
*mm
;
207 struct mem_cgroup
*from
;
208 struct mem_cgroup
*to
;
210 unsigned long precharge
;
211 unsigned long moved_charge
;
212 unsigned long moved_swap
;
213 struct task_struct
*moving_task
; /* a task moving charges */
214 wait_queue_head_t waitq
; /* a waitq for other context */
216 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
217 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
221 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
222 * limit reclaim to prevent infinite loops, if they ever occur.
224 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
225 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
228 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
229 MEM_CGROUP_CHARGE_TYPE_ANON
,
230 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
231 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
235 /* for encoding cft->private value on file */
244 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
245 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
246 #define MEMFILE_ATTR(val) ((val) & 0xffff)
247 /* Used for OOM nofiier */
248 #define OOM_CONTROL (0)
250 /* Some nice accessors for the vmpressure. */
251 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
254 memcg
= root_mem_cgroup
;
255 return &memcg
->vmpressure
;
258 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
260 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
263 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
265 return (memcg
== root_mem_cgroup
);
270 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
271 * The main reason for not using cgroup id for this:
272 * this works better in sparse environments, where we have a lot of memcgs,
273 * but only a few kmem-limited. Or also, if we have, for instance, 200
274 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
275 * 200 entry array for that.
277 * The current size of the caches array is stored in memcg_nr_cache_ids. It
278 * will double each time we have to increase it.
280 static DEFINE_IDA(memcg_cache_ida
);
281 int memcg_nr_cache_ids
;
283 /* Protects memcg_nr_cache_ids */
284 static DECLARE_RWSEM(memcg_cache_ids_sem
);
286 void memcg_get_cache_ids(void)
288 down_read(&memcg_cache_ids_sem
);
291 void memcg_put_cache_ids(void)
293 up_read(&memcg_cache_ids_sem
);
297 * MIN_SIZE is different than 1, because we would like to avoid going through
298 * the alloc/free process all the time. In a small machine, 4 kmem-limited
299 * cgroups is a reasonable guess. In the future, it could be a parameter or
300 * tunable, but that is strictly not necessary.
302 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
303 * this constant directly from cgroup, but it is understandable that this is
304 * better kept as an internal representation in cgroup.c. In any case, the
305 * cgrp_id space is not getting any smaller, and we don't have to necessarily
306 * increase ours as well if it increases.
308 #define MEMCG_CACHES_MIN_SIZE 4
309 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
312 * A lot of the calls to the cache allocation functions are expected to be
313 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
314 * conditional to this static branch, we'll have to allow modules that does
315 * kmem_cache_alloc and the such to see this symbol as well
317 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
318 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
320 #endif /* !CONFIG_SLOB */
323 * mem_cgroup_css_from_page - css of the memcg associated with a page
324 * @page: page of interest
326 * If memcg is bound to the default hierarchy, css of the memcg associated
327 * with @page is returned. The returned css remains associated with @page
328 * until it is released.
330 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
333 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
335 struct mem_cgroup
*memcg
;
337 memcg
= page
->mem_cgroup
;
339 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
340 memcg
= root_mem_cgroup
;
346 * page_cgroup_ino - return inode number of the memcg a page is charged to
349 * Look up the closest online ancestor of the memory cgroup @page is charged to
350 * and return its inode number or 0 if @page is not charged to any cgroup. It
351 * is safe to call this function without holding a reference to @page.
353 * Note, this function is inherently racy, because there is nothing to prevent
354 * the cgroup inode from getting torn down and potentially reallocated a moment
355 * after page_cgroup_ino() returns, so it only should be used by callers that
356 * do not care (such as procfs interfaces).
358 ino_t
page_cgroup_ino(struct page
*page
)
360 struct mem_cgroup
*memcg
;
361 unsigned long ino
= 0;
364 memcg
= READ_ONCE(page
->mem_cgroup
);
365 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
366 memcg
= parent_mem_cgroup(memcg
);
368 ino
= cgroup_ino(memcg
->css
.cgroup
);
373 static struct mem_cgroup_per_node
*
374 mem_cgroup_page_nodeinfo(struct mem_cgroup
*memcg
, struct page
*page
)
376 int nid
= page_to_nid(page
);
378 return memcg
->nodeinfo
[nid
];
381 static struct mem_cgroup_tree_per_node
*
382 soft_limit_tree_node(int nid
)
384 return soft_limit_tree
.rb_tree_per_node
[nid
];
387 static struct mem_cgroup_tree_per_node
*
388 soft_limit_tree_from_page(struct page
*page
)
390 int nid
= page_to_nid(page
);
392 return soft_limit_tree
.rb_tree_per_node
[nid
];
395 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node
*mz
,
396 struct mem_cgroup_tree_per_node
*mctz
,
397 unsigned long new_usage_in_excess
)
399 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
400 struct rb_node
*parent
= NULL
;
401 struct mem_cgroup_per_node
*mz_node
;
406 mz
->usage_in_excess
= new_usage_in_excess
;
407 if (!mz
->usage_in_excess
)
411 mz_node
= rb_entry(parent
, struct mem_cgroup_per_node
,
413 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
416 * We can't avoid mem cgroups that are over their soft
417 * limit by the same amount
419 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
422 rb_link_node(&mz
->tree_node
, parent
, p
);
423 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
427 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
428 struct mem_cgroup_tree_per_node
*mctz
)
432 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
436 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
437 struct mem_cgroup_tree_per_node
*mctz
)
441 spin_lock_irqsave(&mctz
->lock
, flags
);
442 __mem_cgroup_remove_exceeded(mz
, mctz
);
443 spin_unlock_irqrestore(&mctz
->lock
, flags
);
446 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
448 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
449 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
450 unsigned long excess
= 0;
452 if (nr_pages
> soft_limit
)
453 excess
= nr_pages
- soft_limit
;
458 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
460 unsigned long excess
;
461 struct mem_cgroup_per_node
*mz
;
462 struct mem_cgroup_tree_per_node
*mctz
;
464 mctz
= soft_limit_tree_from_page(page
);
466 * Necessary to update all ancestors when hierarchy is used.
467 * because their event counter is not touched.
469 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
470 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
471 excess
= soft_limit_excess(memcg
);
473 * We have to update the tree if mz is on RB-tree or
474 * mem is over its softlimit.
476 if (excess
|| mz
->on_tree
) {
479 spin_lock_irqsave(&mctz
->lock
, flags
);
480 /* if on-tree, remove it */
482 __mem_cgroup_remove_exceeded(mz
, mctz
);
484 * Insert again. mz->usage_in_excess will be updated.
485 * If excess is 0, no tree ops.
487 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
488 spin_unlock_irqrestore(&mctz
->lock
, flags
);
493 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
495 struct mem_cgroup_tree_per_node
*mctz
;
496 struct mem_cgroup_per_node
*mz
;
500 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
501 mctz
= soft_limit_tree_node(nid
);
502 mem_cgroup_remove_exceeded(mz
, mctz
);
506 static struct mem_cgroup_per_node
*
507 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
509 struct rb_node
*rightmost
= NULL
;
510 struct mem_cgroup_per_node
*mz
;
514 rightmost
= rb_last(&mctz
->rb_root
);
516 goto done
; /* Nothing to reclaim from */
518 mz
= rb_entry(rightmost
, struct mem_cgroup_per_node
, tree_node
);
520 * Remove the node now but someone else can add it back,
521 * we will to add it back at the end of reclaim to its correct
522 * position in the tree.
524 __mem_cgroup_remove_exceeded(mz
, mctz
);
525 if (!soft_limit_excess(mz
->memcg
) ||
526 !css_tryget_online(&mz
->memcg
->css
))
532 static struct mem_cgroup_per_node
*
533 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
535 struct mem_cgroup_per_node
*mz
;
537 spin_lock_irq(&mctz
->lock
);
538 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
539 spin_unlock_irq(&mctz
->lock
);
544 * Return page count for single (non recursive) @memcg.
546 * Implementation Note: reading percpu statistics for memcg.
548 * Both of vmstat[] and percpu_counter has threshold and do periodic
549 * synchronization to implement "quick" read. There are trade-off between
550 * reading cost and precision of value. Then, we may have a chance to implement
551 * a periodic synchronization of counter in memcg's counter.
553 * But this _read() function is used for user interface now. The user accounts
554 * memory usage by memory cgroup and he _always_ requires exact value because
555 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
556 * have to visit all online cpus and make sum. So, for now, unnecessary
557 * synchronization is not implemented. (just implemented for cpu hotplug)
559 * If there are kernel internal actions which can make use of some not-exact
560 * value, and reading all cpu value can be performance bottleneck in some
561 * common workload, threshold and synchronization as vmstat[] should be
565 mem_cgroup_read_stat(struct mem_cgroup
*memcg
, enum mem_cgroup_stat_index idx
)
570 /* Per-cpu values can be negative, use a signed accumulator */
571 for_each_possible_cpu(cpu
)
572 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
574 * Summing races with updates, so val may be negative. Avoid exposing
575 * transient negative values.
582 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
583 enum mem_cgroup_events_index idx
)
585 unsigned long val
= 0;
588 for_each_possible_cpu(cpu
)
589 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
593 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
595 bool compound
, int nr_pages
)
598 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
599 * counted as CACHE even if it's on ANON LRU.
602 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
605 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
609 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
610 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
614 /* pagein of a big page is an event. So, ignore page size */
616 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
618 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
619 nr_pages
= -nr_pages
; /* for event */
622 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
625 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
626 int nid
, unsigned int lru_mask
)
628 unsigned long nr
= 0;
629 struct mem_cgroup_per_node
*mz
;
632 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
635 if (!(BIT(lru
) & lru_mask
))
637 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
638 nr
+= mz
->lru_size
[lru
];
643 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
644 unsigned int lru_mask
)
646 unsigned long nr
= 0;
649 for_each_node_state(nid
, N_MEMORY
)
650 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
654 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
655 enum mem_cgroup_events_target target
)
657 unsigned long val
, next
;
659 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
660 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
661 /* from time_after() in jiffies.h */
662 if ((long)next
- (long)val
< 0) {
664 case MEM_CGROUP_TARGET_THRESH
:
665 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
667 case MEM_CGROUP_TARGET_SOFTLIMIT
:
668 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
670 case MEM_CGROUP_TARGET_NUMAINFO
:
671 next
= val
+ NUMAINFO_EVENTS_TARGET
;
676 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
683 * Check events in order.
686 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
688 /* threshold event is triggered in finer grain than soft limit */
689 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
690 MEM_CGROUP_TARGET_THRESH
))) {
692 bool do_numainfo __maybe_unused
;
694 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
695 MEM_CGROUP_TARGET_SOFTLIMIT
);
697 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
698 MEM_CGROUP_TARGET_NUMAINFO
);
700 mem_cgroup_threshold(memcg
);
701 if (unlikely(do_softlimit
))
702 mem_cgroup_update_tree(memcg
, page
);
704 if (unlikely(do_numainfo
))
705 atomic_inc(&memcg
->numainfo_events
);
710 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
713 * mm_update_next_owner() may clear mm->owner to NULL
714 * if it races with swapoff, page migration, etc.
715 * So this can be called with p == NULL.
720 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
722 EXPORT_SYMBOL(mem_cgroup_from_task
);
724 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
726 struct mem_cgroup
*memcg
= NULL
;
731 * Page cache insertions can happen withou an
732 * actual mm context, e.g. during disk probing
733 * on boot, loopback IO, acct() writes etc.
736 memcg
= root_mem_cgroup
;
738 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
739 if (unlikely(!memcg
))
740 memcg
= root_mem_cgroup
;
742 } while (!css_tryget_online(&memcg
->css
));
748 * mem_cgroup_iter - iterate over memory cgroup hierarchy
749 * @root: hierarchy root
750 * @prev: previously returned memcg, NULL on first invocation
751 * @reclaim: cookie for shared reclaim walks, NULL for full walks
753 * Returns references to children of the hierarchy below @root, or
754 * @root itself, or %NULL after a full round-trip.
756 * Caller must pass the return value in @prev on subsequent
757 * invocations for reference counting, or use mem_cgroup_iter_break()
758 * to cancel a hierarchy walk before the round-trip is complete.
760 * Reclaimers can specify a zone and a priority level in @reclaim to
761 * divide up the memcgs in the hierarchy among all concurrent
762 * reclaimers operating on the same zone and priority.
764 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
765 struct mem_cgroup
*prev
,
766 struct mem_cgroup_reclaim_cookie
*reclaim
)
768 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
769 struct cgroup_subsys_state
*css
= NULL
;
770 struct mem_cgroup
*memcg
= NULL
;
771 struct mem_cgroup
*pos
= NULL
;
773 if (mem_cgroup_disabled())
777 root
= root_mem_cgroup
;
779 if (prev
&& !reclaim
)
782 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
791 struct mem_cgroup_per_node
*mz
;
793 mz
= mem_cgroup_nodeinfo(root
, reclaim
->pgdat
->node_id
);
794 iter
= &mz
->iter
[reclaim
->priority
];
796 if (prev
&& reclaim
->generation
!= iter
->generation
)
800 pos
= READ_ONCE(iter
->position
);
801 if (!pos
|| css_tryget(&pos
->css
))
804 * css reference reached zero, so iter->position will
805 * be cleared by ->css_released. However, we should not
806 * rely on this happening soon, because ->css_released
807 * is called from a work queue, and by busy-waiting we
808 * might block it. So we clear iter->position right
811 (void)cmpxchg(&iter
->position
, pos
, NULL
);
819 css
= css_next_descendant_pre(css
, &root
->css
);
822 * Reclaimers share the hierarchy walk, and a
823 * new one might jump in right at the end of
824 * the hierarchy - make sure they see at least
825 * one group and restart from the beginning.
833 * Verify the css and acquire a reference. The root
834 * is provided by the caller, so we know it's alive
835 * and kicking, and don't take an extra reference.
837 memcg
= mem_cgroup_from_css(css
);
839 if (css
== &root
->css
)
850 * The position could have already been updated by a competing
851 * thread, so check that the value hasn't changed since we read
852 * it to avoid reclaiming from the same cgroup twice.
854 (void)cmpxchg(&iter
->position
, pos
, memcg
);
862 reclaim
->generation
= iter
->generation
;
868 if (prev
&& prev
!= root
)
875 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
876 * @root: hierarchy root
877 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
879 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
880 struct mem_cgroup
*prev
)
883 root
= root_mem_cgroup
;
884 if (prev
&& prev
!= root
)
888 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
890 struct mem_cgroup
*memcg
= dead_memcg
;
891 struct mem_cgroup_reclaim_iter
*iter
;
892 struct mem_cgroup_per_node
*mz
;
896 while ((memcg
= parent_mem_cgroup(memcg
))) {
898 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
899 for (i
= 0; i
<= DEF_PRIORITY
; i
++) {
901 cmpxchg(&iter
->position
,
909 * Iteration constructs for visiting all cgroups (under a tree). If
910 * loops are exited prematurely (break), mem_cgroup_iter_break() must
911 * be used for reference counting.
913 #define for_each_mem_cgroup_tree(iter, root) \
914 for (iter = mem_cgroup_iter(root, NULL, NULL); \
916 iter = mem_cgroup_iter(root, iter, NULL))
918 #define for_each_mem_cgroup(iter) \
919 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
921 iter = mem_cgroup_iter(NULL, iter, NULL))
924 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
925 * @memcg: hierarchy root
926 * @fn: function to call for each task
927 * @arg: argument passed to @fn
929 * This function iterates over tasks attached to @memcg or to any of its
930 * descendants and calls @fn for each task. If @fn returns a non-zero
931 * value, the function breaks the iteration loop and returns the value.
932 * Otherwise, it will iterate over all tasks and return 0.
934 * This function must not be called for the root memory cgroup.
936 int mem_cgroup_scan_tasks(struct mem_cgroup
*memcg
,
937 int (*fn
)(struct task_struct
*, void *), void *arg
)
939 struct mem_cgroup
*iter
;
942 BUG_ON(memcg
== root_mem_cgroup
);
944 for_each_mem_cgroup_tree(iter
, memcg
) {
945 struct css_task_iter it
;
946 struct task_struct
*task
;
948 css_task_iter_start(&iter
->css
, &it
);
949 while (!ret
&& (task
= css_task_iter_next(&it
)))
951 css_task_iter_end(&it
);
953 mem_cgroup_iter_break(memcg
, iter
);
961 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
963 * @zone: zone of the page
965 * This function is only safe when following the LRU page isolation
966 * and putback protocol: the LRU lock must be held, and the page must
967 * either be PageLRU() or the caller must have isolated/allocated it.
969 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct pglist_data
*pgdat
)
971 struct mem_cgroup_per_node
*mz
;
972 struct mem_cgroup
*memcg
;
973 struct lruvec
*lruvec
;
975 if (mem_cgroup_disabled()) {
976 lruvec
= &pgdat
->lruvec
;
980 memcg
= page
->mem_cgroup
;
982 * Swapcache readahead pages are added to the LRU - and
983 * possibly migrated - before they are charged.
986 memcg
= root_mem_cgroup
;
988 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
989 lruvec
= &mz
->lruvec
;
992 * Since a node can be onlined after the mem_cgroup was created,
993 * we have to be prepared to initialize lruvec->zone here;
994 * and if offlined then reonlined, we need to reinitialize it.
996 if (unlikely(lruvec
->pgdat
!= pgdat
))
997 lruvec
->pgdat
= pgdat
;
1002 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1003 * @lruvec: mem_cgroup per zone lru vector
1004 * @lru: index of lru list the page is sitting on
1005 * @nr_pages: positive when adding or negative when removing
1007 * This function must be called under lru_lock, just before a page is added
1008 * to or just after a page is removed from an lru list (that ordering being
1009 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1011 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1014 struct mem_cgroup_per_node
*mz
;
1015 unsigned long *lru_size
;
1019 if (mem_cgroup_disabled())
1022 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
1023 lru_size
= mz
->lru_size
+ lru
;
1024 empty
= list_empty(lruvec
->lists
+ lru
);
1027 *lru_size
+= nr_pages
;
1030 if (WARN_ONCE(size
< 0 || empty
!= !size
,
1031 "%s(%p, %d, %d): lru_size %ld but %sempty\n",
1032 __func__
, lruvec
, lru
, nr_pages
, size
, empty
? "" : "not ")) {
1038 *lru_size
+= nr_pages
;
1041 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1043 struct mem_cgroup
*task_memcg
;
1044 struct task_struct
*p
;
1047 p
= find_lock_task_mm(task
);
1049 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1053 * All threads may have already detached their mm's, but the oom
1054 * killer still needs to detect if they have already been oom
1055 * killed to prevent needlessly killing additional tasks.
1058 task_memcg
= mem_cgroup_from_task(task
);
1059 css_get(&task_memcg
->css
);
1062 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1063 css_put(&task_memcg
->css
);
1068 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1069 * @memcg: the memory cgroup
1071 * Returns the maximum amount of memory @mem can be charged with, in
1074 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1076 unsigned long margin
= 0;
1077 unsigned long count
;
1078 unsigned long limit
;
1080 count
= page_counter_read(&memcg
->memory
);
1081 limit
= READ_ONCE(memcg
->memory
.limit
);
1083 margin
= limit
- count
;
1085 if (do_memsw_account()) {
1086 count
= page_counter_read(&memcg
->memsw
);
1087 limit
= READ_ONCE(memcg
->memsw
.limit
);
1089 margin
= min(margin
, limit
- count
);
1098 * A routine for checking "mem" is under move_account() or not.
1100 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1101 * moving cgroups. This is for waiting at high-memory pressure
1104 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1106 struct mem_cgroup
*from
;
1107 struct mem_cgroup
*to
;
1110 * Unlike task_move routines, we access mc.to, mc.from not under
1111 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1113 spin_lock(&mc
.lock
);
1119 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1120 mem_cgroup_is_descendant(to
, memcg
);
1122 spin_unlock(&mc
.lock
);
1126 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1128 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1129 if (mem_cgroup_under_move(memcg
)) {
1131 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1132 /* moving charge context might have finished. */
1135 finish_wait(&mc
.waitq
, &wait
);
1142 #define K(x) ((x) << (PAGE_SHIFT-10))
1144 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1145 * @memcg: The memory cgroup that went over limit
1146 * @p: Task that is going to be killed
1148 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1151 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1153 struct mem_cgroup
*iter
;
1159 pr_info("Task in ");
1160 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1161 pr_cont(" killed as a result of limit of ");
1163 pr_info("Memory limit reached of cgroup ");
1166 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1171 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1172 K((u64
)page_counter_read(&memcg
->memory
)),
1173 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1174 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1175 K((u64
)page_counter_read(&memcg
->memsw
)),
1176 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1177 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1178 K((u64
)page_counter_read(&memcg
->kmem
)),
1179 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1181 for_each_mem_cgroup_tree(iter
, memcg
) {
1182 pr_info("Memory cgroup stats for ");
1183 pr_cont_cgroup_path(iter
->css
.cgroup
);
1186 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1187 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1189 pr_cont(" %s:%luKB", mem_cgroup_stat_names
[i
],
1190 K(mem_cgroup_read_stat(iter
, i
)));
1193 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1194 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1195 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1202 * This function returns the number of memcg under hierarchy tree. Returns
1203 * 1(self count) if no children.
1205 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1208 struct mem_cgroup
*iter
;
1210 for_each_mem_cgroup_tree(iter
, memcg
)
1216 * Return the memory (and swap, if configured) limit for a memcg.
1218 unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1220 unsigned long limit
;
1222 limit
= memcg
->memory
.limit
;
1223 if (mem_cgroup_swappiness(memcg
)) {
1224 unsigned long memsw_limit
;
1225 unsigned long swap_limit
;
1227 memsw_limit
= memcg
->memsw
.limit
;
1228 swap_limit
= memcg
->swap
.limit
;
1229 swap_limit
= min(swap_limit
, (unsigned long)total_swap_pages
);
1230 limit
= min(limit
+ swap_limit
, memsw_limit
);
1235 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1238 struct oom_control oc
= {
1242 .gfp_mask
= gfp_mask
,
1247 mutex_lock(&oom_lock
);
1248 ret
= out_of_memory(&oc
);
1249 mutex_unlock(&oom_lock
);
1253 #if MAX_NUMNODES > 1
1256 * test_mem_cgroup_node_reclaimable
1257 * @memcg: the target memcg
1258 * @nid: the node ID to be checked.
1259 * @noswap : specify true here if the user wants flle only information.
1261 * This function returns whether the specified memcg contains any
1262 * reclaimable pages on a node. Returns true if there are any reclaimable
1263 * pages in the node.
1265 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1266 int nid
, bool noswap
)
1268 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1270 if (noswap
|| !total_swap_pages
)
1272 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1279 * Always updating the nodemask is not very good - even if we have an empty
1280 * list or the wrong list here, we can start from some node and traverse all
1281 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1284 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1288 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1289 * pagein/pageout changes since the last update.
1291 if (!atomic_read(&memcg
->numainfo_events
))
1293 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1296 /* make a nodemask where this memcg uses memory from */
1297 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1299 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1301 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1302 node_clear(nid
, memcg
->scan_nodes
);
1305 atomic_set(&memcg
->numainfo_events
, 0);
1306 atomic_set(&memcg
->numainfo_updating
, 0);
1310 * Selecting a node where we start reclaim from. Because what we need is just
1311 * reducing usage counter, start from anywhere is O,K. Considering
1312 * memory reclaim from current node, there are pros. and cons.
1314 * Freeing memory from current node means freeing memory from a node which
1315 * we'll use or we've used. So, it may make LRU bad. And if several threads
1316 * hit limits, it will see a contention on a node. But freeing from remote
1317 * node means more costs for memory reclaim because of memory latency.
1319 * Now, we use round-robin. Better algorithm is welcomed.
1321 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1325 mem_cgroup_may_update_nodemask(memcg
);
1326 node
= memcg
->last_scanned_node
;
1328 node
= next_node_in(node
, memcg
->scan_nodes
);
1330 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1331 * last time it really checked all the LRUs due to rate limiting.
1332 * Fallback to the current node in that case for simplicity.
1334 if (unlikely(node
== MAX_NUMNODES
))
1335 node
= numa_node_id();
1337 memcg
->last_scanned_node
= node
;
1341 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1347 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1350 unsigned long *total_scanned
)
1352 struct mem_cgroup
*victim
= NULL
;
1355 unsigned long excess
;
1356 unsigned long nr_scanned
;
1357 struct mem_cgroup_reclaim_cookie reclaim
= {
1362 excess
= soft_limit_excess(root_memcg
);
1365 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1370 * If we have not been able to reclaim
1371 * anything, it might because there are
1372 * no reclaimable pages under this hierarchy
1377 * We want to do more targeted reclaim.
1378 * excess >> 2 is not to excessive so as to
1379 * reclaim too much, nor too less that we keep
1380 * coming back to reclaim from this cgroup
1382 if (total
>= (excess
>> 2) ||
1383 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1388 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1389 pgdat
, &nr_scanned
);
1390 *total_scanned
+= nr_scanned
;
1391 if (!soft_limit_excess(root_memcg
))
1394 mem_cgroup_iter_break(root_memcg
, victim
);
1398 #ifdef CONFIG_LOCKDEP
1399 static struct lockdep_map memcg_oom_lock_dep_map
= {
1400 .name
= "memcg_oom_lock",
1404 static DEFINE_SPINLOCK(memcg_oom_lock
);
1407 * Check OOM-Killer is already running under our hierarchy.
1408 * If someone is running, return false.
1410 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1412 struct mem_cgroup
*iter
, *failed
= NULL
;
1414 spin_lock(&memcg_oom_lock
);
1416 for_each_mem_cgroup_tree(iter
, memcg
) {
1417 if (iter
->oom_lock
) {
1419 * this subtree of our hierarchy is already locked
1420 * so we cannot give a lock.
1423 mem_cgroup_iter_break(memcg
, iter
);
1426 iter
->oom_lock
= true;
1431 * OK, we failed to lock the whole subtree so we have
1432 * to clean up what we set up to the failing subtree
1434 for_each_mem_cgroup_tree(iter
, memcg
) {
1435 if (iter
== failed
) {
1436 mem_cgroup_iter_break(memcg
, iter
);
1439 iter
->oom_lock
= false;
1442 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1444 spin_unlock(&memcg_oom_lock
);
1449 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1451 struct mem_cgroup
*iter
;
1453 spin_lock(&memcg_oom_lock
);
1454 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1455 for_each_mem_cgroup_tree(iter
, memcg
)
1456 iter
->oom_lock
= false;
1457 spin_unlock(&memcg_oom_lock
);
1460 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1462 struct mem_cgroup
*iter
;
1464 spin_lock(&memcg_oom_lock
);
1465 for_each_mem_cgroup_tree(iter
, memcg
)
1467 spin_unlock(&memcg_oom_lock
);
1470 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1472 struct mem_cgroup
*iter
;
1475 * When a new child is created while the hierarchy is under oom,
1476 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1478 spin_lock(&memcg_oom_lock
);
1479 for_each_mem_cgroup_tree(iter
, memcg
)
1480 if (iter
->under_oom
> 0)
1482 spin_unlock(&memcg_oom_lock
);
1485 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1487 struct oom_wait_info
{
1488 struct mem_cgroup
*memcg
;
1492 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1493 unsigned mode
, int sync
, void *arg
)
1495 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1496 struct mem_cgroup
*oom_wait_memcg
;
1497 struct oom_wait_info
*oom_wait_info
;
1499 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1500 oom_wait_memcg
= oom_wait_info
->memcg
;
1502 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1503 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1505 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1508 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1511 * For the following lockless ->under_oom test, the only required
1512 * guarantee is that it must see the state asserted by an OOM when
1513 * this function is called as a result of userland actions
1514 * triggered by the notification of the OOM. This is trivially
1515 * achieved by invoking mem_cgroup_mark_under_oom() before
1516 * triggering notification.
1518 if (memcg
&& memcg
->under_oom
)
1519 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1522 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1524 if (!current
->memcg_may_oom
)
1527 * We are in the middle of the charge context here, so we
1528 * don't want to block when potentially sitting on a callstack
1529 * that holds all kinds of filesystem and mm locks.
1531 * Also, the caller may handle a failed allocation gracefully
1532 * (like optional page cache readahead) and so an OOM killer
1533 * invocation might not even be necessary.
1535 * That's why we don't do anything here except remember the
1536 * OOM context and then deal with it at the end of the page
1537 * fault when the stack is unwound, the locks are released,
1538 * and when we know whether the fault was overall successful.
1540 css_get(&memcg
->css
);
1541 current
->memcg_in_oom
= memcg
;
1542 current
->memcg_oom_gfp_mask
= mask
;
1543 current
->memcg_oom_order
= order
;
1547 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1548 * @handle: actually kill/wait or just clean up the OOM state
1550 * This has to be called at the end of a page fault if the memcg OOM
1551 * handler was enabled.
1553 * Memcg supports userspace OOM handling where failed allocations must
1554 * sleep on a waitqueue until the userspace task resolves the
1555 * situation. Sleeping directly in the charge context with all kinds
1556 * of locks held is not a good idea, instead we remember an OOM state
1557 * in the task and mem_cgroup_oom_synchronize() has to be called at
1558 * the end of the page fault to complete the OOM handling.
1560 * Returns %true if an ongoing memcg OOM situation was detected and
1561 * completed, %false otherwise.
1563 bool mem_cgroup_oom_synchronize(bool handle
)
1565 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1566 struct oom_wait_info owait
;
1569 /* OOM is global, do not handle */
1576 owait
.memcg
= memcg
;
1577 owait
.wait
.flags
= 0;
1578 owait
.wait
.func
= memcg_oom_wake_function
;
1579 owait
.wait
.private = current
;
1580 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1582 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1583 mem_cgroup_mark_under_oom(memcg
);
1585 locked
= mem_cgroup_oom_trylock(memcg
);
1588 mem_cgroup_oom_notify(memcg
);
1590 if (locked
&& !memcg
->oom_kill_disable
) {
1591 mem_cgroup_unmark_under_oom(memcg
);
1592 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1593 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1594 current
->memcg_oom_order
);
1597 mem_cgroup_unmark_under_oom(memcg
);
1598 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1602 mem_cgroup_oom_unlock(memcg
);
1604 * There is no guarantee that an OOM-lock contender
1605 * sees the wakeups triggered by the OOM kill
1606 * uncharges. Wake any sleepers explicitely.
1608 memcg_oom_recover(memcg
);
1611 current
->memcg_in_oom
= NULL
;
1612 css_put(&memcg
->css
);
1617 * lock_page_memcg - lock a page->mem_cgroup binding
1620 * This function protects unlocked LRU pages from being moved to
1621 * another cgroup and stabilizes their page->mem_cgroup binding.
1623 void lock_page_memcg(struct page
*page
)
1625 struct mem_cgroup
*memcg
;
1626 unsigned long flags
;
1629 * The RCU lock is held throughout the transaction. The fast
1630 * path can get away without acquiring the memcg->move_lock
1631 * because page moving starts with an RCU grace period.
1635 if (mem_cgroup_disabled())
1638 memcg
= page
->mem_cgroup
;
1639 if (unlikely(!memcg
))
1642 if (atomic_read(&memcg
->moving_account
) <= 0)
1645 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1646 if (memcg
!= page
->mem_cgroup
) {
1647 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1652 * When charge migration first begins, we can have locked and
1653 * unlocked page stat updates happening concurrently. Track
1654 * the task who has the lock for unlock_page_memcg().
1656 memcg
->move_lock_task
= current
;
1657 memcg
->move_lock_flags
= flags
;
1661 EXPORT_SYMBOL(lock_page_memcg
);
1664 * unlock_page_memcg - unlock a page->mem_cgroup binding
1667 void unlock_page_memcg(struct page
*page
)
1669 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
1671 if (memcg
&& memcg
->move_lock_task
== current
) {
1672 unsigned long flags
= memcg
->move_lock_flags
;
1674 memcg
->move_lock_task
= NULL
;
1675 memcg
->move_lock_flags
= 0;
1677 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1682 EXPORT_SYMBOL(unlock_page_memcg
);
1685 * size of first charge trial. "32" comes from vmscan.c's magic value.
1686 * TODO: maybe necessary to use big numbers in big irons.
1688 #define CHARGE_BATCH 32U
1689 struct memcg_stock_pcp
{
1690 struct mem_cgroup
*cached
; /* this never be root cgroup */
1691 unsigned int nr_pages
;
1692 struct work_struct work
;
1693 unsigned long flags
;
1694 #define FLUSHING_CACHED_CHARGE 0
1696 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1697 static DEFINE_MUTEX(percpu_charge_mutex
);
1700 * consume_stock: Try to consume stocked charge on this cpu.
1701 * @memcg: memcg to consume from.
1702 * @nr_pages: how many pages to charge.
1704 * The charges will only happen if @memcg matches the current cpu's memcg
1705 * stock, and at least @nr_pages are available in that stock. Failure to
1706 * service an allocation will refill the stock.
1708 * returns true if successful, false otherwise.
1710 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1712 struct memcg_stock_pcp
*stock
;
1713 unsigned long flags
;
1716 if (nr_pages
> CHARGE_BATCH
)
1719 local_irq_save(flags
);
1721 stock
= this_cpu_ptr(&memcg_stock
);
1722 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
1723 stock
->nr_pages
-= nr_pages
;
1727 local_irq_restore(flags
);
1733 * Returns stocks cached in percpu and reset cached information.
1735 static void drain_stock(struct memcg_stock_pcp
*stock
)
1737 struct mem_cgroup
*old
= stock
->cached
;
1739 if (stock
->nr_pages
) {
1740 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
1741 if (do_memsw_account())
1742 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
1743 css_put_many(&old
->css
, stock
->nr_pages
);
1744 stock
->nr_pages
= 0;
1746 stock
->cached
= NULL
;
1749 static void drain_local_stock(struct work_struct
*dummy
)
1751 struct memcg_stock_pcp
*stock
;
1752 unsigned long flags
;
1754 local_irq_save(flags
);
1756 stock
= this_cpu_ptr(&memcg_stock
);
1758 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
1760 local_irq_restore(flags
);
1764 * Cache charges(val) to local per_cpu area.
1765 * This will be consumed by consume_stock() function, later.
1767 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1769 struct memcg_stock_pcp
*stock
;
1770 unsigned long flags
;
1772 local_irq_save(flags
);
1774 stock
= this_cpu_ptr(&memcg_stock
);
1775 if (stock
->cached
!= memcg
) { /* reset if necessary */
1777 stock
->cached
= memcg
;
1779 stock
->nr_pages
+= nr_pages
;
1781 local_irq_restore(flags
);
1785 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1786 * of the hierarchy under it.
1788 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
1792 /* If someone's already draining, avoid adding running more workers. */
1793 if (!mutex_trylock(&percpu_charge_mutex
))
1795 /* Notify other cpus that system-wide "drain" is running */
1798 for_each_online_cpu(cpu
) {
1799 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1800 struct mem_cgroup
*memcg
;
1802 memcg
= stock
->cached
;
1803 if (!memcg
|| !stock
->nr_pages
)
1805 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
1807 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
1809 drain_local_stock(&stock
->work
);
1811 schedule_work_on(cpu
, &stock
->work
);
1816 mutex_unlock(&percpu_charge_mutex
);
1819 static int memcg_hotplug_cpu_dead(unsigned int cpu
)
1821 struct memcg_stock_pcp
*stock
;
1823 stock
= &per_cpu(memcg_stock
, cpu
);
1828 static void reclaim_high(struct mem_cgroup
*memcg
,
1829 unsigned int nr_pages
,
1833 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
1835 mem_cgroup_events(memcg
, MEMCG_HIGH
, 1);
1836 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
1837 } while ((memcg
= parent_mem_cgroup(memcg
)));
1840 static void high_work_func(struct work_struct
*work
)
1842 struct mem_cgroup
*memcg
;
1844 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
1845 reclaim_high(memcg
, CHARGE_BATCH
, GFP_KERNEL
);
1849 * Scheduled by try_charge() to be executed from the userland return path
1850 * and reclaims memory over the high limit.
1852 void mem_cgroup_handle_over_high(void)
1854 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
1855 struct mem_cgroup
*memcg
;
1857 if (likely(!nr_pages
))
1860 memcg
= get_mem_cgroup_from_mm(current
->mm
);
1861 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
1862 css_put(&memcg
->css
);
1863 current
->memcg_nr_pages_over_high
= 0;
1866 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1867 unsigned int nr_pages
)
1869 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
1870 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1871 struct mem_cgroup
*mem_over_limit
;
1872 struct page_counter
*counter
;
1873 unsigned long nr_reclaimed
;
1874 bool may_swap
= true;
1875 bool drained
= false;
1877 if (mem_cgroup_is_root(memcg
))
1880 if (consume_stock(memcg
, nr_pages
))
1883 if (!do_memsw_account() ||
1884 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
1885 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
1887 if (do_memsw_account())
1888 page_counter_uncharge(&memcg
->memsw
, batch
);
1889 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
1891 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
1895 if (batch
> nr_pages
) {
1901 * Unlike in global OOM situations, memcg is not in a physical
1902 * memory shortage. Allow dying and OOM-killed tasks to
1903 * bypass the last charges so that they can exit quickly and
1904 * free their memory.
1906 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
1907 fatal_signal_pending(current
) ||
1908 current
->flags
& PF_EXITING
))
1912 * Prevent unbounded recursion when reclaim operations need to
1913 * allocate memory. This might exceed the limits temporarily,
1914 * but we prefer facilitating memory reclaim and getting back
1915 * under the limit over triggering OOM kills in these cases.
1917 if (unlikely(current
->flags
& PF_MEMALLOC
))
1920 if (unlikely(task_in_memcg_oom(current
)))
1923 if (!gfpflags_allow_blocking(gfp_mask
))
1926 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
1928 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
1929 gfp_mask
, may_swap
);
1931 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
1935 drain_all_stock(mem_over_limit
);
1940 if (gfp_mask
& __GFP_NORETRY
)
1943 * Even though the limit is exceeded at this point, reclaim
1944 * may have been able to free some pages. Retry the charge
1945 * before killing the task.
1947 * Only for regular pages, though: huge pages are rather
1948 * unlikely to succeed so close to the limit, and we fall back
1949 * to regular pages anyway in case of failure.
1951 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
1954 * At task move, charge accounts can be doubly counted. So, it's
1955 * better to wait until the end of task_move if something is going on.
1957 if (mem_cgroup_wait_acct_move(mem_over_limit
))
1963 if (gfp_mask
& __GFP_NOFAIL
)
1966 if (fatal_signal_pending(current
))
1969 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
1971 mem_cgroup_oom(mem_over_limit
, gfp_mask
,
1972 get_order(nr_pages
* PAGE_SIZE
));
1974 if (!(gfp_mask
& __GFP_NOFAIL
))
1978 * The allocation either can't fail or will lead to more memory
1979 * being freed very soon. Allow memory usage go over the limit
1980 * temporarily by force charging it.
1982 page_counter_charge(&memcg
->memory
, nr_pages
);
1983 if (do_memsw_account())
1984 page_counter_charge(&memcg
->memsw
, nr_pages
);
1985 css_get_many(&memcg
->css
, nr_pages
);
1990 css_get_many(&memcg
->css
, batch
);
1991 if (batch
> nr_pages
)
1992 refill_stock(memcg
, batch
- nr_pages
);
1995 * If the hierarchy is above the normal consumption range, schedule
1996 * reclaim on returning to userland. We can perform reclaim here
1997 * if __GFP_RECLAIM but let's always punt for simplicity and so that
1998 * GFP_KERNEL can consistently be used during reclaim. @memcg is
1999 * not recorded as it most likely matches current's and won't
2000 * change in the meantime. As high limit is checked again before
2001 * reclaim, the cost of mismatch is negligible.
2004 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2005 /* Don't bother a random interrupted task */
2006 if (in_interrupt()) {
2007 schedule_work(&memcg
->high_work
);
2010 current
->memcg_nr_pages_over_high
+= batch
;
2011 set_notify_resume(current
);
2014 } while ((memcg
= parent_mem_cgroup(memcg
)));
2019 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2021 if (mem_cgroup_is_root(memcg
))
2024 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2025 if (do_memsw_account())
2026 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2028 css_put_many(&memcg
->css
, nr_pages
);
2031 static void lock_page_lru(struct page
*page
, int *isolated
)
2033 struct zone
*zone
= page_zone(page
);
2035 spin_lock_irq(zone_lru_lock(zone
));
2036 if (PageLRU(page
)) {
2037 struct lruvec
*lruvec
;
2039 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
2041 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2047 static void unlock_page_lru(struct page
*page
, int isolated
)
2049 struct zone
*zone
= page_zone(page
);
2052 struct lruvec
*lruvec
;
2054 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
2055 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2057 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2059 spin_unlock_irq(zone_lru_lock(zone
));
2062 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2067 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2070 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2071 * may already be on some other mem_cgroup's LRU. Take care of it.
2074 lock_page_lru(page
, &isolated
);
2077 * Nobody should be changing or seriously looking at
2078 * page->mem_cgroup at this point:
2080 * - the page is uncharged
2082 * - the page is off-LRU
2084 * - an anonymous fault has exclusive page access, except for
2085 * a locked page table
2087 * - a page cache insertion, a swapin fault, or a migration
2088 * have the page locked
2090 page
->mem_cgroup
= memcg
;
2093 unlock_page_lru(page
, isolated
);
2097 static int memcg_alloc_cache_id(void)
2102 id
= ida_simple_get(&memcg_cache_ida
,
2103 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2107 if (id
< memcg_nr_cache_ids
)
2111 * There's no space for the new id in memcg_caches arrays,
2112 * so we have to grow them.
2114 down_write(&memcg_cache_ids_sem
);
2116 size
= 2 * (id
+ 1);
2117 if (size
< MEMCG_CACHES_MIN_SIZE
)
2118 size
= MEMCG_CACHES_MIN_SIZE
;
2119 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2120 size
= MEMCG_CACHES_MAX_SIZE
;
2122 err
= memcg_update_all_caches(size
);
2124 err
= memcg_update_all_list_lrus(size
);
2126 memcg_nr_cache_ids
= size
;
2128 up_write(&memcg_cache_ids_sem
);
2131 ida_simple_remove(&memcg_cache_ida
, id
);
2137 static void memcg_free_cache_id(int id
)
2139 ida_simple_remove(&memcg_cache_ida
, id
);
2142 struct memcg_kmem_cache_create_work
{
2143 struct mem_cgroup
*memcg
;
2144 struct kmem_cache
*cachep
;
2145 struct work_struct work
;
2148 static struct workqueue_struct
*memcg_kmem_cache_create_wq
;
2150 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2152 struct memcg_kmem_cache_create_work
*cw
=
2153 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2154 struct mem_cgroup
*memcg
= cw
->memcg
;
2155 struct kmem_cache
*cachep
= cw
->cachep
;
2157 memcg_create_kmem_cache(memcg
, cachep
);
2159 css_put(&memcg
->css
);
2164 * Enqueue the creation of a per-memcg kmem_cache.
2166 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2167 struct kmem_cache
*cachep
)
2169 struct memcg_kmem_cache_create_work
*cw
;
2171 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2175 css_get(&memcg
->css
);
2178 cw
->cachep
= cachep
;
2179 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2181 queue_work(memcg_kmem_cache_create_wq
, &cw
->work
);
2184 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2185 struct kmem_cache
*cachep
)
2188 * We need to stop accounting when we kmalloc, because if the
2189 * corresponding kmalloc cache is not yet created, the first allocation
2190 * in __memcg_schedule_kmem_cache_create will recurse.
2192 * However, it is better to enclose the whole function. Depending on
2193 * the debugging options enabled, INIT_WORK(), for instance, can
2194 * trigger an allocation. This too, will make us recurse. Because at
2195 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2196 * the safest choice is to do it like this, wrapping the whole function.
2198 current
->memcg_kmem_skip_account
= 1;
2199 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2200 current
->memcg_kmem_skip_account
= 0;
2203 static inline bool memcg_kmem_bypass(void)
2205 if (in_interrupt() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
2211 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2212 * @cachep: the original global kmem cache
2214 * Return the kmem_cache we're supposed to use for a slab allocation.
2215 * We try to use the current memcg's version of the cache.
2217 * If the cache does not exist yet, if we are the first user of it, we
2218 * create it asynchronously in a workqueue and let the current allocation
2219 * go through with the original cache.
2221 * This function takes a reference to the cache it returns to assure it
2222 * won't get destroyed while we are working with it. Once the caller is
2223 * done with it, memcg_kmem_put_cache() must be called to release the
2226 struct kmem_cache
*memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2228 struct mem_cgroup
*memcg
;
2229 struct kmem_cache
*memcg_cachep
;
2232 VM_BUG_ON(!is_root_cache(cachep
));
2234 if (memcg_kmem_bypass())
2237 if (current
->memcg_kmem_skip_account
)
2240 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2241 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2245 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2246 if (likely(memcg_cachep
))
2247 return memcg_cachep
;
2250 * If we are in a safe context (can wait, and not in interrupt
2251 * context), we could be be predictable and return right away.
2252 * This would guarantee that the allocation being performed
2253 * already belongs in the new cache.
2255 * However, there are some clashes that can arrive from locking.
2256 * For instance, because we acquire the slab_mutex while doing
2257 * memcg_create_kmem_cache, this means no further allocation
2258 * could happen with the slab_mutex held. So it's better to
2261 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2263 css_put(&memcg
->css
);
2268 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2269 * @cachep: the cache returned by memcg_kmem_get_cache
2271 void memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2273 if (!is_root_cache(cachep
))
2274 css_put(&cachep
->memcg_params
.memcg
->css
);
2278 * memcg_kmem_charge: charge a kmem page
2279 * @page: page to charge
2280 * @gfp: reclaim mode
2281 * @order: allocation order
2282 * @memcg: memory cgroup to charge
2284 * Returns 0 on success, an error code on failure.
2286 int memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2287 struct mem_cgroup
*memcg
)
2289 unsigned int nr_pages
= 1 << order
;
2290 struct page_counter
*counter
;
2293 ret
= try_charge(memcg
, gfp
, nr_pages
);
2297 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2298 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2299 cancel_charge(memcg
, nr_pages
);
2303 page
->mem_cgroup
= memcg
;
2309 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2310 * @page: page to charge
2311 * @gfp: reclaim mode
2312 * @order: allocation order
2314 * Returns 0 on success, an error code on failure.
2316 int memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2318 struct mem_cgroup
*memcg
;
2321 if (memcg_kmem_bypass())
2324 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2325 if (!mem_cgroup_is_root(memcg
)) {
2326 ret
= memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2328 __SetPageKmemcg(page
);
2330 css_put(&memcg
->css
);
2334 * memcg_kmem_uncharge: uncharge a kmem page
2335 * @page: page to uncharge
2336 * @order: allocation order
2338 void memcg_kmem_uncharge(struct page
*page
, int order
)
2340 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2341 unsigned int nr_pages
= 1 << order
;
2346 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2348 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2349 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2351 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2352 if (do_memsw_account())
2353 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2355 page
->mem_cgroup
= NULL
;
2357 /* slab pages do not have PageKmemcg flag set */
2358 if (PageKmemcg(page
))
2359 __ClearPageKmemcg(page
);
2361 css_put_many(&memcg
->css
, nr_pages
);
2363 #endif /* !CONFIG_SLOB */
2365 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2368 * Because tail pages are not marked as "used", set it. We're under
2369 * zone_lru_lock and migration entries setup in all page mappings.
2371 void mem_cgroup_split_huge_fixup(struct page
*head
)
2375 if (mem_cgroup_disabled())
2378 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2379 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2381 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2384 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2386 #ifdef CONFIG_MEMCG_SWAP
2387 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2390 int val
= (charge
) ? 1 : -1;
2391 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2395 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2396 * @entry: swap entry to be moved
2397 * @from: mem_cgroup which the entry is moved from
2398 * @to: mem_cgroup which the entry is moved to
2400 * It succeeds only when the swap_cgroup's record for this entry is the same
2401 * as the mem_cgroup's id of @from.
2403 * Returns 0 on success, -EINVAL on failure.
2405 * The caller must have charged to @to, IOW, called page_counter_charge() about
2406 * both res and memsw, and called css_get().
2408 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2409 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2411 unsigned short old_id
, new_id
;
2413 old_id
= mem_cgroup_id(from
);
2414 new_id
= mem_cgroup_id(to
);
2416 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2417 mem_cgroup_swap_statistics(from
, false);
2418 mem_cgroup_swap_statistics(to
, true);
2424 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2425 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2431 static DEFINE_MUTEX(memcg_limit_mutex
);
2433 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2434 unsigned long limit
)
2436 unsigned long curusage
;
2437 unsigned long oldusage
;
2438 bool enlarge
= false;
2443 * For keeping hierarchical_reclaim simple, how long we should retry
2444 * is depends on callers. We set our retry-count to be function
2445 * of # of children which we should visit in this loop.
2447 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2448 mem_cgroup_count_children(memcg
);
2450 oldusage
= page_counter_read(&memcg
->memory
);
2453 if (signal_pending(current
)) {
2458 mutex_lock(&memcg_limit_mutex
);
2459 if (limit
> memcg
->memsw
.limit
) {
2460 mutex_unlock(&memcg_limit_mutex
);
2464 if (limit
> memcg
->memory
.limit
)
2466 ret
= page_counter_limit(&memcg
->memory
, limit
);
2467 mutex_unlock(&memcg_limit_mutex
);
2472 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2474 curusage
= page_counter_read(&memcg
->memory
);
2475 /* Usage is reduced ? */
2476 if (curusage
>= oldusage
)
2479 oldusage
= curusage
;
2480 } while (retry_count
);
2482 if (!ret
&& enlarge
)
2483 memcg_oom_recover(memcg
);
2488 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2489 unsigned long limit
)
2491 unsigned long curusage
;
2492 unsigned long oldusage
;
2493 bool enlarge
= false;
2497 /* see mem_cgroup_resize_res_limit */
2498 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2499 mem_cgroup_count_children(memcg
);
2501 oldusage
= page_counter_read(&memcg
->memsw
);
2504 if (signal_pending(current
)) {
2509 mutex_lock(&memcg_limit_mutex
);
2510 if (limit
< memcg
->memory
.limit
) {
2511 mutex_unlock(&memcg_limit_mutex
);
2515 if (limit
> memcg
->memsw
.limit
)
2517 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2518 mutex_unlock(&memcg_limit_mutex
);
2523 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2525 curusage
= page_counter_read(&memcg
->memsw
);
2526 /* Usage is reduced ? */
2527 if (curusage
>= oldusage
)
2530 oldusage
= curusage
;
2531 } while (retry_count
);
2533 if (!ret
&& enlarge
)
2534 memcg_oom_recover(memcg
);
2539 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
2541 unsigned long *total_scanned
)
2543 unsigned long nr_reclaimed
= 0;
2544 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
2545 unsigned long reclaimed
;
2547 struct mem_cgroup_tree_per_node
*mctz
;
2548 unsigned long excess
;
2549 unsigned long nr_scanned
;
2554 mctz
= soft_limit_tree_node(pgdat
->node_id
);
2557 * Do not even bother to check the largest node if the root
2558 * is empty. Do it lockless to prevent lock bouncing. Races
2559 * are acceptable as soft limit is best effort anyway.
2561 if (RB_EMPTY_ROOT(&mctz
->rb_root
))
2565 * This loop can run a while, specially if mem_cgroup's continuously
2566 * keep exceeding their soft limit and putting the system under
2573 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2578 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
2579 gfp_mask
, &nr_scanned
);
2580 nr_reclaimed
+= reclaimed
;
2581 *total_scanned
+= nr_scanned
;
2582 spin_lock_irq(&mctz
->lock
);
2583 __mem_cgroup_remove_exceeded(mz
, mctz
);
2586 * If we failed to reclaim anything from this memory cgroup
2587 * it is time to move on to the next cgroup
2591 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2593 excess
= soft_limit_excess(mz
->memcg
);
2595 * One school of thought says that we should not add
2596 * back the node to the tree if reclaim returns 0.
2597 * But our reclaim could return 0, simply because due
2598 * to priority we are exposing a smaller subset of
2599 * memory to reclaim from. Consider this as a longer
2602 /* If excess == 0, no tree ops */
2603 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2604 spin_unlock_irq(&mctz
->lock
);
2605 css_put(&mz
->memcg
->css
);
2608 * Could not reclaim anything and there are no more
2609 * mem cgroups to try or we seem to be looping without
2610 * reclaiming anything.
2612 if (!nr_reclaimed
&&
2614 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2616 } while (!nr_reclaimed
);
2618 css_put(&next_mz
->memcg
->css
);
2619 return nr_reclaimed
;
2623 * Test whether @memcg has children, dead or alive. Note that this
2624 * function doesn't care whether @memcg has use_hierarchy enabled and
2625 * returns %true if there are child csses according to the cgroup
2626 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2628 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2633 ret
= css_next_child(NULL
, &memcg
->css
);
2639 * Reclaims as many pages from the given memcg as possible.
2641 * Caller is responsible for holding css reference for memcg.
2643 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2645 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2647 /* we call try-to-free pages for make this cgroup empty */
2648 lru_add_drain_all();
2649 /* try to free all pages in this cgroup */
2650 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2653 if (signal_pending(current
))
2656 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2660 /* maybe some writeback is necessary */
2661 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2669 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
2670 char *buf
, size_t nbytes
,
2673 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2675 if (mem_cgroup_is_root(memcg
))
2677 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
2680 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
2683 return mem_cgroup_from_css(css
)->use_hierarchy
;
2686 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
2687 struct cftype
*cft
, u64 val
)
2690 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2691 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
2693 if (memcg
->use_hierarchy
== val
)
2697 * If parent's use_hierarchy is set, we can't make any modifications
2698 * in the child subtrees. If it is unset, then the change can
2699 * occur, provided the current cgroup has no children.
2701 * For the root cgroup, parent_mem is NULL, we allow value to be
2702 * set if there are no children.
2704 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
2705 (val
== 1 || val
== 0)) {
2706 if (!memcg_has_children(memcg
))
2707 memcg
->use_hierarchy
= val
;
2716 static void tree_stat(struct mem_cgroup
*memcg
, unsigned long *stat
)
2718 struct mem_cgroup
*iter
;
2721 memset(stat
, 0, sizeof(*stat
) * MEMCG_NR_STAT
);
2723 for_each_mem_cgroup_tree(iter
, memcg
) {
2724 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
2725 stat
[i
] += mem_cgroup_read_stat(iter
, i
);
2729 static void tree_events(struct mem_cgroup
*memcg
, unsigned long *events
)
2731 struct mem_cgroup
*iter
;
2734 memset(events
, 0, sizeof(*events
) * MEMCG_NR_EVENTS
);
2736 for_each_mem_cgroup_tree(iter
, memcg
) {
2737 for (i
= 0; i
< MEMCG_NR_EVENTS
; i
++)
2738 events
[i
] += mem_cgroup_read_events(iter
, i
);
2742 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
2744 unsigned long val
= 0;
2746 if (mem_cgroup_is_root(memcg
)) {
2747 struct mem_cgroup
*iter
;
2749 for_each_mem_cgroup_tree(iter
, memcg
) {
2750 val
+= mem_cgroup_read_stat(iter
,
2751 MEM_CGROUP_STAT_CACHE
);
2752 val
+= mem_cgroup_read_stat(iter
,
2753 MEM_CGROUP_STAT_RSS
);
2755 val
+= mem_cgroup_read_stat(iter
,
2756 MEM_CGROUP_STAT_SWAP
);
2760 val
= page_counter_read(&memcg
->memory
);
2762 val
= page_counter_read(&memcg
->memsw
);
2775 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
2778 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2779 struct page_counter
*counter
;
2781 switch (MEMFILE_TYPE(cft
->private)) {
2783 counter
= &memcg
->memory
;
2786 counter
= &memcg
->memsw
;
2789 counter
= &memcg
->kmem
;
2792 counter
= &memcg
->tcpmem
;
2798 switch (MEMFILE_ATTR(cft
->private)) {
2800 if (counter
== &memcg
->memory
)
2801 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
2802 if (counter
== &memcg
->memsw
)
2803 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
2804 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
2806 return (u64
)counter
->limit
* PAGE_SIZE
;
2808 return (u64
)counter
->watermark
* PAGE_SIZE
;
2810 return counter
->failcnt
;
2811 case RES_SOFT_LIMIT
:
2812 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
2819 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2823 if (cgroup_memory_nokmem
)
2826 BUG_ON(memcg
->kmemcg_id
>= 0);
2827 BUG_ON(memcg
->kmem_state
);
2829 memcg_id
= memcg_alloc_cache_id();
2833 static_branch_inc(&memcg_kmem_enabled_key
);
2835 * A memory cgroup is considered kmem-online as soon as it gets
2836 * kmemcg_id. Setting the id after enabling static branching will
2837 * guarantee no one starts accounting before all call sites are
2840 memcg
->kmemcg_id
= memcg_id
;
2841 memcg
->kmem_state
= KMEM_ONLINE
;
2846 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2848 struct cgroup_subsys_state
*css
;
2849 struct mem_cgroup
*parent
, *child
;
2852 if (memcg
->kmem_state
!= KMEM_ONLINE
)
2855 * Clear the online state before clearing memcg_caches array
2856 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2857 * guarantees that no cache will be created for this cgroup
2858 * after we are done (see memcg_create_kmem_cache()).
2860 memcg
->kmem_state
= KMEM_ALLOCATED
;
2862 memcg_deactivate_kmem_caches(memcg
);
2864 kmemcg_id
= memcg
->kmemcg_id
;
2865 BUG_ON(kmemcg_id
< 0);
2867 parent
= parent_mem_cgroup(memcg
);
2869 parent
= root_mem_cgroup
;
2872 * Change kmemcg_id of this cgroup and all its descendants to the
2873 * parent's id, and then move all entries from this cgroup's list_lrus
2874 * to ones of the parent. After we have finished, all list_lrus
2875 * corresponding to this cgroup are guaranteed to remain empty. The
2876 * ordering is imposed by list_lru_node->lock taken by
2877 * memcg_drain_all_list_lrus().
2879 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2880 css_for_each_descendant_pre(css
, &memcg
->css
) {
2881 child
= mem_cgroup_from_css(css
);
2882 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
2883 child
->kmemcg_id
= parent
->kmemcg_id
;
2884 if (!memcg
->use_hierarchy
)
2889 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
2891 memcg_free_cache_id(kmemcg_id
);
2894 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2896 /* css_alloc() failed, offlining didn't happen */
2897 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
2898 memcg_offline_kmem(memcg
);
2900 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
2901 memcg_destroy_kmem_caches(memcg
);
2902 static_branch_dec(&memcg_kmem_enabled_key
);
2903 WARN_ON(page_counter_read(&memcg
->kmem
));
2907 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2911 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2914 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2917 #endif /* !CONFIG_SLOB */
2919 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2920 unsigned long limit
)
2924 mutex_lock(&memcg_limit_mutex
);
2925 ret
= page_counter_limit(&memcg
->kmem
, limit
);
2926 mutex_unlock(&memcg_limit_mutex
);
2930 static int memcg_update_tcp_limit(struct mem_cgroup
*memcg
, unsigned long limit
)
2934 mutex_lock(&memcg_limit_mutex
);
2936 ret
= page_counter_limit(&memcg
->tcpmem
, limit
);
2940 if (!memcg
->tcpmem_active
) {
2942 * The active flag needs to be written after the static_key
2943 * update. This is what guarantees that the socket activation
2944 * function is the last one to run. See mem_cgroup_sk_alloc()
2945 * for details, and note that we don't mark any socket as
2946 * belonging to this memcg until that flag is up.
2948 * We need to do this, because static_keys will span multiple
2949 * sites, but we can't control their order. If we mark a socket
2950 * as accounted, but the accounting functions are not patched in
2951 * yet, we'll lose accounting.
2953 * We never race with the readers in mem_cgroup_sk_alloc(),
2954 * because when this value change, the code to process it is not
2957 static_branch_inc(&memcg_sockets_enabled_key
);
2958 memcg
->tcpmem_active
= true;
2961 mutex_unlock(&memcg_limit_mutex
);
2966 * The user of this function is...
2969 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
2970 char *buf
, size_t nbytes
, loff_t off
)
2972 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2973 unsigned long nr_pages
;
2976 buf
= strstrip(buf
);
2977 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
2981 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
2983 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
2987 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
2989 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
2992 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
2995 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
2998 ret
= memcg_update_tcp_limit(memcg
, nr_pages
);
3002 case RES_SOFT_LIMIT
:
3003 memcg
->soft_limit
= nr_pages
;
3007 return ret
?: nbytes
;
3010 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3011 size_t nbytes
, loff_t off
)
3013 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3014 struct page_counter
*counter
;
3016 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3018 counter
= &memcg
->memory
;
3021 counter
= &memcg
->memsw
;
3024 counter
= &memcg
->kmem
;
3027 counter
= &memcg
->tcpmem
;
3033 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3035 page_counter_reset_watermark(counter
);
3038 counter
->failcnt
= 0;
3047 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3050 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3054 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3055 struct cftype
*cft
, u64 val
)
3057 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3059 if (val
& ~MOVE_MASK
)
3063 * No kind of locking is needed in here, because ->can_attach() will
3064 * check this value once in the beginning of the process, and then carry
3065 * on with stale data. This means that changes to this value will only
3066 * affect task migrations starting after the change.
3068 memcg
->move_charge_at_immigrate
= val
;
3072 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3073 struct cftype
*cft
, u64 val
)
3080 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3084 unsigned int lru_mask
;
3087 static const struct numa_stat stats
[] = {
3088 { "total", LRU_ALL
},
3089 { "file", LRU_ALL_FILE
},
3090 { "anon", LRU_ALL_ANON
},
3091 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3093 const struct numa_stat
*stat
;
3096 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3098 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3099 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3100 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3101 for_each_node_state(nid
, N_MEMORY
) {
3102 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3104 seq_printf(m
, " N%d=%lu", nid
, nr
);
3109 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3110 struct mem_cgroup
*iter
;
3113 for_each_mem_cgroup_tree(iter
, memcg
)
3114 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3115 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3116 for_each_node_state(nid
, N_MEMORY
) {
3118 for_each_mem_cgroup_tree(iter
, memcg
)
3119 nr
+= mem_cgroup_node_nr_lru_pages(
3120 iter
, nid
, stat
->lru_mask
);
3121 seq_printf(m
, " N%d=%lu", nid
, nr
);
3128 #endif /* CONFIG_NUMA */
3130 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3132 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3133 unsigned long memory
, memsw
;
3134 struct mem_cgroup
*mi
;
3137 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3138 MEM_CGROUP_STAT_NSTATS
);
3139 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3140 MEM_CGROUP_EVENTS_NSTATS
);
3141 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3143 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3144 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3146 seq_printf(m
, "%s %lu\n", mem_cgroup_stat_names
[i
],
3147 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3150 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3151 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3152 mem_cgroup_read_events(memcg
, i
));
3154 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3155 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3156 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3158 /* Hierarchical information */
3159 memory
= memsw
= PAGE_COUNTER_MAX
;
3160 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3161 memory
= min(memory
, mi
->memory
.limit
);
3162 memsw
= min(memsw
, mi
->memsw
.limit
);
3164 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3165 (u64
)memory
* PAGE_SIZE
);
3166 if (do_memsw_account())
3167 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3168 (u64
)memsw
* PAGE_SIZE
);
3170 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3171 unsigned long long val
= 0;
3173 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3175 for_each_mem_cgroup_tree(mi
, memcg
)
3176 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3177 seq_printf(m
, "total_%s %llu\n", mem_cgroup_stat_names
[i
], val
);
3180 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3181 unsigned long long val
= 0;
3183 for_each_mem_cgroup_tree(mi
, memcg
)
3184 val
+= mem_cgroup_read_events(mi
, i
);
3185 seq_printf(m
, "total_%s %llu\n",
3186 mem_cgroup_events_names
[i
], val
);
3189 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3190 unsigned long long val
= 0;
3192 for_each_mem_cgroup_tree(mi
, memcg
)
3193 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3194 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3197 #ifdef CONFIG_DEBUG_VM
3200 struct mem_cgroup_per_node
*mz
;
3201 struct zone_reclaim_stat
*rstat
;
3202 unsigned long recent_rotated
[2] = {0, 0};
3203 unsigned long recent_scanned
[2] = {0, 0};
3205 for_each_online_pgdat(pgdat
) {
3206 mz
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
3207 rstat
= &mz
->lruvec
.reclaim_stat
;
3209 recent_rotated
[0] += rstat
->recent_rotated
[0];
3210 recent_rotated
[1] += rstat
->recent_rotated
[1];
3211 recent_scanned
[0] += rstat
->recent_scanned
[0];
3212 recent_scanned
[1] += rstat
->recent_scanned
[1];
3214 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3215 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3216 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3217 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3224 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3227 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3229 return mem_cgroup_swappiness(memcg
);
3232 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3233 struct cftype
*cft
, u64 val
)
3235 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3241 memcg
->swappiness
= val
;
3243 vm_swappiness
= val
;
3248 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3250 struct mem_cgroup_threshold_ary
*t
;
3251 unsigned long usage
;
3256 t
= rcu_dereference(memcg
->thresholds
.primary
);
3258 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3263 usage
= mem_cgroup_usage(memcg
, swap
);
3266 * current_threshold points to threshold just below or equal to usage.
3267 * If it's not true, a threshold was crossed after last
3268 * call of __mem_cgroup_threshold().
3270 i
= t
->current_threshold
;
3273 * Iterate backward over array of thresholds starting from
3274 * current_threshold and check if a threshold is crossed.
3275 * If none of thresholds below usage is crossed, we read
3276 * only one element of the array here.
3278 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3279 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3281 /* i = current_threshold + 1 */
3285 * Iterate forward over array of thresholds starting from
3286 * current_threshold+1 and check if a threshold is crossed.
3287 * If none of thresholds above usage is crossed, we read
3288 * only one element of the array here.
3290 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3291 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3293 /* Update current_threshold */
3294 t
->current_threshold
= i
- 1;
3299 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3302 __mem_cgroup_threshold(memcg
, false);
3303 if (do_memsw_account())
3304 __mem_cgroup_threshold(memcg
, true);
3306 memcg
= parent_mem_cgroup(memcg
);
3310 static int compare_thresholds(const void *a
, const void *b
)
3312 const struct mem_cgroup_threshold
*_a
= a
;
3313 const struct mem_cgroup_threshold
*_b
= b
;
3315 if (_a
->threshold
> _b
->threshold
)
3318 if (_a
->threshold
< _b
->threshold
)
3324 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3326 struct mem_cgroup_eventfd_list
*ev
;
3328 spin_lock(&memcg_oom_lock
);
3330 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3331 eventfd_signal(ev
->eventfd
, 1);
3333 spin_unlock(&memcg_oom_lock
);
3337 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3339 struct mem_cgroup
*iter
;
3341 for_each_mem_cgroup_tree(iter
, memcg
)
3342 mem_cgroup_oom_notify_cb(iter
);
3345 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3346 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3348 struct mem_cgroup_thresholds
*thresholds
;
3349 struct mem_cgroup_threshold_ary
*new;
3350 unsigned long threshold
;
3351 unsigned long usage
;
3354 ret
= page_counter_memparse(args
, "-1", &threshold
);
3358 mutex_lock(&memcg
->thresholds_lock
);
3361 thresholds
= &memcg
->thresholds
;
3362 usage
= mem_cgroup_usage(memcg
, false);
3363 } else if (type
== _MEMSWAP
) {
3364 thresholds
= &memcg
->memsw_thresholds
;
3365 usage
= mem_cgroup_usage(memcg
, true);
3369 /* Check if a threshold crossed before adding a new one */
3370 if (thresholds
->primary
)
3371 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3373 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3375 /* Allocate memory for new array of thresholds */
3376 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3384 /* Copy thresholds (if any) to new array */
3385 if (thresholds
->primary
) {
3386 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3387 sizeof(struct mem_cgroup_threshold
));
3390 /* Add new threshold */
3391 new->entries
[size
- 1].eventfd
= eventfd
;
3392 new->entries
[size
- 1].threshold
= threshold
;
3394 /* Sort thresholds. Registering of new threshold isn't time-critical */
3395 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3396 compare_thresholds
, NULL
);
3398 /* Find current threshold */
3399 new->current_threshold
= -1;
3400 for (i
= 0; i
< size
; i
++) {
3401 if (new->entries
[i
].threshold
<= usage
) {
3403 * new->current_threshold will not be used until
3404 * rcu_assign_pointer(), so it's safe to increment
3407 ++new->current_threshold
;
3412 /* Free old spare buffer and save old primary buffer as spare */
3413 kfree(thresholds
->spare
);
3414 thresholds
->spare
= thresholds
->primary
;
3416 rcu_assign_pointer(thresholds
->primary
, new);
3418 /* To be sure that nobody uses thresholds */
3422 mutex_unlock(&memcg
->thresholds_lock
);
3427 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3428 struct eventfd_ctx
*eventfd
, const char *args
)
3430 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3433 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3434 struct eventfd_ctx
*eventfd
, const char *args
)
3436 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3439 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3440 struct eventfd_ctx
*eventfd
, enum res_type type
)
3442 struct mem_cgroup_thresholds
*thresholds
;
3443 struct mem_cgroup_threshold_ary
*new;
3444 unsigned long usage
;
3447 mutex_lock(&memcg
->thresholds_lock
);
3450 thresholds
= &memcg
->thresholds
;
3451 usage
= mem_cgroup_usage(memcg
, false);
3452 } else if (type
== _MEMSWAP
) {
3453 thresholds
= &memcg
->memsw_thresholds
;
3454 usage
= mem_cgroup_usage(memcg
, true);
3458 if (!thresholds
->primary
)
3461 /* Check if a threshold crossed before removing */
3462 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3464 /* Calculate new number of threshold */
3466 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3467 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3471 new = thresholds
->spare
;
3473 /* Set thresholds array to NULL if we don't have thresholds */
3482 /* Copy thresholds and find current threshold */
3483 new->current_threshold
= -1;
3484 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3485 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3488 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3489 if (new->entries
[j
].threshold
<= usage
) {
3491 * new->current_threshold will not be used
3492 * until rcu_assign_pointer(), so it's safe to increment
3495 ++new->current_threshold
;
3501 /* Swap primary and spare array */
3502 thresholds
->spare
= thresholds
->primary
;
3504 rcu_assign_pointer(thresholds
->primary
, new);
3506 /* To be sure that nobody uses thresholds */
3509 /* If all events are unregistered, free the spare array */
3511 kfree(thresholds
->spare
);
3512 thresholds
->spare
= NULL
;
3515 mutex_unlock(&memcg
->thresholds_lock
);
3518 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3519 struct eventfd_ctx
*eventfd
)
3521 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3524 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3525 struct eventfd_ctx
*eventfd
)
3527 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3530 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3531 struct eventfd_ctx
*eventfd
, const char *args
)
3533 struct mem_cgroup_eventfd_list
*event
;
3535 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3539 spin_lock(&memcg_oom_lock
);
3541 event
->eventfd
= eventfd
;
3542 list_add(&event
->list
, &memcg
->oom_notify
);
3544 /* already in OOM ? */
3545 if (memcg
->under_oom
)
3546 eventfd_signal(eventfd
, 1);
3547 spin_unlock(&memcg_oom_lock
);
3552 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3553 struct eventfd_ctx
*eventfd
)
3555 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3557 spin_lock(&memcg_oom_lock
);
3559 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3560 if (ev
->eventfd
== eventfd
) {
3561 list_del(&ev
->list
);
3566 spin_unlock(&memcg_oom_lock
);
3569 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3571 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3573 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3574 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3578 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3579 struct cftype
*cft
, u64 val
)
3581 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3583 /* cannot set to root cgroup and only 0 and 1 are allowed */
3584 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3587 memcg
->oom_kill_disable
= val
;
3589 memcg_oom_recover(memcg
);
3594 #ifdef CONFIG_CGROUP_WRITEBACK
3596 struct list_head
*mem_cgroup_cgwb_list(struct mem_cgroup
*memcg
)
3598 return &memcg
->cgwb_list
;
3601 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3603 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3606 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3608 wb_domain_exit(&memcg
->cgwb_domain
);
3611 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3613 wb_domain_size_changed(&memcg
->cgwb_domain
);
3616 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3618 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3620 if (!memcg
->css
.parent
)
3623 return &memcg
->cgwb_domain
;
3627 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3628 * @wb: bdi_writeback in question
3629 * @pfilepages: out parameter for number of file pages
3630 * @pheadroom: out parameter for number of allocatable pages according to memcg
3631 * @pdirty: out parameter for number of dirty pages
3632 * @pwriteback: out parameter for number of pages under writeback
3634 * Determine the numbers of file, headroom, dirty, and writeback pages in
3635 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3636 * is a bit more involved.
3638 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3639 * headroom is calculated as the lowest headroom of itself and the
3640 * ancestors. Note that this doesn't consider the actual amount of
3641 * available memory in the system. The caller should further cap
3642 * *@pheadroom accordingly.
3644 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
3645 unsigned long *pheadroom
, unsigned long *pdirty
,
3646 unsigned long *pwriteback
)
3648 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3649 struct mem_cgroup
*parent
;
3651 *pdirty
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
3653 /* this should eventually include NR_UNSTABLE_NFS */
3654 *pwriteback
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
3655 *pfilepages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
3656 (1 << LRU_ACTIVE_FILE
));
3657 *pheadroom
= PAGE_COUNTER_MAX
;
3659 while ((parent
= parent_mem_cgroup(memcg
))) {
3660 unsigned long ceiling
= min(memcg
->memory
.limit
, memcg
->high
);
3661 unsigned long used
= page_counter_read(&memcg
->memory
);
3663 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
3668 #else /* CONFIG_CGROUP_WRITEBACK */
3670 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3675 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3679 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3683 #endif /* CONFIG_CGROUP_WRITEBACK */
3686 * DO NOT USE IN NEW FILES.
3688 * "cgroup.event_control" implementation.
3690 * This is way over-engineered. It tries to support fully configurable
3691 * events for each user. Such level of flexibility is completely
3692 * unnecessary especially in the light of the planned unified hierarchy.
3694 * Please deprecate this and replace with something simpler if at all
3699 * Unregister event and free resources.
3701 * Gets called from workqueue.
3703 static void memcg_event_remove(struct work_struct
*work
)
3705 struct mem_cgroup_event
*event
=
3706 container_of(work
, struct mem_cgroup_event
, remove
);
3707 struct mem_cgroup
*memcg
= event
->memcg
;
3709 remove_wait_queue(event
->wqh
, &event
->wait
);
3711 event
->unregister_event(memcg
, event
->eventfd
);
3713 /* Notify userspace the event is going away. */
3714 eventfd_signal(event
->eventfd
, 1);
3716 eventfd_ctx_put(event
->eventfd
);
3718 css_put(&memcg
->css
);
3722 * Gets called on POLLHUP on eventfd when user closes it.
3724 * Called with wqh->lock held and interrupts disabled.
3726 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
3727 int sync
, void *key
)
3729 struct mem_cgroup_event
*event
=
3730 container_of(wait
, struct mem_cgroup_event
, wait
);
3731 struct mem_cgroup
*memcg
= event
->memcg
;
3732 unsigned long flags
= (unsigned long)key
;
3734 if (flags
& POLLHUP
) {
3736 * If the event has been detached at cgroup removal, we
3737 * can simply return knowing the other side will cleanup
3740 * We can't race against event freeing since the other
3741 * side will require wqh->lock via remove_wait_queue(),
3744 spin_lock(&memcg
->event_list_lock
);
3745 if (!list_empty(&event
->list
)) {
3746 list_del_init(&event
->list
);
3748 * We are in atomic context, but cgroup_event_remove()
3749 * may sleep, so we have to call it in workqueue.
3751 schedule_work(&event
->remove
);
3753 spin_unlock(&memcg
->event_list_lock
);
3759 static void memcg_event_ptable_queue_proc(struct file
*file
,
3760 wait_queue_head_t
*wqh
, poll_table
*pt
)
3762 struct mem_cgroup_event
*event
=
3763 container_of(pt
, struct mem_cgroup_event
, pt
);
3766 add_wait_queue(wqh
, &event
->wait
);
3770 * DO NOT USE IN NEW FILES.
3772 * Parse input and register new cgroup event handler.
3774 * Input must be in format '<event_fd> <control_fd> <args>'.
3775 * Interpretation of args is defined by control file implementation.
3777 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
3778 char *buf
, size_t nbytes
, loff_t off
)
3780 struct cgroup_subsys_state
*css
= of_css(of
);
3781 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3782 struct mem_cgroup_event
*event
;
3783 struct cgroup_subsys_state
*cfile_css
;
3784 unsigned int efd
, cfd
;
3791 buf
= strstrip(buf
);
3793 efd
= simple_strtoul(buf
, &endp
, 10);
3798 cfd
= simple_strtoul(buf
, &endp
, 10);
3799 if ((*endp
!= ' ') && (*endp
!= '\0'))
3803 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
3807 event
->memcg
= memcg
;
3808 INIT_LIST_HEAD(&event
->list
);
3809 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
3810 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
3811 INIT_WORK(&event
->remove
, memcg_event_remove
);
3819 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
3820 if (IS_ERR(event
->eventfd
)) {
3821 ret
= PTR_ERR(event
->eventfd
);
3828 goto out_put_eventfd
;
3831 /* the process need read permission on control file */
3832 /* AV: shouldn't we check that it's been opened for read instead? */
3833 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
3838 * Determine the event callbacks and set them in @event. This used
3839 * to be done via struct cftype but cgroup core no longer knows
3840 * about these events. The following is crude but the whole thing
3841 * is for compatibility anyway.
3843 * DO NOT ADD NEW FILES.
3845 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
3847 if (!strcmp(name
, "memory.usage_in_bytes")) {
3848 event
->register_event
= mem_cgroup_usage_register_event
;
3849 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
3850 } else if (!strcmp(name
, "memory.oom_control")) {
3851 event
->register_event
= mem_cgroup_oom_register_event
;
3852 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
3853 } else if (!strcmp(name
, "memory.pressure_level")) {
3854 event
->register_event
= vmpressure_register_event
;
3855 event
->unregister_event
= vmpressure_unregister_event
;
3856 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
3857 event
->register_event
= memsw_cgroup_usage_register_event
;
3858 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
3865 * Verify @cfile should belong to @css. Also, remaining events are
3866 * automatically removed on cgroup destruction but the removal is
3867 * asynchronous, so take an extra ref on @css.
3869 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
3870 &memory_cgrp_subsys
);
3872 if (IS_ERR(cfile_css
))
3874 if (cfile_css
!= css
) {
3879 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
3883 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
3885 spin_lock(&memcg
->event_list_lock
);
3886 list_add(&event
->list
, &memcg
->event_list
);
3887 spin_unlock(&memcg
->event_list_lock
);
3899 eventfd_ctx_put(event
->eventfd
);
3908 static struct cftype mem_cgroup_legacy_files
[] = {
3910 .name
= "usage_in_bytes",
3911 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
3912 .read_u64
= mem_cgroup_read_u64
,
3915 .name
= "max_usage_in_bytes",
3916 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
3917 .write
= mem_cgroup_reset
,
3918 .read_u64
= mem_cgroup_read_u64
,
3921 .name
= "limit_in_bytes",
3922 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
3923 .write
= mem_cgroup_write
,
3924 .read_u64
= mem_cgroup_read_u64
,
3927 .name
= "soft_limit_in_bytes",
3928 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
3929 .write
= mem_cgroup_write
,
3930 .read_u64
= mem_cgroup_read_u64
,
3934 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
3935 .write
= mem_cgroup_reset
,
3936 .read_u64
= mem_cgroup_read_u64
,
3940 .seq_show
= memcg_stat_show
,
3943 .name
= "force_empty",
3944 .write
= mem_cgroup_force_empty_write
,
3947 .name
= "use_hierarchy",
3948 .write_u64
= mem_cgroup_hierarchy_write
,
3949 .read_u64
= mem_cgroup_hierarchy_read
,
3952 .name
= "cgroup.event_control", /* XXX: for compat */
3953 .write
= memcg_write_event_control
,
3954 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
3957 .name
= "swappiness",
3958 .read_u64
= mem_cgroup_swappiness_read
,
3959 .write_u64
= mem_cgroup_swappiness_write
,
3962 .name
= "move_charge_at_immigrate",
3963 .read_u64
= mem_cgroup_move_charge_read
,
3964 .write_u64
= mem_cgroup_move_charge_write
,
3967 .name
= "oom_control",
3968 .seq_show
= mem_cgroup_oom_control_read
,
3969 .write_u64
= mem_cgroup_oom_control_write
,
3970 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
3973 .name
= "pressure_level",
3977 .name
= "numa_stat",
3978 .seq_show
= memcg_numa_stat_show
,
3982 .name
= "kmem.limit_in_bytes",
3983 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
3984 .write
= mem_cgroup_write
,
3985 .read_u64
= mem_cgroup_read_u64
,
3988 .name
= "kmem.usage_in_bytes",
3989 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
3990 .read_u64
= mem_cgroup_read_u64
,
3993 .name
= "kmem.failcnt",
3994 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
3995 .write
= mem_cgroup_reset
,
3996 .read_u64
= mem_cgroup_read_u64
,
3999 .name
= "kmem.max_usage_in_bytes",
4000 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4001 .write
= mem_cgroup_reset
,
4002 .read_u64
= mem_cgroup_read_u64
,
4004 #ifdef CONFIG_SLABINFO
4006 .name
= "kmem.slabinfo",
4007 .seq_start
= slab_start
,
4008 .seq_next
= slab_next
,
4009 .seq_stop
= slab_stop
,
4010 .seq_show
= memcg_slab_show
,
4014 .name
= "kmem.tcp.limit_in_bytes",
4015 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4016 .write
= mem_cgroup_write
,
4017 .read_u64
= mem_cgroup_read_u64
,
4020 .name
= "kmem.tcp.usage_in_bytes",
4021 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4022 .read_u64
= mem_cgroup_read_u64
,
4025 .name
= "kmem.tcp.failcnt",
4026 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4027 .write
= mem_cgroup_reset
,
4028 .read_u64
= mem_cgroup_read_u64
,
4031 .name
= "kmem.tcp.max_usage_in_bytes",
4032 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4033 .write
= mem_cgroup_reset
,
4034 .read_u64
= mem_cgroup_read_u64
,
4036 { }, /* terminate */
4040 * Private memory cgroup IDR
4042 * Swap-out records and page cache shadow entries need to store memcg
4043 * references in constrained space, so we maintain an ID space that is
4044 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4045 * memory-controlled cgroups to 64k.
4047 * However, there usually are many references to the oflline CSS after
4048 * the cgroup has been destroyed, such as page cache or reclaimable
4049 * slab objects, that don't need to hang on to the ID. We want to keep
4050 * those dead CSS from occupying IDs, or we might quickly exhaust the
4051 * relatively small ID space and prevent the creation of new cgroups
4052 * even when there are much fewer than 64k cgroups - possibly none.
4054 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4055 * be freed and recycled when it's no longer needed, which is usually
4056 * when the CSS is offlined.
4058 * The only exception to that are records of swapped out tmpfs/shmem
4059 * pages that need to be attributed to live ancestors on swapin. But
4060 * those references are manageable from userspace.
4063 static DEFINE_IDR(mem_cgroup_idr
);
4065 static void mem_cgroup_id_get_many(struct mem_cgroup
*memcg
, unsigned int n
)
4067 VM_BUG_ON(atomic_read(&memcg
->id
.ref
) <= 0);
4068 atomic_add(n
, &memcg
->id
.ref
);
4071 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
4073 VM_BUG_ON(atomic_read(&memcg
->id
.ref
) < n
);
4074 if (atomic_sub_and_test(n
, &memcg
->id
.ref
)) {
4075 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4078 /* Memcg ID pins CSS */
4079 css_put(&memcg
->css
);
4083 static inline void mem_cgroup_id_get(struct mem_cgroup
*memcg
)
4085 mem_cgroup_id_get_many(memcg
, 1);
4088 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
4090 mem_cgroup_id_put_many(memcg
, 1);
4094 * mem_cgroup_from_id - look up a memcg from a memcg id
4095 * @id: the memcg id to look up
4097 * Caller must hold rcu_read_lock().
4099 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
4101 WARN_ON_ONCE(!rcu_read_lock_held());
4102 return idr_find(&mem_cgroup_idr
, id
);
4105 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4107 struct mem_cgroup_per_node
*pn
;
4110 * This routine is called against possible nodes.
4111 * But it's BUG to call kmalloc() against offline node.
4113 * TODO: this routine can waste much memory for nodes which will
4114 * never be onlined. It's better to use memory hotplug callback
4117 if (!node_state(node
, N_NORMAL_MEMORY
))
4119 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4123 lruvec_init(&pn
->lruvec
);
4124 pn
->usage_in_excess
= 0;
4125 pn
->on_tree
= false;
4128 memcg
->nodeinfo
[node
] = pn
;
4132 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4134 kfree(memcg
->nodeinfo
[node
]);
4137 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4141 memcg_wb_domain_exit(memcg
);
4143 free_mem_cgroup_per_node_info(memcg
, node
);
4144 free_percpu(memcg
->stat
);
4148 static struct mem_cgroup
*mem_cgroup_alloc(void)
4150 struct mem_cgroup
*memcg
;
4154 size
= sizeof(struct mem_cgroup
);
4155 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4157 memcg
= kzalloc(size
, GFP_KERNEL
);
4161 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
4162 1, MEM_CGROUP_ID_MAX
,
4164 if (memcg
->id
.id
< 0)
4167 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4172 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
4175 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4178 INIT_WORK(&memcg
->high_work
, high_work_func
);
4179 memcg
->last_scanned_node
= MAX_NUMNODES
;
4180 INIT_LIST_HEAD(&memcg
->oom_notify
);
4181 mutex_init(&memcg
->thresholds_lock
);
4182 spin_lock_init(&memcg
->move_lock
);
4183 vmpressure_init(&memcg
->vmpressure
);
4184 INIT_LIST_HEAD(&memcg
->event_list
);
4185 spin_lock_init(&memcg
->event_list_lock
);
4186 memcg
->socket_pressure
= jiffies
;
4188 memcg
->kmemcg_id
= -1;
4190 #ifdef CONFIG_CGROUP_WRITEBACK
4191 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4193 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
4196 if (memcg
->id
.id
> 0)
4197 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4198 mem_cgroup_free(memcg
);
4202 static struct cgroup_subsys_state
* __ref
4203 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4205 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
4206 struct mem_cgroup
*memcg
;
4207 long error
= -ENOMEM
;
4209 memcg
= mem_cgroup_alloc();
4211 return ERR_PTR(error
);
4213 memcg
->high
= PAGE_COUNTER_MAX
;
4214 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4216 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4217 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4219 if (parent
&& parent
->use_hierarchy
) {
4220 memcg
->use_hierarchy
= true;
4221 page_counter_init(&memcg
->memory
, &parent
->memory
);
4222 page_counter_init(&memcg
->swap
, &parent
->swap
);
4223 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4224 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4225 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
4227 page_counter_init(&memcg
->memory
, NULL
);
4228 page_counter_init(&memcg
->swap
, NULL
);
4229 page_counter_init(&memcg
->memsw
, NULL
);
4230 page_counter_init(&memcg
->kmem
, NULL
);
4231 page_counter_init(&memcg
->tcpmem
, NULL
);
4233 * Deeper hierachy with use_hierarchy == false doesn't make
4234 * much sense so let cgroup subsystem know about this
4235 * unfortunate state in our controller.
4237 if (parent
!= root_mem_cgroup
)
4238 memory_cgrp_subsys
.broken_hierarchy
= true;
4241 /* The following stuff does not apply to the root */
4243 root_mem_cgroup
= memcg
;
4247 error
= memcg_online_kmem(memcg
);
4251 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4252 static_branch_inc(&memcg_sockets_enabled_key
);
4256 mem_cgroup_free(memcg
);
4257 return ERR_PTR(-ENOMEM
);
4260 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4262 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4264 /* Online state pins memcg ID, memcg ID pins CSS */
4265 atomic_set(&memcg
->id
.ref
, 1);
4270 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4272 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4273 struct mem_cgroup_event
*event
, *tmp
;
4276 * Unregister events and notify userspace.
4277 * Notify userspace about cgroup removing only after rmdir of cgroup
4278 * directory to avoid race between userspace and kernelspace.
4280 spin_lock(&memcg
->event_list_lock
);
4281 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4282 list_del_init(&event
->list
);
4283 schedule_work(&event
->remove
);
4285 spin_unlock(&memcg
->event_list_lock
);
4287 memcg_offline_kmem(memcg
);
4288 wb_memcg_offline(memcg
);
4290 mem_cgroup_id_put(memcg
);
4293 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4295 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4297 invalidate_reclaim_iterators(memcg
);
4300 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4302 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4304 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4305 static_branch_dec(&memcg_sockets_enabled_key
);
4307 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
4308 static_branch_dec(&memcg_sockets_enabled_key
);
4310 vmpressure_cleanup(&memcg
->vmpressure
);
4311 cancel_work_sync(&memcg
->high_work
);
4312 mem_cgroup_remove_from_trees(memcg
);
4313 memcg_free_kmem(memcg
);
4314 mem_cgroup_free(memcg
);
4318 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4319 * @css: the target css
4321 * Reset the states of the mem_cgroup associated with @css. This is
4322 * invoked when the userland requests disabling on the default hierarchy
4323 * but the memcg is pinned through dependency. The memcg should stop
4324 * applying policies and should revert to the vanilla state as it may be
4325 * made visible again.
4327 * The current implementation only resets the essential configurations.
4328 * This needs to be expanded to cover all the visible parts.
4330 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4332 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4334 page_counter_limit(&memcg
->memory
, PAGE_COUNTER_MAX
);
4335 page_counter_limit(&memcg
->swap
, PAGE_COUNTER_MAX
);
4336 page_counter_limit(&memcg
->memsw
, PAGE_COUNTER_MAX
);
4337 page_counter_limit(&memcg
->kmem
, PAGE_COUNTER_MAX
);
4338 page_counter_limit(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
4340 memcg
->high
= PAGE_COUNTER_MAX
;
4341 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4342 memcg_wb_domain_size_changed(memcg
);
4346 /* Handlers for move charge at task migration. */
4347 static int mem_cgroup_do_precharge(unsigned long count
)
4351 /* Try a single bulk charge without reclaim first, kswapd may wake */
4352 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4354 mc
.precharge
+= count
;
4358 /* Try charges one by one with reclaim */
4360 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4374 enum mc_target_type
{
4380 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4381 unsigned long addr
, pte_t ptent
)
4383 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4385 if (!page
|| !page_mapped(page
))
4387 if (PageAnon(page
)) {
4388 if (!(mc
.flags
& MOVE_ANON
))
4391 if (!(mc
.flags
& MOVE_FILE
))
4394 if (!get_page_unless_zero(page
))
4401 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4402 pte_t ptent
, swp_entry_t
*entry
)
4404 struct page
*page
= NULL
;
4405 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4407 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4410 * Because lookup_swap_cache() updates some statistics counter,
4411 * we call find_get_page() with swapper_space directly.
4413 page
= find_get_page(swap_address_space(ent
), swp_offset(ent
));
4414 if (do_memsw_account())
4415 entry
->val
= ent
.val
;
4420 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4421 pte_t ptent
, swp_entry_t
*entry
)
4427 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4428 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4430 struct page
*page
= NULL
;
4431 struct address_space
*mapping
;
4434 if (!vma
->vm_file
) /* anonymous vma */
4436 if (!(mc
.flags
& MOVE_FILE
))
4439 mapping
= vma
->vm_file
->f_mapping
;
4440 pgoff
= linear_page_index(vma
, addr
);
4442 /* page is moved even if it's not RSS of this task(page-faulted). */
4444 /* shmem/tmpfs may report page out on swap: account for that too. */
4445 if (shmem_mapping(mapping
)) {
4446 page
= find_get_entry(mapping
, pgoff
);
4447 if (radix_tree_exceptional_entry(page
)) {
4448 swp_entry_t swp
= radix_to_swp_entry(page
);
4449 if (do_memsw_account())
4451 page
= find_get_page(swap_address_space(swp
),
4455 page
= find_get_page(mapping
, pgoff
);
4457 page
= find_get_page(mapping
, pgoff
);
4463 * mem_cgroup_move_account - move account of the page
4465 * @compound: charge the page as compound or small page
4466 * @from: mem_cgroup which the page is moved from.
4467 * @to: mem_cgroup which the page is moved to. @from != @to.
4469 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4471 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4474 static int mem_cgroup_move_account(struct page
*page
,
4476 struct mem_cgroup
*from
,
4477 struct mem_cgroup
*to
)
4479 unsigned long flags
;
4480 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
4484 VM_BUG_ON(from
== to
);
4485 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4486 VM_BUG_ON(compound
&& !PageTransHuge(page
));
4489 * Prevent mem_cgroup_migrate() from looking at
4490 * page->mem_cgroup of its source page while we change it.
4493 if (!trylock_page(page
))
4497 if (page
->mem_cgroup
!= from
)
4500 anon
= PageAnon(page
);
4502 spin_lock_irqsave(&from
->move_lock
, flags
);
4504 if (!anon
&& page_mapped(page
)) {
4505 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4507 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4512 * move_lock grabbed above and caller set from->moving_account, so
4513 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4514 * So mapping should be stable for dirty pages.
4516 if (!anon
&& PageDirty(page
)) {
4517 struct address_space
*mapping
= page_mapping(page
);
4519 if (mapping_cap_account_dirty(mapping
)) {
4520 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4522 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4527 if (PageWriteback(page
)) {
4528 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4530 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4535 * It is safe to change page->mem_cgroup here because the page
4536 * is referenced, charged, and isolated - we can't race with
4537 * uncharging, charging, migration, or LRU putback.
4540 /* caller should have done css_get */
4541 page
->mem_cgroup
= to
;
4542 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4546 local_irq_disable();
4547 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
4548 memcg_check_events(to
, page
);
4549 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
4550 memcg_check_events(from
, page
);
4559 * get_mctgt_type - get target type of moving charge
4560 * @vma: the vma the pte to be checked belongs
4561 * @addr: the address corresponding to the pte to be checked
4562 * @ptent: the pte to be checked
4563 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4566 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4567 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4568 * move charge. if @target is not NULL, the page is stored in target->page
4569 * with extra refcnt got(Callers should handle it).
4570 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4571 * target for charge migration. if @target is not NULL, the entry is stored
4574 * Called with pte lock held.
4577 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4578 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4580 struct page
*page
= NULL
;
4581 enum mc_target_type ret
= MC_TARGET_NONE
;
4582 swp_entry_t ent
= { .val
= 0 };
4584 if (pte_present(ptent
))
4585 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4586 else if (is_swap_pte(ptent
))
4587 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
4588 else if (pte_none(ptent
))
4589 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4591 if (!page
&& !ent
.val
)
4595 * Do only loose check w/o serialization.
4596 * mem_cgroup_move_account() checks the page is valid or
4597 * not under LRU exclusion.
4599 if (page
->mem_cgroup
== mc
.from
) {
4600 ret
= MC_TARGET_PAGE
;
4602 target
->page
= page
;
4604 if (!ret
|| !target
)
4607 /* There is a swap entry and a page doesn't exist or isn't charged */
4608 if (ent
.val
&& !ret
&&
4609 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4610 ret
= MC_TARGET_SWAP
;
4617 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4619 * We don't consider swapping or file mapped pages because THP does not
4620 * support them for now.
4621 * Caller should make sure that pmd_trans_huge(pmd) is true.
4623 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4624 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4626 struct page
*page
= NULL
;
4627 enum mc_target_type ret
= MC_TARGET_NONE
;
4629 page
= pmd_page(pmd
);
4630 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4631 if (!(mc
.flags
& MOVE_ANON
))
4633 if (page
->mem_cgroup
== mc
.from
) {
4634 ret
= MC_TARGET_PAGE
;
4637 target
->page
= page
;
4643 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4644 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4646 return MC_TARGET_NONE
;
4650 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4651 unsigned long addr
, unsigned long end
,
4652 struct mm_walk
*walk
)
4654 struct vm_area_struct
*vma
= walk
->vma
;
4658 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4660 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4661 mc
.precharge
+= HPAGE_PMD_NR
;
4666 if (pmd_trans_unstable(pmd
))
4668 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4669 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4670 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4671 mc
.precharge
++; /* increment precharge temporarily */
4672 pte_unmap_unlock(pte
- 1, ptl
);
4678 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4680 unsigned long precharge
;
4682 struct mm_walk mem_cgroup_count_precharge_walk
= {
4683 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4686 down_read(&mm
->mmap_sem
);
4687 walk_page_range(0, mm
->highest_vm_end
,
4688 &mem_cgroup_count_precharge_walk
);
4689 up_read(&mm
->mmap_sem
);
4691 precharge
= mc
.precharge
;
4697 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4699 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4701 VM_BUG_ON(mc
.moving_task
);
4702 mc
.moving_task
= current
;
4703 return mem_cgroup_do_precharge(precharge
);
4706 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4707 static void __mem_cgroup_clear_mc(void)
4709 struct mem_cgroup
*from
= mc
.from
;
4710 struct mem_cgroup
*to
= mc
.to
;
4712 /* we must uncharge all the leftover precharges from mc.to */
4714 cancel_charge(mc
.to
, mc
.precharge
);
4718 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4719 * we must uncharge here.
4721 if (mc
.moved_charge
) {
4722 cancel_charge(mc
.from
, mc
.moved_charge
);
4723 mc
.moved_charge
= 0;
4725 /* we must fixup refcnts and charges */
4726 if (mc
.moved_swap
) {
4727 /* uncharge swap account from the old cgroup */
4728 if (!mem_cgroup_is_root(mc
.from
))
4729 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4731 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
4734 * we charged both to->memory and to->memsw, so we
4735 * should uncharge to->memory.
4737 if (!mem_cgroup_is_root(mc
.to
))
4738 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4740 mem_cgroup_id_get_many(mc
.to
, mc
.moved_swap
);
4741 css_put_many(&mc
.to
->css
, mc
.moved_swap
);
4745 memcg_oom_recover(from
);
4746 memcg_oom_recover(to
);
4747 wake_up_all(&mc
.waitq
);
4750 static void mem_cgroup_clear_mc(void)
4752 struct mm_struct
*mm
= mc
.mm
;
4755 * we must clear moving_task before waking up waiters at the end of
4758 mc
.moving_task
= NULL
;
4759 __mem_cgroup_clear_mc();
4760 spin_lock(&mc
.lock
);
4764 spin_unlock(&mc
.lock
);
4769 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4771 struct cgroup_subsys_state
*css
;
4772 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
4773 struct mem_cgroup
*from
;
4774 struct task_struct
*leader
, *p
;
4775 struct mm_struct
*mm
;
4776 unsigned long move_flags
;
4779 /* charge immigration isn't supported on the default hierarchy */
4780 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4784 * Multi-process migrations only happen on the default hierarchy
4785 * where charge immigration is not used. Perform charge
4786 * immigration if @tset contains a leader and whine if there are
4790 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
4793 memcg
= mem_cgroup_from_css(css
);
4799 * We are now commited to this value whatever it is. Changes in this
4800 * tunable will only affect upcoming migrations, not the current one.
4801 * So we need to save it, and keep it going.
4803 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
4807 from
= mem_cgroup_from_task(p
);
4809 VM_BUG_ON(from
== memcg
);
4811 mm
= get_task_mm(p
);
4814 /* We move charges only when we move a owner of the mm */
4815 if (mm
->owner
== p
) {
4818 VM_BUG_ON(mc
.precharge
);
4819 VM_BUG_ON(mc
.moved_charge
);
4820 VM_BUG_ON(mc
.moved_swap
);
4822 spin_lock(&mc
.lock
);
4826 mc
.flags
= move_flags
;
4827 spin_unlock(&mc
.lock
);
4828 /* We set mc.moving_task later */
4830 ret
= mem_cgroup_precharge_mc(mm
);
4832 mem_cgroup_clear_mc();
4839 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4842 mem_cgroup_clear_mc();
4845 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4846 unsigned long addr
, unsigned long end
,
4847 struct mm_walk
*walk
)
4850 struct vm_area_struct
*vma
= walk
->vma
;
4853 enum mc_target_type target_type
;
4854 union mc_target target
;
4857 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4859 if (mc
.precharge
< HPAGE_PMD_NR
) {
4863 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
4864 if (target_type
== MC_TARGET_PAGE
) {
4866 if (!isolate_lru_page(page
)) {
4867 if (!mem_cgroup_move_account(page
, true,
4869 mc
.precharge
-= HPAGE_PMD_NR
;
4870 mc
.moved_charge
+= HPAGE_PMD_NR
;
4872 putback_lru_page(page
);
4880 if (pmd_trans_unstable(pmd
))
4883 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4884 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4885 pte_t ptent
= *(pte
++);
4891 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
4892 case MC_TARGET_PAGE
:
4895 * We can have a part of the split pmd here. Moving it
4896 * can be done but it would be too convoluted so simply
4897 * ignore such a partial THP and keep it in original
4898 * memcg. There should be somebody mapping the head.
4900 if (PageTransCompound(page
))
4902 if (isolate_lru_page(page
))
4904 if (!mem_cgroup_move_account(page
, false,
4907 /* we uncharge from mc.from later. */
4910 putback_lru_page(page
);
4911 put
: /* get_mctgt_type() gets the page */
4914 case MC_TARGET_SWAP
:
4916 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
4918 /* we fixup refcnts and charges later. */
4926 pte_unmap_unlock(pte
- 1, ptl
);
4931 * We have consumed all precharges we got in can_attach().
4932 * We try charge one by one, but don't do any additional
4933 * charges to mc.to if we have failed in charge once in attach()
4936 ret
= mem_cgroup_do_precharge(1);
4944 static void mem_cgroup_move_charge(void)
4946 struct mm_walk mem_cgroup_move_charge_walk
= {
4947 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4951 lru_add_drain_all();
4953 * Signal lock_page_memcg() to take the memcg's move_lock
4954 * while we're moving its pages to another memcg. Then wait
4955 * for already started RCU-only updates to finish.
4957 atomic_inc(&mc
.from
->moving_account
);
4960 if (unlikely(!down_read_trylock(&mc
.mm
->mmap_sem
))) {
4962 * Someone who are holding the mmap_sem might be waiting in
4963 * waitq. So we cancel all extra charges, wake up all waiters,
4964 * and retry. Because we cancel precharges, we might not be able
4965 * to move enough charges, but moving charge is a best-effort
4966 * feature anyway, so it wouldn't be a big problem.
4968 __mem_cgroup_clear_mc();
4973 * When we have consumed all precharges and failed in doing
4974 * additional charge, the page walk just aborts.
4976 walk_page_range(0, mc
.mm
->highest_vm_end
, &mem_cgroup_move_charge_walk
);
4978 up_read(&mc
.mm
->mmap_sem
);
4979 atomic_dec(&mc
.from
->moving_account
);
4982 static void mem_cgroup_move_task(void)
4985 mem_cgroup_move_charge();
4986 mem_cgroup_clear_mc();
4989 #else /* !CONFIG_MMU */
4990 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4994 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4997 static void mem_cgroup_move_task(void)
5003 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5004 * to verify whether we're attached to the default hierarchy on each mount
5007 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5010 * use_hierarchy is forced on the default hierarchy. cgroup core
5011 * guarantees that @root doesn't have any children, so turning it
5012 * on for the root memcg is enough.
5014 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5015 root_mem_cgroup
->use_hierarchy
= true;
5017 root_mem_cgroup
->use_hierarchy
= false;
5020 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5023 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5025 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
5028 static int memory_low_show(struct seq_file
*m
, void *v
)
5030 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5031 unsigned long low
= READ_ONCE(memcg
->low
);
5033 if (low
== PAGE_COUNTER_MAX
)
5034 seq_puts(m
, "max\n");
5036 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
5041 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5042 char *buf
, size_t nbytes
, loff_t off
)
5044 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5048 buf
= strstrip(buf
);
5049 err
= page_counter_memparse(buf
, "max", &low
);
5058 static int memory_high_show(struct seq_file
*m
, void *v
)
5060 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5061 unsigned long high
= READ_ONCE(memcg
->high
);
5063 if (high
== PAGE_COUNTER_MAX
)
5064 seq_puts(m
, "max\n");
5066 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
5071 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5072 char *buf
, size_t nbytes
, loff_t off
)
5074 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5075 unsigned long nr_pages
;
5079 buf
= strstrip(buf
);
5080 err
= page_counter_memparse(buf
, "max", &high
);
5086 nr_pages
= page_counter_read(&memcg
->memory
);
5087 if (nr_pages
> high
)
5088 try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
5091 memcg_wb_domain_size_changed(memcg
);
5095 static int memory_max_show(struct seq_file
*m
, void *v
)
5097 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5098 unsigned long max
= READ_ONCE(memcg
->memory
.limit
);
5100 if (max
== PAGE_COUNTER_MAX
)
5101 seq_puts(m
, "max\n");
5103 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5108 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5109 char *buf
, size_t nbytes
, loff_t off
)
5111 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5112 unsigned int nr_reclaims
= MEM_CGROUP_RECLAIM_RETRIES
;
5113 bool drained
= false;
5117 buf
= strstrip(buf
);
5118 err
= page_counter_memparse(buf
, "max", &max
);
5122 xchg(&memcg
->memory
.limit
, max
);
5125 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
5127 if (nr_pages
<= max
)
5130 if (signal_pending(current
)) {
5136 drain_all_stock(memcg
);
5142 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
5148 mem_cgroup_events(memcg
, MEMCG_OOM
, 1);
5149 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
5153 memcg_wb_domain_size_changed(memcg
);
5157 static int memory_events_show(struct seq_file
*m
, void *v
)
5159 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5161 seq_printf(m
, "low %lu\n", mem_cgroup_read_events(memcg
, MEMCG_LOW
));
5162 seq_printf(m
, "high %lu\n", mem_cgroup_read_events(memcg
, MEMCG_HIGH
));
5163 seq_printf(m
, "max %lu\n", mem_cgroup_read_events(memcg
, MEMCG_MAX
));
5164 seq_printf(m
, "oom %lu\n", mem_cgroup_read_events(memcg
, MEMCG_OOM
));
5169 static int memory_stat_show(struct seq_file
*m
, void *v
)
5171 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5172 unsigned long stat
[MEMCG_NR_STAT
];
5173 unsigned long events
[MEMCG_NR_EVENTS
];
5177 * Provide statistics on the state of the memory subsystem as
5178 * well as cumulative event counters that show past behavior.
5180 * This list is ordered following a combination of these gradients:
5181 * 1) generic big picture -> specifics and details
5182 * 2) reflecting userspace activity -> reflecting kernel heuristics
5184 * Current memory state:
5187 tree_stat(memcg
, stat
);
5188 tree_events(memcg
, events
);
5190 seq_printf(m
, "anon %llu\n",
5191 (u64
)stat
[MEM_CGROUP_STAT_RSS
] * PAGE_SIZE
);
5192 seq_printf(m
, "file %llu\n",
5193 (u64
)stat
[MEM_CGROUP_STAT_CACHE
] * PAGE_SIZE
);
5194 seq_printf(m
, "kernel_stack %llu\n",
5195 (u64
)stat
[MEMCG_KERNEL_STACK_KB
] * 1024);
5196 seq_printf(m
, "slab %llu\n",
5197 (u64
)(stat
[MEMCG_SLAB_RECLAIMABLE
] +
5198 stat
[MEMCG_SLAB_UNRECLAIMABLE
]) * PAGE_SIZE
);
5199 seq_printf(m
, "sock %llu\n",
5200 (u64
)stat
[MEMCG_SOCK
] * PAGE_SIZE
);
5202 seq_printf(m
, "file_mapped %llu\n",
5203 (u64
)stat
[MEM_CGROUP_STAT_FILE_MAPPED
] * PAGE_SIZE
);
5204 seq_printf(m
, "file_dirty %llu\n",
5205 (u64
)stat
[MEM_CGROUP_STAT_DIRTY
] * PAGE_SIZE
);
5206 seq_printf(m
, "file_writeback %llu\n",
5207 (u64
)stat
[MEM_CGROUP_STAT_WRITEBACK
] * PAGE_SIZE
);
5209 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5210 struct mem_cgroup
*mi
;
5211 unsigned long val
= 0;
5213 for_each_mem_cgroup_tree(mi
, memcg
)
5214 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
));
5215 seq_printf(m
, "%s %llu\n",
5216 mem_cgroup_lru_names
[i
], (u64
)val
* PAGE_SIZE
);
5219 seq_printf(m
, "slab_reclaimable %llu\n",
5220 (u64
)stat
[MEMCG_SLAB_RECLAIMABLE
] * PAGE_SIZE
);
5221 seq_printf(m
, "slab_unreclaimable %llu\n",
5222 (u64
)stat
[MEMCG_SLAB_UNRECLAIMABLE
] * PAGE_SIZE
);
5224 /* Accumulated memory events */
5226 seq_printf(m
, "pgfault %lu\n",
5227 events
[MEM_CGROUP_EVENTS_PGFAULT
]);
5228 seq_printf(m
, "pgmajfault %lu\n",
5229 events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
5234 static struct cftype memory_files
[] = {
5237 .flags
= CFTYPE_NOT_ON_ROOT
,
5238 .read_u64
= memory_current_read
,
5242 .flags
= CFTYPE_NOT_ON_ROOT
,
5243 .seq_show
= memory_low_show
,
5244 .write
= memory_low_write
,
5248 .flags
= CFTYPE_NOT_ON_ROOT
,
5249 .seq_show
= memory_high_show
,
5250 .write
= memory_high_write
,
5254 .flags
= CFTYPE_NOT_ON_ROOT
,
5255 .seq_show
= memory_max_show
,
5256 .write
= memory_max_write
,
5260 .flags
= CFTYPE_NOT_ON_ROOT
,
5261 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
5262 .seq_show
= memory_events_show
,
5266 .flags
= CFTYPE_NOT_ON_ROOT
,
5267 .seq_show
= memory_stat_show
,
5272 struct cgroup_subsys memory_cgrp_subsys
= {
5273 .css_alloc
= mem_cgroup_css_alloc
,
5274 .css_online
= mem_cgroup_css_online
,
5275 .css_offline
= mem_cgroup_css_offline
,
5276 .css_released
= mem_cgroup_css_released
,
5277 .css_free
= mem_cgroup_css_free
,
5278 .css_reset
= mem_cgroup_css_reset
,
5279 .can_attach
= mem_cgroup_can_attach
,
5280 .cancel_attach
= mem_cgroup_cancel_attach
,
5281 .post_attach
= mem_cgroup_move_task
,
5282 .bind
= mem_cgroup_bind
,
5283 .dfl_cftypes
= memory_files
,
5284 .legacy_cftypes
= mem_cgroup_legacy_files
,
5289 * mem_cgroup_low - check if memory consumption is below the normal range
5290 * @root: the highest ancestor to consider
5291 * @memcg: the memory cgroup to check
5293 * Returns %true if memory consumption of @memcg, and that of all
5294 * configurable ancestors up to @root, is below the normal range.
5296 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5298 if (mem_cgroup_disabled())
5302 * The toplevel group doesn't have a configurable range, so
5303 * it's never low when looked at directly, and it is not
5304 * considered an ancestor when assessing the hierarchy.
5307 if (memcg
== root_mem_cgroup
)
5310 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5313 while (memcg
!= root
) {
5314 memcg
= parent_mem_cgroup(memcg
);
5316 if (memcg
== root_mem_cgroup
)
5319 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5326 * mem_cgroup_try_charge - try charging a page
5327 * @page: page to charge
5328 * @mm: mm context of the victim
5329 * @gfp_mask: reclaim mode
5330 * @memcgp: charged memcg return
5331 * @compound: charge the page as compound or small page
5333 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5334 * pages according to @gfp_mask if necessary.
5336 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5337 * Otherwise, an error code is returned.
5339 * After page->mapping has been set up, the caller must finalize the
5340 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5341 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5343 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5344 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
5347 struct mem_cgroup
*memcg
= NULL
;
5348 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5351 if (mem_cgroup_disabled())
5354 if (PageSwapCache(page
)) {
5356 * Every swap fault against a single page tries to charge the
5357 * page, bail as early as possible. shmem_unuse() encounters
5358 * already charged pages, too. The USED bit is protected by
5359 * the page lock, which serializes swap cache removal, which
5360 * in turn serializes uncharging.
5362 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5363 if (page
->mem_cgroup
)
5366 if (do_swap_account
) {
5367 swp_entry_t ent
= { .val
= page_private(page
), };
5368 unsigned short id
= lookup_swap_cgroup_id(ent
);
5371 memcg
= mem_cgroup_from_id(id
);
5372 if (memcg
&& !css_tryget_online(&memcg
->css
))
5379 memcg
= get_mem_cgroup_from_mm(mm
);
5381 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5383 css_put(&memcg
->css
);
5390 * mem_cgroup_commit_charge - commit a page charge
5391 * @page: page to charge
5392 * @memcg: memcg to charge the page to
5393 * @lrucare: page might be on LRU already
5394 * @compound: charge the page as compound or small page
5396 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5397 * after page->mapping has been set up. This must happen atomically
5398 * as part of the page instantiation, i.e. under the page table lock
5399 * for anonymous pages, under the page lock for page and swap cache.
5401 * In addition, the page must not be on the LRU during the commit, to
5402 * prevent racing with task migration. If it might be, use @lrucare.
5404 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5406 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5407 bool lrucare
, bool compound
)
5409 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5411 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5412 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5414 if (mem_cgroup_disabled())
5417 * Swap faults will attempt to charge the same page multiple
5418 * times. But reuse_swap_page() might have removed the page
5419 * from swapcache already, so we can't check PageSwapCache().
5424 commit_charge(page
, memcg
, lrucare
);
5426 local_irq_disable();
5427 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
5428 memcg_check_events(memcg
, page
);
5431 if (do_memsw_account() && PageSwapCache(page
)) {
5432 swp_entry_t entry
= { .val
= page_private(page
) };
5434 * The swap entry might not get freed for a long time,
5435 * let's not wait for it. The page already received a
5436 * memory+swap charge, drop the swap entry duplicate.
5438 mem_cgroup_uncharge_swap(entry
);
5443 * mem_cgroup_cancel_charge - cancel a page charge
5444 * @page: page to charge
5445 * @memcg: memcg to charge the page to
5446 * @compound: charge the page as compound or small page
5448 * Cancel a charge transaction started by mem_cgroup_try_charge().
5450 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5453 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5455 if (mem_cgroup_disabled())
5458 * Swap faults will attempt to charge the same page multiple
5459 * times. But reuse_swap_page() might have removed the page
5460 * from swapcache already, so we can't check PageSwapCache().
5465 cancel_charge(memcg
, nr_pages
);
5468 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5469 unsigned long nr_anon
, unsigned long nr_file
,
5470 unsigned long nr_huge
, unsigned long nr_kmem
,
5471 struct page
*dummy_page
)
5473 unsigned long nr_pages
= nr_anon
+ nr_file
+ nr_kmem
;
5474 unsigned long flags
;
5476 if (!mem_cgroup_is_root(memcg
)) {
5477 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5478 if (do_memsw_account())
5479 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5480 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && nr_kmem
)
5481 page_counter_uncharge(&memcg
->kmem
, nr_kmem
);
5482 memcg_oom_recover(memcg
);
5485 local_irq_save(flags
);
5486 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5487 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5488 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5489 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5490 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5491 memcg_check_events(memcg
, dummy_page
);
5492 local_irq_restore(flags
);
5494 if (!mem_cgroup_is_root(memcg
))
5495 css_put_many(&memcg
->css
, nr_pages
);
5498 static void uncharge_list(struct list_head
*page_list
)
5500 struct mem_cgroup
*memcg
= NULL
;
5501 unsigned long nr_anon
= 0;
5502 unsigned long nr_file
= 0;
5503 unsigned long nr_huge
= 0;
5504 unsigned long nr_kmem
= 0;
5505 unsigned long pgpgout
= 0;
5506 struct list_head
*next
;
5510 * Note that the list can be a single page->lru; hence the
5511 * do-while loop instead of a simple list_for_each_entry().
5513 next
= page_list
->next
;
5515 page
= list_entry(next
, struct page
, lru
);
5516 next
= page
->lru
.next
;
5518 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5519 VM_BUG_ON_PAGE(page_count(page
), page
);
5521 if (!page
->mem_cgroup
)
5525 * Nobody should be changing or seriously looking at
5526 * page->mem_cgroup at this point, we have fully
5527 * exclusive access to the page.
5530 if (memcg
!= page
->mem_cgroup
) {
5532 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5533 nr_huge
, nr_kmem
, page
);
5534 pgpgout
= nr_anon
= nr_file
=
5535 nr_huge
= nr_kmem
= 0;
5537 memcg
= page
->mem_cgroup
;
5540 if (!PageKmemcg(page
)) {
5541 unsigned int nr_pages
= 1;
5543 if (PageTransHuge(page
)) {
5544 nr_pages
<<= compound_order(page
);
5545 nr_huge
+= nr_pages
;
5548 nr_anon
+= nr_pages
;
5550 nr_file
+= nr_pages
;
5553 nr_kmem
+= 1 << compound_order(page
);
5554 __ClearPageKmemcg(page
);
5557 page
->mem_cgroup
= NULL
;
5558 } while (next
!= page_list
);
5561 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5562 nr_huge
, nr_kmem
, page
);
5566 * mem_cgroup_uncharge - uncharge a page
5567 * @page: page to uncharge
5569 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5570 * mem_cgroup_commit_charge().
5572 void mem_cgroup_uncharge(struct page
*page
)
5574 if (mem_cgroup_disabled())
5577 /* Don't touch page->lru of any random page, pre-check: */
5578 if (!page
->mem_cgroup
)
5581 INIT_LIST_HEAD(&page
->lru
);
5582 uncharge_list(&page
->lru
);
5586 * mem_cgroup_uncharge_list - uncharge a list of page
5587 * @page_list: list of pages to uncharge
5589 * Uncharge a list of pages previously charged with
5590 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5592 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5594 if (mem_cgroup_disabled())
5597 if (!list_empty(page_list
))
5598 uncharge_list(page_list
);
5602 * mem_cgroup_migrate - charge a page's replacement
5603 * @oldpage: currently circulating page
5604 * @newpage: replacement page
5606 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5607 * be uncharged upon free.
5609 * Both pages must be locked, @newpage->mapping must be set up.
5611 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
5613 struct mem_cgroup
*memcg
;
5614 unsigned int nr_pages
;
5616 unsigned long flags
;
5618 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5619 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5620 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5621 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5624 if (mem_cgroup_disabled())
5627 /* Page cache replacement: new page already charged? */
5628 if (newpage
->mem_cgroup
)
5631 /* Swapcache readahead pages can get replaced before being charged */
5632 memcg
= oldpage
->mem_cgroup
;
5636 /* Force-charge the new page. The old one will be freed soon */
5637 compound
= PageTransHuge(newpage
);
5638 nr_pages
= compound
? hpage_nr_pages(newpage
) : 1;
5640 page_counter_charge(&memcg
->memory
, nr_pages
);
5641 if (do_memsw_account())
5642 page_counter_charge(&memcg
->memsw
, nr_pages
);
5643 css_get_many(&memcg
->css
, nr_pages
);
5645 commit_charge(newpage
, memcg
, false);
5647 local_irq_save(flags
);
5648 mem_cgroup_charge_statistics(memcg
, newpage
, compound
, nr_pages
);
5649 memcg_check_events(memcg
, newpage
);
5650 local_irq_restore(flags
);
5653 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
5654 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
5656 void mem_cgroup_sk_alloc(struct sock
*sk
)
5658 struct mem_cgroup
*memcg
;
5660 if (!mem_cgroup_sockets_enabled
)
5664 * Socket cloning can throw us here with sk_memcg already
5665 * filled. It won't however, necessarily happen from
5666 * process context. So the test for root memcg given
5667 * the current task's memcg won't help us in this case.
5669 * Respecting the original socket's memcg is a better
5670 * decision in this case.
5673 BUG_ON(mem_cgroup_is_root(sk
->sk_memcg
));
5674 css_get(&sk
->sk_memcg
->css
);
5679 memcg
= mem_cgroup_from_task(current
);
5680 if (memcg
== root_mem_cgroup
)
5682 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
5684 if (css_tryget_online(&memcg
->css
))
5685 sk
->sk_memcg
= memcg
;
5690 void mem_cgroup_sk_free(struct sock
*sk
)
5693 css_put(&sk
->sk_memcg
->css
);
5697 * mem_cgroup_charge_skmem - charge socket memory
5698 * @memcg: memcg to charge
5699 * @nr_pages: number of pages to charge
5701 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5702 * @memcg's configured limit, %false if the charge had to be forced.
5704 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5706 gfp_t gfp_mask
= GFP_KERNEL
;
5708 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5709 struct page_counter
*fail
;
5711 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
5712 memcg
->tcpmem_pressure
= 0;
5715 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
5716 memcg
->tcpmem_pressure
= 1;
5720 /* Don't block in the packet receive path */
5722 gfp_mask
= GFP_NOWAIT
;
5724 this_cpu_add(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5726 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
5729 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
5734 * mem_cgroup_uncharge_skmem - uncharge socket memory
5735 * @memcg - memcg to uncharge
5736 * @nr_pages - number of pages to uncharge
5738 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5740 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5741 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
5745 this_cpu_sub(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5747 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5748 css_put_many(&memcg
->css
, nr_pages
);
5751 static int __init
cgroup_memory(char *s
)
5755 while ((token
= strsep(&s
, ",")) != NULL
) {
5758 if (!strcmp(token
, "nosocket"))
5759 cgroup_memory_nosocket
= true;
5760 if (!strcmp(token
, "nokmem"))
5761 cgroup_memory_nokmem
= true;
5765 __setup("cgroup.memory=", cgroup_memory
);
5768 * subsys_initcall() for memory controller.
5770 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
5771 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
5772 * basically everything that doesn't depend on a specific mem_cgroup structure
5773 * should be initialized from here.
5775 static int __init
mem_cgroup_init(void)
5781 * Kmem cache creation is mostly done with the slab_mutex held,
5782 * so use a special workqueue to avoid stalling all worker
5783 * threads in case lots of cgroups are created simultaneously.
5785 memcg_kmem_cache_create_wq
=
5786 alloc_ordered_workqueue("memcg_kmem_cache_create", 0);
5787 BUG_ON(!memcg_kmem_cache_create_wq
);
5790 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD
, "mm/memctrl:dead", NULL
,
5791 memcg_hotplug_cpu_dead
);
5793 for_each_possible_cpu(cpu
)
5794 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5797 for_each_node(node
) {
5798 struct mem_cgroup_tree_per_node
*rtpn
;
5800 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5801 node_online(node
) ? node
: NUMA_NO_NODE
);
5803 rtpn
->rb_root
= RB_ROOT
;
5804 spin_lock_init(&rtpn
->lock
);
5805 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5810 subsys_initcall(mem_cgroup_init
);
5812 #ifdef CONFIG_MEMCG_SWAP
5813 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
5815 while (!atomic_inc_not_zero(&memcg
->id
.ref
)) {
5817 * The root cgroup cannot be destroyed, so it's refcount must
5820 if (WARN_ON_ONCE(memcg
== root_mem_cgroup
)) {
5824 memcg
= parent_mem_cgroup(memcg
);
5826 memcg
= root_mem_cgroup
;
5832 * mem_cgroup_swapout - transfer a memsw charge to swap
5833 * @page: page whose memsw charge to transfer
5834 * @entry: swap entry to move the charge to
5836 * Transfer the memsw charge of @page to @entry.
5838 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5840 struct mem_cgroup
*memcg
, *swap_memcg
;
5841 unsigned short oldid
;
5843 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5844 VM_BUG_ON_PAGE(page_count(page
), page
);
5846 if (!do_memsw_account())
5849 memcg
= page
->mem_cgroup
;
5851 /* Readahead page, never charged */
5856 * In case the memcg owning these pages has been offlined and doesn't
5857 * have an ID allocated to it anymore, charge the closest online
5858 * ancestor for the swap instead and transfer the memory+swap charge.
5860 swap_memcg
= mem_cgroup_id_get_online(memcg
);
5861 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
));
5862 VM_BUG_ON_PAGE(oldid
, page
);
5863 mem_cgroup_swap_statistics(swap_memcg
, true);
5865 page
->mem_cgroup
= NULL
;
5867 if (!mem_cgroup_is_root(memcg
))
5868 page_counter_uncharge(&memcg
->memory
, 1);
5870 if (memcg
!= swap_memcg
) {
5871 if (!mem_cgroup_is_root(swap_memcg
))
5872 page_counter_charge(&swap_memcg
->memsw
, 1);
5873 page_counter_uncharge(&memcg
->memsw
, 1);
5877 * Interrupts should be disabled here because the caller holds the
5878 * mapping->tree_lock lock which is taken with interrupts-off. It is
5879 * important here to have the interrupts disabled because it is the
5880 * only synchronisation we have for udpating the per-CPU variables.
5882 VM_BUG_ON(!irqs_disabled());
5883 mem_cgroup_charge_statistics(memcg
, page
, false, -1);
5884 memcg_check_events(memcg
, page
);
5886 if (!mem_cgroup_is_root(memcg
))
5887 css_put(&memcg
->css
);
5891 * mem_cgroup_try_charge_swap - try charging a swap entry
5892 * @page: page being added to swap
5893 * @entry: swap entry to charge
5895 * Try to charge @entry to the memcg that @page belongs to.
5897 * Returns 0 on success, -ENOMEM on failure.
5899 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
5901 struct mem_cgroup
*memcg
;
5902 struct page_counter
*counter
;
5903 unsigned short oldid
;
5905 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
5908 memcg
= page
->mem_cgroup
;
5910 /* Readahead page, never charged */
5914 memcg
= mem_cgroup_id_get_online(memcg
);
5916 if (!mem_cgroup_is_root(memcg
) &&
5917 !page_counter_try_charge(&memcg
->swap
, 1, &counter
)) {
5918 mem_cgroup_id_put(memcg
);
5922 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5923 VM_BUG_ON_PAGE(oldid
, page
);
5924 mem_cgroup_swap_statistics(memcg
, true);
5930 * mem_cgroup_uncharge_swap - uncharge a swap entry
5931 * @entry: swap entry to uncharge
5933 * Drop the swap charge associated with @entry.
5935 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5937 struct mem_cgroup
*memcg
;
5940 if (!do_swap_account
)
5943 id
= swap_cgroup_record(entry
, 0);
5945 memcg
= mem_cgroup_from_id(id
);
5947 if (!mem_cgroup_is_root(memcg
)) {
5948 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5949 page_counter_uncharge(&memcg
->swap
, 1);
5951 page_counter_uncharge(&memcg
->memsw
, 1);
5953 mem_cgroup_swap_statistics(memcg
, false);
5954 mem_cgroup_id_put(memcg
);
5959 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
5961 long nr_swap_pages
= get_nr_swap_pages();
5963 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5964 return nr_swap_pages
;
5965 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5966 nr_swap_pages
= min_t(long, nr_swap_pages
,
5967 READ_ONCE(memcg
->swap
.limit
) -
5968 page_counter_read(&memcg
->swap
));
5969 return nr_swap_pages
;
5972 bool mem_cgroup_swap_full(struct page
*page
)
5974 struct mem_cgroup
*memcg
;
5976 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5980 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5983 memcg
= page
->mem_cgroup
;
5987 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5988 if (page_counter_read(&memcg
->swap
) * 2 >= memcg
->swap
.limit
)
5994 /* for remember boot option*/
5995 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5996 static int really_do_swap_account __initdata
= 1;
5998 static int really_do_swap_account __initdata
;
6001 static int __init
enable_swap_account(char *s
)
6003 if (!strcmp(s
, "1"))
6004 really_do_swap_account
= 1;
6005 else if (!strcmp(s
, "0"))
6006 really_do_swap_account
= 0;
6009 __setup("swapaccount=", enable_swap_account
);
6011 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
6014 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6016 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
6019 static int swap_max_show(struct seq_file
*m
, void *v
)
6021 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
6022 unsigned long max
= READ_ONCE(memcg
->swap
.limit
);
6024 if (max
== PAGE_COUNTER_MAX
)
6025 seq_puts(m
, "max\n");
6027 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
6032 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
6033 char *buf
, size_t nbytes
, loff_t off
)
6035 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6039 buf
= strstrip(buf
);
6040 err
= page_counter_memparse(buf
, "max", &max
);
6044 mutex_lock(&memcg_limit_mutex
);
6045 err
= page_counter_limit(&memcg
->swap
, max
);
6046 mutex_unlock(&memcg_limit_mutex
);
6053 static struct cftype swap_files
[] = {
6055 .name
= "swap.current",
6056 .flags
= CFTYPE_NOT_ON_ROOT
,
6057 .read_u64
= swap_current_read
,
6061 .flags
= CFTYPE_NOT_ON_ROOT
,
6062 .seq_show
= swap_max_show
,
6063 .write
= swap_max_write
,
6068 static struct cftype memsw_cgroup_files
[] = {
6070 .name
= "memsw.usage_in_bytes",
6071 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6072 .read_u64
= mem_cgroup_read_u64
,
6075 .name
= "memsw.max_usage_in_bytes",
6076 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6077 .write
= mem_cgroup_reset
,
6078 .read_u64
= mem_cgroup_read_u64
,
6081 .name
= "memsw.limit_in_bytes",
6082 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6083 .write
= mem_cgroup_write
,
6084 .read_u64
= mem_cgroup_read_u64
,
6087 .name
= "memsw.failcnt",
6088 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6089 .write
= mem_cgroup_reset
,
6090 .read_u64
= mem_cgroup_read_u64
,
6092 { }, /* terminate */
6095 static int __init
mem_cgroup_swap_init(void)
6097 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6098 do_swap_account
= 1;
6099 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
6101 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
6102 memsw_cgroup_files
));
6106 subsys_initcall(mem_cgroup_swap_init
);
6108 #endif /* CONFIG_MEMCG_SWAP */