1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
65 #include <linux/tracehook.h>
71 #include <asm/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
76 EXPORT_SYMBOL(memory_cgrp_subsys
);
78 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
80 #define MEM_CGROUP_RECLAIM_RETRIES 5
82 /* Socket memory accounting disabled? */
83 static bool cgroup_memory_nosocket
;
85 /* Kernel memory accounting disabled? */
86 static bool cgroup_memory_nokmem
;
88 /* Whether the swap controller is active */
89 #ifdef CONFIG_MEMCG_SWAP
90 int do_swap_account __read_mostly
;
92 #define do_swap_account 0
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys
) && do_swap_account
;
101 static const char * const mem_cgroup_stat_names
[] = {
111 static const char * const mem_cgroup_events_names
[] = {
118 static const char * const mem_cgroup_lru_names
[] = {
126 #define THRESHOLDS_EVENTS_TARGET 128
127 #define SOFTLIMIT_EVENTS_TARGET 1024
128 #define NUMAINFO_EVENTS_TARGET 1024
131 * Cgroups above their limits are maintained in a RB-Tree, independent of
132 * their hierarchy representation
135 struct mem_cgroup_tree_per_zone
{
136 struct rb_root rb_root
;
140 struct mem_cgroup_tree_per_node
{
141 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
144 struct mem_cgroup_tree
{
145 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
148 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
151 struct mem_cgroup_eventfd_list
{
152 struct list_head list
;
153 struct eventfd_ctx
*eventfd
;
157 * cgroup_event represents events which userspace want to receive.
159 struct mem_cgroup_event
{
161 * memcg which the event belongs to.
163 struct mem_cgroup
*memcg
;
165 * eventfd to signal userspace about the event.
167 struct eventfd_ctx
*eventfd
;
169 * Each of these stored in a list by the cgroup.
171 struct list_head list
;
173 * register_event() callback will be used to add new userspace
174 * waiter for changes related to this event. Use eventfd_signal()
175 * on eventfd to send notification to userspace.
177 int (*register_event
)(struct mem_cgroup
*memcg
,
178 struct eventfd_ctx
*eventfd
, const char *args
);
180 * unregister_event() callback will be called when userspace closes
181 * the eventfd or on cgroup removing. This callback must be set,
182 * if you want provide notification functionality.
184 void (*unregister_event
)(struct mem_cgroup
*memcg
,
185 struct eventfd_ctx
*eventfd
);
187 * All fields below needed to unregister event when
188 * userspace closes eventfd.
191 wait_queue_head_t
*wqh
;
193 struct work_struct remove
;
196 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
197 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
199 /* Stuffs for move charges at task migration. */
201 * Types of charges to be moved.
203 #define MOVE_ANON 0x1U
204 #define MOVE_FILE 0x2U
205 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
207 /* "mc" and its members are protected by cgroup_mutex */
208 static struct move_charge_struct
{
209 spinlock_t lock
; /* for from, to */
210 struct mm_struct
*mm
;
211 struct mem_cgroup
*from
;
212 struct mem_cgroup
*to
;
214 unsigned long precharge
;
215 unsigned long moved_charge
;
216 unsigned long moved_swap
;
217 struct task_struct
*moving_task
; /* a task moving charges */
218 wait_queue_head_t waitq
; /* a waitq for other context */
220 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
221 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
225 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
226 * limit reclaim to prevent infinite loops, if they ever occur.
228 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
229 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
232 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
233 MEM_CGROUP_CHARGE_TYPE_ANON
,
234 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
235 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
239 /* for encoding cft->private value on file */
248 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
249 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
250 #define MEMFILE_ATTR(val) ((val) & 0xffff)
251 /* Used for OOM nofiier */
252 #define OOM_CONTROL (0)
254 /* Some nice accessors for the vmpressure. */
255 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
258 memcg
= root_mem_cgroup
;
259 return &memcg
->vmpressure
;
262 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
264 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
267 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
269 return (memcg
== root_mem_cgroup
);
274 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
275 * The main reason for not using cgroup id for this:
276 * this works better in sparse environments, where we have a lot of memcgs,
277 * but only a few kmem-limited. Or also, if we have, for instance, 200
278 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
279 * 200 entry array for that.
281 * The current size of the caches array is stored in memcg_nr_cache_ids. It
282 * will double each time we have to increase it.
284 static DEFINE_IDA(memcg_cache_ida
);
285 int memcg_nr_cache_ids
;
287 /* Protects memcg_nr_cache_ids */
288 static DECLARE_RWSEM(memcg_cache_ids_sem
);
290 void memcg_get_cache_ids(void)
292 down_read(&memcg_cache_ids_sem
);
295 void memcg_put_cache_ids(void)
297 up_read(&memcg_cache_ids_sem
);
301 * MIN_SIZE is different than 1, because we would like to avoid going through
302 * the alloc/free process all the time. In a small machine, 4 kmem-limited
303 * cgroups is a reasonable guess. In the future, it could be a parameter or
304 * tunable, but that is strictly not necessary.
306 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
307 * this constant directly from cgroup, but it is understandable that this is
308 * better kept as an internal representation in cgroup.c. In any case, the
309 * cgrp_id space is not getting any smaller, and we don't have to necessarily
310 * increase ours as well if it increases.
312 #define MEMCG_CACHES_MIN_SIZE 4
313 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
316 * A lot of the calls to the cache allocation functions are expected to be
317 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
318 * conditional to this static branch, we'll have to allow modules that does
319 * kmem_cache_alloc and the such to see this symbol as well
321 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
322 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
324 #endif /* !CONFIG_SLOB */
327 * mem_cgroup_css_from_page - css of the memcg associated with a page
328 * @page: page of interest
330 * If memcg is bound to the default hierarchy, css of the memcg associated
331 * with @page is returned. The returned css remains associated with @page
332 * until it is released.
334 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
337 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
339 struct mem_cgroup
*memcg
;
341 memcg
= page
->mem_cgroup
;
343 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
344 memcg
= root_mem_cgroup
;
350 * page_cgroup_ino - return inode number of the memcg a page is charged to
353 * Look up the closest online ancestor of the memory cgroup @page is charged to
354 * and return its inode number or 0 if @page is not charged to any cgroup. It
355 * is safe to call this function without holding a reference to @page.
357 * Note, this function is inherently racy, because there is nothing to prevent
358 * the cgroup inode from getting torn down and potentially reallocated a moment
359 * after page_cgroup_ino() returns, so it only should be used by callers that
360 * do not care (such as procfs interfaces).
362 ino_t
page_cgroup_ino(struct page
*page
)
364 struct mem_cgroup
*memcg
;
365 unsigned long ino
= 0;
368 memcg
= READ_ONCE(page
->mem_cgroup
);
369 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
370 memcg
= parent_mem_cgroup(memcg
);
372 ino
= cgroup_ino(memcg
->css
.cgroup
);
377 static struct mem_cgroup_per_zone
*
378 mem_cgroup_page_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
380 int nid
= page_to_nid(page
);
381 int zid
= page_zonenum(page
);
383 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
386 static struct mem_cgroup_tree_per_zone
*
387 soft_limit_tree_node_zone(int nid
, int zid
)
389 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
392 static struct mem_cgroup_tree_per_zone
*
393 soft_limit_tree_from_page(struct page
*page
)
395 int nid
= page_to_nid(page
);
396 int zid
= page_zonenum(page
);
398 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
401 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone
*mz
,
402 struct mem_cgroup_tree_per_zone
*mctz
,
403 unsigned long new_usage_in_excess
)
405 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
406 struct rb_node
*parent
= NULL
;
407 struct mem_cgroup_per_zone
*mz_node
;
412 mz
->usage_in_excess
= new_usage_in_excess
;
413 if (!mz
->usage_in_excess
)
417 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
419 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
422 * We can't avoid mem cgroups that are over their soft
423 * limit by the same amount
425 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
428 rb_link_node(&mz
->tree_node
, parent
, p
);
429 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
433 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
434 struct mem_cgroup_tree_per_zone
*mctz
)
438 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
442 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
443 struct mem_cgroup_tree_per_zone
*mctz
)
447 spin_lock_irqsave(&mctz
->lock
, flags
);
448 __mem_cgroup_remove_exceeded(mz
, mctz
);
449 spin_unlock_irqrestore(&mctz
->lock
, flags
);
452 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
454 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
455 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
456 unsigned long excess
= 0;
458 if (nr_pages
> soft_limit
)
459 excess
= nr_pages
- soft_limit
;
464 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
466 unsigned long excess
;
467 struct mem_cgroup_per_zone
*mz
;
468 struct mem_cgroup_tree_per_zone
*mctz
;
470 mctz
= soft_limit_tree_from_page(page
);
472 * Necessary to update all ancestors when hierarchy is used.
473 * because their event counter is not touched.
475 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
476 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
477 excess
= soft_limit_excess(memcg
);
479 * We have to update the tree if mz is on RB-tree or
480 * mem is over its softlimit.
482 if (excess
|| mz
->on_tree
) {
485 spin_lock_irqsave(&mctz
->lock
, flags
);
486 /* if on-tree, remove it */
488 __mem_cgroup_remove_exceeded(mz
, mctz
);
490 * Insert again. mz->usage_in_excess will be updated.
491 * If excess is 0, no tree ops.
493 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
494 spin_unlock_irqrestore(&mctz
->lock
, flags
);
499 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
501 struct mem_cgroup_tree_per_zone
*mctz
;
502 struct mem_cgroup_per_zone
*mz
;
506 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
507 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
508 mctz
= soft_limit_tree_node_zone(nid
, zid
);
509 mem_cgroup_remove_exceeded(mz
, mctz
);
514 static struct mem_cgroup_per_zone
*
515 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
517 struct rb_node
*rightmost
= NULL
;
518 struct mem_cgroup_per_zone
*mz
;
522 rightmost
= rb_last(&mctz
->rb_root
);
524 goto done
; /* Nothing to reclaim from */
526 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
528 * Remove the node now but someone else can add it back,
529 * we will to add it back at the end of reclaim to its correct
530 * position in the tree.
532 __mem_cgroup_remove_exceeded(mz
, mctz
);
533 if (!soft_limit_excess(mz
->memcg
) ||
534 !css_tryget_online(&mz
->memcg
->css
))
540 static struct mem_cgroup_per_zone
*
541 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
543 struct mem_cgroup_per_zone
*mz
;
545 spin_lock_irq(&mctz
->lock
);
546 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
547 spin_unlock_irq(&mctz
->lock
);
552 * Return page count for single (non recursive) @memcg.
554 * Implementation Note: reading percpu statistics for memcg.
556 * Both of vmstat[] and percpu_counter has threshold and do periodic
557 * synchronization to implement "quick" read. There are trade-off between
558 * reading cost and precision of value. Then, we may have a chance to implement
559 * a periodic synchronization of counter in memcg's counter.
561 * But this _read() function is used for user interface now. The user accounts
562 * memory usage by memory cgroup and he _always_ requires exact value because
563 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
564 * have to visit all online cpus and make sum. So, for now, unnecessary
565 * synchronization is not implemented. (just implemented for cpu hotplug)
567 * If there are kernel internal actions which can make use of some not-exact
568 * value, and reading all cpu value can be performance bottleneck in some
569 * common workload, threshold and synchronization as vmstat[] should be
573 mem_cgroup_read_stat(struct mem_cgroup
*memcg
, enum mem_cgroup_stat_index idx
)
578 /* Per-cpu values can be negative, use a signed accumulator */
579 for_each_possible_cpu(cpu
)
580 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
582 * Summing races with updates, so val may be negative. Avoid exposing
583 * transient negative values.
590 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
591 enum mem_cgroup_events_index idx
)
593 unsigned long val
= 0;
596 for_each_possible_cpu(cpu
)
597 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
601 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
603 bool compound
, int nr_pages
)
606 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
607 * counted as CACHE even if it's on ANON LRU.
610 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
613 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
617 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
618 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
622 /* pagein of a big page is an event. So, ignore page size */
624 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
626 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
627 nr_pages
= -nr_pages
; /* for event */
630 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
633 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
634 int nid
, unsigned int lru_mask
)
636 unsigned long nr
= 0;
639 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
641 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
642 struct mem_cgroup_per_zone
*mz
;
646 if (!(BIT(lru
) & lru_mask
))
648 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
649 nr
+= mz
->lru_size
[lru
];
655 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
656 unsigned int lru_mask
)
658 unsigned long nr
= 0;
661 for_each_node_state(nid
, N_MEMORY
)
662 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
666 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
667 enum mem_cgroup_events_target target
)
669 unsigned long val
, next
;
671 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
672 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
673 /* from time_after() in jiffies.h */
674 if ((long)next
- (long)val
< 0) {
676 case MEM_CGROUP_TARGET_THRESH
:
677 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
679 case MEM_CGROUP_TARGET_SOFTLIMIT
:
680 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
682 case MEM_CGROUP_TARGET_NUMAINFO
:
683 next
= val
+ NUMAINFO_EVENTS_TARGET
;
688 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
695 * Check events in order.
698 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
700 /* threshold event is triggered in finer grain than soft limit */
701 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
702 MEM_CGROUP_TARGET_THRESH
))) {
704 bool do_numainfo __maybe_unused
;
706 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
707 MEM_CGROUP_TARGET_SOFTLIMIT
);
709 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
710 MEM_CGROUP_TARGET_NUMAINFO
);
712 mem_cgroup_threshold(memcg
);
713 if (unlikely(do_softlimit
))
714 mem_cgroup_update_tree(memcg
, page
);
716 if (unlikely(do_numainfo
))
717 atomic_inc(&memcg
->numainfo_events
);
722 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
725 * mm_update_next_owner() may clear mm->owner to NULL
726 * if it races with swapoff, page migration, etc.
727 * So this can be called with p == NULL.
732 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
734 EXPORT_SYMBOL(mem_cgroup_from_task
);
736 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
738 struct mem_cgroup
*memcg
= NULL
;
743 * Page cache insertions can happen withou an
744 * actual mm context, e.g. during disk probing
745 * on boot, loopback IO, acct() writes etc.
748 memcg
= root_mem_cgroup
;
750 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
751 if (unlikely(!memcg
))
752 memcg
= root_mem_cgroup
;
754 } while (!css_tryget_online(&memcg
->css
));
760 * mem_cgroup_iter - iterate over memory cgroup hierarchy
761 * @root: hierarchy root
762 * @prev: previously returned memcg, NULL on first invocation
763 * @reclaim: cookie for shared reclaim walks, NULL for full walks
765 * Returns references to children of the hierarchy below @root, or
766 * @root itself, or %NULL after a full round-trip.
768 * Caller must pass the return value in @prev on subsequent
769 * invocations for reference counting, or use mem_cgroup_iter_break()
770 * to cancel a hierarchy walk before the round-trip is complete.
772 * Reclaimers can specify a zone and a priority level in @reclaim to
773 * divide up the memcgs in the hierarchy among all concurrent
774 * reclaimers operating on the same zone and priority.
776 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
777 struct mem_cgroup
*prev
,
778 struct mem_cgroup_reclaim_cookie
*reclaim
)
780 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
781 struct cgroup_subsys_state
*css
= NULL
;
782 struct mem_cgroup
*memcg
= NULL
;
783 struct mem_cgroup
*pos
= NULL
;
785 if (mem_cgroup_disabled())
789 root
= root_mem_cgroup
;
791 if (prev
&& !reclaim
)
794 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
803 struct mem_cgroup_per_zone
*mz
;
805 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
806 iter
= &mz
->iter
[reclaim
->priority
];
808 if (prev
&& reclaim
->generation
!= iter
->generation
)
812 pos
= READ_ONCE(iter
->position
);
813 if (!pos
|| css_tryget(&pos
->css
))
816 * css reference reached zero, so iter->position will
817 * be cleared by ->css_released. However, we should not
818 * rely on this happening soon, because ->css_released
819 * is called from a work queue, and by busy-waiting we
820 * might block it. So we clear iter->position right
823 (void)cmpxchg(&iter
->position
, pos
, NULL
);
831 css
= css_next_descendant_pre(css
, &root
->css
);
834 * Reclaimers share the hierarchy walk, and a
835 * new one might jump in right at the end of
836 * the hierarchy - make sure they see at least
837 * one group and restart from the beginning.
845 * Verify the css and acquire a reference. The root
846 * is provided by the caller, so we know it's alive
847 * and kicking, and don't take an extra reference.
849 memcg
= mem_cgroup_from_css(css
);
851 if (css
== &root
->css
)
862 * The position could have already been updated by a competing
863 * thread, so check that the value hasn't changed since we read
864 * it to avoid reclaiming from the same cgroup twice.
866 (void)cmpxchg(&iter
->position
, pos
, memcg
);
874 reclaim
->generation
= iter
->generation
;
880 if (prev
&& prev
!= root
)
887 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
888 * @root: hierarchy root
889 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
891 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
892 struct mem_cgroup
*prev
)
895 root
= root_mem_cgroup
;
896 if (prev
&& prev
!= root
)
900 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
902 struct mem_cgroup
*memcg
= dead_memcg
;
903 struct mem_cgroup_reclaim_iter
*iter
;
904 struct mem_cgroup_per_zone
*mz
;
908 while ((memcg
= parent_mem_cgroup(memcg
))) {
910 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
911 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
912 for (i
= 0; i
<= DEF_PRIORITY
; i
++) {
914 cmpxchg(&iter
->position
,
923 * Iteration constructs for visiting all cgroups (under a tree). If
924 * loops are exited prematurely (break), mem_cgroup_iter_break() must
925 * be used for reference counting.
927 #define for_each_mem_cgroup_tree(iter, root) \
928 for (iter = mem_cgroup_iter(root, NULL, NULL); \
930 iter = mem_cgroup_iter(root, iter, NULL))
932 #define for_each_mem_cgroup(iter) \
933 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
935 iter = mem_cgroup_iter(NULL, iter, NULL))
938 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
940 * @zone: zone of the page
942 * This function is only safe when following the LRU page isolation
943 * and putback protocol: the LRU lock must be held, and the page must
944 * either be PageLRU() or the caller must have isolated/allocated it.
946 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct pglist_data
*pgdat
)
948 struct mem_cgroup_per_zone
*mz
;
949 struct mem_cgroup
*memcg
;
950 struct lruvec
*lruvec
;
952 if (mem_cgroup_disabled()) {
953 lruvec
= &pgdat
->lruvec
;
957 memcg
= page
->mem_cgroup
;
959 * Swapcache readahead pages are added to the LRU - and
960 * possibly migrated - before they are charged.
963 memcg
= root_mem_cgroup
;
965 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
966 lruvec
= &mz
->lruvec
;
969 * Since a node can be onlined after the mem_cgroup was created,
970 * we have to be prepared to initialize lruvec->zone here;
971 * and if offlined then reonlined, we need to reinitialize it.
973 if (unlikely(lruvec
->pgdat
!= pgdat
))
974 lruvec
->pgdat
= pgdat
;
979 * mem_cgroup_update_lru_size - account for adding or removing an lru page
980 * @lruvec: mem_cgroup per zone lru vector
981 * @lru: index of lru list the page is sitting on
982 * @zid: Zone ID of the zone pages have been added to
983 * @nr_pages: positive when adding or negative when removing
985 * This function must be called under lru_lock, just before a page is added
986 * to or just after a page is removed from an lru list (that ordering being
987 * so as to allow it to check that lru_size 0 is consistent with list_empty).
989 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
990 enum zone_type zid
, int nr_pages
)
992 struct mem_cgroup_per_zone
*mz
;
993 unsigned long *lru_size
;
997 __update_lru_size(lruvec
, lru
, zid
, nr_pages
);
999 if (mem_cgroup_disabled())
1002 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1003 lru_size
= mz
->lru_size
+ lru
;
1004 empty
= list_empty(lruvec
->lists
+ lru
);
1007 *lru_size
+= nr_pages
;
1010 if (WARN_ONCE(size
< 0 || empty
!= !size
,
1011 "%s(%p, %d, %d): lru_size %ld but %sempty\n",
1012 __func__
, lruvec
, lru
, nr_pages
, size
, empty
? "" : "not ")) {
1018 *lru_size
+= nr_pages
;
1021 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1023 struct mem_cgroup
*task_memcg
;
1024 struct task_struct
*p
;
1027 p
= find_lock_task_mm(task
);
1029 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1033 * All threads may have already detached their mm's, but the oom
1034 * killer still needs to detect if they have already been oom
1035 * killed to prevent needlessly killing additional tasks.
1038 task_memcg
= mem_cgroup_from_task(task
);
1039 css_get(&task_memcg
->css
);
1042 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1043 css_put(&task_memcg
->css
);
1048 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1049 * @memcg: the memory cgroup
1051 * Returns the maximum amount of memory @mem can be charged with, in
1054 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1056 unsigned long margin
= 0;
1057 unsigned long count
;
1058 unsigned long limit
;
1060 count
= page_counter_read(&memcg
->memory
);
1061 limit
= READ_ONCE(memcg
->memory
.limit
);
1063 margin
= limit
- count
;
1065 if (do_memsw_account()) {
1066 count
= page_counter_read(&memcg
->memsw
);
1067 limit
= READ_ONCE(memcg
->memsw
.limit
);
1069 margin
= min(margin
, limit
- count
);
1078 * A routine for checking "mem" is under move_account() or not.
1080 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1081 * moving cgroups. This is for waiting at high-memory pressure
1084 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1086 struct mem_cgroup
*from
;
1087 struct mem_cgroup
*to
;
1090 * Unlike task_move routines, we access mc.to, mc.from not under
1091 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1093 spin_lock(&mc
.lock
);
1099 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1100 mem_cgroup_is_descendant(to
, memcg
);
1102 spin_unlock(&mc
.lock
);
1106 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1108 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1109 if (mem_cgroup_under_move(memcg
)) {
1111 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1112 /* moving charge context might have finished. */
1115 finish_wait(&mc
.waitq
, &wait
);
1122 #define K(x) ((x) << (PAGE_SHIFT-10))
1124 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1125 * @memcg: The memory cgroup that went over limit
1126 * @p: Task that is going to be killed
1128 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1131 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1133 struct mem_cgroup
*iter
;
1139 pr_info("Task in ");
1140 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1141 pr_cont(" killed as a result of limit of ");
1143 pr_info("Memory limit reached of cgroup ");
1146 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1151 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1152 K((u64
)page_counter_read(&memcg
->memory
)),
1153 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1154 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1155 K((u64
)page_counter_read(&memcg
->memsw
)),
1156 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1157 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1158 K((u64
)page_counter_read(&memcg
->kmem
)),
1159 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1161 for_each_mem_cgroup_tree(iter
, memcg
) {
1162 pr_info("Memory cgroup stats for ");
1163 pr_cont_cgroup_path(iter
->css
.cgroup
);
1166 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1167 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1169 pr_cont(" %s:%luKB", mem_cgroup_stat_names
[i
],
1170 K(mem_cgroup_read_stat(iter
, i
)));
1173 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1174 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1175 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1182 * This function returns the number of memcg under hierarchy tree. Returns
1183 * 1(self count) if no children.
1185 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1188 struct mem_cgroup
*iter
;
1190 for_each_mem_cgroup_tree(iter
, memcg
)
1196 * Return the memory (and swap, if configured) limit for a memcg.
1198 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1200 unsigned long limit
;
1202 limit
= memcg
->memory
.limit
;
1203 if (mem_cgroup_swappiness(memcg
)) {
1204 unsigned long memsw_limit
;
1205 unsigned long swap_limit
;
1207 memsw_limit
= memcg
->memsw
.limit
;
1208 swap_limit
= memcg
->swap
.limit
;
1209 swap_limit
= min(swap_limit
, (unsigned long)total_swap_pages
);
1210 limit
= min(limit
+ swap_limit
, memsw_limit
);
1215 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1218 struct oom_control oc
= {
1222 .gfp_mask
= gfp_mask
,
1225 struct mem_cgroup
*iter
;
1226 unsigned long chosen_points
= 0;
1227 unsigned long totalpages
;
1228 unsigned int points
= 0;
1229 struct task_struct
*chosen
= NULL
;
1231 mutex_lock(&oom_lock
);
1234 * If current has a pending SIGKILL or is exiting, then automatically
1235 * select it. The goal is to allow it to allocate so that it may
1236 * quickly exit and free its memory.
1238 if (task_will_free_mem(current
)) {
1239 mark_oom_victim(current
);
1240 wake_oom_reaper(current
);
1244 check_panic_on_oom(&oc
, CONSTRAINT_MEMCG
);
1245 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1246 for_each_mem_cgroup_tree(iter
, memcg
) {
1247 struct css_task_iter it
;
1248 struct task_struct
*task
;
1250 css_task_iter_start(&iter
->css
, &it
);
1251 while ((task
= css_task_iter_next(&it
))) {
1252 switch (oom_scan_process_thread(&oc
, task
)) {
1253 case OOM_SCAN_SELECT
:
1255 put_task_struct(chosen
);
1257 chosen_points
= ULONG_MAX
;
1258 get_task_struct(chosen
);
1260 case OOM_SCAN_CONTINUE
:
1262 case OOM_SCAN_ABORT
:
1263 css_task_iter_end(&it
);
1264 mem_cgroup_iter_break(memcg
, iter
);
1266 put_task_struct(chosen
);
1267 /* Set a dummy value to return "true". */
1268 chosen
= (void *) 1;
1273 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1274 if (!points
|| points
< chosen_points
)
1276 /* Prefer thread group leaders for display purposes */
1277 if (points
== chosen_points
&&
1278 thread_group_leader(chosen
))
1282 put_task_struct(chosen
);
1284 chosen_points
= points
;
1285 get_task_struct(chosen
);
1287 css_task_iter_end(&it
);
1291 points
= chosen_points
* 1000 / totalpages
;
1292 oom_kill_process(&oc
, chosen
, points
, totalpages
,
1293 "Memory cgroup out of memory");
1296 mutex_unlock(&oom_lock
);
1300 #if MAX_NUMNODES > 1
1303 * test_mem_cgroup_node_reclaimable
1304 * @memcg: the target memcg
1305 * @nid: the node ID to be checked.
1306 * @noswap : specify true here if the user wants flle only information.
1308 * This function returns whether the specified memcg contains any
1309 * reclaimable pages on a node. Returns true if there are any reclaimable
1310 * pages in the node.
1312 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1313 int nid
, bool noswap
)
1315 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1317 if (noswap
|| !total_swap_pages
)
1319 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1326 * Always updating the nodemask is not very good - even if we have an empty
1327 * list or the wrong list here, we can start from some node and traverse all
1328 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1331 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1335 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1336 * pagein/pageout changes since the last update.
1338 if (!atomic_read(&memcg
->numainfo_events
))
1340 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1343 /* make a nodemask where this memcg uses memory from */
1344 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1346 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1348 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1349 node_clear(nid
, memcg
->scan_nodes
);
1352 atomic_set(&memcg
->numainfo_events
, 0);
1353 atomic_set(&memcg
->numainfo_updating
, 0);
1357 * Selecting a node where we start reclaim from. Because what we need is just
1358 * reducing usage counter, start from anywhere is O,K. Considering
1359 * memory reclaim from current node, there are pros. and cons.
1361 * Freeing memory from current node means freeing memory from a node which
1362 * we'll use or we've used. So, it may make LRU bad. And if several threads
1363 * hit limits, it will see a contention on a node. But freeing from remote
1364 * node means more costs for memory reclaim because of memory latency.
1366 * Now, we use round-robin. Better algorithm is welcomed.
1368 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1372 mem_cgroup_may_update_nodemask(memcg
);
1373 node
= memcg
->last_scanned_node
;
1375 node
= next_node_in(node
, memcg
->scan_nodes
);
1377 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1378 * last time it really checked all the LRUs due to rate limiting.
1379 * Fallback to the current node in that case for simplicity.
1381 if (unlikely(node
== MAX_NUMNODES
))
1382 node
= numa_node_id();
1384 memcg
->last_scanned_node
= node
;
1388 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1394 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1397 unsigned long *total_scanned
)
1399 struct mem_cgroup
*victim
= NULL
;
1402 unsigned long excess
;
1403 unsigned long nr_scanned
;
1404 struct mem_cgroup_reclaim_cookie reclaim
= {
1409 excess
= soft_limit_excess(root_memcg
);
1412 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1417 * If we have not been able to reclaim
1418 * anything, it might because there are
1419 * no reclaimable pages under this hierarchy
1424 * We want to do more targeted reclaim.
1425 * excess >> 2 is not to excessive so as to
1426 * reclaim too much, nor too less that we keep
1427 * coming back to reclaim from this cgroup
1429 if (total
>= (excess
>> 2) ||
1430 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1435 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1437 *total_scanned
+= nr_scanned
;
1438 if (!soft_limit_excess(root_memcg
))
1441 mem_cgroup_iter_break(root_memcg
, victim
);
1445 #ifdef CONFIG_LOCKDEP
1446 static struct lockdep_map memcg_oom_lock_dep_map
= {
1447 .name
= "memcg_oom_lock",
1451 static DEFINE_SPINLOCK(memcg_oom_lock
);
1454 * Check OOM-Killer is already running under our hierarchy.
1455 * If someone is running, return false.
1457 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1459 struct mem_cgroup
*iter
, *failed
= NULL
;
1461 spin_lock(&memcg_oom_lock
);
1463 for_each_mem_cgroup_tree(iter
, memcg
) {
1464 if (iter
->oom_lock
) {
1466 * this subtree of our hierarchy is already locked
1467 * so we cannot give a lock.
1470 mem_cgroup_iter_break(memcg
, iter
);
1473 iter
->oom_lock
= true;
1478 * OK, we failed to lock the whole subtree so we have
1479 * to clean up what we set up to the failing subtree
1481 for_each_mem_cgroup_tree(iter
, memcg
) {
1482 if (iter
== failed
) {
1483 mem_cgroup_iter_break(memcg
, iter
);
1486 iter
->oom_lock
= false;
1489 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1491 spin_unlock(&memcg_oom_lock
);
1496 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1498 struct mem_cgroup
*iter
;
1500 spin_lock(&memcg_oom_lock
);
1501 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1502 for_each_mem_cgroup_tree(iter
, memcg
)
1503 iter
->oom_lock
= false;
1504 spin_unlock(&memcg_oom_lock
);
1507 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1509 struct mem_cgroup
*iter
;
1511 spin_lock(&memcg_oom_lock
);
1512 for_each_mem_cgroup_tree(iter
, memcg
)
1514 spin_unlock(&memcg_oom_lock
);
1517 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1519 struct mem_cgroup
*iter
;
1522 * When a new child is created while the hierarchy is under oom,
1523 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1525 spin_lock(&memcg_oom_lock
);
1526 for_each_mem_cgroup_tree(iter
, memcg
)
1527 if (iter
->under_oom
> 0)
1529 spin_unlock(&memcg_oom_lock
);
1532 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1534 struct oom_wait_info
{
1535 struct mem_cgroup
*memcg
;
1539 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1540 unsigned mode
, int sync
, void *arg
)
1542 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1543 struct mem_cgroup
*oom_wait_memcg
;
1544 struct oom_wait_info
*oom_wait_info
;
1546 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1547 oom_wait_memcg
= oom_wait_info
->memcg
;
1549 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1550 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1552 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1555 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1558 * For the following lockless ->under_oom test, the only required
1559 * guarantee is that it must see the state asserted by an OOM when
1560 * this function is called as a result of userland actions
1561 * triggered by the notification of the OOM. This is trivially
1562 * achieved by invoking mem_cgroup_mark_under_oom() before
1563 * triggering notification.
1565 if (memcg
&& memcg
->under_oom
)
1566 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1569 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1571 if (!current
->memcg_may_oom
)
1574 * We are in the middle of the charge context here, so we
1575 * don't want to block when potentially sitting on a callstack
1576 * that holds all kinds of filesystem and mm locks.
1578 * Also, the caller may handle a failed allocation gracefully
1579 * (like optional page cache readahead) and so an OOM killer
1580 * invocation might not even be necessary.
1582 * That's why we don't do anything here except remember the
1583 * OOM context and then deal with it at the end of the page
1584 * fault when the stack is unwound, the locks are released,
1585 * and when we know whether the fault was overall successful.
1587 css_get(&memcg
->css
);
1588 current
->memcg_in_oom
= memcg
;
1589 current
->memcg_oom_gfp_mask
= mask
;
1590 current
->memcg_oom_order
= order
;
1594 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1595 * @handle: actually kill/wait or just clean up the OOM state
1597 * This has to be called at the end of a page fault if the memcg OOM
1598 * handler was enabled.
1600 * Memcg supports userspace OOM handling where failed allocations must
1601 * sleep on a waitqueue until the userspace task resolves the
1602 * situation. Sleeping directly in the charge context with all kinds
1603 * of locks held is not a good idea, instead we remember an OOM state
1604 * in the task and mem_cgroup_oom_synchronize() has to be called at
1605 * the end of the page fault to complete the OOM handling.
1607 * Returns %true if an ongoing memcg OOM situation was detected and
1608 * completed, %false otherwise.
1610 bool mem_cgroup_oom_synchronize(bool handle
)
1612 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1613 struct oom_wait_info owait
;
1616 /* OOM is global, do not handle */
1620 if (!handle
|| oom_killer_disabled
)
1623 owait
.memcg
= memcg
;
1624 owait
.wait
.flags
= 0;
1625 owait
.wait
.func
= memcg_oom_wake_function
;
1626 owait
.wait
.private = current
;
1627 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1629 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1630 mem_cgroup_mark_under_oom(memcg
);
1632 locked
= mem_cgroup_oom_trylock(memcg
);
1635 mem_cgroup_oom_notify(memcg
);
1637 if (locked
&& !memcg
->oom_kill_disable
) {
1638 mem_cgroup_unmark_under_oom(memcg
);
1639 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1640 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1641 current
->memcg_oom_order
);
1644 mem_cgroup_unmark_under_oom(memcg
);
1645 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1649 mem_cgroup_oom_unlock(memcg
);
1651 * There is no guarantee that an OOM-lock contender
1652 * sees the wakeups triggered by the OOM kill
1653 * uncharges. Wake any sleepers explicitely.
1655 memcg_oom_recover(memcg
);
1658 current
->memcg_in_oom
= NULL
;
1659 css_put(&memcg
->css
);
1664 * lock_page_memcg - lock a page->mem_cgroup binding
1667 * This function protects unlocked LRU pages from being moved to
1668 * another cgroup and stabilizes their page->mem_cgroup binding.
1670 void lock_page_memcg(struct page
*page
)
1672 struct mem_cgroup
*memcg
;
1673 unsigned long flags
;
1676 * The RCU lock is held throughout the transaction. The fast
1677 * path can get away without acquiring the memcg->move_lock
1678 * because page moving starts with an RCU grace period.
1682 if (mem_cgroup_disabled())
1685 memcg
= page
->mem_cgroup
;
1686 if (unlikely(!memcg
))
1689 if (atomic_read(&memcg
->moving_account
) <= 0)
1692 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1693 if (memcg
!= page
->mem_cgroup
) {
1694 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1699 * When charge migration first begins, we can have locked and
1700 * unlocked page stat updates happening concurrently. Track
1701 * the task who has the lock for unlock_page_memcg().
1703 memcg
->move_lock_task
= current
;
1704 memcg
->move_lock_flags
= flags
;
1708 EXPORT_SYMBOL(lock_page_memcg
);
1711 * unlock_page_memcg - unlock a page->mem_cgroup binding
1714 void unlock_page_memcg(struct page
*page
)
1716 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
1718 if (memcg
&& memcg
->move_lock_task
== current
) {
1719 unsigned long flags
= memcg
->move_lock_flags
;
1721 memcg
->move_lock_task
= NULL
;
1722 memcg
->move_lock_flags
= 0;
1724 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1729 EXPORT_SYMBOL(unlock_page_memcg
);
1732 * size of first charge trial. "32" comes from vmscan.c's magic value.
1733 * TODO: maybe necessary to use big numbers in big irons.
1735 #define CHARGE_BATCH 32U
1736 struct memcg_stock_pcp
{
1737 struct mem_cgroup
*cached
; /* this never be root cgroup */
1738 unsigned int nr_pages
;
1739 struct work_struct work
;
1740 unsigned long flags
;
1741 #define FLUSHING_CACHED_CHARGE 0
1743 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1744 static DEFINE_MUTEX(percpu_charge_mutex
);
1747 * consume_stock: Try to consume stocked charge on this cpu.
1748 * @memcg: memcg to consume from.
1749 * @nr_pages: how many pages to charge.
1751 * The charges will only happen if @memcg matches the current cpu's memcg
1752 * stock, and at least @nr_pages are available in that stock. Failure to
1753 * service an allocation will refill the stock.
1755 * returns true if successful, false otherwise.
1757 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1759 struct memcg_stock_pcp
*stock
;
1762 if (nr_pages
> CHARGE_BATCH
)
1765 stock
= &get_cpu_var(memcg_stock
);
1766 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
1767 stock
->nr_pages
-= nr_pages
;
1770 put_cpu_var(memcg_stock
);
1775 * Returns stocks cached in percpu and reset cached information.
1777 static void drain_stock(struct memcg_stock_pcp
*stock
)
1779 struct mem_cgroup
*old
= stock
->cached
;
1781 if (stock
->nr_pages
) {
1782 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
1783 if (do_memsw_account())
1784 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
1785 css_put_many(&old
->css
, stock
->nr_pages
);
1786 stock
->nr_pages
= 0;
1788 stock
->cached
= NULL
;
1792 * This must be called under preempt disabled or must be called by
1793 * a thread which is pinned to local cpu.
1795 static void drain_local_stock(struct work_struct
*dummy
)
1797 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
1799 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
1803 * Cache charges(val) to local per_cpu area.
1804 * This will be consumed by consume_stock() function, later.
1806 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1808 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
1810 if (stock
->cached
!= memcg
) { /* reset if necessary */
1812 stock
->cached
= memcg
;
1814 stock
->nr_pages
+= nr_pages
;
1815 put_cpu_var(memcg_stock
);
1819 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1820 * of the hierarchy under it.
1822 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
1826 /* If someone's already draining, avoid adding running more workers. */
1827 if (!mutex_trylock(&percpu_charge_mutex
))
1829 /* Notify other cpus that system-wide "drain" is running */
1832 for_each_online_cpu(cpu
) {
1833 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1834 struct mem_cgroup
*memcg
;
1836 memcg
= stock
->cached
;
1837 if (!memcg
|| !stock
->nr_pages
)
1839 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
1841 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
1843 drain_local_stock(&stock
->work
);
1845 schedule_work_on(cpu
, &stock
->work
);
1850 mutex_unlock(&percpu_charge_mutex
);
1853 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
1854 unsigned long action
,
1857 int cpu
= (unsigned long)hcpu
;
1858 struct memcg_stock_pcp
*stock
;
1860 if (action
== CPU_ONLINE
)
1863 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
1866 stock
= &per_cpu(memcg_stock
, cpu
);
1871 static void reclaim_high(struct mem_cgroup
*memcg
,
1872 unsigned int nr_pages
,
1876 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
1878 mem_cgroup_events(memcg
, MEMCG_HIGH
, 1);
1879 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
1880 } while ((memcg
= parent_mem_cgroup(memcg
)));
1883 static void high_work_func(struct work_struct
*work
)
1885 struct mem_cgroup
*memcg
;
1887 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
1888 reclaim_high(memcg
, CHARGE_BATCH
, GFP_KERNEL
);
1892 * Scheduled by try_charge() to be executed from the userland return path
1893 * and reclaims memory over the high limit.
1895 void mem_cgroup_handle_over_high(void)
1897 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
1898 struct mem_cgroup
*memcg
;
1900 if (likely(!nr_pages
))
1903 memcg
= get_mem_cgroup_from_mm(current
->mm
);
1904 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
1905 css_put(&memcg
->css
);
1906 current
->memcg_nr_pages_over_high
= 0;
1909 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1910 unsigned int nr_pages
)
1912 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
1913 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1914 struct mem_cgroup
*mem_over_limit
;
1915 struct page_counter
*counter
;
1916 unsigned long nr_reclaimed
;
1917 bool may_swap
= true;
1918 bool drained
= false;
1920 if (mem_cgroup_is_root(memcg
))
1923 if (consume_stock(memcg
, nr_pages
))
1926 if (!do_memsw_account() ||
1927 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
1928 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
1930 if (do_memsw_account())
1931 page_counter_uncharge(&memcg
->memsw
, batch
);
1932 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
1934 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
1938 if (batch
> nr_pages
) {
1944 * Unlike in global OOM situations, memcg is not in a physical
1945 * memory shortage. Allow dying and OOM-killed tasks to
1946 * bypass the last charges so that they can exit quickly and
1947 * free their memory.
1949 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
1950 fatal_signal_pending(current
) ||
1951 current
->flags
& PF_EXITING
))
1954 if (unlikely(task_in_memcg_oom(current
)))
1957 if (!gfpflags_allow_blocking(gfp_mask
))
1960 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
1962 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
1963 gfp_mask
, may_swap
);
1965 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
1969 drain_all_stock(mem_over_limit
);
1974 if (gfp_mask
& __GFP_NORETRY
)
1977 * Even though the limit is exceeded at this point, reclaim
1978 * may have been able to free some pages. Retry the charge
1979 * before killing the task.
1981 * Only for regular pages, though: huge pages are rather
1982 * unlikely to succeed so close to the limit, and we fall back
1983 * to regular pages anyway in case of failure.
1985 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
1988 * At task move, charge accounts can be doubly counted. So, it's
1989 * better to wait until the end of task_move if something is going on.
1991 if (mem_cgroup_wait_acct_move(mem_over_limit
))
1997 if (gfp_mask
& __GFP_NOFAIL
)
2000 if (fatal_signal_pending(current
))
2003 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
2005 mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2006 get_order(nr_pages
* PAGE_SIZE
));
2008 if (!(gfp_mask
& __GFP_NOFAIL
))
2012 * The allocation either can't fail or will lead to more memory
2013 * being freed very soon. Allow memory usage go over the limit
2014 * temporarily by force charging it.
2016 page_counter_charge(&memcg
->memory
, nr_pages
);
2017 if (do_memsw_account())
2018 page_counter_charge(&memcg
->memsw
, nr_pages
);
2019 css_get_many(&memcg
->css
, nr_pages
);
2024 css_get_many(&memcg
->css
, batch
);
2025 if (batch
> nr_pages
)
2026 refill_stock(memcg
, batch
- nr_pages
);
2029 * If the hierarchy is above the normal consumption range, schedule
2030 * reclaim on returning to userland. We can perform reclaim here
2031 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2032 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2033 * not recorded as it most likely matches current's and won't
2034 * change in the meantime. As high limit is checked again before
2035 * reclaim, the cost of mismatch is negligible.
2038 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2039 /* Don't bother a random interrupted task */
2040 if (in_interrupt()) {
2041 schedule_work(&memcg
->high_work
);
2044 current
->memcg_nr_pages_over_high
+= batch
;
2045 set_notify_resume(current
);
2048 } while ((memcg
= parent_mem_cgroup(memcg
)));
2053 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2055 if (mem_cgroup_is_root(memcg
))
2058 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2059 if (do_memsw_account())
2060 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2062 css_put_many(&memcg
->css
, nr_pages
);
2065 static void lock_page_lru(struct page
*page
, int *isolated
)
2067 struct zone
*zone
= page_zone(page
);
2069 spin_lock_irq(zone_lru_lock(zone
));
2070 if (PageLRU(page
)) {
2071 struct lruvec
*lruvec
;
2073 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
2075 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2081 static void unlock_page_lru(struct page
*page
, int isolated
)
2083 struct zone
*zone
= page_zone(page
);
2086 struct lruvec
*lruvec
;
2088 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
2089 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2091 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2093 spin_unlock_irq(zone_lru_lock(zone
));
2096 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2101 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2104 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2105 * may already be on some other mem_cgroup's LRU. Take care of it.
2108 lock_page_lru(page
, &isolated
);
2111 * Nobody should be changing or seriously looking at
2112 * page->mem_cgroup at this point:
2114 * - the page is uncharged
2116 * - the page is off-LRU
2118 * - an anonymous fault has exclusive page access, except for
2119 * a locked page table
2121 * - a page cache insertion, a swapin fault, or a migration
2122 * have the page locked
2124 page
->mem_cgroup
= memcg
;
2127 unlock_page_lru(page
, isolated
);
2131 static int memcg_alloc_cache_id(void)
2136 id
= ida_simple_get(&memcg_cache_ida
,
2137 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2141 if (id
< memcg_nr_cache_ids
)
2145 * There's no space for the new id in memcg_caches arrays,
2146 * so we have to grow them.
2148 down_write(&memcg_cache_ids_sem
);
2150 size
= 2 * (id
+ 1);
2151 if (size
< MEMCG_CACHES_MIN_SIZE
)
2152 size
= MEMCG_CACHES_MIN_SIZE
;
2153 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2154 size
= MEMCG_CACHES_MAX_SIZE
;
2156 err
= memcg_update_all_caches(size
);
2158 err
= memcg_update_all_list_lrus(size
);
2160 memcg_nr_cache_ids
= size
;
2162 up_write(&memcg_cache_ids_sem
);
2165 ida_simple_remove(&memcg_cache_ida
, id
);
2171 static void memcg_free_cache_id(int id
)
2173 ida_simple_remove(&memcg_cache_ida
, id
);
2176 struct memcg_kmem_cache_create_work
{
2177 struct mem_cgroup
*memcg
;
2178 struct kmem_cache
*cachep
;
2179 struct work_struct work
;
2182 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2184 struct memcg_kmem_cache_create_work
*cw
=
2185 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2186 struct mem_cgroup
*memcg
= cw
->memcg
;
2187 struct kmem_cache
*cachep
= cw
->cachep
;
2189 memcg_create_kmem_cache(memcg
, cachep
);
2191 css_put(&memcg
->css
);
2196 * Enqueue the creation of a per-memcg kmem_cache.
2198 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2199 struct kmem_cache
*cachep
)
2201 struct memcg_kmem_cache_create_work
*cw
;
2203 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2207 css_get(&memcg
->css
);
2210 cw
->cachep
= cachep
;
2211 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2213 schedule_work(&cw
->work
);
2216 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2217 struct kmem_cache
*cachep
)
2220 * We need to stop accounting when we kmalloc, because if the
2221 * corresponding kmalloc cache is not yet created, the first allocation
2222 * in __memcg_schedule_kmem_cache_create will recurse.
2224 * However, it is better to enclose the whole function. Depending on
2225 * the debugging options enabled, INIT_WORK(), for instance, can
2226 * trigger an allocation. This too, will make us recurse. Because at
2227 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2228 * the safest choice is to do it like this, wrapping the whole function.
2230 current
->memcg_kmem_skip_account
= 1;
2231 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2232 current
->memcg_kmem_skip_account
= 0;
2235 static inline bool memcg_kmem_bypass(void)
2237 if (in_interrupt() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
2243 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2244 * @cachep: the original global kmem cache
2246 * Return the kmem_cache we're supposed to use for a slab allocation.
2247 * We try to use the current memcg's version of the cache.
2249 * If the cache does not exist yet, if we are the first user of it, we
2250 * create it asynchronously in a workqueue and let the current allocation
2251 * go through with the original cache.
2253 * This function takes a reference to the cache it returns to assure it
2254 * won't get destroyed while we are working with it. Once the caller is
2255 * done with it, memcg_kmem_put_cache() must be called to release the
2258 struct kmem_cache
*memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2260 struct mem_cgroup
*memcg
;
2261 struct kmem_cache
*memcg_cachep
;
2264 VM_BUG_ON(!is_root_cache(cachep
));
2266 if (memcg_kmem_bypass())
2269 if (current
->memcg_kmem_skip_account
)
2272 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2273 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2277 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2278 if (likely(memcg_cachep
))
2279 return memcg_cachep
;
2282 * If we are in a safe context (can wait, and not in interrupt
2283 * context), we could be be predictable and return right away.
2284 * This would guarantee that the allocation being performed
2285 * already belongs in the new cache.
2287 * However, there are some clashes that can arrive from locking.
2288 * For instance, because we acquire the slab_mutex while doing
2289 * memcg_create_kmem_cache, this means no further allocation
2290 * could happen with the slab_mutex held. So it's better to
2293 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2295 css_put(&memcg
->css
);
2300 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2301 * @cachep: the cache returned by memcg_kmem_get_cache
2303 void memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2305 if (!is_root_cache(cachep
))
2306 css_put(&cachep
->memcg_params
.memcg
->css
);
2310 * memcg_kmem_charge: charge a kmem page
2311 * @page: page to charge
2312 * @gfp: reclaim mode
2313 * @order: allocation order
2314 * @memcg: memory cgroup to charge
2316 * Returns 0 on success, an error code on failure.
2318 int memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2319 struct mem_cgroup
*memcg
)
2321 unsigned int nr_pages
= 1 << order
;
2322 struct page_counter
*counter
;
2325 ret
= try_charge(memcg
, gfp
, nr_pages
);
2329 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2330 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2331 cancel_charge(memcg
, nr_pages
);
2335 page
->mem_cgroup
= memcg
;
2341 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2342 * @page: page to charge
2343 * @gfp: reclaim mode
2344 * @order: allocation order
2346 * Returns 0 on success, an error code on failure.
2348 int memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2350 struct mem_cgroup
*memcg
;
2353 if (memcg_kmem_bypass())
2356 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2357 if (!mem_cgroup_is_root(memcg
))
2358 ret
= memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2359 css_put(&memcg
->css
);
2363 * memcg_kmem_uncharge: uncharge a kmem page
2364 * @page: page to uncharge
2365 * @order: allocation order
2367 void memcg_kmem_uncharge(struct page
*page
, int order
)
2369 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2370 unsigned int nr_pages
= 1 << order
;
2375 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2377 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2378 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2380 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2381 if (do_memsw_account())
2382 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2384 page
->mem_cgroup
= NULL
;
2385 css_put_many(&memcg
->css
, nr_pages
);
2387 #endif /* !CONFIG_SLOB */
2389 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2392 * Because tail pages are not marked as "used", set it. We're under
2393 * zone_lru_lock and migration entries setup in all page mappings.
2395 void mem_cgroup_split_huge_fixup(struct page
*head
)
2399 if (mem_cgroup_disabled())
2402 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2403 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2405 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2408 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2410 #ifdef CONFIG_MEMCG_SWAP
2411 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2414 int val
= (charge
) ? 1 : -1;
2415 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2419 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2420 * @entry: swap entry to be moved
2421 * @from: mem_cgroup which the entry is moved from
2422 * @to: mem_cgroup which the entry is moved to
2424 * It succeeds only when the swap_cgroup's record for this entry is the same
2425 * as the mem_cgroup's id of @from.
2427 * Returns 0 on success, -EINVAL on failure.
2429 * The caller must have charged to @to, IOW, called page_counter_charge() about
2430 * both res and memsw, and called css_get().
2432 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2433 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2435 unsigned short old_id
, new_id
;
2437 old_id
= mem_cgroup_id(from
);
2438 new_id
= mem_cgroup_id(to
);
2440 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2441 mem_cgroup_swap_statistics(from
, false);
2442 mem_cgroup_swap_statistics(to
, true);
2448 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2449 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2455 static DEFINE_MUTEX(memcg_limit_mutex
);
2457 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2458 unsigned long limit
)
2460 unsigned long curusage
;
2461 unsigned long oldusage
;
2462 bool enlarge
= false;
2467 * For keeping hierarchical_reclaim simple, how long we should retry
2468 * is depends on callers. We set our retry-count to be function
2469 * of # of children which we should visit in this loop.
2471 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2472 mem_cgroup_count_children(memcg
);
2474 oldusage
= page_counter_read(&memcg
->memory
);
2477 if (signal_pending(current
)) {
2482 mutex_lock(&memcg_limit_mutex
);
2483 if (limit
> memcg
->memsw
.limit
) {
2484 mutex_unlock(&memcg_limit_mutex
);
2488 if (limit
> memcg
->memory
.limit
)
2490 ret
= page_counter_limit(&memcg
->memory
, limit
);
2491 mutex_unlock(&memcg_limit_mutex
);
2496 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2498 curusage
= page_counter_read(&memcg
->memory
);
2499 /* Usage is reduced ? */
2500 if (curusage
>= oldusage
)
2503 oldusage
= curusage
;
2504 } while (retry_count
);
2506 if (!ret
&& enlarge
)
2507 memcg_oom_recover(memcg
);
2512 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2513 unsigned long limit
)
2515 unsigned long curusage
;
2516 unsigned long oldusage
;
2517 bool enlarge
= false;
2521 /* see mem_cgroup_resize_res_limit */
2522 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2523 mem_cgroup_count_children(memcg
);
2525 oldusage
= page_counter_read(&memcg
->memsw
);
2528 if (signal_pending(current
)) {
2533 mutex_lock(&memcg_limit_mutex
);
2534 if (limit
< memcg
->memory
.limit
) {
2535 mutex_unlock(&memcg_limit_mutex
);
2539 if (limit
> memcg
->memsw
.limit
)
2541 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2542 mutex_unlock(&memcg_limit_mutex
);
2547 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2549 curusage
= page_counter_read(&memcg
->memsw
);
2550 /* Usage is reduced ? */
2551 if (curusage
>= oldusage
)
2554 oldusage
= curusage
;
2555 } while (retry_count
);
2557 if (!ret
&& enlarge
)
2558 memcg_oom_recover(memcg
);
2563 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
2565 unsigned long *total_scanned
)
2567 unsigned long nr_reclaimed
= 0;
2568 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
2569 unsigned long reclaimed
;
2571 struct mem_cgroup_tree_per_zone
*mctz
;
2572 unsigned long excess
;
2573 unsigned long nr_scanned
;
2578 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
2580 * This loop can run a while, specially if mem_cgroup's continuously
2581 * keep exceeding their soft limit and putting the system under
2588 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2593 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
2594 gfp_mask
, &nr_scanned
);
2595 nr_reclaimed
+= reclaimed
;
2596 *total_scanned
+= nr_scanned
;
2597 spin_lock_irq(&mctz
->lock
);
2598 __mem_cgroup_remove_exceeded(mz
, mctz
);
2601 * If we failed to reclaim anything from this memory cgroup
2602 * it is time to move on to the next cgroup
2606 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2608 excess
= soft_limit_excess(mz
->memcg
);
2610 * One school of thought says that we should not add
2611 * back the node to the tree if reclaim returns 0.
2612 * But our reclaim could return 0, simply because due
2613 * to priority we are exposing a smaller subset of
2614 * memory to reclaim from. Consider this as a longer
2617 /* If excess == 0, no tree ops */
2618 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2619 spin_unlock_irq(&mctz
->lock
);
2620 css_put(&mz
->memcg
->css
);
2623 * Could not reclaim anything and there are no more
2624 * mem cgroups to try or we seem to be looping without
2625 * reclaiming anything.
2627 if (!nr_reclaimed
&&
2629 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2631 } while (!nr_reclaimed
);
2633 css_put(&next_mz
->memcg
->css
);
2634 return nr_reclaimed
;
2638 * Test whether @memcg has children, dead or alive. Note that this
2639 * function doesn't care whether @memcg has use_hierarchy enabled and
2640 * returns %true if there are child csses according to the cgroup
2641 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2643 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2648 ret
= css_next_child(NULL
, &memcg
->css
);
2654 * Reclaims as many pages from the given memcg as possible.
2656 * Caller is responsible for holding css reference for memcg.
2658 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2660 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2662 /* we call try-to-free pages for make this cgroup empty */
2663 lru_add_drain_all();
2664 /* try to free all pages in this cgroup */
2665 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2668 if (signal_pending(current
))
2671 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2675 /* maybe some writeback is necessary */
2676 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2684 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
2685 char *buf
, size_t nbytes
,
2688 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2690 if (mem_cgroup_is_root(memcg
))
2692 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
2695 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
2698 return mem_cgroup_from_css(css
)->use_hierarchy
;
2701 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
2702 struct cftype
*cft
, u64 val
)
2705 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2706 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
2708 if (memcg
->use_hierarchy
== val
)
2712 * If parent's use_hierarchy is set, we can't make any modifications
2713 * in the child subtrees. If it is unset, then the change can
2714 * occur, provided the current cgroup has no children.
2716 * For the root cgroup, parent_mem is NULL, we allow value to be
2717 * set if there are no children.
2719 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
2720 (val
== 1 || val
== 0)) {
2721 if (!memcg_has_children(memcg
))
2722 memcg
->use_hierarchy
= val
;
2731 static void tree_stat(struct mem_cgroup
*memcg
, unsigned long *stat
)
2733 struct mem_cgroup
*iter
;
2736 memset(stat
, 0, sizeof(*stat
) * MEMCG_NR_STAT
);
2738 for_each_mem_cgroup_tree(iter
, memcg
) {
2739 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
2740 stat
[i
] += mem_cgroup_read_stat(iter
, i
);
2744 static void tree_events(struct mem_cgroup
*memcg
, unsigned long *events
)
2746 struct mem_cgroup
*iter
;
2749 memset(events
, 0, sizeof(*events
) * MEMCG_NR_EVENTS
);
2751 for_each_mem_cgroup_tree(iter
, memcg
) {
2752 for (i
= 0; i
< MEMCG_NR_EVENTS
; i
++)
2753 events
[i
] += mem_cgroup_read_events(iter
, i
);
2757 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
2759 unsigned long val
= 0;
2761 if (mem_cgroup_is_root(memcg
)) {
2762 struct mem_cgroup
*iter
;
2764 for_each_mem_cgroup_tree(iter
, memcg
) {
2765 val
+= mem_cgroup_read_stat(iter
,
2766 MEM_CGROUP_STAT_CACHE
);
2767 val
+= mem_cgroup_read_stat(iter
,
2768 MEM_CGROUP_STAT_RSS
);
2770 val
+= mem_cgroup_read_stat(iter
,
2771 MEM_CGROUP_STAT_SWAP
);
2775 val
= page_counter_read(&memcg
->memory
);
2777 val
= page_counter_read(&memcg
->memsw
);
2790 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
2793 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2794 struct page_counter
*counter
;
2796 switch (MEMFILE_TYPE(cft
->private)) {
2798 counter
= &memcg
->memory
;
2801 counter
= &memcg
->memsw
;
2804 counter
= &memcg
->kmem
;
2807 counter
= &memcg
->tcpmem
;
2813 switch (MEMFILE_ATTR(cft
->private)) {
2815 if (counter
== &memcg
->memory
)
2816 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
2817 if (counter
== &memcg
->memsw
)
2818 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
2819 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
2821 return (u64
)counter
->limit
* PAGE_SIZE
;
2823 return (u64
)counter
->watermark
* PAGE_SIZE
;
2825 return counter
->failcnt
;
2826 case RES_SOFT_LIMIT
:
2827 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
2834 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2838 if (cgroup_memory_nokmem
)
2841 BUG_ON(memcg
->kmemcg_id
>= 0);
2842 BUG_ON(memcg
->kmem_state
);
2844 memcg_id
= memcg_alloc_cache_id();
2848 static_branch_inc(&memcg_kmem_enabled_key
);
2850 * A memory cgroup is considered kmem-online as soon as it gets
2851 * kmemcg_id. Setting the id after enabling static branching will
2852 * guarantee no one starts accounting before all call sites are
2855 memcg
->kmemcg_id
= memcg_id
;
2856 memcg
->kmem_state
= KMEM_ONLINE
;
2861 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2863 struct cgroup_subsys_state
*css
;
2864 struct mem_cgroup
*parent
, *child
;
2867 if (memcg
->kmem_state
!= KMEM_ONLINE
)
2870 * Clear the online state before clearing memcg_caches array
2871 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2872 * guarantees that no cache will be created for this cgroup
2873 * after we are done (see memcg_create_kmem_cache()).
2875 memcg
->kmem_state
= KMEM_ALLOCATED
;
2877 memcg_deactivate_kmem_caches(memcg
);
2879 kmemcg_id
= memcg
->kmemcg_id
;
2880 BUG_ON(kmemcg_id
< 0);
2882 parent
= parent_mem_cgroup(memcg
);
2884 parent
= root_mem_cgroup
;
2887 * Change kmemcg_id of this cgroup and all its descendants to the
2888 * parent's id, and then move all entries from this cgroup's list_lrus
2889 * to ones of the parent. After we have finished, all list_lrus
2890 * corresponding to this cgroup are guaranteed to remain empty. The
2891 * ordering is imposed by list_lru_node->lock taken by
2892 * memcg_drain_all_list_lrus().
2894 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2895 css_for_each_descendant_pre(css
, &memcg
->css
) {
2896 child
= mem_cgroup_from_css(css
);
2897 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
2898 child
->kmemcg_id
= parent
->kmemcg_id
;
2899 if (!memcg
->use_hierarchy
)
2904 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
2906 memcg_free_cache_id(kmemcg_id
);
2909 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2911 /* css_alloc() failed, offlining didn't happen */
2912 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
2913 memcg_offline_kmem(memcg
);
2915 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
2916 memcg_destroy_kmem_caches(memcg
);
2917 static_branch_dec(&memcg_kmem_enabled_key
);
2918 WARN_ON(page_counter_read(&memcg
->kmem
));
2922 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2926 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2929 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2932 #endif /* !CONFIG_SLOB */
2934 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2935 unsigned long limit
)
2939 mutex_lock(&memcg_limit_mutex
);
2940 ret
= page_counter_limit(&memcg
->kmem
, limit
);
2941 mutex_unlock(&memcg_limit_mutex
);
2945 static int memcg_update_tcp_limit(struct mem_cgroup
*memcg
, unsigned long limit
)
2949 mutex_lock(&memcg_limit_mutex
);
2951 ret
= page_counter_limit(&memcg
->tcpmem
, limit
);
2955 if (!memcg
->tcpmem_active
) {
2957 * The active flag needs to be written after the static_key
2958 * update. This is what guarantees that the socket activation
2959 * function is the last one to run. See sock_update_memcg() for
2960 * details, and note that we don't mark any socket as belonging
2961 * to this memcg until that flag is up.
2963 * We need to do this, because static_keys will span multiple
2964 * sites, but we can't control their order. If we mark a socket
2965 * as accounted, but the accounting functions are not patched in
2966 * yet, we'll lose accounting.
2968 * We never race with the readers in sock_update_memcg(),
2969 * because when this value change, the code to process it is not
2972 static_branch_inc(&memcg_sockets_enabled_key
);
2973 memcg
->tcpmem_active
= true;
2976 mutex_unlock(&memcg_limit_mutex
);
2981 * The user of this function is...
2984 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
2985 char *buf
, size_t nbytes
, loff_t off
)
2987 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2988 unsigned long nr_pages
;
2991 buf
= strstrip(buf
);
2992 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
2996 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
2998 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3002 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3004 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
3007 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
3010 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
3013 ret
= memcg_update_tcp_limit(memcg
, nr_pages
);
3017 case RES_SOFT_LIMIT
:
3018 memcg
->soft_limit
= nr_pages
;
3022 return ret
?: nbytes
;
3025 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3026 size_t nbytes
, loff_t off
)
3028 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3029 struct page_counter
*counter
;
3031 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3033 counter
= &memcg
->memory
;
3036 counter
= &memcg
->memsw
;
3039 counter
= &memcg
->kmem
;
3042 counter
= &memcg
->tcpmem
;
3048 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3050 page_counter_reset_watermark(counter
);
3053 counter
->failcnt
= 0;
3062 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3065 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3069 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3070 struct cftype
*cft
, u64 val
)
3072 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3074 if (val
& ~MOVE_MASK
)
3078 * No kind of locking is needed in here, because ->can_attach() will
3079 * check this value once in the beginning of the process, and then carry
3080 * on with stale data. This means that changes to this value will only
3081 * affect task migrations starting after the change.
3083 memcg
->move_charge_at_immigrate
= val
;
3087 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3088 struct cftype
*cft
, u64 val
)
3095 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3099 unsigned int lru_mask
;
3102 static const struct numa_stat stats
[] = {
3103 { "total", LRU_ALL
},
3104 { "file", LRU_ALL_FILE
},
3105 { "anon", LRU_ALL_ANON
},
3106 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3108 const struct numa_stat
*stat
;
3111 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3113 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3114 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3115 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3116 for_each_node_state(nid
, N_MEMORY
) {
3117 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3119 seq_printf(m
, " N%d=%lu", nid
, nr
);
3124 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3125 struct mem_cgroup
*iter
;
3128 for_each_mem_cgroup_tree(iter
, memcg
)
3129 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3130 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3131 for_each_node_state(nid
, N_MEMORY
) {
3133 for_each_mem_cgroup_tree(iter
, memcg
)
3134 nr
+= mem_cgroup_node_nr_lru_pages(
3135 iter
, nid
, stat
->lru_mask
);
3136 seq_printf(m
, " N%d=%lu", nid
, nr
);
3143 #endif /* CONFIG_NUMA */
3145 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3147 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3148 unsigned long memory
, memsw
;
3149 struct mem_cgroup
*mi
;
3152 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3153 MEM_CGROUP_STAT_NSTATS
);
3154 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3155 MEM_CGROUP_EVENTS_NSTATS
);
3156 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3158 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3159 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3161 seq_printf(m
, "%s %lu\n", mem_cgroup_stat_names
[i
],
3162 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3165 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3166 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3167 mem_cgroup_read_events(memcg
, i
));
3169 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3170 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3171 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3173 /* Hierarchical information */
3174 memory
= memsw
= PAGE_COUNTER_MAX
;
3175 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3176 memory
= min(memory
, mi
->memory
.limit
);
3177 memsw
= min(memsw
, mi
->memsw
.limit
);
3179 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3180 (u64
)memory
* PAGE_SIZE
);
3181 if (do_memsw_account())
3182 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3183 (u64
)memsw
* PAGE_SIZE
);
3185 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3186 unsigned long long val
= 0;
3188 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3190 for_each_mem_cgroup_tree(mi
, memcg
)
3191 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3192 seq_printf(m
, "total_%s %llu\n", mem_cgroup_stat_names
[i
], val
);
3195 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3196 unsigned long long val
= 0;
3198 for_each_mem_cgroup_tree(mi
, memcg
)
3199 val
+= mem_cgroup_read_events(mi
, i
);
3200 seq_printf(m
, "total_%s %llu\n",
3201 mem_cgroup_events_names
[i
], val
);
3204 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3205 unsigned long long val
= 0;
3207 for_each_mem_cgroup_tree(mi
, memcg
)
3208 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3209 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3212 #ifdef CONFIG_DEBUG_VM
3215 struct mem_cgroup_per_zone
*mz
;
3216 struct zone_reclaim_stat
*rstat
;
3217 unsigned long recent_rotated
[2] = {0, 0};
3218 unsigned long recent_scanned
[2] = {0, 0};
3220 for_each_online_node(nid
)
3221 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3222 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
3223 rstat
= &mz
->lruvec
.reclaim_stat
;
3225 recent_rotated
[0] += rstat
->recent_rotated
[0];
3226 recent_rotated
[1] += rstat
->recent_rotated
[1];
3227 recent_scanned
[0] += rstat
->recent_scanned
[0];
3228 recent_scanned
[1] += rstat
->recent_scanned
[1];
3230 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3231 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3232 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3233 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3240 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3243 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3245 return mem_cgroup_swappiness(memcg
);
3248 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3249 struct cftype
*cft
, u64 val
)
3251 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3257 memcg
->swappiness
= val
;
3259 vm_swappiness
= val
;
3264 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3266 struct mem_cgroup_threshold_ary
*t
;
3267 unsigned long usage
;
3272 t
= rcu_dereference(memcg
->thresholds
.primary
);
3274 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3279 usage
= mem_cgroup_usage(memcg
, swap
);
3282 * current_threshold points to threshold just below or equal to usage.
3283 * If it's not true, a threshold was crossed after last
3284 * call of __mem_cgroup_threshold().
3286 i
= t
->current_threshold
;
3289 * Iterate backward over array of thresholds starting from
3290 * current_threshold and check if a threshold is crossed.
3291 * If none of thresholds below usage is crossed, we read
3292 * only one element of the array here.
3294 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3295 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3297 /* i = current_threshold + 1 */
3301 * Iterate forward over array of thresholds starting from
3302 * current_threshold+1 and check if a threshold is crossed.
3303 * If none of thresholds above usage is crossed, we read
3304 * only one element of the array here.
3306 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3307 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3309 /* Update current_threshold */
3310 t
->current_threshold
= i
- 1;
3315 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3318 __mem_cgroup_threshold(memcg
, false);
3319 if (do_memsw_account())
3320 __mem_cgroup_threshold(memcg
, true);
3322 memcg
= parent_mem_cgroup(memcg
);
3326 static int compare_thresholds(const void *a
, const void *b
)
3328 const struct mem_cgroup_threshold
*_a
= a
;
3329 const struct mem_cgroup_threshold
*_b
= b
;
3331 if (_a
->threshold
> _b
->threshold
)
3334 if (_a
->threshold
< _b
->threshold
)
3340 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3342 struct mem_cgroup_eventfd_list
*ev
;
3344 spin_lock(&memcg_oom_lock
);
3346 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3347 eventfd_signal(ev
->eventfd
, 1);
3349 spin_unlock(&memcg_oom_lock
);
3353 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3355 struct mem_cgroup
*iter
;
3357 for_each_mem_cgroup_tree(iter
, memcg
)
3358 mem_cgroup_oom_notify_cb(iter
);
3361 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3362 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3364 struct mem_cgroup_thresholds
*thresholds
;
3365 struct mem_cgroup_threshold_ary
*new;
3366 unsigned long threshold
;
3367 unsigned long usage
;
3370 ret
= page_counter_memparse(args
, "-1", &threshold
);
3374 mutex_lock(&memcg
->thresholds_lock
);
3377 thresholds
= &memcg
->thresholds
;
3378 usage
= mem_cgroup_usage(memcg
, false);
3379 } else if (type
== _MEMSWAP
) {
3380 thresholds
= &memcg
->memsw_thresholds
;
3381 usage
= mem_cgroup_usage(memcg
, true);
3385 /* Check if a threshold crossed before adding a new one */
3386 if (thresholds
->primary
)
3387 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3389 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3391 /* Allocate memory for new array of thresholds */
3392 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3400 /* Copy thresholds (if any) to new array */
3401 if (thresholds
->primary
) {
3402 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3403 sizeof(struct mem_cgroup_threshold
));
3406 /* Add new threshold */
3407 new->entries
[size
- 1].eventfd
= eventfd
;
3408 new->entries
[size
- 1].threshold
= threshold
;
3410 /* Sort thresholds. Registering of new threshold isn't time-critical */
3411 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3412 compare_thresholds
, NULL
);
3414 /* Find current threshold */
3415 new->current_threshold
= -1;
3416 for (i
= 0; i
< size
; i
++) {
3417 if (new->entries
[i
].threshold
<= usage
) {
3419 * new->current_threshold will not be used until
3420 * rcu_assign_pointer(), so it's safe to increment
3423 ++new->current_threshold
;
3428 /* Free old spare buffer and save old primary buffer as spare */
3429 kfree(thresholds
->spare
);
3430 thresholds
->spare
= thresholds
->primary
;
3432 rcu_assign_pointer(thresholds
->primary
, new);
3434 /* To be sure that nobody uses thresholds */
3438 mutex_unlock(&memcg
->thresholds_lock
);
3443 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3444 struct eventfd_ctx
*eventfd
, const char *args
)
3446 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3449 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3450 struct eventfd_ctx
*eventfd
, const char *args
)
3452 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3455 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3456 struct eventfd_ctx
*eventfd
, enum res_type type
)
3458 struct mem_cgroup_thresholds
*thresholds
;
3459 struct mem_cgroup_threshold_ary
*new;
3460 unsigned long usage
;
3463 mutex_lock(&memcg
->thresholds_lock
);
3466 thresholds
= &memcg
->thresholds
;
3467 usage
= mem_cgroup_usage(memcg
, false);
3468 } else if (type
== _MEMSWAP
) {
3469 thresholds
= &memcg
->memsw_thresholds
;
3470 usage
= mem_cgroup_usage(memcg
, true);
3474 if (!thresholds
->primary
)
3477 /* Check if a threshold crossed before removing */
3478 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3480 /* Calculate new number of threshold */
3482 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3483 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3487 new = thresholds
->spare
;
3489 /* Set thresholds array to NULL if we don't have thresholds */
3498 /* Copy thresholds and find current threshold */
3499 new->current_threshold
= -1;
3500 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3501 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3504 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3505 if (new->entries
[j
].threshold
<= usage
) {
3507 * new->current_threshold will not be used
3508 * until rcu_assign_pointer(), so it's safe to increment
3511 ++new->current_threshold
;
3517 /* Swap primary and spare array */
3518 thresholds
->spare
= thresholds
->primary
;
3520 rcu_assign_pointer(thresholds
->primary
, new);
3522 /* To be sure that nobody uses thresholds */
3525 /* If all events are unregistered, free the spare array */
3527 kfree(thresholds
->spare
);
3528 thresholds
->spare
= NULL
;
3531 mutex_unlock(&memcg
->thresholds_lock
);
3534 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3535 struct eventfd_ctx
*eventfd
)
3537 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3540 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3541 struct eventfd_ctx
*eventfd
)
3543 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3546 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3547 struct eventfd_ctx
*eventfd
, const char *args
)
3549 struct mem_cgroup_eventfd_list
*event
;
3551 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3555 spin_lock(&memcg_oom_lock
);
3557 event
->eventfd
= eventfd
;
3558 list_add(&event
->list
, &memcg
->oom_notify
);
3560 /* already in OOM ? */
3561 if (memcg
->under_oom
)
3562 eventfd_signal(eventfd
, 1);
3563 spin_unlock(&memcg_oom_lock
);
3568 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3569 struct eventfd_ctx
*eventfd
)
3571 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3573 spin_lock(&memcg_oom_lock
);
3575 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3576 if (ev
->eventfd
== eventfd
) {
3577 list_del(&ev
->list
);
3582 spin_unlock(&memcg_oom_lock
);
3585 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3587 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3589 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3590 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3594 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3595 struct cftype
*cft
, u64 val
)
3597 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3599 /* cannot set to root cgroup and only 0 and 1 are allowed */
3600 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3603 memcg
->oom_kill_disable
= val
;
3605 memcg_oom_recover(memcg
);
3610 #ifdef CONFIG_CGROUP_WRITEBACK
3612 struct list_head
*mem_cgroup_cgwb_list(struct mem_cgroup
*memcg
)
3614 return &memcg
->cgwb_list
;
3617 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3619 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3622 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3624 wb_domain_exit(&memcg
->cgwb_domain
);
3627 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3629 wb_domain_size_changed(&memcg
->cgwb_domain
);
3632 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3634 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3636 if (!memcg
->css
.parent
)
3639 return &memcg
->cgwb_domain
;
3643 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3644 * @wb: bdi_writeback in question
3645 * @pfilepages: out parameter for number of file pages
3646 * @pheadroom: out parameter for number of allocatable pages according to memcg
3647 * @pdirty: out parameter for number of dirty pages
3648 * @pwriteback: out parameter for number of pages under writeback
3650 * Determine the numbers of file, headroom, dirty, and writeback pages in
3651 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3652 * is a bit more involved.
3654 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3655 * headroom is calculated as the lowest headroom of itself and the
3656 * ancestors. Note that this doesn't consider the actual amount of
3657 * available memory in the system. The caller should further cap
3658 * *@pheadroom accordingly.
3660 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
3661 unsigned long *pheadroom
, unsigned long *pdirty
,
3662 unsigned long *pwriteback
)
3664 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3665 struct mem_cgroup
*parent
;
3667 *pdirty
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
3669 /* this should eventually include NR_UNSTABLE_NFS */
3670 *pwriteback
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
3671 *pfilepages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
3672 (1 << LRU_ACTIVE_FILE
));
3673 *pheadroom
= PAGE_COUNTER_MAX
;
3675 while ((parent
= parent_mem_cgroup(memcg
))) {
3676 unsigned long ceiling
= min(memcg
->memory
.limit
, memcg
->high
);
3677 unsigned long used
= page_counter_read(&memcg
->memory
);
3679 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
3684 #else /* CONFIG_CGROUP_WRITEBACK */
3686 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3691 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3695 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3699 #endif /* CONFIG_CGROUP_WRITEBACK */
3702 * DO NOT USE IN NEW FILES.
3704 * "cgroup.event_control" implementation.
3706 * This is way over-engineered. It tries to support fully configurable
3707 * events for each user. Such level of flexibility is completely
3708 * unnecessary especially in the light of the planned unified hierarchy.
3710 * Please deprecate this and replace with something simpler if at all
3715 * Unregister event and free resources.
3717 * Gets called from workqueue.
3719 static void memcg_event_remove(struct work_struct
*work
)
3721 struct mem_cgroup_event
*event
=
3722 container_of(work
, struct mem_cgroup_event
, remove
);
3723 struct mem_cgroup
*memcg
= event
->memcg
;
3725 remove_wait_queue(event
->wqh
, &event
->wait
);
3727 event
->unregister_event(memcg
, event
->eventfd
);
3729 /* Notify userspace the event is going away. */
3730 eventfd_signal(event
->eventfd
, 1);
3732 eventfd_ctx_put(event
->eventfd
);
3734 css_put(&memcg
->css
);
3738 * Gets called on POLLHUP on eventfd when user closes it.
3740 * Called with wqh->lock held and interrupts disabled.
3742 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
3743 int sync
, void *key
)
3745 struct mem_cgroup_event
*event
=
3746 container_of(wait
, struct mem_cgroup_event
, wait
);
3747 struct mem_cgroup
*memcg
= event
->memcg
;
3748 unsigned long flags
= (unsigned long)key
;
3750 if (flags
& POLLHUP
) {
3752 * If the event has been detached at cgroup removal, we
3753 * can simply return knowing the other side will cleanup
3756 * We can't race against event freeing since the other
3757 * side will require wqh->lock via remove_wait_queue(),
3760 spin_lock(&memcg
->event_list_lock
);
3761 if (!list_empty(&event
->list
)) {
3762 list_del_init(&event
->list
);
3764 * We are in atomic context, but cgroup_event_remove()
3765 * may sleep, so we have to call it in workqueue.
3767 schedule_work(&event
->remove
);
3769 spin_unlock(&memcg
->event_list_lock
);
3775 static void memcg_event_ptable_queue_proc(struct file
*file
,
3776 wait_queue_head_t
*wqh
, poll_table
*pt
)
3778 struct mem_cgroup_event
*event
=
3779 container_of(pt
, struct mem_cgroup_event
, pt
);
3782 add_wait_queue(wqh
, &event
->wait
);
3786 * DO NOT USE IN NEW FILES.
3788 * Parse input and register new cgroup event handler.
3790 * Input must be in format '<event_fd> <control_fd> <args>'.
3791 * Interpretation of args is defined by control file implementation.
3793 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
3794 char *buf
, size_t nbytes
, loff_t off
)
3796 struct cgroup_subsys_state
*css
= of_css(of
);
3797 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3798 struct mem_cgroup_event
*event
;
3799 struct cgroup_subsys_state
*cfile_css
;
3800 unsigned int efd
, cfd
;
3807 buf
= strstrip(buf
);
3809 efd
= simple_strtoul(buf
, &endp
, 10);
3814 cfd
= simple_strtoul(buf
, &endp
, 10);
3815 if ((*endp
!= ' ') && (*endp
!= '\0'))
3819 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
3823 event
->memcg
= memcg
;
3824 INIT_LIST_HEAD(&event
->list
);
3825 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
3826 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
3827 INIT_WORK(&event
->remove
, memcg_event_remove
);
3835 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
3836 if (IS_ERR(event
->eventfd
)) {
3837 ret
= PTR_ERR(event
->eventfd
);
3844 goto out_put_eventfd
;
3847 /* the process need read permission on control file */
3848 /* AV: shouldn't we check that it's been opened for read instead? */
3849 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
3854 * Determine the event callbacks and set them in @event. This used
3855 * to be done via struct cftype but cgroup core no longer knows
3856 * about these events. The following is crude but the whole thing
3857 * is for compatibility anyway.
3859 * DO NOT ADD NEW FILES.
3861 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
3863 if (!strcmp(name
, "memory.usage_in_bytes")) {
3864 event
->register_event
= mem_cgroup_usage_register_event
;
3865 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
3866 } else if (!strcmp(name
, "memory.oom_control")) {
3867 event
->register_event
= mem_cgroup_oom_register_event
;
3868 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
3869 } else if (!strcmp(name
, "memory.pressure_level")) {
3870 event
->register_event
= vmpressure_register_event
;
3871 event
->unregister_event
= vmpressure_unregister_event
;
3872 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
3873 event
->register_event
= memsw_cgroup_usage_register_event
;
3874 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
3881 * Verify @cfile should belong to @css. Also, remaining events are
3882 * automatically removed on cgroup destruction but the removal is
3883 * asynchronous, so take an extra ref on @css.
3885 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
3886 &memory_cgrp_subsys
);
3888 if (IS_ERR(cfile_css
))
3890 if (cfile_css
!= css
) {
3895 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
3899 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
3901 spin_lock(&memcg
->event_list_lock
);
3902 list_add(&event
->list
, &memcg
->event_list
);
3903 spin_unlock(&memcg
->event_list_lock
);
3915 eventfd_ctx_put(event
->eventfd
);
3924 static struct cftype mem_cgroup_legacy_files
[] = {
3926 .name
= "usage_in_bytes",
3927 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
3928 .read_u64
= mem_cgroup_read_u64
,
3931 .name
= "max_usage_in_bytes",
3932 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
3933 .write
= mem_cgroup_reset
,
3934 .read_u64
= mem_cgroup_read_u64
,
3937 .name
= "limit_in_bytes",
3938 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
3939 .write
= mem_cgroup_write
,
3940 .read_u64
= mem_cgroup_read_u64
,
3943 .name
= "soft_limit_in_bytes",
3944 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
3945 .write
= mem_cgroup_write
,
3946 .read_u64
= mem_cgroup_read_u64
,
3950 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
3951 .write
= mem_cgroup_reset
,
3952 .read_u64
= mem_cgroup_read_u64
,
3956 .seq_show
= memcg_stat_show
,
3959 .name
= "force_empty",
3960 .write
= mem_cgroup_force_empty_write
,
3963 .name
= "use_hierarchy",
3964 .write_u64
= mem_cgroup_hierarchy_write
,
3965 .read_u64
= mem_cgroup_hierarchy_read
,
3968 .name
= "cgroup.event_control", /* XXX: for compat */
3969 .write
= memcg_write_event_control
,
3970 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
3973 .name
= "swappiness",
3974 .read_u64
= mem_cgroup_swappiness_read
,
3975 .write_u64
= mem_cgroup_swappiness_write
,
3978 .name
= "move_charge_at_immigrate",
3979 .read_u64
= mem_cgroup_move_charge_read
,
3980 .write_u64
= mem_cgroup_move_charge_write
,
3983 .name
= "oom_control",
3984 .seq_show
= mem_cgroup_oom_control_read
,
3985 .write_u64
= mem_cgroup_oom_control_write
,
3986 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
3989 .name
= "pressure_level",
3993 .name
= "numa_stat",
3994 .seq_show
= memcg_numa_stat_show
,
3998 .name
= "kmem.limit_in_bytes",
3999 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4000 .write
= mem_cgroup_write
,
4001 .read_u64
= mem_cgroup_read_u64
,
4004 .name
= "kmem.usage_in_bytes",
4005 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4006 .read_u64
= mem_cgroup_read_u64
,
4009 .name
= "kmem.failcnt",
4010 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4011 .write
= mem_cgroup_reset
,
4012 .read_u64
= mem_cgroup_read_u64
,
4015 .name
= "kmem.max_usage_in_bytes",
4016 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4017 .write
= mem_cgroup_reset
,
4018 .read_u64
= mem_cgroup_read_u64
,
4020 #ifdef CONFIG_SLABINFO
4022 .name
= "kmem.slabinfo",
4023 .seq_start
= slab_start
,
4024 .seq_next
= slab_next
,
4025 .seq_stop
= slab_stop
,
4026 .seq_show
= memcg_slab_show
,
4030 .name
= "kmem.tcp.limit_in_bytes",
4031 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4032 .write
= mem_cgroup_write
,
4033 .read_u64
= mem_cgroup_read_u64
,
4036 .name
= "kmem.tcp.usage_in_bytes",
4037 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4038 .read_u64
= mem_cgroup_read_u64
,
4041 .name
= "kmem.tcp.failcnt",
4042 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4043 .write
= mem_cgroup_reset
,
4044 .read_u64
= mem_cgroup_read_u64
,
4047 .name
= "kmem.tcp.max_usage_in_bytes",
4048 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4049 .write
= mem_cgroup_reset
,
4050 .read_u64
= mem_cgroup_read_u64
,
4052 { }, /* terminate */
4056 * Private memory cgroup IDR
4058 * Swap-out records and page cache shadow entries need to store memcg
4059 * references in constrained space, so we maintain an ID space that is
4060 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4061 * memory-controlled cgroups to 64k.
4063 * However, there usually are many references to the oflline CSS after
4064 * the cgroup has been destroyed, such as page cache or reclaimable
4065 * slab objects, that don't need to hang on to the ID. We want to keep
4066 * those dead CSS from occupying IDs, or we might quickly exhaust the
4067 * relatively small ID space and prevent the creation of new cgroups
4068 * even when there are much fewer than 64k cgroups - possibly none.
4070 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4071 * be freed and recycled when it's no longer needed, which is usually
4072 * when the CSS is offlined.
4074 * The only exception to that are records of swapped out tmpfs/shmem
4075 * pages that need to be attributed to live ancestors on swapin. But
4076 * those references are manageable from userspace.
4079 static DEFINE_IDR(mem_cgroup_idr
);
4081 static void mem_cgroup_id_get(struct mem_cgroup
*memcg
)
4083 atomic_inc(&memcg
->id
.ref
);
4086 static void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
4088 if (atomic_dec_and_test(&memcg
->id
.ref
)) {
4089 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4092 /* Memcg ID pins CSS */
4093 css_put(&memcg
->css
);
4098 * mem_cgroup_from_id - look up a memcg from a memcg id
4099 * @id: the memcg id to look up
4101 * Caller must hold rcu_read_lock().
4103 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
4105 WARN_ON_ONCE(!rcu_read_lock_held());
4106 return idr_find(&mem_cgroup_idr
, id
);
4109 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4111 struct mem_cgroup_per_node
*pn
;
4112 struct mem_cgroup_per_zone
*mz
;
4113 int zone
, tmp
= node
;
4115 * This routine is called against possible nodes.
4116 * But it's BUG to call kmalloc() against offline node.
4118 * TODO: this routine can waste much memory for nodes which will
4119 * never be onlined. It's better to use memory hotplug callback
4122 if (!node_state(node
, N_NORMAL_MEMORY
))
4124 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4128 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4129 mz
= &pn
->zoneinfo
[zone
];
4130 lruvec_init(&mz
->lruvec
);
4131 mz
->usage_in_excess
= 0;
4132 mz
->on_tree
= false;
4135 memcg
->nodeinfo
[node
] = pn
;
4139 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4141 kfree(memcg
->nodeinfo
[node
]);
4144 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4148 memcg_wb_domain_exit(memcg
);
4150 free_mem_cgroup_per_zone_info(memcg
, node
);
4151 free_percpu(memcg
->stat
);
4155 static struct mem_cgroup
*mem_cgroup_alloc(void)
4157 struct mem_cgroup
*memcg
;
4161 size
= sizeof(struct mem_cgroup
);
4162 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4164 memcg
= kzalloc(size
, GFP_KERNEL
);
4168 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
4169 1, MEM_CGROUP_ID_MAX
,
4171 if (memcg
->id
.id
< 0)
4174 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4179 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
4182 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4185 INIT_WORK(&memcg
->high_work
, high_work_func
);
4186 memcg
->last_scanned_node
= MAX_NUMNODES
;
4187 INIT_LIST_HEAD(&memcg
->oom_notify
);
4188 mutex_init(&memcg
->thresholds_lock
);
4189 spin_lock_init(&memcg
->move_lock
);
4190 vmpressure_init(&memcg
->vmpressure
);
4191 INIT_LIST_HEAD(&memcg
->event_list
);
4192 spin_lock_init(&memcg
->event_list_lock
);
4193 memcg
->socket_pressure
= jiffies
;
4195 memcg
->kmemcg_id
= -1;
4197 #ifdef CONFIG_CGROUP_WRITEBACK
4198 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4200 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
4203 if (memcg
->id
.id
> 0)
4204 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4205 mem_cgroup_free(memcg
);
4209 static struct cgroup_subsys_state
* __ref
4210 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4212 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
4213 struct mem_cgroup
*memcg
;
4214 long error
= -ENOMEM
;
4216 memcg
= mem_cgroup_alloc();
4218 return ERR_PTR(error
);
4220 memcg
->high
= PAGE_COUNTER_MAX
;
4221 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4223 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4224 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4226 if (parent
&& parent
->use_hierarchy
) {
4227 memcg
->use_hierarchy
= true;
4228 page_counter_init(&memcg
->memory
, &parent
->memory
);
4229 page_counter_init(&memcg
->swap
, &parent
->swap
);
4230 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4231 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4232 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
4234 page_counter_init(&memcg
->memory
, NULL
);
4235 page_counter_init(&memcg
->swap
, NULL
);
4236 page_counter_init(&memcg
->memsw
, NULL
);
4237 page_counter_init(&memcg
->kmem
, NULL
);
4238 page_counter_init(&memcg
->tcpmem
, NULL
);
4240 * Deeper hierachy with use_hierarchy == false doesn't make
4241 * much sense so let cgroup subsystem know about this
4242 * unfortunate state in our controller.
4244 if (parent
!= root_mem_cgroup
)
4245 memory_cgrp_subsys
.broken_hierarchy
= true;
4248 /* The following stuff does not apply to the root */
4250 root_mem_cgroup
= memcg
;
4254 error
= memcg_online_kmem(memcg
);
4258 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4259 static_branch_inc(&memcg_sockets_enabled_key
);
4263 mem_cgroup_free(memcg
);
4264 return ERR_PTR(-ENOMEM
);
4267 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4269 /* Online state pins memcg ID, memcg ID pins CSS */
4270 mem_cgroup_id_get(mem_cgroup_from_css(css
));
4275 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4277 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4278 struct mem_cgroup_event
*event
, *tmp
;
4281 * Unregister events and notify userspace.
4282 * Notify userspace about cgroup removing only after rmdir of cgroup
4283 * directory to avoid race between userspace and kernelspace.
4285 spin_lock(&memcg
->event_list_lock
);
4286 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4287 list_del_init(&event
->list
);
4288 schedule_work(&event
->remove
);
4290 spin_unlock(&memcg
->event_list_lock
);
4292 memcg_offline_kmem(memcg
);
4293 wb_memcg_offline(memcg
);
4295 mem_cgroup_id_put(memcg
);
4298 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4300 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4302 invalidate_reclaim_iterators(memcg
);
4305 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4307 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4309 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4310 static_branch_dec(&memcg_sockets_enabled_key
);
4312 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
4313 static_branch_dec(&memcg_sockets_enabled_key
);
4315 vmpressure_cleanup(&memcg
->vmpressure
);
4316 cancel_work_sync(&memcg
->high_work
);
4317 mem_cgroup_remove_from_trees(memcg
);
4318 memcg_free_kmem(memcg
);
4319 mem_cgroup_free(memcg
);
4323 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4324 * @css: the target css
4326 * Reset the states of the mem_cgroup associated with @css. This is
4327 * invoked when the userland requests disabling on the default hierarchy
4328 * but the memcg is pinned through dependency. The memcg should stop
4329 * applying policies and should revert to the vanilla state as it may be
4330 * made visible again.
4332 * The current implementation only resets the essential configurations.
4333 * This needs to be expanded to cover all the visible parts.
4335 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4337 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4339 page_counter_limit(&memcg
->memory
, PAGE_COUNTER_MAX
);
4340 page_counter_limit(&memcg
->swap
, PAGE_COUNTER_MAX
);
4341 page_counter_limit(&memcg
->memsw
, PAGE_COUNTER_MAX
);
4342 page_counter_limit(&memcg
->kmem
, PAGE_COUNTER_MAX
);
4343 page_counter_limit(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
4345 memcg
->high
= PAGE_COUNTER_MAX
;
4346 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4347 memcg_wb_domain_size_changed(memcg
);
4351 /* Handlers for move charge at task migration. */
4352 static int mem_cgroup_do_precharge(unsigned long count
)
4356 /* Try a single bulk charge without reclaim first, kswapd may wake */
4357 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4359 mc
.precharge
+= count
;
4363 /* Try charges one by one with reclaim */
4365 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4379 enum mc_target_type
{
4385 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4386 unsigned long addr
, pte_t ptent
)
4388 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4390 if (!page
|| !page_mapped(page
))
4392 if (PageAnon(page
)) {
4393 if (!(mc
.flags
& MOVE_ANON
))
4396 if (!(mc
.flags
& MOVE_FILE
))
4399 if (!get_page_unless_zero(page
))
4406 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4407 pte_t ptent
, swp_entry_t
*entry
)
4409 struct page
*page
= NULL
;
4410 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4412 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4415 * Because lookup_swap_cache() updates some statistics counter,
4416 * we call find_get_page() with swapper_space directly.
4418 page
= find_get_page(swap_address_space(ent
), ent
.val
);
4419 if (do_memsw_account())
4420 entry
->val
= ent
.val
;
4425 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4426 pte_t ptent
, swp_entry_t
*entry
)
4432 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4433 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4435 struct page
*page
= NULL
;
4436 struct address_space
*mapping
;
4439 if (!vma
->vm_file
) /* anonymous vma */
4441 if (!(mc
.flags
& MOVE_FILE
))
4444 mapping
= vma
->vm_file
->f_mapping
;
4445 pgoff
= linear_page_index(vma
, addr
);
4447 /* page is moved even if it's not RSS of this task(page-faulted). */
4449 /* shmem/tmpfs may report page out on swap: account for that too. */
4450 if (shmem_mapping(mapping
)) {
4451 page
= find_get_entry(mapping
, pgoff
);
4452 if (radix_tree_exceptional_entry(page
)) {
4453 swp_entry_t swp
= radix_to_swp_entry(page
);
4454 if (do_memsw_account())
4456 page
= find_get_page(swap_address_space(swp
), swp
.val
);
4459 page
= find_get_page(mapping
, pgoff
);
4461 page
= find_get_page(mapping
, pgoff
);
4467 * mem_cgroup_move_account - move account of the page
4469 * @compound: charge the page as compound or small page
4470 * @from: mem_cgroup which the page is moved from.
4471 * @to: mem_cgroup which the page is moved to. @from != @to.
4473 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4475 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4478 static int mem_cgroup_move_account(struct page
*page
,
4480 struct mem_cgroup
*from
,
4481 struct mem_cgroup
*to
)
4483 unsigned long flags
;
4484 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
4488 VM_BUG_ON(from
== to
);
4489 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4490 VM_BUG_ON(compound
&& !PageTransHuge(page
));
4493 * Prevent mem_cgroup_migrate() from looking at
4494 * page->mem_cgroup of its source page while we change it.
4497 if (!trylock_page(page
))
4501 if (page
->mem_cgroup
!= from
)
4504 anon
= PageAnon(page
);
4506 spin_lock_irqsave(&from
->move_lock
, flags
);
4508 if (!anon
&& page_mapped(page
)) {
4509 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4511 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4516 * move_lock grabbed above and caller set from->moving_account, so
4517 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4518 * So mapping should be stable for dirty pages.
4520 if (!anon
&& PageDirty(page
)) {
4521 struct address_space
*mapping
= page_mapping(page
);
4523 if (mapping_cap_account_dirty(mapping
)) {
4524 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4526 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4531 if (PageWriteback(page
)) {
4532 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4534 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4539 * It is safe to change page->mem_cgroup here because the page
4540 * is referenced, charged, and isolated - we can't race with
4541 * uncharging, charging, migration, or LRU putback.
4544 /* caller should have done css_get */
4545 page
->mem_cgroup
= to
;
4546 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4550 local_irq_disable();
4551 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
4552 memcg_check_events(to
, page
);
4553 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
4554 memcg_check_events(from
, page
);
4563 * get_mctgt_type - get target type of moving charge
4564 * @vma: the vma the pte to be checked belongs
4565 * @addr: the address corresponding to the pte to be checked
4566 * @ptent: the pte to be checked
4567 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4570 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4571 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4572 * move charge. if @target is not NULL, the page is stored in target->page
4573 * with extra refcnt got(Callers should handle it).
4574 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4575 * target for charge migration. if @target is not NULL, the entry is stored
4578 * Called with pte lock held.
4581 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4582 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4584 struct page
*page
= NULL
;
4585 enum mc_target_type ret
= MC_TARGET_NONE
;
4586 swp_entry_t ent
= { .val
= 0 };
4588 if (pte_present(ptent
))
4589 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4590 else if (is_swap_pte(ptent
))
4591 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
4592 else if (pte_none(ptent
))
4593 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4595 if (!page
&& !ent
.val
)
4599 * Do only loose check w/o serialization.
4600 * mem_cgroup_move_account() checks the page is valid or
4601 * not under LRU exclusion.
4603 if (page
->mem_cgroup
== mc
.from
) {
4604 ret
= MC_TARGET_PAGE
;
4606 target
->page
= page
;
4608 if (!ret
|| !target
)
4611 /* There is a swap entry and a page doesn't exist or isn't charged */
4612 if (ent
.val
&& !ret
&&
4613 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4614 ret
= MC_TARGET_SWAP
;
4621 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4623 * We don't consider swapping or file mapped pages because THP does not
4624 * support them for now.
4625 * Caller should make sure that pmd_trans_huge(pmd) is true.
4627 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4628 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4630 struct page
*page
= NULL
;
4631 enum mc_target_type ret
= MC_TARGET_NONE
;
4633 page
= pmd_page(pmd
);
4634 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4635 if (!(mc
.flags
& MOVE_ANON
))
4637 if (page
->mem_cgroup
== mc
.from
) {
4638 ret
= MC_TARGET_PAGE
;
4641 target
->page
= page
;
4647 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4648 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4650 return MC_TARGET_NONE
;
4654 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4655 unsigned long addr
, unsigned long end
,
4656 struct mm_walk
*walk
)
4658 struct vm_area_struct
*vma
= walk
->vma
;
4662 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4664 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4665 mc
.precharge
+= HPAGE_PMD_NR
;
4670 if (pmd_trans_unstable(pmd
))
4672 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4673 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4674 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4675 mc
.precharge
++; /* increment precharge temporarily */
4676 pte_unmap_unlock(pte
- 1, ptl
);
4682 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4684 unsigned long precharge
;
4686 struct mm_walk mem_cgroup_count_precharge_walk
= {
4687 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4690 down_read(&mm
->mmap_sem
);
4691 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk
);
4692 up_read(&mm
->mmap_sem
);
4694 precharge
= mc
.precharge
;
4700 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4702 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4704 VM_BUG_ON(mc
.moving_task
);
4705 mc
.moving_task
= current
;
4706 return mem_cgroup_do_precharge(precharge
);
4709 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4710 static void __mem_cgroup_clear_mc(void)
4712 struct mem_cgroup
*from
= mc
.from
;
4713 struct mem_cgroup
*to
= mc
.to
;
4715 /* we must uncharge all the leftover precharges from mc.to */
4717 cancel_charge(mc
.to
, mc
.precharge
);
4721 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4722 * we must uncharge here.
4724 if (mc
.moved_charge
) {
4725 cancel_charge(mc
.from
, mc
.moved_charge
);
4726 mc
.moved_charge
= 0;
4728 /* we must fixup refcnts and charges */
4729 if (mc
.moved_swap
) {
4730 /* uncharge swap account from the old cgroup */
4731 if (!mem_cgroup_is_root(mc
.from
))
4732 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4735 * we charged both to->memory and to->memsw, so we
4736 * should uncharge to->memory.
4738 if (!mem_cgroup_is_root(mc
.to
))
4739 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4741 css_put_many(&mc
.from
->css
, mc
.moved_swap
);
4743 /* we've already done css_get(mc.to) */
4746 memcg_oom_recover(from
);
4747 memcg_oom_recover(to
);
4748 wake_up_all(&mc
.waitq
);
4751 static void mem_cgroup_clear_mc(void)
4753 struct mm_struct
*mm
= mc
.mm
;
4756 * we must clear moving_task before waking up waiters at the end of
4759 mc
.moving_task
= NULL
;
4760 __mem_cgroup_clear_mc();
4761 spin_lock(&mc
.lock
);
4765 spin_unlock(&mc
.lock
);
4770 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4772 struct cgroup_subsys_state
*css
;
4773 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
4774 struct mem_cgroup
*from
;
4775 struct task_struct
*leader
, *p
;
4776 struct mm_struct
*mm
;
4777 unsigned long move_flags
;
4780 /* charge immigration isn't supported on the default hierarchy */
4781 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4785 * Multi-process migrations only happen on the default hierarchy
4786 * where charge immigration is not used. Perform charge
4787 * immigration if @tset contains a leader and whine if there are
4791 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
4794 memcg
= mem_cgroup_from_css(css
);
4800 * We are now commited to this value whatever it is. Changes in this
4801 * tunable will only affect upcoming migrations, not the current one.
4802 * So we need to save it, and keep it going.
4804 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
4808 from
= mem_cgroup_from_task(p
);
4810 VM_BUG_ON(from
== memcg
);
4812 mm
= get_task_mm(p
);
4815 /* We move charges only when we move a owner of the mm */
4816 if (mm
->owner
== p
) {
4819 VM_BUG_ON(mc
.precharge
);
4820 VM_BUG_ON(mc
.moved_charge
);
4821 VM_BUG_ON(mc
.moved_swap
);
4823 spin_lock(&mc
.lock
);
4827 mc
.flags
= move_flags
;
4828 spin_unlock(&mc
.lock
);
4829 /* We set mc.moving_task later */
4831 ret
= mem_cgroup_precharge_mc(mm
);
4833 mem_cgroup_clear_mc();
4840 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4843 mem_cgroup_clear_mc();
4846 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4847 unsigned long addr
, unsigned long end
,
4848 struct mm_walk
*walk
)
4851 struct vm_area_struct
*vma
= walk
->vma
;
4854 enum mc_target_type target_type
;
4855 union mc_target target
;
4858 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4860 if (mc
.precharge
< HPAGE_PMD_NR
) {
4864 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
4865 if (target_type
== MC_TARGET_PAGE
) {
4867 if (!isolate_lru_page(page
)) {
4868 if (!mem_cgroup_move_account(page
, true,
4870 mc
.precharge
-= HPAGE_PMD_NR
;
4871 mc
.moved_charge
+= HPAGE_PMD_NR
;
4873 putback_lru_page(page
);
4881 if (pmd_trans_unstable(pmd
))
4884 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4885 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4886 pte_t ptent
= *(pte
++);
4892 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
4893 case MC_TARGET_PAGE
:
4896 * We can have a part of the split pmd here. Moving it
4897 * can be done but it would be too convoluted so simply
4898 * ignore such a partial THP and keep it in original
4899 * memcg. There should be somebody mapping the head.
4901 if (PageTransCompound(page
))
4903 if (isolate_lru_page(page
))
4905 if (!mem_cgroup_move_account(page
, false,
4908 /* we uncharge from mc.from later. */
4911 putback_lru_page(page
);
4912 put
: /* get_mctgt_type() gets the page */
4915 case MC_TARGET_SWAP
:
4917 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
4919 /* we fixup refcnts and charges later. */
4927 pte_unmap_unlock(pte
- 1, ptl
);
4932 * We have consumed all precharges we got in can_attach().
4933 * We try charge one by one, but don't do any additional
4934 * charges to mc.to if we have failed in charge once in attach()
4937 ret
= mem_cgroup_do_precharge(1);
4945 static void mem_cgroup_move_charge(void)
4947 struct mm_walk mem_cgroup_move_charge_walk
= {
4948 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4952 lru_add_drain_all();
4954 * Signal lock_page_memcg() to take the memcg's move_lock
4955 * while we're moving its pages to another memcg. Then wait
4956 * for already started RCU-only updates to finish.
4958 atomic_inc(&mc
.from
->moving_account
);
4961 if (unlikely(!down_read_trylock(&mc
.mm
->mmap_sem
))) {
4963 * Someone who are holding the mmap_sem might be waiting in
4964 * waitq. So we cancel all extra charges, wake up all waiters,
4965 * and retry. Because we cancel precharges, we might not be able
4966 * to move enough charges, but moving charge is a best-effort
4967 * feature anyway, so it wouldn't be a big problem.
4969 __mem_cgroup_clear_mc();
4974 * When we have consumed all precharges and failed in doing
4975 * additional charge, the page walk just aborts.
4977 walk_page_range(0, ~0UL, &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
] * PAGE_SIZE
);
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
);
5555 page
->mem_cgroup
= NULL
;
5556 } while (next
!= page_list
);
5559 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5560 nr_huge
, nr_kmem
, page
);
5564 * mem_cgroup_uncharge - uncharge a page
5565 * @page: page to uncharge
5567 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5568 * mem_cgroup_commit_charge().
5570 void mem_cgroup_uncharge(struct page
*page
)
5572 if (mem_cgroup_disabled())
5575 /* Don't touch page->lru of any random page, pre-check: */
5576 if (!page
->mem_cgroup
)
5579 INIT_LIST_HEAD(&page
->lru
);
5580 uncharge_list(&page
->lru
);
5584 * mem_cgroup_uncharge_list - uncharge a list of page
5585 * @page_list: list of pages to uncharge
5587 * Uncharge a list of pages previously charged with
5588 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5590 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5592 if (mem_cgroup_disabled())
5595 if (!list_empty(page_list
))
5596 uncharge_list(page_list
);
5600 * mem_cgroup_migrate - charge a page's replacement
5601 * @oldpage: currently circulating page
5602 * @newpage: replacement page
5604 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5605 * be uncharged upon free.
5607 * Both pages must be locked, @newpage->mapping must be set up.
5609 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
5611 struct mem_cgroup
*memcg
;
5612 unsigned int nr_pages
;
5614 unsigned long flags
;
5616 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5617 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5618 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5619 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5622 if (mem_cgroup_disabled())
5625 /* Page cache replacement: new page already charged? */
5626 if (newpage
->mem_cgroup
)
5629 /* Swapcache readahead pages can get replaced before being charged */
5630 memcg
= oldpage
->mem_cgroup
;
5634 /* Force-charge the new page. The old one will be freed soon */
5635 compound
= PageTransHuge(newpage
);
5636 nr_pages
= compound
? hpage_nr_pages(newpage
) : 1;
5638 page_counter_charge(&memcg
->memory
, nr_pages
);
5639 if (do_memsw_account())
5640 page_counter_charge(&memcg
->memsw
, nr_pages
);
5641 css_get_many(&memcg
->css
, nr_pages
);
5643 commit_charge(newpage
, memcg
, false);
5645 local_irq_save(flags
);
5646 mem_cgroup_charge_statistics(memcg
, newpage
, compound
, nr_pages
);
5647 memcg_check_events(memcg
, newpage
);
5648 local_irq_restore(flags
);
5651 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
5652 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
5654 void sock_update_memcg(struct sock
*sk
)
5656 struct mem_cgroup
*memcg
;
5658 /* Socket cloning can throw us here with sk_cgrp already
5659 * filled. It won't however, necessarily happen from
5660 * process context. So the test for root memcg given
5661 * the current task's memcg won't help us in this case.
5663 * Respecting the original socket's memcg is a better
5664 * decision in this case.
5667 BUG_ON(mem_cgroup_is_root(sk
->sk_memcg
));
5668 css_get(&sk
->sk_memcg
->css
);
5673 memcg
= mem_cgroup_from_task(current
);
5674 if (memcg
== root_mem_cgroup
)
5676 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
5678 if (css_tryget_online(&memcg
->css
))
5679 sk
->sk_memcg
= memcg
;
5683 EXPORT_SYMBOL(sock_update_memcg
);
5685 void sock_release_memcg(struct sock
*sk
)
5687 WARN_ON(!sk
->sk_memcg
);
5688 css_put(&sk
->sk_memcg
->css
);
5692 * mem_cgroup_charge_skmem - charge socket memory
5693 * @memcg: memcg to charge
5694 * @nr_pages: number of pages to charge
5696 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5697 * @memcg's configured limit, %false if the charge had to be forced.
5699 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5701 gfp_t gfp_mask
= GFP_KERNEL
;
5703 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5704 struct page_counter
*fail
;
5706 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
5707 memcg
->tcpmem_pressure
= 0;
5710 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
5711 memcg
->tcpmem_pressure
= 1;
5715 /* Don't block in the packet receive path */
5717 gfp_mask
= GFP_NOWAIT
;
5719 this_cpu_add(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5721 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
5724 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
5729 * mem_cgroup_uncharge_skmem - uncharge socket memory
5730 * @memcg - memcg to uncharge
5731 * @nr_pages - number of pages to uncharge
5733 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5735 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5736 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
5740 this_cpu_sub(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5742 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5743 css_put_many(&memcg
->css
, nr_pages
);
5746 static int __init
cgroup_memory(char *s
)
5750 while ((token
= strsep(&s
, ",")) != NULL
) {
5753 if (!strcmp(token
, "nosocket"))
5754 cgroup_memory_nosocket
= true;
5755 if (!strcmp(token
, "nokmem"))
5756 cgroup_memory_nokmem
= true;
5760 __setup("cgroup.memory=", cgroup_memory
);
5763 * subsys_initcall() for memory controller.
5765 * Some parts like hotcpu_notifier() have to be initialized from this context
5766 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5767 * everything that doesn't depend on a specific mem_cgroup structure should
5768 * be initialized from here.
5770 static int __init
mem_cgroup_init(void)
5774 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5776 for_each_possible_cpu(cpu
)
5777 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5780 for_each_node(node
) {
5781 struct mem_cgroup_tree_per_node
*rtpn
;
5784 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5785 node_online(node
) ? node
: NUMA_NO_NODE
);
5787 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5788 struct mem_cgroup_tree_per_zone
*rtpz
;
5790 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
5791 rtpz
->rb_root
= RB_ROOT
;
5792 spin_lock_init(&rtpz
->lock
);
5794 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5799 subsys_initcall(mem_cgroup_init
);
5801 #ifdef CONFIG_MEMCG_SWAP
5803 * mem_cgroup_swapout - transfer a memsw charge to swap
5804 * @page: page whose memsw charge to transfer
5805 * @entry: swap entry to move the charge to
5807 * Transfer the memsw charge of @page to @entry.
5809 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5811 struct mem_cgroup
*memcg
;
5812 unsigned short oldid
;
5814 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5815 VM_BUG_ON_PAGE(page_count(page
), page
);
5817 if (!do_memsw_account())
5820 memcg
= page
->mem_cgroup
;
5822 /* Readahead page, never charged */
5826 mem_cgroup_id_get(memcg
);
5827 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5828 VM_BUG_ON_PAGE(oldid
, page
);
5829 mem_cgroup_swap_statistics(memcg
, true);
5831 page
->mem_cgroup
= NULL
;
5833 if (!mem_cgroup_is_root(memcg
))
5834 page_counter_uncharge(&memcg
->memory
, 1);
5837 * Interrupts should be disabled here because the caller holds the
5838 * mapping->tree_lock lock which is taken with interrupts-off. It is
5839 * important here to have the interrupts disabled because it is the
5840 * only synchronisation we have for udpating the per-CPU variables.
5842 VM_BUG_ON(!irqs_disabled());
5843 mem_cgroup_charge_statistics(memcg
, page
, false, -1);
5844 memcg_check_events(memcg
, page
);
5846 if (!mem_cgroup_is_root(memcg
))
5847 css_put(&memcg
->css
);
5851 * mem_cgroup_try_charge_swap - try charging a swap entry
5852 * @page: page being added to swap
5853 * @entry: swap entry to charge
5855 * Try to charge @entry to the memcg that @page belongs to.
5857 * Returns 0 on success, -ENOMEM on failure.
5859 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
5861 struct mem_cgroup
*memcg
;
5862 struct page_counter
*counter
;
5863 unsigned short oldid
;
5865 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
5868 memcg
= page
->mem_cgroup
;
5870 /* Readahead page, never charged */
5874 if (!mem_cgroup_is_root(memcg
) &&
5875 !page_counter_try_charge(&memcg
->swap
, 1, &counter
))
5878 mem_cgroup_id_get(memcg
);
5879 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5880 VM_BUG_ON_PAGE(oldid
, page
);
5881 mem_cgroup_swap_statistics(memcg
, true);
5887 * mem_cgroup_uncharge_swap - uncharge a swap entry
5888 * @entry: swap entry to uncharge
5890 * Drop the swap charge associated with @entry.
5892 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5894 struct mem_cgroup
*memcg
;
5897 if (!do_swap_account
)
5900 id
= swap_cgroup_record(entry
, 0);
5902 memcg
= mem_cgroup_from_id(id
);
5904 if (!mem_cgroup_is_root(memcg
)) {
5905 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5906 page_counter_uncharge(&memcg
->swap
, 1);
5908 page_counter_uncharge(&memcg
->memsw
, 1);
5910 mem_cgroup_swap_statistics(memcg
, false);
5911 mem_cgroup_id_put(memcg
);
5916 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
5918 long nr_swap_pages
= get_nr_swap_pages();
5920 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5921 return nr_swap_pages
;
5922 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5923 nr_swap_pages
= min_t(long, nr_swap_pages
,
5924 READ_ONCE(memcg
->swap
.limit
) -
5925 page_counter_read(&memcg
->swap
));
5926 return nr_swap_pages
;
5929 bool mem_cgroup_swap_full(struct page
*page
)
5931 struct mem_cgroup
*memcg
;
5933 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5937 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5940 memcg
= page
->mem_cgroup
;
5944 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5945 if (page_counter_read(&memcg
->swap
) * 2 >= memcg
->swap
.limit
)
5951 /* for remember boot option*/
5952 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5953 static int really_do_swap_account __initdata
= 1;
5955 static int really_do_swap_account __initdata
;
5958 static int __init
enable_swap_account(char *s
)
5960 if (!strcmp(s
, "1"))
5961 really_do_swap_account
= 1;
5962 else if (!strcmp(s
, "0"))
5963 really_do_swap_account
= 0;
5966 __setup("swapaccount=", enable_swap_account
);
5968 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
5971 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5973 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
5976 static int swap_max_show(struct seq_file
*m
, void *v
)
5978 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5979 unsigned long max
= READ_ONCE(memcg
->swap
.limit
);
5981 if (max
== PAGE_COUNTER_MAX
)
5982 seq_puts(m
, "max\n");
5984 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5989 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
5990 char *buf
, size_t nbytes
, loff_t off
)
5992 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5996 buf
= strstrip(buf
);
5997 err
= page_counter_memparse(buf
, "max", &max
);
6001 mutex_lock(&memcg_limit_mutex
);
6002 err
= page_counter_limit(&memcg
->swap
, max
);
6003 mutex_unlock(&memcg_limit_mutex
);
6010 static struct cftype swap_files
[] = {
6012 .name
= "swap.current",
6013 .flags
= CFTYPE_NOT_ON_ROOT
,
6014 .read_u64
= swap_current_read
,
6018 .flags
= CFTYPE_NOT_ON_ROOT
,
6019 .seq_show
= swap_max_show
,
6020 .write
= swap_max_write
,
6025 static struct cftype memsw_cgroup_files
[] = {
6027 .name
= "memsw.usage_in_bytes",
6028 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6029 .read_u64
= mem_cgroup_read_u64
,
6032 .name
= "memsw.max_usage_in_bytes",
6033 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6034 .write
= mem_cgroup_reset
,
6035 .read_u64
= mem_cgroup_read_u64
,
6038 .name
= "memsw.limit_in_bytes",
6039 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6040 .write
= mem_cgroup_write
,
6041 .read_u64
= mem_cgroup_read_u64
,
6044 .name
= "memsw.failcnt",
6045 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6046 .write
= mem_cgroup_reset
,
6047 .read_u64
= mem_cgroup_read_u64
,
6049 { }, /* terminate */
6052 static int __init
mem_cgroup_swap_init(void)
6054 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6055 do_swap_account
= 1;
6056 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
6058 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
6059 memsw_cgroup_files
));
6063 subsys_initcall(mem_cgroup_swap_init
);
6065 #endif /* CONFIG_MEMCG_SWAP */