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 mem_cgroup
*from
;
211 struct mem_cgroup
*to
;
213 unsigned long precharge
;
214 unsigned long moved_charge
;
215 unsigned long moved_swap
;
216 struct task_struct
*moving_task
; /* a task moving charges */
217 wait_queue_head_t waitq
; /* a waitq for other context */
219 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
220 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
224 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
225 * limit reclaim to prevent infinite loops, if they ever occur.
227 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
228 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
231 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
232 MEM_CGROUP_CHARGE_TYPE_ANON
,
233 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
234 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
238 /* for encoding cft->private value on file */
247 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
248 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
249 #define MEMFILE_ATTR(val) ((val) & 0xffff)
250 /* Used for OOM nofiier */
251 #define OOM_CONTROL (0)
254 * The memcg_create_mutex will be held whenever a new cgroup is created.
255 * As a consequence, any change that needs to protect against new child cgroups
256 * appearing has to hold it as well.
258 static DEFINE_MUTEX(memcg_create_mutex
);
260 /* Some nice accessors for the vmpressure. */
261 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
264 memcg
= root_mem_cgroup
;
265 return &memcg
->vmpressure
;
268 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
270 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
273 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
275 return (memcg
== root_mem_cgroup
);
279 * We restrict the id in the range of [1, 65535], so it can fit into
282 #define MEM_CGROUP_ID_MAX USHRT_MAX
284 static inline unsigned short mem_cgroup_id(struct mem_cgroup
*memcg
)
286 return memcg
->css
.id
;
290 * A helper function to get mem_cgroup from ID. must be called under
291 * rcu_read_lock(). The caller is responsible for calling
292 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
293 * refcnt from swap can be called against removed memcg.)
295 static inline struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
297 struct cgroup_subsys_state
*css
;
299 css
= css_from_id(id
, &memory_cgrp_subsys
);
300 return mem_cgroup_from_css(css
);
305 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
306 * The main reason for not using cgroup id for this:
307 * this works better in sparse environments, where we have a lot of memcgs,
308 * but only a few kmem-limited. Or also, if we have, for instance, 200
309 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
310 * 200 entry array for that.
312 * The current size of the caches array is stored in memcg_nr_cache_ids. It
313 * will double each time we have to increase it.
315 static DEFINE_IDA(memcg_cache_ida
);
316 int memcg_nr_cache_ids
;
318 /* Protects memcg_nr_cache_ids */
319 static DECLARE_RWSEM(memcg_cache_ids_sem
);
321 void memcg_get_cache_ids(void)
323 down_read(&memcg_cache_ids_sem
);
326 void memcg_put_cache_ids(void)
328 up_read(&memcg_cache_ids_sem
);
332 * MIN_SIZE is different than 1, because we would like to avoid going through
333 * the alloc/free process all the time. In a small machine, 4 kmem-limited
334 * cgroups is a reasonable guess. In the future, it could be a parameter or
335 * tunable, but that is strictly not necessary.
337 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
338 * this constant directly from cgroup, but it is understandable that this is
339 * better kept as an internal representation in cgroup.c. In any case, the
340 * cgrp_id space is not getting any smaller, and we don't have to necessarily
341 * increase ours as well if it increases.
343 #define MEMCG_CACHES_MIN_SIZE 4
344 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
347 * A lot of the calls to the cache allocation functions are expected to be
348 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
349 * conditional to this static branch, we'll have to allow modules that does
350 * kmem_cache_alloc and the such to see this symbol as well
352 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
353 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
355 #endif /* !CONFIG_SLOB */
357 static struct mem_cgroup_per_zone
*
358 mem_cgroup_zone_zoneinfo(struct mem_cgroup
*memcg
, struct zone
*zone
)
360 int nid
= zone_to_nid(zone
);
361 int zid
= zone_idx(zone
);
363 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
367 * mem_cgroup_css_from_page - css of the memcg associated with a page
368 * @page: page of interest
370 * If memcg is bound to the default hierarchy, css of the memcg associated
371 * with @page is returned. The returned css remains associated with @page
372 * until it is released.
374 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
377 * XXX: The above description of behavior on the default hierarchy isn't
378 * strictly true yet as replace_page_cache_page() can modify the
379 * association before @page is released even on the default hierarchy;
380 * however, the current and planned usages don't mix the the two functions
381 * and replace_page_cache_page() will soon be updated to make the invariant
384 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
386 struct mem_cgroup
*memcg
;
388 memcg
= page
->mem_cgroup
;
390 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
391 memcg
= root_mem_cgroup
;
397 * page_cgroup_ino - return inode number of the memcg a page is charged to
400 * Look up the closest online ancestor of the memory cgroup @page is charged to
401 * and return its inode number or 0 if @page is not charged to any cgroup. It
402 * is safe to call this function without holding a reference to @page.
404 * Note, this function is inherently racy, because there is nothing to prevent
405 * the cgroup inode from getting torn down and potentially reallocated a moment
406 * after page_cgroup_ino() returns, so it only should be used by callers that
407 * do not care (such as procfs interfaces).
409 ino_t
page_cgroup_ino(struct page
*page
)
411 struct mem_cgroup
*memcg
;
412 unsigned long ino
= 0;
415 memcg
= READ_ONCE(page
->mem_cgroup
);
416 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
417 memcg
= parent_mem_cgroup(memcg
);
419 ino
= cgroup_ino(memcg
->css
.cgroup
);
424 static struct mem_cgroup_per_zone
*
425 mem_cgroup_page_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
427 int nid
= page_to_nid(page
);
428 int zid
= page_zonenum(page
);
430 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
433 static struct mem_cgroup_tree_per_zone
*
434 soft_limit_tree_node_zone(int nid
, int zid
)
436 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
439 static struct mem_cgroup_tree_per_zone
*
440 soft_limit_tree_from_page(struct page
*page
)
442 int nid
= page_to_nid(page
);
443 int zid
= page_zonenum(page
);
445 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
448 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone
*mz
,
449 struct mem_cgroup_tree_per_zone
*mctz
,
450 unsigned long new_usage_in_excess
)
452 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
453 struct rb_node
*parent
= NULL
;
454 struct mem_cgroup_per_zone
*mz_node
;
459 mz
->usage_in_excess
= new_usage_in_excess
;
460 if (!mz
->usage_in_excess
)
464 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
466 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
469 * We can't avoid mem cgroups that are over their soft
470 * limit by the same amount
472 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
475 rb_link_node(&mz
->tree_node
, parent
, p
);
476 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
480 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
481 struct mem_cgroup_tree_per_zone
*mctz
)
485 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
489 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
490 struct mem_cgroup_tree_per_zone
*mctz
)
494 spin_lock_irqsave(&mctz
->lock
, flags
);
495 __mem_cgroup_remove_exceeded(mz
, mctz
);
496 spin_unlock_irqrestore(&mctz
->lock
, flags
);
499 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
501 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
502 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
503 unsigned long excess
= 0;
505 if (nr_pages
> soft_limit
)
506 excess
= nr_pages
- soft_limit
;
511 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
513 unsigned long excess
;
514 struct mem_cgroup_per_zone
*mz
;
515 struct mem_cgroup_tree_per_zone
*mctz
;
517 mctz
= soft_limit_tree_from_page(page
);
519 * Necessary to update all ancestors when hierarchy is used.
520 * because their event counter is not touched.
522 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
523 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
524 excess
= soft_limit_excess(memcg
);
526 * We have to update the tree if mz is on RB-tree or
527 * mem is over its softlimit.
529 if (excess
|| mz
->on_tree
) {
532 spin_lock_irqsave(&mctz
->lock
, flags
);
533 /* if on-tree, remove it */
535 __mem_cgroup_remove_exceeded(mz
, mctz
);
537 * Insert again. mz->usage_in_excess will be updated.
538 * If excess is 0, no tree ops.
540 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
541 spin_unlock_irqrestore(&mctz
->lock
, flags
);
546 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
548 struct mem_cgroup_tree_per_zone
*mctz
;
549 struct mem_cgroup_per_zone
*mz
;
553 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
554 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
555 mctz
= soft_limit_tree_node_zone(nid
, zid
);
556 mem_cgroup_remove_exceeded(mz
, mctz
);
561 static struct mem_cgroup_per_zone
*
562 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
564 struct rb_node
*rightmost
= NULL
;
565 struct mem_cgroup_per_zone
*mz
;
569 rightmost
= rb_last(&mctz
->rb_root
);
571 goto done
; /* Nothing to reclaim from */
573 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
575 * Remove the node now but someone else can add it back,
576 * we will to add it back at the end of reclaim to its correct
577 * position in the tree.
579 __mem_cgroup_remove_exceeded(mz
, mctz
);
580 if (!soft_limit_excess(mz
->memcg
) ||
581 !css_tryget_online(&mz
->memcg
->css
))
587 static struct mem_cgroup_per_zone
*
588 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
590 struct mem_cgroup_per_zone
*mz
;
592 spin_lock_irq(&mctz
->lock
);
593 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
594 spin_unlock_irq(&mctz
->lock
);
599 * Return page count for single (non recursive) @memcg.
601 * Implementation Note: reading percpu statistics for memcg.
603 * Both of vmstat[] and percpu_counter has threshold and do periodic
604 * synchronization to implement "quick" read. There are trade-off between
605 * reading cost and precision of value. Then, we may have a chance to implement
606 * a periodic synchronization of counter in memcg's counter.
608 * But this _read() function is used for user interface now. The user accounts
609 * memory usage by memory cgroup and he _always_ requires exact value because
610 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
611 * have to visit all online cpus and make sum. So, for now, unnecessary
612 * synchronization is not implemented. (just implemented for cpu hotplug)
614 * If there are kernel internal actions which can make use of some not-exact
615 * value, and reading all cpu value can be performance bottleneck in some
616 * common workload, threshold and synchronization as vmstat[] should be
620 mem_cgroup_read_stat(struct mem_cgroup
*memcg
, enum mem_cgroup_stat_index idx
)
625 /* Per-cpu values can be negative, use a signed accumulator */
626 for_each_possible_cpu(cpu
)
627 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
629 * Summing races with updates, so val may be negative. Avoid exposing
630 * transient negative values.
637 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
638 enum mem_cgroup_events_index idx
)
640 unsigned long val
= 0;
643 for_each_possible_cpu(cpu
)
644 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
648 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
650 bool compound
, int nr_pages
)
653 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
654 * counted as CACHE even if it's on ANON LRU.
657 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
660 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
664 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
665 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
669 /* pagein of a big page is an event. So, ignore page size */
671 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
673 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
674 nr_pages
= -nr_pages
; /* for event */
677 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
680 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
682 unsigned int lru_mask
)
684 unsigned long nr
= 0;
687 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
689 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
690 struct mem_cgroup_per_zone
*mz
;
694 if (!(BIT(lru
) & lru_mask
))
696 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
697 nr
+= mz
->lru_size
[lru
];
703 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
704 unsigned int lru_mask
)
706 unsigned long nr
= 0;
709 for_each_node_state(nid
, N_MEMORY
)
710 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
714 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
715 enum mem_cgroup_events_target target
)
717 unsigned long val
, next
;
719 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
720 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
721 /* from time_after() in jiffies.h */
722 if ((long)next
- (long)val
< 0) {
724 case MEM_CGROUP_TARGET_THRESH
:
725 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
727 case MEM_CGROUP_TARGET_SOFTLIMIT
:
728 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
730 case MEM_CGROUP_TARGET_NUMAINFO
:
731 next
= val
+ NUMAINFO_EVENTS_TARGET
;
736 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
743 * Check events in order.
746 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
748 /* threshold event is triggered in finer grain than soft limit */
749 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
750 MEM_CGROUP_TARGET_THRESH
))) {
752 bool do_numainfo __maybe_unused
;
754 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
755 MEM_CGROUP_TARGET_SOFTLIMIT
);
757 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
758 MEM_CGROUP_TARGET_NUMAINFO
);
760 mem_cgroup_threshold(memcg
);
761 if (unlikely(do_softlimit
))
762 mem_cgroup_update_tree(memcg
, page
);
764 if (unlikely(do_numainfo
))
765 atomic_inc(&memcg
->numainfo_events
);
770 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
773 * mm_update_next_owner() may clear mm->owner to NULL
774 * if it races with swapoff, page migration, etc.
775 * So this can be called with p == NULL.
780 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
782 EXPORT_SYMBOL(mem_cgroup_from_task
);
784 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
786 struct mem_cgroup
*memcg
= NULL
;
791 * Page cache insertions can happen withou an
792 * actual mm context, e.g. during disk probing
793 * on boot, loopback IO, acct() writes etc.
796 memcg
= root_mem_cgroup
;
798 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
799 if (unlikely(!memcg
))
800 memcg
= root_mem_cgroup
;
802 } while (!css_tryget_online(&memcg
->css
));
808 * mem_cgroup_iter - iterate over memory cgroup hierarchy
809 * @root: hierarchy root
810 * @prev: previously returned memcg, NULL on first invocation
811 * @reclaim: cookie for shared reclaim walks, NULL for full walks
813 * Returns references to children of the hierarchy below @root, or
814 * @root itself, or %NULL after a full round-trip.
816 * Caller must pass the return value in @prev on subsequent
817 * invocations for reference counting, or use mem_cgroup_iter_break()
818 * to cancel a hierarchy walk before the round-trip is complete.
820 * Reclaimers can specify a zone and a priority level in @reclaim to
821 * divide up the memcgs in the hierarchy among all concurrent
822 * reclaimers operating on the same zone and priority.
824 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
825 struct mem_cgroup
*prev
,
826 struct mem_cgroup_reclaim_cookie
*reclaim
)
828 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
829 struct cgroup_subsys_state
*css
= NULL
;
830 struct mem_cgroup
*memcg
= NULL
;
831 struct mem_cgroup
*pos
= NULL
;
833 if (mem_cgroup_disabled())
837 root
= root_mem_cgroup
;
839 if (prev
&& !reclaim
)
842 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
851 struct mem_cgroup_per_zone
*mz
;
853 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
854 iter
= &mz
->iter
[reclaim
->priority
];
856 if (prev
&& reclaim
->generation
!= iter
->generation
)
860 pos
= READ_ONCE(iter
->position
);
861 if (!pos
|| css_tryget(&pos
->css
))
864 * css reference reached zero, so iter->position will
865 * be cleared by ->css_released. However, we should not
866 * rely on this happening soon, because ->css_released
867 * is called from a work queue, and by busy-waiting we
868 * might block it. So we clear iter->position right
871 (void)cmpxchg(&iter
->position
, pos
, NULL
);
879 css
= css_next_descendant_pre(css
, &root
->css
);
882 * Reclaimers share the hierarchy walk, and a
883 * new one might jump in right at the end of
884 * the hierarchy - make sure they see at least
885 * one group and restart from the beginning.
893 * Verify the css and acquire a reference. The root
894 * is provided by the caller, so we know it's alive
895 * and kicking, and don't take an extra reference.
897 memcg
= mem_cgroup_from_css(css
);
899 if (css
== &root
->css
)
902 if (css_tryget(css
)) {
904 * Make sure the memcg is initialized:
905 * mem_cgroup_css_online() orders the the
906 * initialization against setting the flag.
908 if (smp_load_acquire(&memcg
->initialized
))
919 * The position could have already been updated by a competing
920 * thread, so check that the value hasn't changed since we read
921 * it to avoid reclaiming from the same cgroup twice.
923 (void)cmpxchg(&iter
->position
, pos
, memcg
);
931 reclaim
->generation
= iter
->generation
;
937 if (prev
&& prev
!= root
)
944 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
945 * @root: hierarchy root
946 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
948 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
949 struct mem_cgroup
*prev
)
952 root
= root_mem_cgroup
;
953 if (prev
&& prev
!= root
)
957 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
959 struct mem_cgroup
*memcg
= dead_memcg
;
960 struct mem_cgroup_reclaim_iter
*iter
;
961 struct mem_cgroup_per_zone
*mz
;
965 while ((memcg
= parent_mem_cgroup(memcg
))) {
967 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
968 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
969 for (i
= 0; i
<= DEF_PRIORITY
; i
++) {
971 cmpxchg(&iter
->position
,
980 * Iteration constructs for visiting all cgroups (under a tree). If
981 * loops are exited prematurely (break), mem_cgroup_iter_break() must
982 * be used for reference counting.
984 #define for_each_mem_cgroup_tree(iter, root) \
985 for (iter = mem_cgroup_iter(root, NULL, NULL); \
987 iter = mem_cgroup_iter(root, iter, NULL))
989 #define for_each_mem_cgroup(iter) \
990 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
992 iter = mem_cgroup_iter(NULL, iter, NULL))
995 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
996 * @zone: zone of the wanted lruvec
997 * @memcg: memcg of the wanted lruvec
999 * Returns the lru list vector holding pages for the given @zone and
1000 * @mem. This can be the global zone lruvec, if the memory controller
1003 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1004 struct mem_cgroup
*memcg
)
1006 struct mem_cgroup_per_zone
*mz
;
1007 struct lruvec
*lruvec
;
1009 if (mem_cgroup_disabled()) {
1010 lruvec
= &zone
->lruvec
;
1014 mz
= mem_cgroup_zone_zoneinfo(memcg
, zone
);
1015 lruvec
= &mz
->lruvec
;
1018 * Since a node can be onlined after the mem_cgroup was created,
1019 * we have to be prepared to initialize lruvec->zone here;
1020 * and if offlined then reonlined, we need to reinitialize it.
1022 if (unlikely(lruvec
->zone
!= zone
))
1023 lruvec
->zone
= zone
;
1028 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1030 * @zone: zone of the page
1032 * This function is only safe when following the LRU page isolation
1033 * and putback protocol: the LRU lock must be held, and the page must
1034 * either be PageLRU() or the caller must have isolated/allocated it.
1036 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1038 struct mem_cgroup_per_zone
*mz
;
1039 struct mem_cgroup
*memcg
;
1040 struct lruvec
*lruvec
;
1042 if (mem_cgroup_disabled()) {
1043 lruvec
= &zone
->lruvec
;
1047 memcg
= page
->mem_cgroup
;
1049 * Swapcache readahead pages are added to the LRU - and
1050 * possibly migrated - before they are charged.
1053 memcg
= root_mem_cgroup
;
1055 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1056 lruvec
= &mz
->lruvec
;
1059 * Since a node can be onlined after the mem_cgroup was created,
1060 * we have to be prepared to initialize lruvec->zone here;
1061 * and if offlined then reonlined, we need to reinitialize it.
1063 if (unlikely(lruvec
->zone
!= zone
))
1064 lruvec
->zone
= zone
;
1069 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1070 * @lruvec: mem_cgroup per zone lru vector
1071 * @lru: index of lru list the page is sitting on
1072 * @nr_pages: positive when adding or negative when removing
1074 * This function must be called when a page is added to or removed from an
1077 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1080 struct mem_cgroup_per_zone
*mz
;
1081 unsigned long *lru_size
;
1083 if (mem_cgroup_disabled())
1086 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1087 lru_size
= mz
->lru_size
+ lru
;
1088 *lru_size
+= nr_pages
;
1089 VM_BUG_ON((long)(*lru_size
) < 0);
1092 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1094 struct mem_cgroup
*task_memcg
;
1095 struct task_struct
*p
;
1098 p
= find_lock_task_mm(task
);
1100 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1104 * All threads may have already detached their mm's, but the oom
1105 * killer still needs to detect if they have already been oom
1106 * killed to prevent needlessly killing additional tasks.
1109 task_memcg
= mem_cgroup_from_task(task
);
1110 css_get(&task_memcg
->css
);
1113 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1114 css_put(&task_memcg
->css
);
1119 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1120 * @memcg: the memory cgroup
1122 * Returns the maximum amount of memory @mem can be charged with, in
1125 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1127 unsigned long margin
= 0;
1128 unsigned long count
;
1129 unsigned long limit
;
1131 count
= page_counter_read(&memcg
->memory
);
1132 limit
= READ_ONCE(memcg
->memory
.limit
);
1134 margin
= limit
- count
;
1136 if (do_memsw_account()) {
1137 count
= page_counter_read(&memcg
->memsw
);
1138 limit
= READ_ONCE(memcg
->memsw
.limit
);
1140 margin
= min(margin
, limit
- count
);
1147 * A routine for checking "mem" is under move_account() or not.
1149 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1150 * moving cgroups. This is for waiting at high-memory pressure
1153 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1155 struct mem_cgroup
*from
;
1156 struct mem_cgroup
*to
;
1159 * Unlike task_move routines, we access mc.to, mc.from not under
1160 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1162 spin_lock(&mc
.lock
);
1168 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1169 mem_cgroup_is_descendant(to
, memcg
);
1171 spin_unlock(&mc
.lock
);
1175 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1177 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1178 if (mem_cgroup_under_move(memcg
)) {
1180 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1181 /* moving charge context might have finished. */
1184 finish_wait(&mc
.waitq
, &wait
);
1191 #define K(x) ((x) << (PAGE_SHIFT-10))
1193 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1194 * @memcg: The memory cgroup that went over limit
1195 * @p: Task that is going to be killed
1197 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1200 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1202 /* oom_info_lock ensures that parallel ooms do not interleave */
1203 static DEFINE_MUTEX(oom_info_lock
);
1204 struct mem_cgroup
*iter
;
1207 mutex_lock(&oom_info_lock
);
1211 pr_info("Task in ");
1212 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1213 pr_cont(" killed as a result of limit of ");
1215 pr_info("Memory limit reached of cgroup ");
1218 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1223 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1224 K((u64
)page_counter_read(&memcg
->memory
)),
1225 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1226 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1227 K((u64
)page_counter_read(&memcg
->memsw
)),
1228 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1229 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1230 K((u64
)page_counter_read(&memcg
->kmem
)),
1231 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1233 for_each_mem_cgroup_tree(iter
, memcg
) {
1234 pr_info("Memory cgroup stats for ");
1235 pr_cont_cgroup_path(iter
->css
.cgroup
);
1238 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1239 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
1241 pr_cont(" %s:%luKB", mem_cgroup_stat_names
[i
],
1242 K(mem_cgroup_read_stat(iter
, i
)));
1245 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1246 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1247 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1251 mutex_unlock(&oom_info_lock
);
1255 * This function returns the number of memcg under hierarchy tree. Returns
1256 * 1(self count) if no children.
1258 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1261 struct mem_cgroup
*iter
;
1263 for_each_mem_cgroup_tree(iter
, memcg
)
1269 * Return the memory (and swap, if configured) limit for a memcg.
1271 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1273 unsigned long limit
;
1275 limit
= memcg
->memory
.limit
;
1276 if (mem_cgroup_swappiness(memcg
)) {
1277 unsigned long memsw_limit
;
1279 memsw_limit
= memcg
->memsw
.limit
;
1280 limit
= min(limit
+ total_swap_pages
, memsw_limit
);
1285 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1288 struct oom_control oc
= {
1291 .gfp_mask
= gfp_mask
,
1294 struct mem_cgroup
*iter
;
1295 unsigned long chosen_points
= 0;
1296 unsigned long totalpages
;
1297 unsigned int points
= 0;
1298 struct task_struct
*chosen
= NULL
;
1300 mutex_lock(&oom_lock
);
1303 * If current has a pending SIGKILL or is exiting, then automatically
1304 * select it. The goal is to allow it to allocate so that it may
1305 * quickly exit and free its memory.
1307 if (fatal_signal_pending(current
) || task_will_free_mem(current
)) {
1308 mark_oom_victim(current
);
1312 check_panic_on_oom(&oc
, CONSTRAINT_MEMCG
, memcg
);
1313 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1314 for_each_mem_cgroup_tree(iter
, memcg
) {
1315 struct css_task_iter it
;
1316 struct task_struct
*task
;
1318 css_task_iter_start(&iter
->css
, &it
);
1319 while ((task
= css_task_iter_next(&it
))) {
1320 switch (oom_scan_process_thread(&oc
, task
, totalpages
)) {
1321 case OOM_SCAN_SELECT
:
1323 put_task_struct(chosen
);
1325 chosen_points
= ULONG_MAX
;
1326 get_task_struct(chosen
);
1328 case OOM_SCAN_CONTINUE
:
1330 case OOM_SCAN_ABORT
:
1331 css_task_iter_end(&it
);
1332 mem_cgroup_iter_break(memcg
, iter
);
1334 put_task_struct(chosen
);
1339 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1340 if (!points
|| points
< chosen_points
)
1342 /* Prefer thread group leaders for display purposes */
1343 if (points
== chosen_points
&&
1344 thread_group_leader(chosen
))
1348 put_task_struct(chosen
);
1350 chosen_points
= points
;
1351 get_task_struct(chosen
);
1353 css_task_iter_end(&it
);
1357 points
= chosen_points
* 1000 / totalpages
;
1358 oom_kill_process(&oc
, chosen
, points
, totalpages
, memcg
,
1359 "Memory cgroup out of memory");
1362 mutex_unlock(&oom_lock
);
1365 #if MAX_NUMNODES > 1
1368 * test_mem_cgroup_node_reclaimable
1369 * @memcg: the target memcg
1370 * @nid: the node ID to be checked.
1371 * @noswap : specify true here if the user wants flle only information.
1373 * This function returns whether the specified memcg contains any
1374 * reclaimable pages on a node. Returns true if there are any reclaimable
1375 * pages in the node.
1377 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1378 int nid
, bool noswap
)
1380 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1382 if (noswap
|| !total_swap_pages
)
1384 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1391 * Always updating the nodemask is not very good - even if we have an empty
1392 * list or the wrong list here, we can start from some node and traverse all
1393 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1396 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1400 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1401 * pagein/pageout changes since the last update.
1403 if (!atomic_read(&memcg
->numainfo_events
))
1405 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1408 /* make a nodemask where this memcg uses memory from */
1409 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1411 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1413 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1414 node_clear(nid
, memcg
->scan_nodes
);
1417 atomic_set(&memcg
->numainfo_events
, 0);
1418 atomic_set(&memcg
->numainfo_updating
, 0);
1422 * Selecting a node where we start reclaim from. Because what we need is just
1423 * reducing usage counter, start from anywhere is O,K. Considering
1424 * memory reclaim from current node, there are pros. and cons.
1426 * Freeing memory from current node means freeing memory from a node which
1427 * we'll use or we've used. So, it may make LRU bad. And if several threads
1428 * hit limits, it will see a contention on a node. But freeing from remote
1429 * node means more costs for memory reclaim because of memory latency.
1431 * Now, we use round-robin. Better algorithm is welcomed.
1433 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1437 mem_cgroup_may_update_nodemask(memcg
);
1438 node
= memcg
->last_scanned_node
;
1440 node
= next_node(node
, memcg
->scan_nodes
);
1441 if (node
== MAX_NUMNODES
)
1442 node
= first_node(memcg
->scan_nodes
);
1444 * We call this when we hit limit, not when pages are added to LRU.
1445 * No LRU may hold pages because all pages are UNEVICTABLE or
1446 * memcg is too small and all pages are not on LRU. In that case,
1447 * we use curret node.
1449 if (unlikely(node
== MAX_NUMNODES
))
1450 node
= numa_node_id();
1452 memcg
->last_scanned_node
= node
;
1456 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1462 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1465 unsigned long *total_scanned
)
1467 struct mem_cgroup
*victim
= NULL
;
1470 unsigned long excess
;
1471 unsigned long nr_scanned
;
1472 struct mem_cgroup_reclaim_cookie reclaim
= {
1477 excess
= soft_limit_excess(root_memcg
);
1480 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1485 * If we have not been able to reclaim
1486 * anything, it might because there are
1487 * no reclaimable pages under this hierarchy
1492 * We want to do more targeted reclaim.
1493 * excess >> 2 is not to excessive so as to
1494 * reclaim too much, nor too less that we keep
1495 * coming back to reclaim from this cgroup
1497 if (total
>= (excess
>> 2) ||
1498 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1503 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1505 *total_scanned
+= nr_scanned
;
1506 if (!soft_limit_excess(root_memcg
))
1509 mem_cgroup_iter_break(root_memcg
, victim
);
1513 #ifdef CONFIG_LOCKDEP
1514 static struct lockdep_map memcg_oom_lock_dep_map
= {
1515 .name
= "memcg_oom_lock",
1519 static DEFINE_SPINLOCK(memcg_oom_lock
);
1522 * Check OOM-Killer is already running under our hierarchy.
1523 * If someone is running, return false.
1525 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1527 struct mem_cgroup
*iter
, *failed
= NULL
;
1529 spin_lock(&memcg_oom_lock
);
1531 for_each_mem_cgroup_tree(iter
, memcg
) {
1532 if (iter
->oom_lock
) {
1534 * this subtree of our hierarchy is already locked
1535 * so we cannot give a lock.
1538 mem_cgroup_iter_break(memcg
, iter
);
1541 iter
->oom_lock
= true;
1546 * OK, we failed to lock the whole subtree so we have
1547 * to clean up what we set up to the failing subtree
1549 for_each_mem_cgroup_tree(iter
, memcg
) {
1550 if (iter
== failed
) {
1551 mem_cgroup_iter_break(memcg
, iter
);
1554 iter
->oom_lock
= false;
1557 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1559 spin_unlock(&memcg_oom_lock
);
1564 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1566 struct mem_cgroup
*iter
;
1568 spin_lock(&memcg_oom_lock
);
1569 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1570 for_each_mem_cgroup_tree(iter
, memcg
)
1571 iter
->oom_lock
= false;
1572 spin_unlock(&memcg_oom_lock
);
1575 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1577 struct mem_cgroup
*iter
;
1579 spin_lock(&memcg_oom_lock
);
1580 for_each_mem_cgroup_tree(iter
, memcg
)
1582 spin_unlock(&memcg_oom_lock
);
1585 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1587 struct mem_cgroup
*iter
;
1590 * When a new child is created while the hierarchy is under oom,
1591 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1593 spin_lock(&memcg_oom_lock
);
1594 for_each_mem_cgroup_tree(iter
, memcg
)
1595 if (iter
->under_oom
> 0)
1597 spin_unlock(&memcg_oom_lock
);
1600 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1602 struct oom_wait_info
{
1603 struct mem_cgroup
*memcg
;
1607 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1608 unsigned mode
, int sync
, void *arg
)
1610 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1611 struct mem_cgroup
*oom_wait_memcg
;
1612 struct oom_wait_info
*oom_wait_info
;
1614 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1615 oom_wait_memcg
= oom_wait_info
->memcg
;
1617 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1618 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1620 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1623 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1626 * For the following lockless ->under_oom test, the only required
1627 * guarantee is that it must see the state asserted by an OOM when
1628 * this function is called as a result of userland actions
1629 * triggered by the notification of the OOM. This is trivially
1630 * achieved by invoking mem_cgroup_mark_under_oom() before
1631 * triggering notification.
1633 if (memcg
&& memcg
->under_oom
)
1634 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1637 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1639 if (!current
->memcg_may_oom
)
1642 * We are in the middle of the charge context here, so we
1643 * don't want to block when potentially sitting on a callstack
1644 * that holds all kinds of filesystem and mm locks.
1646 * Also, the caller may handle a failed allocation gracefully
1647 * (like optional page cache readahead) and so an OOM killer
1648 * invocation might not even be necessary.
1650 * That's why we don't do anything here except remember the
1651 * OOM context and then deal with it at the end of the page
1652 * fault when the stack is unwound, the locks are released,
1653 * and when we know whether the fault was overall successful.
1655 css_get(&memcg
->css
);
1656 current
->memcg_in_oom
= memcg
;
1657 current
->memcg_oom_gfp_mask
= mask
;
1658 current
->memcg_oom_order
= order
;
1662 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1663 * @handle: actually kill/wait or just clean up the OOM state
1665 * This has to be called at the end of a page fault if the memcg OOM
1666 * handler was enabled.
1668 * Memcg supports userspace OOM handling where failed allocations must
1669 * sleep on a waitqueue until the userspace task resolves the
1670 * situation. Sleeping directly in the charge context with all kinds
1671 * of locks held is not a good idea, instead we remember an OOM state
1672 * in the task and mem_cgroup_oom_synchronize() has to be called at
1673 * the end of the page fault to complete the OOM handling.
1675 * Returns %true if an ongoing memcg OOM situation was detected and
1676 * completed, %false otherwise.
1678 bool mem_cgroup_oom_synchronize(bool handle
)
1680 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1681 struct oom_wait_info owait
;
1684 /* OOM is global, do not handle */
1688 if (!handle
|| oom_killer_disabled
)
1691 owait
.memcg
= memcg
;
1692 owait
.wait
.flags
= 0;
1693 owait
.wait
.func
= memcg_oom_wake_function
;
1694 owait
.wait
.private = current
;
1695 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1697 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1698 mem_cgroup_mark_under_oom(memcg
);
1700 locked
= mem_cgroup_oom_trylock(memcg
);
1703 mem_cgroup_oom_notify(memcg
);
1705 if (locked
&& !memcg
->oom_kill_disable
) {
1706 mem_cgroup_unmark_under_oom(memcg
);
1707 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1708 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1709 current
->memcg_oom_order
);
1712 mem_cgroup_unmark_under_oom(memcg
);
1713 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1717 mem_cgroup_oom_unlock(memcg
);
1719 * There is no guarantee that an OOM-lock contender
1720 * sees the wakeups triggered by the OOM kill
1721 * uncharges. Wake any sleepers explicitely.
1723 memcg_oom_recover(memcg
);
1726 current
->memcg_in_oom
= NULL
;
1727 css_put(&memcg
->css
);
1732 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1733 * @page: page that is going to change accounted state
1735 * This function must mark the beginning of an accounted page state
1736 * change to prevent double accounting when the page is concurrently
1737 * being moved to another memcg:
1739 * memcg = mem_cgroup_begin_page_stat(page);
1740 * if (TestClearPageState(page))
1741 * mem_cgroup_update_page_stat(memcg, state, -1);
1742 * mem_cgroup_end_page_stat(memcg);
1744 struct mem_cgroup
*mem_cgroup_begin_page_stat(struct page
*page
)
1746 struct mem_cgroup
*memcg
;
1747 unsigned long flags
;
1750 * The RCU lock is held throughout the transaction. The fast
1751 * path can get away without acquiring the memcg->move_lock
1752 * because page moving starts with an RCU grace period.
1754 * The RCU lock also protects the memcg from being freed when
1755 * the page state that is going to change is the only thing
1756 * preventing the page from being uncharged.
1757 * E.g. end-writeback clearing PageWriteback(), which allows
1758 * migration to go ahead and uncharge the page before the
1759 * account transaction might be complete.
1763 if (mem_cgroup_disabled())
1766 memcg
= page
->mem_cgroup
;
1767 if (unlikely(!memcg
))
1770 if (atomic_read(&memcg
->moving_account
) <= 0)
1773 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1774 if (memcg
!= page
->mem_cgroup
) {
1775 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1780 * When charge migration first begins, we can have locked and
1781 * unlocked page stat updates happening concurrently. Track
1782 * the task who has the lock for mem_cgroup_end_page_stat().
1784 memcg
->move_lock_task
= current
;
1785 memcg
->move_lock_flags
= flags
;
1789 EXPORT_SYMBOL(mem_cgroup_begin_page_stat
);
1792 * mem_cgroup_end_page_stat - finish a page state statistics transaction
1793 * @memcg: the memcg that was accounted against
1795 void mem_cgroup_end_page_stat(struct mem_cgroup
*memcg
)
1797 if (memcg
&& memcg
->move_lock_task
== current
) {
1798 unsigned long flags
= memcg
->move_lock_flags
;
1800 memcg
->move_lock_task
= NULL
;
1801 memcg
->move_lock_flags
= 0;
1803 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1808 EXPORT_SYMBOL(mem_cgroup_end_page_stat
);
1811 * size of first charge trial. "32" comes from vmscan.c's magic value.
1812 * TODO: maybe necessary to use big numbers in big irons.
1814 #define CHARGE_BATCH 32U
1815 struct memcg_stock_pcp
{
1816 struct mem_cgroup
*cached
; /* this never be root cgroup */
1817 unsigned int nr_pages
;
1818 struct work_struct work
;
1819 unsigned long flags
;
1820 #define FLUSHING_CACHED_CHARGE 0
1822 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1823 static DEFINE_MUTEX(percpu_charge_mutex
);
1826 * consume_stock: Try to consume stocked charge on this cpu.
1827 * @memcg: memcg to consume from.
1828 * @nr_pages: how many pages to charge.
1830 * The charges will only happen if @memcg matches the current cpu's memcg
1831 * stock, and at least @nr_pages are available in that stock. Failure to
1832 * service an allocation will refill the stock.
1834 * returns true if successful, false otherwise.
1836 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1838 struct memcg_stock_pcp
*stock
;
1841 if (nr_pages
> CHARGE_BATCH
)
1844 stock
= &get_cpu_var(memcg_stock
);
1845 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
1846 stock
->nr_pages
-= nr_pages
;
1849 put_cpu_var(memcg_stock
);
1854 * Returns stocks cached in percpu and reset cached information.
1856 static void drain_stock(struct memcg_stock_pcp
*stock
)
1858 struct mem_cgroup
*old
= stock
->cached
;
1860 if (stock
->nr_pages
) {
1861 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
1862 if (do_memsw_account())
1863 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
1864 css_put_many(&old
->css
, stock
->nr_pages
);
1865 stock
->nr_pages
= 0;
1867 stock
->cached
= NULL
;
1871 * This must be called under preempt disabled or must be called by
1872 * a thread which is pinned to local cpu.
1874 static void drain_local_stock(struct work_struct
*dummy
)
1876 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
1878 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
1882 * Cache charges(val) to local per_cpu area.
1883 * This will be consumed by consume_stock() function, later.
1885 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1887 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
1889 if (stock
->cached
!= memcg
) { /* reset if necessary */
1891 stock
->cached
= memcg
;
1893 stock
->nr_pages
+= nr_pages
;
1894 put_cpu_var(memcg_stock
);
1898 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1899 * of the hierarchy under it.
1901 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
1905 /* If someone's already draining, avoid adding running more workers. */
1906 if (!mutex_trylock(&percpu_charge_mutex
))
1908 /* Notify other cpus that system-wide "drain" is running */
1911 for_each_online_cpu(cpu
) {
1912 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1913 struct mem_cgroup
*memcg
;
1915 memcg
= stock
->cached
;
1916 if (!memcg
|| !stock
->nr_pages
)
1918 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
1920 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
1922 drain_local_stock(&stock
->work
);
1924 schedule_work_on(cpu
, &stock
->work
);
1929 mutex_unlock(&percpu_charge_mutex
);
1932 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
1933 unsigned long action
,
1936 int cpu
= (unsigned long)hcpu
;
1937 struct memcg_stock_pcp
*stock
;
1939 if (action
== CPU_ONLINE
)
1942 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
1945 stock
= &per_cpu(memcg_stock
, cpu
);
1950 static void reclaim_high(struct mem_cgroup
*memcg
,
1951 unsigned int nr_pages
,
1955 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
1957 mem_cgroup_events(memcg
, MEMCG_HIGH
, 1);
1958 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
1959 } while ((memcg
= parent_mem_cgroup(memcg
)));
1962 static void high_work_func(struct work_struct
*work
)
1964 struct mem_cgroup
*memcg
;
1966 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
1967 reclaim_high(memcg
, CHARGE_BATCH
, GFP_KERNEL
);
1971 * Scheduled by try_charge() to be executed from the userland return path
1972 * and reclaims memory over the high limit.
1974 void mem_cgroup_handle_over_high(void)
1976 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
1977 struct mem_cgroup
*memcg
;
1979 if (likely(!nr_pages
))
1982 memcg
= get_mem_cgroup_from_mm(current
->mm
);
1983 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
1984 css_put(&memcg
->css
);
1985 current
->memcg_nr_pages_over_high
= 0;
1988 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1989 unsigned int nr_pages
)
1991 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
1992 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1993 struct mem_cgroup
*mem_over_limit
;
1994 struct page_counter
*counter
;
1995 unsigned long nr_reclaimed
;
1996 bool may_swap
= true;
1997 bool drained
= false;
1999 if (mem_cgroup_is_root(memcg
))
2002 if (consume_stock(memcg
, nr_pages
))
2005 if (!do_memsw_account() ||
2006 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2007 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2009 if (do_memsw_account())
2010 page_counter_uncharge(&memcg
->memsw
, batch
);
2011 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2013 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2017 if (batch
> nr_pages
) {
2023 * Unlike in global OOM situations, memcg is not in a physical
2024 * memory shortage. Allow dying and OOM-killed tasks to
2025 * bypass the last charges so that they can exit quickly and
2026 * free their memory.
2028 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
2029 fatal_signal_pending(current
) ||
2030 current
->flags
& PF_EXITING
))
2033 if (unlikely(task_in_memcg_oom(current
)))
2036 if (!gfpflags_allow_blocking(gfp_mask
))
2039 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
2041 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2042 gfp_mask
, may_swap
);
2044 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2048 drain_all_stock(mem_over_limit
);
2053 if (gfp_mask
& __GFP_NORETRY
)
2056 * Even though the limit is exceeded at this point, reclaim
2057 * may have been able to free some pages. Retry the charge
2058 * before killing the task.
2060 * Only for regular pages, though: huge pages are rather
2061 * unlikely to succeed so close to the limit, and we fall back
2062 * to regular pages anyway in case of failure.
2064 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2067 * At task move, charge accounts can be doubly counted. So, it's
2068 * better to wait until the end of task_move if something is going on.
2070 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2076 if (gfp_mask
& __GFP_NOFAIL
)
2079 if (fatal_signal_pending(current
))
2082 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
2084 mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2085 get_order(nr_pages
* PAGE_SIZE
));
2087 if (!(gfp_mask
& __GFP_NOFAIL
))
2091 * The allocation either can't fail or will lead to more memory
2092 * being freed very soon. Allow memory usage go over the limit
2093 * temporarily by force charging it.
2095 page_counter_charge(&memcg
->memory
, nr_pages
);
2096 if (do_memsw_account())
2097 page_counter_charge(&memcg
->memsw
, nr_pages
);
2098 css_get_many(&memcg
->css
, nr_pages
);
2103 css_get_many(&memcg
->css
, batch
);
2104 if (batch
> nr_pages
)
2105 refill_stock(memcg
, batch
- nr_pages
);
2108 * If the hierarchy is above the normal consumption range, schedule
2109 * reclaim on returning to userland. We can perform reclaim here
2110 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2111 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2112 * not recorded as it most likely matches current's and won't
2113 * change in the meantime. As high limit is checked again before
2114 * reclaim, the cost of mismatch is negligible.
2117 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2118 /* Don't bother a random interrupted task */
2119 if (in_interrupt()) {
2120 schedule_work(&memcg
->high_work
);
2123 current
->memcg_nr_pages_over_high
+= batch
;
2124 set_notify_resume(current
);
2127 } while ((memcg
= parent_mem_cgroup(memcg
)));
2132 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2134 if (mem_cgroup_is_root(memcg
))
2137 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2138 if (do_memsw_account())
2139 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2141 css_put_many(&memcg
->css
, nr_pages
);
2144 static void lock_page_lru(struct page
*page
, int *isolated
)
2146 struct zone
*zone
= page_zone(page
);
2148 spin_lock_irq(&zone
->lru_lock
);
2149 if (PageLRU(page
)) {
2150 struct lruvec
*lruvec
;
2152 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2154 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2160 static void unlock_page_lru(struct page
*page
, int isolated
)
2162 struct zone
*zone
= page_zone(page
);
2165 struct lruvec
*lruvec
;
2167 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2168 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2170 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2172 spin_unlock_irq(&zone
->lru_lock
);
2175 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2180 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2183 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2184 * may already be on some other mem_cgroup's LRU. Take care of it.
2187 lock_page_lru(page
, &isolated
);
2190 * Nobody should be changing or seriously looking at
2191 * page->mem_cgroup at this point:
2193 * - the page is uncharged
2195 * - the page is off-LRU
2197 * - an anonymous fault has exclusive page access, except for
2198 * a locked page table
2200 * - a page cache insertion, a swapin fault, or a migration
2201 * have the page locked
2203 page
->mem_cgroup
= memcg
;
2206 unlock_page_lru(page
, isolated
);
2210 static int memcg_alloc_cache_id(void)
2215 id
= ida_simple_get(&memcg_cache_ida
,
2216 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2220 if (id
< memcg_nr_cache_ids
)
2224 * There's no space for the new id in memcg_caches arrays,
2225 * so we have to grow them.
2227 down_write(&memcg_cache_ids_sem
);
2229 size
= 2 * (id
+ 1);
2230 if (size
< MEMCG_CACHES_MIN_SIZE
)
2231 size
= MEMCG_CACHES_MIN_SIZE
;
2232 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2233 size
= MEMCG_CACHES_MAX_SIZE
;
2235 err
= memcg_update_all_caches(size
);
2237 err
= memcg_update_all_list_lrus(size
);
2239 memcg_nr_cache_ids
= size
;
2241 up_write(&memcg_cache_ids_sem
);
2244 ida_simple_remove(&memcg_cache_ida
, id
);
2250 static void memcg_free_cache_id(int id
)
2252 ida_simple_remove(&memcg_cache_ida
, id
);
2255 struct memcg_kmem_cache_create_work
{
2256 struct mem_cgroup
*memcg
;
2257 struct kmem_cache
*cachep
;
2258 struct work_struct work
;
2261 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2263 struct memcg_kmem_cache_create_work
*cw
=
2264 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2265 struct mem_cgroup
*memcg
= cw
->memcg
;
2266 struct kmem_cache
*cachep
= cw
->cachep
;
2268 memcg_create_kmem_cache(memcg
, cachep
);
2270 css_put(&memcg
->css
);
2275 * Enqueue the creation of a per-memcg kmem_cache.
2277 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2278 struct kmem_cache
*cachep
)
2280 struct memcg_kmem_cache_create_work
*cw
;
2282 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2286 css_get(&memcg
->css
);
2289 cw
->cachep
= cachep
;
2290 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2292 schedule_work(&cw
->work
);
2295 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2296 struct kmem_cache
*cachep
)
2299 * We need to stop accounting when we kmalloc, because if the
2300 * corresponding kmalloc cache is not yet created, the first allocation
2301 * in __memcg_schedule_kmem_cache_create will recurse.
2303 * However, it is better to enclose the whole function. Depending on
2304 * the debugging options enabled, INIT_WORK(), for instance, can
2305 * trigger an allocation. This too, will make us recurse. Because at
2306 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2307 * the safest choice is to do it like this, wrapping the whole function.
2309 current
->memcg_kmem_skip_account
= 1;
2310 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2311 current
->memcg_kmem_skip_account
= 0;
2315 * Return the kmem_cache we're supposed to use for a slab allocation.
2316 * We try to use the current memcg's version of the cache.
2318 * If the cache does not exist yet, if we are the first user of it,
2319 * we either create it immediately, if possible, or create it asynchronously
2321 * In the latter case, we will let the current allocation go through with
2322 * the original cache.
2324 * Can't be called in interrupt context or from kernel threads.
2325 * This function needs to be called with rcu_read_lock() held.
2327 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2329 struct mem_cgroup
*memcg
;
2330 struct kmem_cache
*memcg_cachep
;
2333 VM_BUG_ON(!is_root_cache(cachep
));
2335 if (cachep
->flags
& SLAB_ACCOUNT
)
2336 gfp
|= __GFP_ACCOUNT
;
2338 if (!(gfp
& __GFP_ACCOUNT
))
2341 if (current
->memcg_kmem_skip_account
)
2344 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2345 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2349 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2350 if (likely(memcg_cachep
))
2351 return memcg_cachep
;
2354 * If we are in a safe context (can wait, and not in interrupt
2355 * context), we could be be predictable and return right away.
2356 * This would guarantee that the allocation being performed
2357 * already belongs in the new cache.
2359 * However, there are some clashes that can arrive from locking.
2360 * For instance, because we acquire the slab_mutex while doing
2361 * memcg_create_kmem_cache, this means no further allocation
2362 * could happen with the slab_mutex held. So it's better to
2365 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2367 css_put(&memcg
->css
);
2371 void __memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2373 if (!is_root_cache(cachep
))
2374 css_put(&cachep
->memcg_params
.memcg
->css
);
2377 int __memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2378 struct mem_cgroup
*memcg
)
2380 unsigned int nr_pages
= 1 << order
;
2381 struct page_counter
*counter
;
2384 if (!memcg_kmem_online(memcg
))
2387 ret
= try_charge(memcg
, gfp
, nr_pages
);
2391 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2392 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2393 cancel_charge(memcg
, nr_pages
);
2397 page
->mem_cgroup
= memcg
;
2402 int __memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2404 struct mem_cgroup
*memcg
;
2407 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2408 ret
= __memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2409 css_put(&memcg
->css
);
2413 void __memcg_kmem_uncharge(struct page
*page
, int order
)
2415 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2416 unsigned int nr_pages
= 1 << order
;
2421 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2423 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2424 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2426 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2427 if (do_memsw_account())
2428 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2430 page
->mem_cgroup
= NULL
;
2431 css_put_many(&memcg
->css
, nr_pages
);
2433 #endif /* !CONFIG_SLOB */
2435 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2438 * Because tail pages are not marked as "used", set it. We're under
2439 * zone->lru_lock and migration entries setup in all page mappings.
2441 void mem_cgroup_split_huge_fixup(struct page
*head
)
2445 if (mem_cgroup_disabled())
2448 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2449 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2451 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2454 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2456 #ifdef CONFIG_MEMCG_SWAP
2457 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2460 int val
= (charge
) ? 1 : -1;
2461 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2465 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2466 * @entry: swap entry to be moved
2467 * @from: mem_cgroup which the entry is moved from
2468 * @to: mem_cgroup which the entry is moved to
2470 * It succeeds only when the swap_cgroup's record for this entry is the same
2471 * as the mem_cgroup's id of @from.
2473 * Returns 0 on success, -EINVAL on failure.
2475 * The caller must have charged to @to, IOW, called page_counter_charge() about
2476 * both res and memsw, and called css_get().
2478 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2479 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2481 unsigned short old_id
, new_id
;
2483 old_id
= mem_cgroup_id(from
);
2484 new_id
= mem_cgroup_id(to
);
2486 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2487 mem_cgroup_swap_statistics(from
, false);
2488 mem_cgroup_swap_statistics(to
, true);
2494 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2495 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2501 static DEFINE_MUTEX(memcg_limit_mutex
);
2503 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2504 unsigned long limit
)
2506 unsigned long curusage
;
2507 unsigned long oldusage
;
2508 bool enlarge
= false;
2513 * For keeping hierarchical_reclaim simple, how long we should retry
2514 * is depends on callers. We set our retry-count to be function
2515 * of # of children which we should visit in this loop.
2517 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2518 mem_cgroup_count_children(memcg
);
2520 oldusage
= page_counter_read(&memcg
->memory
);
2523 if (signal_pending(current
)) {
2528 mutex_lock(&memcg_limit_mutex
);
2529 if (limit
> memcg
->memsw
.limit
) {
2530 mutex_unlock(&memcg_limit_mutex
);
2534 if (limit
> memcg
->memory
.limit
)
2536 ret
= page_counter_limit(&memcg
->memory
, limit
);
2537 mutex_unlock(&memcg_limit_mutex
);
2542 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2544 curusage
= page_counter_read(&memcg
->memory
);
2545 /* Usage is reduced ? */
2546 if (curusage
>= oldusage
)
2549 oldusage
= curusage
;
2550 } while (retry_count
);
2552 if (!ret
&& enlarge
)
2553 memcg_oom_recover(memcg
);
2558 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2559 unsigned long limit
)
2561 unsigned long curusage
;
2562 unsigned long oldusage
;
2563 bool enlarge
= false;
2567 /* see mem_cgroup_resize_res_limit */
2568 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2569 mem_cgroup_count_children(memcg
);
2571 oldusage
= page_counter_read(&memcg
->memsw
);
2574 if (signal_pending(current
)) {
2579 mutex_lock(&memcg_limit_mutex
);
2580 if (limit
< memcg
->memory
.limit
) {
2581 mutex_unlock(&memcg_limit_mutex
);
2585 if (limit
> memcg
->memsw
.limit
)
2587 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2588 mutex_unlock(&memcg_limit_mutex
);
2593 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2595 curusage
= page_counter_read(&memcg
->memsw
);
2596 /* Usage is reduced ? */
2597 if (curusage
>= oldusage
)
2600 oldusage
= curusage
;
2601 } while (retry_count
);
2603 if (!ret
&& enlarge
)
2604 memcg_oom_recover(memcg
);
2609 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
2611 unsigned long *total_scanned
)
2613 unsigned long nr_reclaimed
= 0;
2614 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
2615 unsigned long reclaimed
;
2617 struct mem_cgroup_tree_per_zone
*mctz
;
2618 unsigned long excess
;
2619 unsigned long nr_scanned
;
2624 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
2626 * This loop can run a while, specially if mem_cgroup's continuously
2627 * keep exceeding their soft limit and putting the system under
2634 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2639 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
2640 gfp_mask
, &nr_scanned
);
2641 nr_reclaimed
+= reclaimed
;
2642 *total_scanned
+= nr_scanned
;
2643 spin_lock_irq(&mctz
->lock
);
2644 __mem_cgroup_remove_exceeded(mz
, mctz
);
2647 * If we failed to reclaim anything from this memory cgroup
2648 * it is time to move on to the next cgroup
2652 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2654 excess
= soft_limit_excess(mz
->memcg
);
2656 * One school of thought says that we should not add
2657 * back the node to the tree if reclaim returns 0.
2658 * But our reclaim could return 0, simply because due
2659 * to priority we are exposing a smaller subset of
2660 * memory to reclaim from. Consider this as a longer
2663 /* If excess == 0, no tree ops */
2664 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2665 spin_unlock_irq(&mctz
->lock
);
2666 css_put(&mz
->memcg
->css
);
2669 * Could not reclaim anything and there are no more
2670 * mem cgroups to try or we seem to be looping without
2671 * reclaiming anything.
2673 if (!nr_reclaimed
&&
2675 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2677 } while (!nr_reclaimed
);
2679 css_put(&next_mz
->memcg
->css
);
2680 return nr_reclaimed
;
2684 * Test whether @memcg has children, dead or alive. Note that this
2685 * function doesn't care whether @memcg has use_hierarchy enabled and
2686 * returns %true if there are child csses according to the cgroup
2687 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2689 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2694 * The lock does not prevent addition or deletion of children, but
2695 * it prevents a new child from being initialized based on this
2696 * parent in css_online(), so it's enough to decide whether
2697 * hierarchically inherited attributes can still be changed or not.
2699 lockdep_assert_held(&memcg_create_mutex
);
2702 ret
= css_next_child(NULL
, &memcg
->css
);
2708 * Reclaims as many pages from the given memcg as possible and moves
2709 * the rest to the parent.
2711 * Caller is responsible for holding css reference for memcg.
2713 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2715 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2717 /* we call try-to-free pages for make this cgroup empty */
2718 lru_add_drain_all();
2719 /* try to free all pages in this cgroup */
2720 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2723 if (signal_pending(current
))
2726 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2730 /* maybe some writeback is necessary */
2731 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2739 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
2740 char *buf
, size_t nbytes
,
2743 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2745 if (mem_cgroup_is_root(memcg
))
2747 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
2750 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
2753 return mem_cgroup_from_css(css
)->use_hierarchy
;
2756 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
2757 struct cftype
*cft
, u64 val
)
2760 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2761 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
2763 mutex_lock(&memcg_create_mutex
);
2765 if (memcg
->use_hierarchy
== val
)
2769 * If parent's use_hierarchy is set, we can't make any modifications
2770 * in the child subtrees. If it is unset, then the change can
2771 * occur, provided the current cgroup has no children.
2773 * For the root cgroup, parent_mem is NULL, we allow value to be
2774 * set if there are no children.
2776 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
2777 (val
== 1 || val
== 0)) {
2778 if (!memcg_has_children(memcg
))
2779 memcg
->use_hierarchy
= val
;
2786 mutex_unlock(&memcg_create_mutex
);
2791 static unsigned long tree_stat(struct mem_cgroup
*memcg
,
2792 enum mem_cgroup_stat_index idx
)
2794 struct mem_cgroup
*iter
;
2795 unsigned long val
= 0;
2797 for_each_mem_cgroup_tree(iter
, memcg
)
2798 val
+= mem_cgroup_read_stat(iter
, idx
);
2803 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
2807 if (mem_cgroup_is_root(memcg
)) {
2808 val
= tree_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
2809 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_RSS
);
2811 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
2814 val
= page_counter_read(&memcg
->memory
);
2816 val
= page_counter_read(&memcg
->memsw
);
2829 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
2832 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2833 struct page_counter
*counter
;
2835 switch (MEMFILE_TYPE(cft
->private)) {
2837 counter
= &memcg
->memory
;
2840 counter
= &memcg
->memsw
;
2843 counter
= &memcg
->kmem
;
2845 #if defined(CONFIG_MEMCG_LEGACY_KMEM) && defined(CONFIG_INET)
2847 counter
= &memcg
->tcp_mem
.memory_allocated
;
2854 switch (MEMFILE_ATTR(cft
->private)) {
2856 if (counter
== &memcg
->memory
)
2857 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
2858 if (counter
== &memcg
->memsw
)
2859 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
2860 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
2862 return (u64
)counter
->limit
* PAGE_SIZE
;
2864 return (u64
)counter
->watermark
* PAGE_SIZE
;
2866 return counter
->failcnt
;
2867 case RES_SOFT_LIMIT
:
2868 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
2875 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2880 BUG_ON(memcg
->kmemcg_id
>= 0);
2881 BUG_ON(memcg
->kmem_state
);
2884 * For simplicity, we won't allow this to be disabled. It also can't
2885 * be changed if the cgroup has children already, or if tasks had
2888 * If tasks join before we set the limit, a person looking at
2889 * kmem.usage_in_bytes will have no way to determine when it took
2890 * place, which makes the value quite meaningless.
2892 * After it first became limited, changes in the value of the limit are
2893 * of course permitted.
2895 mutex_lock(&memcg_create_mutex
);
2896 if (cgroup_is_populated(memcg
->css
.cgroup
) ||
2897 (memcg
->use_hierarchy
&& memcg_has_children(memcg
)))
2899 mutex_unlock(&memcg_create_mutex
);
2903 memcg_id
= memcg_alloc_cache_id();
2909 static_branch_inc(&memcg_kmem_enabled_key
);
2911 * A memory cgroup is considered kmem-online as soon as it gets
2912 * kmemcg_id. Setting the id after enabling static branching will
2913 * guarantee no one starts accounting before all call sites are
2916 memcg
->kmemcg_id
= memcg_id
;
2917 memcg
->kmem_state
= KMEM_ONLINE
;
2922 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
2925 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
2930 mutex_lock(&memcg_limit_mutex
);
2932 * If the parent cgroup is not kmem-online now, it cannot be
2933 * onlined after this point, because it has at least one child
2936 if (memcg_kmem_online(parent
) ||
2937 (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nokmem
))
2938 ret
= memcg_online_kmem(memcg
);
2939 mutex_unlock(&memcg_limit_mutex
);
2943 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2945 struct cgroup_subsys_state
*css
;
2946 struct mem_cgroup
*parent
, *child
;
2949 if (memcg
->kmem_state
!= KMEM_ONLINE
)
2952 * Clear the online state before clearing memcg_caches array
2953 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2954 * guarantees that no cache will be created for this cgroup
2955 * after we are done (see memcg_create_kmem_cache()).
2957 memcg
->kmem_state
= KMEM_ALLOCATED
;
2959 memcg_deactivate_kmem_caches(memcg
);
2961 kmemcg_id
= memcg
->kmemcg_id
;
2962 BUG_ON(kmemcg_id
< 0);
2964 parent
= parent_mem_cgroup(memcg
);
2966 parent
= root_mem_cgroup
;
2969 * Change kmemcg_id of this cgroup and all its descendants to the
2970 * parent's id, and then move all entries from this cgroup's list_lrus
2971 * to ones of the parent. After we have finished, all list_lrus
2972 * corresponding to this cgroup are guaranteed to remain empty. The
2973 * ordering is imposed by list_lru_node->lock taken by
2974 * memcg_drain_all_list_lrus().
2976 css_for_each_descendant_pre(css
, &memcg
->css
) {
2977 child
= mem_cgroup_from_css(css
);
2978 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
2979 child
->kmemcg_id
= parent
->kmemcg_id
;
2980 if (!memcg
->use_hierarchy
)
2983 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
2985 memcg_free_cache_id(kmemcg_id
);
2988 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2990 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
2991 memcg_destroy_kmem_caches(memcg
);
2992 static_branch_dec(&memcg_kmem_enabled_key
);
2993 WARN_ON(page_counter_read(&memcg
->kmem
));
2997 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
3001 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3004 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3007 #endif /* !CONFIG_SLOB */
3009 #ifdef CONFIG_MEMCG_LEGACY_KMEM
3010 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
3011 unsigned long limit
)
3015 mutex_lock(&memcg_limit_mutex
);
3016 /* Top-level cgroup doesn't propagate from root */
3017 if (!memcg_kmem_online(memcg
)) {
3018 ret
= memcg_online_kmem(memcg
);
3022 ret
= page_counter_limit(&memcg
->kmem
, limit
);
3024 mutex_unlock(&memcg_limit_mutex
);
3028 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
3029 unsigned long limit
)
3033 #endif /* CONFIG_MEMCG_LEGACY_KMEM */
3036 #if defined(CONFIG_MEMCG_LEGACY_KMEM) && defined(CONFIG_INET)
3037 static int memcg_update_tcp_limit(struct mem_cgroup
*memcg
, unsigned long limit
)
3041 mutex_lock(&memcg_limit_mutex
);
3043 ret
= page_counter_limit(&memcg
->tcp_mem
.memory_allocated
, limit
);
3047 if (!memcg
->tcp_mem
.active
) {
3049 * The active flag needs to be written after the static_key
3050 * update. This is what guarantees that the socket activation
3051 * function is the last one to run. See sock_update_memcg() for
3052 * details, and note that we don't mark any socket as belonging
3053 * to this memcg until that flag is up.
3055 * We need to do this, because static_keys will span multiple
3056 * sites, but we can't control their order. If we mark a socket
3057 * as accounted, but the accounting functions are not patched in
3058 * yet, we'll lose accounting.
3060 * We never race with the readers in sock_update_memcg(),
3061 * because when this value change, the code to process it is not
3064 static_branch_inc(&memcg_sockets_enabled_key
);
3065 memcg
->tcp_mem
.active
= true;
3068 mutex_unlock(&memcg_limit_mutex
);
3072 static int memcg_update_tcp_limit(struct mem_cgroup
*memcg
, unsigned long limit
)
3076 #endif /* CONFIG_MEMCG_LEGACY_KMEM && CONFIG_INET */
3079 * The user of this function is...
3082 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3083 char *buf
, size_t nbytes
, loff_t off
)
3085 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3086 unsigned long nr_pages
;
3089 buf
= strstrip(buf
);
3090 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3094 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3096 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3100 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3102 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
3105 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
3108 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
3111 ret
= memcg_update_tcp_limit(memcg
, nr_pages
);
3115 case RES_SOFT_LIMIT
:
3116 memcg
->soft_limit
= nr_pages
;
3120 return ret
?: nbytes
;
3123 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3124 size_t nbytes
, loff_t off
)
3126 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3127 struct page_counter
*counter
;
3129 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3131 counter
= &memcg
->memory
;
3134 counter
= &memcg
->memsw
;
3137 counter
= &memcg
->kmem
;
3139 #if defined(CONFIG_MEMCG_LEGACY_KMEM) && defined(CONFIG_INET)
3141 counter
= &memcg
->tcp_mem
.memory_allocated
;
3148 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3150 page_counter_reset_watermark(counter
);
3153 counter
->failcnt
= 0;
3162 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3165 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3169 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3170 struct cftype
*cft
, u64 val
)
3172 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3174 if (val
& ~MOVE_MASK
)
3178 * No kind of locking is needed in here, because ->can_attach() will
3179 * check this value once in the beginning of the process, and then carry
3180 * on with stale data. This means that changes to this value will only
3181 * affect task migrations starting after the change.
3183 memcg
->move_charge_at_immigrate
= val
;
3187 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3188 struct cftype
*cft
, u64 val
)
3195 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3199 unsigned int lru_mask
;
3202 static const struct numa_stat stats
[] = {
3203 { "total", LRU_ALL
},
3204 { "file", LRU_ALL_FILE
},
3205 { "anon", LRU_ALL_ANON
},
3206 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3208 const struct numa_stat
*stat
;
3211 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3213 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3214 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3215 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3216 for_each_node_state(nid
, N_MEMORY
) {
3217 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3219 seq_printf(m
, " N%d=%lu", nid
, nr
);
3224 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3225 struct mem_cgroup
*iter
;
3228 for_each_mem_cgroup_tree(iter
, memcg
)
3229 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3230 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3231 for_each_node_state(nid
, N_MEMORY
) {
3233 for_each_mem_cgroup_tree(iter
, memcg
)
3234 nr
+= mem_cgroup_node_nr_lru_pages(
3235 iter
, nid
, stat
->lru_mask
);
3236 seq_printf(m
, " N%d=%lu", nid
, nr
);
3243 #endif /* CONFIG_NUMA */
3245 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3247 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3248 unsigned long memory
, memsw
;
3249 struct mem_cgroup
*mi
;
3252 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3253 MEM_CGROUP_STAT_NSTATS
);
3254 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3255 MEM_CGROUP_EVENTS_NSTATS
);
3256 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3258 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3259 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3261 seq_printf(m
, "%s %lu\n", mem_cgroup_stat_names
[i
],
3262 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3265 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3266 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3267 mem_cgroup_read_events(memcg
, i
));
3269 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3270 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3271 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3273 /* Hierarchical information */
3274 memory
= memsw
= PAGE_COUNTER_MAX
;
3275 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3276 memory
= min(memory
, mi
->memory
.limit
);
3277 memsw
= min(memsw
, mi
->memsw
.limit
);
3279 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3280 (u64
)memory
* PAGE_SIZE
);
3281 if (do_memsw_account())
3282 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3283 (u64
)memsw
* PAGE_SIZE
);
3285 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3286 unsigned long long val
= 0;
3288 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3290 for_each_mem_cgroup_tree(mi
, memcg
)
3291 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3292 seq_printf(m
, "total_%s %llu\n", mem_cgroup_stat_names
[i
], val
);
3295 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3296 unsigned long long val
= 0;
3298 for_each_mem_cgroup_tree(mi
, memcg
)
3299 val
+= mem_cgroup_read_events(mi
, i
);
3300 seq_printf(m
, "total_%s %llu\n",
3301 mem_cgroup_events_names
[i
], val
);
3304 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3305 unsigned long long val
= 0;
3307 for_each_mem_cgroup_tree(mi
, memcg
)
3308 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3309 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3312 #ifdef CONFIG_DEBUG_VM
3315 struct mem_cgroup_per_zone
*mz
;
3316 struct zone_reclaim_stat
*rstat
;
3317 unsigned long recent_rotated
[2] = {0, 0};
3318 unsigned long recent_scanned
[2] = {0, 0};
3320 for_each_online_node(nid
)
3321 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3322 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
3323 rstat
= &mz
->lruvec
.reclaim_stat
;
3325 recent_rotated
[0] += rstat
->recent_rotated
[0];
3326 recent_rotated
[1] += rstat
->recent_rotated
[1];
3327 recent_scanned
[0] += rstat
->recent_scanned
[0];
3328 recent_scanned
[1] += rstat
->recent_scanned
[1];
3330 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3331 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3332 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3333 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3340 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3343 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3345 return mem_cgroup_swappiness(memcg
);
3348 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3349 struct cftype
*cft
, u64 val
)
3351 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3357 memcg
->swappiness
= val
;
3359 vm_swappiness
= val
;
3364 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3366 struct mem_cgroup_threshold_ary
*t
;
3367 unsigned long usage
;
3372 t
= rcu_dereference(memcg
->thresholds
.primary
);
3374 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3379 usage
= mem_cgroup_usage(memcg
, swap
);
3382 * current_threshold points to threshold just below or equal to usage.
3383 * If it's not true, a threshold was crossed after last
3384 * call of __mem_cgroup_threshold().
3386 i
= t
->current_threshold
;
3389 * Iterate backward over array of thresholds starting from
3390 * current_threshold and check if a threshold is crossed.
3391 * If none of thresholds below usage is crossed, we read
3392 * only one element of the array here.
3394 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3395 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3397 /* i = current_threshold + 1 */
3401 * Iterate forward over array of thresholds starting from
3402 * current_threshold+1 and check if a threshold is crossed.
3403 * If none of thresholds above usage is crossed, we read
3404 * only one element of the array here.
3406 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3407 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3409 /* Update current_threshold */
3410 t
->current_threshold
= i
- 1;
3415 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3418 __mem_cgroup_threshold(memcg
, false);
3419 if (do_memsw_account())
3420 __mem_cgroup_threshold(memcg
, true);
3422 memcg
= parent_mem_cgroup(memcg
);
3426 static int compare_thresholds(const void *a
, const void *b
)
3428 const struct mem_cgroup_threshold
*_a
= a
;
3429 const struct mem_cgroup_threshold
*_b
= b
;
3431 if (_a
->threshold
> _b
->threshold
)
3434 if (_a
->threshold
< _b
->threshold
)
3440 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3442 struct mem_cgroup_eventfd_list
*ev
;
3444 spin_lock(&memcg_oom_lock
);
3446 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3447 eventfd_signal(ev
->eventfd
, 1);
3449 spin_unlock(&memcg_oom_lock
);
3453 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3455 struct mem_cgroup
*iter
;
3457 for_each_mem_cgroup_tree(iter
, memcg
)
3458 mem_cgroup_oom_notify_cb(iter
);
3461 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3462 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3464 struct mem_cgroup_thresholds
*thresholds
;
3465 struct mem_cgroup_threshold_ary
*new;
3466 unsigned long threshold
;
3467 unsigned long usage
;
3470 ret
= page_counter_memparse(args
, "-1", &threshold
);
3474 mutex_lock(&memcg
->thresholds_lock
);
3477 thresholds
= &memcg
->thresholds
;
3478 usage
= mem_cgroup_usage(memcg
, false);
3479 } else if (type
== _MEMSWAP
) {
3480 thresholds
= &memcg
->memsw_thresholds
;
3481 usage
= mem_cgroup_usage(memcg
, true);
3485 /* Check if a threshold crossed before adding a new one */
3486 if (thresholds
->primary
)
3487 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3489 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3491 /* Allocate memory for new array of thresholds */
3492 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3500 /* Copy thresholds (if any) to new array */
3501 if (thresholds
->primary
) {
3502 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3503 sizeof(struct mem_cgroup_threshold
));
3506 /* Add new threshold */
3507 new->entries
[size
- 1].eventfd
= eventfd
;
3508 new->entries
[size
- 1].threshold
= threshold
;
3510 /* Sort thresholds. Registering of new threshold isn't time-critical */
3511 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3512 compare_thresholds
, NULL
);
3514 /* Find current threshold */
3515 new->current_threshold
= -1;
3516 for (i
= 0; i
< size
; i
++) {
3517 if (new->entries
[i
].threshold
<= usage
) {
3519 * new->current_threshold will not be used until
3520 * rcu_assign_pointer(), so it's safe to increment
3523 ++new->current_threshold
;
3528 /* Free old spare buffer and save old primary buffer as spare */
3529 kfree(thresholds
->spare
);
3530 thresholds
->spare
= thresholds
->primary
;
3532 rcu_assign_pointer(thresholds
->primary
, new);
3534 /* To be sure that nobody uses thresholds */
3538 mutex_unlock(&memcg
->thresholds_lock
);
3543 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3544 struct eventfd_ctx
*eventfd
, const char *args
)
3546 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3549 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3550 struct eventfd_ctx
*eventfd
, const char *args
)
3552 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3555 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3556 struct eventfd_ctx
*eventfd
, enum res_type type
)
3558 struct mem_cgroup_thresholds
*thresholds
;
3559 struct mem_cgroup_threshold_ary
*new;
3560 unsigned long usage
;
3563 mutex_lock(&memcg
->thresholds_lock
);
3566 thresholds
= &memcg
->thresholds
;
3567 usage
= mem_cgroup_usage(memcg
, false);
3568 } else if (type
== _MEMSWAP
) {
3569 thresholds
= &memcg
->memsw_thresholds
;
3570 usage
= mem_cgroup_usage(memcg
, true);
3574 if (!thresholds
->primary
)
3577 /* Check if a threshold crossed before removing */
3578 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3580 /* Calculate new number of threshold */
3582 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3583 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3587 new = thresholds
->spare
;
3589 /* Set thresholds array to NULL if we don't have thresholds */
3598 /* Copy thresholds and find current threshold */
3599 new->current_threshold
= -1;
3600 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3601 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3604 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3605 if (new->entries
[j
].threshold
<= usage
) {
3607 * new->current_threshold will not be used
3608 * until rcu_assign_pointer(), so it's safe to increment
3611 ++new->current_threshold
;
3617 /* Swap primary and spare array */
3618 thresholds
->spare
= thresholds
->primary
;
3620 rcu_assign_pointer(thresholds
->primary
, new);
3622 /* To be sure that nobody uses thresholds */
3625 /* If all events are unregistered, free the spare array */
3627 kfree(thresholds
->spare
);
3628 thresholds
->spare
= NULL
;
3631 mutex_unlock(&memcg
->thresholds_lock
);
3634 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3635 struct eventfd_ctx
*eventfd
)
3637 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3640 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3641 struct eventfd_ctx
*eventfd
)
3643 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3646 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3647 struct eventfd_ctx
*eventfd
, const char *args
)
3649 struct mem_cgroup_eventfd_list
*event
;
3651 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3655 spin_lock(&memcg_oom_lock
);
3657 event
->eventfd
= eventfd
;
3658 list_add(&event
->list
, &memcg
->oom_notify
);
3660 /* already in OOM ? */
3661 if (memcg
->under_oom
)
3662 eventfd_signal(eventfd
, 1);
3663 spin_unlock(&memcg_oom_lock
);
3668 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3669 struct eventfd_ctx
*eventfd
)
3671 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3673 spin_lock(&memcg_oom_lock
);
3675 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3676 if (ev
->eventfd
== eventfd
) {
3677 list_del(&ev
->list
);
3682 spin_unlock(&memcg_oom_lock
);
3685 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3687 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3689 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3690 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3694 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3695 struct cftype
*cft
, u64 val
)
3697 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3699 /* cannot set to root cgroup and only 0 and 1 are allowed */
3700 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3703 memcg
->oom_kill_disable
= val
;
3705 memcg_oom_recover(memcg
);
3710 #ifdef CONFIG_CGROUP_WRITEBACK
3712 struct list_head
*mem_cgroup_cgwb_list(struct mem_cgroup
*memcg
)
3714 return &memcg
->cgwb_list
;
3717 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3719 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3722 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3724 wb_domain_exit(&memcg
->cgwb_domain
);
3727 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3729 wb_domain_size_changed(&memcg
->cgwb_domain
);
3732 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3734 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3736 if (!memcg
->css
.parent
)
3739 return &memcg
->cgwb_domain
;
3743 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3744 * @wb: bdi_writeback in question
3745 * @pfilepages: out parameter for number of file pages
3746 * @pheadroom: out parameter for number of allocatable pages according to memcg
3747 * @pdirty: out parameter for number of dirty pages
3748 * @pwriteback: out parameter for number of pages under writeback
3750 * Determine the numbers of file, headroom, dirty, and writeback pages in
3751 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3752 * is a bit more involved.
3754 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3755 * headroom is calculated as the lowest headroom of itself and the
3756 * ancestors. Note that this doesn't consider the actual amount of
3757 * available memory in the system. The caller should further cap
3758 * *@pheadroom accordingly.
3760 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
3761 unsigned long *pheadroom
, unsigned long *pdirty
,
3762 unsigned long *pwriteback
)
3764 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3765 struct mem_cgroup
*parent
;
3767 *pdirty
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
3769 /* this should eventually include NR_UNSTABLE_NFS */
3770 *pwriteback
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
3771 *pfilepages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
3772 (1 << LRU_ACTIVE_FILE
));
3773 *pheadroom
= PAGE_COUNTER_MAX
;
3775 while ((parent
= parent_mem_cgroup(memcg
))) {
3776 unsigned long ceiling
= min(memcg
->memory
.limit
, memcg
->high
);
3777 unsigned long used
= page_counter_read(&memcg
->memory
);
3779 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
3784 #else /* CONFIG_CGROUP_WRITEBACK */
3786 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3791 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3795 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3799 #endif /* CONFIG_CGROUP_WRITEBACK */
3802 * DO NOT USE IN NEW FILES.
3804 * "cgroup.event_control" implementation.
3806 * This is way over-engineered. It tries to support fully configurable
3807 * events for each user. Such level of flexibility is completely
3808 * unnecessary especially in the light of the planned unified hierarchy.
3810 * Please deprecate this and replace with something simpler if at all
3815 * Unregister event and free resources.
3817 * Gets called from workqueue.
3819 static void memcg_event_remove(struct work_struct
*work
)
3821 struct mem_cgroup_event
*event
=
3822 container_of(work
, struct mem_cgroup_event
, remove
);
3823 struct mem_cgroup
*memcg
= event
->memcg
;
3825 remove_wait_queue(event
->wqh
, &event
->wait
);
3827 event
->unregister_event(memcg
, event
->eventfd
);
3829 /* Notify userspace the event is going away. */
3830 eventfd_signal(event
->eventfd
, 1);
3832 eventfd_ctx_put(event
->eventfd
);
3834 css_put(&memcg
->css
);
3838 * Gets called on POLLHUP on eventfd when user closes it.
3840 * Called with wqh->lock held and interrupts disabled.
3842 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
3843 int sync
, void *key
)
3845 struct mem_cgroup_event
*event
=
3846 container_of(wait
, struct mem_cgroup_event
, wait
);
3847 struct mem_cgroup
*memcg
= event
->memcg
;
3848 unsigned long flags
= (unsigned long)key
;
3850 if (flags
& POLLHUP
) {
3852 * If the event has been detached at cgroup removal, we
3853 * can simply return knowing the other side will cleanup
3856 * We can't race against event freeing since the other
3857 * side will require wqh->lock via remove_wait_queue(),
3860 spin_lock(&memcg
->event_list_lock
);
3861 if (!list_empty(&event
->list
)) {
3862 list_del_init(&event
->list
);
3864 * We are in atomic context, but cgroup_event_remove()
3865 * may sleep, so we have to call it in workqueue.
3867 schedule_work(&event
->remove
);
3869 spin_unlock(&memcg
->event_list_lock
);
3875 static void memcg_event_ptable_queue_proc(struct file
*file
,
3876 wait_queue_head_t
*wqh
, poll_table
*pt
)
3878 struct mem_cgroup_event
*event
=
3879 container_of(pt
, struct mem_cgroup_event
, pt
);
3882 add_wait_queue(wqh
, &event
->wait
);
3886 * DO NOT USE IN NEW FILES.
3888 * Parse input and register new cgroup event handler.
3890 * Input must be in format '<event_fd> <control_fd> <args>'.
3891 * Interpretation of args is defined by control file implementation.
3893 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
3894 char *buf
, size_t nbytes
, loff_t off
)
3896 struct cgroup_subsys_state
*css
= of_css(of
);
3897 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3898 struct mem_cgroup_event
*event
;
3899 struct cgroup_subsys_state
*cfile_css
;
3900 unsigned int efd
, cfd
;
3907 buf
= strstrip(buf
);
3909 efd
= simple_strtoul(buf
, &endp
, 10);
3914 cfd
= simple_strtoul(buf
, &endp
, 10);
3915 if ((*endp
!= ' ') && (*endp
!= '\0'))
3919 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
3923 event
->memcg
= memcg
;
3924 INIT_LIST_HEAD(&event
->list
);
3925 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
3926 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
3927 INIT_WORK(&event
->remove
, memcg_event_remove
);
3935 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
3936 if (IS_ERR(event
->eventfd
)) {
3937 ret
= PTR_ERR(event
->eventfd
);
3944 goto out_put_eventfd
;
3947 /* the process need read permission on control file */
3948 /* AV: shouldn't we check that it's been opened for read instead? */
3949 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
3954 * Determine the event callbacks and set them in @event. This used
3955 * to be done via struct cftype but cgroup core no longer knows
3956 * about these events. The following is crude but the whole thing
3957 * is for compatibility anyway.
3959 * DO NOT ADD NEW FILES.
3961 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
3963 if (!strcmp(name
, "memory.usage_in_bytes")) {
3964 event
->register_event
= mem_cgroup_usage_register_event
;
3965 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
3966 } else if (!strcmp(name
, "memory.oom_control")) {
3967 event
->register_event
= mem_cgroup_oom_register_event
;
3968 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
3969 } else if (!strcmp(name
, "memory.pressure_level")) {
3970 event
->register_event
= vmpressure_register_event
;
3971 event
->unregister_event
= vmpressure_unregister_event
;
3972 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
3973 event
->register_event
= memsw_cgroup_usage_register_event
;
3974 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
3981 * Verify @cfile should belong to @css. Also, remaining events are
3982 * automatically removed on cgroup destruction but the removal is
3983 * asynchronous, so take an extra ref on @css.
3985 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
3986 &memory_cgrp_subsys
);
3988 if (IS_ERR(cfile_css
))
3990 if (cfile_css
!= css
) {
3995 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
3999 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
4001 spin_lock(&memcg
->event_list_lock
);
4002 list_add(&event
->list
, &memcg
->event_list
);
4003 spin_unlock(&memcg
->event_list_lock
);
4015 eventfd_ctx_put(event
->eventfd
);
4024 static struct cftype mem_cgroup_legacy_files
[] = {
4026 .name
= "usage_in_bytes",
4027 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4028 .read_u64
= mem_cgroup_read_u64
,
4031 .name
= "max_usage_in_bytes",
4032 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4033 .write
= mem_cgroup_reset
,
4034 .read_u64
= mem_cgroup_read_u64
,
4037 .name
= "limit_in_bytes",
4038 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4039 .write
= mem_cgroup_write
,
4040 .read_u64
= mem_cgroup_read_u64
,
4043 .name
= "soft_limit_in_bytes",
4044 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4045 .write
= mem_cgroup_write
,
4046 .read_u64
= mem_cgroup_read_u64
,
4050 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4051 .write
= mem_cgroup_reset
,
4052 .read_u64
= mem_cgroup_read_u64
,
4056 .seq_show
= memcg_stat_show
,
4059 .name
= "force_empty",
4060 .write
= mem_cgroup_force_empty_write
,
4063 .name
= "use_hierarchy",
4064 .write_u64
= mem_cgroup_hierarchy_write
,
4065 .read_u64
= mem_cgroup_hierarchy_read
,
4068 .name
= "cgroup.event_control", /* XXX: for compat */
4069 .write
= memcg_write_event_control
,
4070 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
4073 .name
= "swappiness",
4074 .read_u64
= mem_cgroup_swappiness_read
,
4075 .write_u64
= mem_cgroup_swappiness_write
,
4078 .name
= "move_charge_at_immigrate",
4079 .read_u64
= mem_cgroup_move_charge_read
,
4080 .write_u64
= mem_cgroup_move_charge_write
,
4083 .name
= "oom_control",
4084 .seq_show
= mem_cgroup_oom_control_read
,
4085 .write_u64
= mem_cgroup_oom_control_write
,
4086 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4089 .name
= "pressure_level",
4093 .name
= "numa_stat",
4094 .seq_show
= memcg_numa_stat_show
,
4097 #ifdef CONFIG_MEMCG_LEGACY_KMEM
4099 .name
= "kmem.limit_in_bytes",
4100 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4101 .write
= mem_cgroup_write
,
4102 .read_u64
= mem_cgroup_read_u64
,
4105 .name
= "kmem.usage_in_bytes",
4106 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4107 .read_u64
= mem_cgroup_read_u64
,
4110 .name
= "kmem.failcnt",
4111 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4112 .write
= mem_cgroup_reset
,
4113 .read_u64
= mem_cgroup_read_u64
,
4116 .name
= "kmem.max_usage_in_bytes",
4117 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4118 .write
= mem_cgroup_reset
,
4119 .read_u64
= mem_cgroup_read_u64
,
4121 #ifdef CONFIG_SLABINFO
4123 .name
= "kmem.slabinfo",
4124 .seq_start
= slab_start
,
4125 .seq_next
= slab_next
,
4126 .seq_stop
= slab_stop
,
4127 .seq_show
= memcg_slab_show
,
4132 .name
= "kmem.tcp.limit_in_bytes",
4133 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4134 .write
= mem_cgroup_write
,
4135 .read_u64
= mem_cgroup_read_u64
,
4138 .name
= "kmem.tcp.usage_in_bytes",
4139 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4140 .read_u64
= mem_cgroup_read_u64
,
4143 .name
= "kmem.tcp.failcnt",
4144 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4145 .write
= mem_cgroup_reset
,
4146 .read_u64
= mem_cgroup_read_u64
,
4149 .name
= "kmem.tcp.max_usage_in_bytes",
4150 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4151 .write
= mem_cgroup_reset
,
4152 .read_u64
= mem_cgroup_read_u64
,
4156 { }, /* terminate */
4159 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4161 struct mem_cgroup_per_node
*pn
;
4162 struct mem_cgroup_per_zone
*mz
;
4163 int zone
, tmp
= node
;
4165 * This routine is called against possible nodes.
4166 * But it's BUG to call kmalloc() against offline node.
4168 * TODO: this routine can waste much memory for nodes which will
4169 * never be onlined. It's better to use memory hotplug callback
4172 if (!node_state(node
, N_NORMAL_MEMORY
))
4174 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4178 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4179 mz
= &pn
->zoneinfo
[zone
];
4180 lruvec_init(&mz
->lruvec
);
4181 mz
->usage_in_excess
= 0;
4182 mz
->on_tree
= false;
4185 memcg
->nodeinfo
[node
] = pn
;
4189 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4191 kfree(memcg
->nodeinfo
[node
]);
4194 static struct mem_cgroup
*mem_cgroup_alloc(void)
4196 struct mem_cgroup
*memcg
;
4199 size
= sizeof(struct mem_cgroup
);
4200 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4202 memcg
= kzalloc(size
, GFP_KERNEL
);
4206 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4210 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4216 free_percpu(memcg
->stat
);
4223 * At destroying mem_cgroup, references from swap_cgroup can remain.
4224 * (scanning all at force_empty is too costly...)
4226 * Instead of clearing all references at force_empty, we remember
4227 * the number of reference from swap_cgroup and free mem_cgroup when
4228 * it goes down to 0.
4230 * Removal of cgroup itself succeeds regardless of refs from swap.
4233 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
4237 cancel_work_sync(&memcg
->high_work
);
4239 mem_cgroup_remove_from_trees(memcg
);
4242 free_mem_cgroup_per_zone_info(memcg
, node
);
4244 free_percpu(memcg
->stat
);
4245 memcg_wb_domain_exit(memcg
);
4249 static struct cgroup_subsys_state
* __ref
4250 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4252 struct mem_cgroup
*memcg
;
4253 long error
= -ENOMEM
;
4256 memcg
= mem_cgroup_alloc();
4258 return ERR_PTR(error
);
4261 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
4265 if (parent_css
== NULL
) {
4266 root_mem_cgroup
= memcg
;
4267 page_counter_init(&memcg
->memory
, NULL
);
4268 memcg
->high
= PAGE_COUNTER_MAX
;
4269 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4270 page_counter_init(&memcg
->memsw
, NULL
);
4271 page_counter_init(&memcg
->kmem
, NULL
);
4274 INIT_WORK(&memcg
->high_work
, high_work_func
);
4275 memcg
->last_scanned_node
= MAX_NUMNODES
;
4276 INIT_LIST_HEAD(&memcg
->oom_notify
);
4277 memcg
->move_charge_at_immigrate
= 0;
4278 mutex_init(&memcg
->thresholds_lock
);
4279 spin_lock_init(&memcg
->move_lock
);
4280 vmpressure_init(&memcg
->vmpressure
);
4281 INIT_LIST_HEAD(&memcg
->event_list
);
4282 spin_lock_init(&memcg
->event_list_lock
);
4284 memcg
->kmemcg_id
= -1;
4286 #ifdef CONFIG_CGROUP_WRITEBACK
4287 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4290 memcg
->socket_pressure
= jiffies
;
4295 __mem_cgroup_free(memcg
);
4296 return ERR_PTR(error
);
4300 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4302 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4303 struct mem_cgroup
*parent
= mem_cgroup_from_css(css
->parent
);
4306 if (css
->id
> MEM_CGROUP_ID_MAX
)
4312 mutex_lock(&memcg_create_mutex
);
4314 memcg
->use_hierarchy
= parent
->use_hierarchy
;
4315 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4316 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4318 if (parent
->use_hierarchy
) {
4319 page_counter_init(&memcg
->memory
, &parent
->memory
);
4320 memcg
->high
= PAGE_COUNTER_MAX
;
4321 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4322 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4323 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4324 #if defined(CONFIG_MEMCG_LEGACY_KMEM) && defined(CONFIG_INET)
4325 page_counter_init(&memcg
->tcp_mem
.memory_allocated
,
4326 &parent
->tcp_mem
.memory_allocated
);
4330 * No need to take a reference to the parent because cgroup
4331 * core guarantees its existence.
4334 page_counter_init(&memcg
->memory
, NULL
);
4335 memcg
->high
= PAGE_COUNTER_MAX
;
4336 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4337 page_counter_init(&memcg
->memsw
, NULL
);
4338 page_counter_init(&memcg
->kmem
, NULL
);
4339 #if defined(CONFIG_MEMCG_LEGACY_KMEM) && defined(CONFIG_INET)
4340 page_counter_init(&memcg
->tcp_mem
.memory_allocated
, NULL
);
4343 * Deeper hierachy with use_hierarchy == false doesn't make
4344 * much sense so let cgroup subsystem know about this
4345 * unfortunate state in our controller.
4347 if (parent
!= root_mem_cgroup
)
4348 memory_cgrp_subsys
.broken_hierarchy
= true;
4350 mutex_unlock(&memcg_create_mutex
);
4352 ret
= memcg_propagate_kmem(memcg
);
4357 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4358 static_branch_inc(&memcg_sockets_enabled_key
);
4362 * Make sure the memcg is initialized: mem_cgroup_iter()
4363 * orders reading memcg->initialized against its callers
4364 * reading the memcg members.
4366 smp_store_release(&memcg
->initialized
, 1);
4371 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4373 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4374 struct mem_cgroup_event
*event
, *tmp
;
4377 * Unregister events and notify userspace.
4378 * Notify userspace about cgroup removing only after rmdir of cgroup
4379 * directory to avoid race between userspace and kernelspace.
4381 spin_lock(&memcg
->event_list_lock
);
4382 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4383 list_del_init(&event
->list
);
4384 schedule_work(&event
->remove
);
4386 spin_unlock(&memcg
->event_list_lock
);
4388 vmpressure_cleanup(&memcg
->vmpressure
);
4390 memcg_offline_kmem(memcg
);
4392 wb_memcg_offline(memcg
);
4395 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4397 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4399 invalidate_reclaim_iterators(memcg
);
4402 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4404 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4407 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4408 static_branch_dec(&memcg_sockets_enabled_key
);
4411 memcg_free_kmem(memcg
);
4413 #if defined(CONFIG_MEMCG_LEGACY_KMEM) && defined(CONFIG_INET)
4414 if (memcg
->tcp_mem
.active
)
4415 static_branch_dec(&memcg_sockets_enabled_key
);
4418 __mem_cgroup_free(memcg
);
4422 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4423 * @css: the target css
4425 * Reset the states of the mem_cgroup associated with @css. This is
4426 * invoked when the userland requests disabling on the default hierarchy
4427 * but the memcg is pinned through dependency. The memcg should stop
4428 * applying policies and should revert to the vanilla state as it may be
4429 * made visible again.
4431 * The current implementation only resets the essential configurations.
4432 * This needs to be expanded to cover all the visible parts.
4434 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4436 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4438 mem_cgroup_resize_limit(memcg
, PAGE_COUNTER_MAX
);
4439 mem_cgroup_resize_memsw_limit(memcg
, PAGE_COUNTER_MAX
);
4440 memcg_update_kmem_limit(memcg
, PAGE_COUNTER_MAX
);
4442 memcg
->high
= PAGE_COUNTER_MAX
;
4443 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4444 memcg_wb_domain_size_changed(memcg
);
4448 /* Handlers for move charge at task migration. */
4449 static int mem_cgroup_do_precharge(unsigned long count
)
4453 /* Try a single bulk charge without reclaim first, kswapd may wake */
4454 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4456 mc
.precharge
+= count
;
4460 /* Try charges one by one with reclaim */
4462 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4472 * get_mctgt_type - get target type of moving charge
4473 * @vma: the vma the pte to be checked belongs
4474 * @addr: the address corresponding to the pte to be checked
4475 * @ptent: the pte to be checked
4476 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4479 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4480 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4481 * move charge. if @target is not NULL, the page is stored in target->page
4482 * with extra refcnt got(Callers should handle it).
4483 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4484 * target for charge migration. if @target is not NULL, the entry is stored
4487 * Called with pte lock held.
4494 enum mc_target_type
{
4500 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4501 unsigned long addr
, pte_t ptent
)
4503 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4505 if (!page
|| !page_mapped(page
))
4507 if (PageAnon(page
)) {
4508 if (!(mc
.flags
& MOVE_ANON
))
4511 if (!(mc
.flags
& MOVE_FILE
))
4514 if (!get_page_unless_zero(page
))
4521 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4522 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4524 struct page
*page
= NULL
;
4525 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4527 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4530 * Because lookup_swap_cache() updates some statistics counter,
4531 * we call find_get_page() with swapper_space directly.
4533 page
= find_get_page(swap_address_space(ent
), ent
.val
);
4534 if (do_memsw_account())
4535 entry
->val
= ent
.val
;
4540 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4541 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4547 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4548 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4550 struct page
*page
= NULL
;
4551 struct address_space
*mapping
;
4554 if (!vma
->vm_file
) /* anonymous vma */
4556 if (!(mc
.flags
& MOVE_FILE
))
4559 mapping
= vma
->vm_file
->f_mapping
;
4560 pgoff
= linear_page_index(vma
, addr
);
4562 /* page is moved even if it's not RSS of this task(page-faulted). */
4564 /* shmem/tmpfs may report page out on swap: account for that too. */
4565 if (shmem_mapping(mapping
)) {
4566 page
= find_get_entry(mapping
, pgoff
);
4567 if (radix_tree_exceptional_entry(page
)) {
4568 swp_entry_t swp
= radix_to_swp_entry(page
);
4569 if (do_memsw_account())
4571 page
= find_get_page(swap_address_space(swp
), swp
.val
);
4574 page
= find_get_page(mapping
, pgoff
);
4576 page
= find_get_page(mapping
, pgoff
);
4582 * mem_cgroup_move_account - move account of the page
4584 * @nr_pages: number of regular pages (>1 for huge pages)
4585 * @from: mem_cgroup which the page is moved from.
4586 * @to: mem_cgroup which the page is moved to. @from != @to.
4588 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4590 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4593 static int mem_cgroup_move_account(struct page
*page
,
4595 struct mem_cgroup
*from
,
4596 struct mem_cgroup
*to
)
4598 unsigned long flags
;
4599 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
4603 VM_BUG_ON(from
== to
);
4604 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4605 VM_BUG_ON(compound
&& !PageTransHuge(page
));
4608 * Prevent mem_cgroup_replace_page() from looking at
4609 * page->mem_cgroup of its source page while we change it.
4612 if (!trylock_page(page
))
4616 if (page
->mem_cgroup
!= from
)
4619 anon
= PageAnon(page
);
4621 spin_lock_irqsave(&from
->move_lock
, flags
);
4623 if (!anon
&& page_mapped(page
)) {
4624 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4626 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4631 * move_lock grabbed above and caller set from->moving_account, so
4632 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4633 * So mapping should be stable for dirty pages.
4635 if (!anon
&& PageDirty(page
)) {
4636 struct address_space
*mapping
= page_mapping(page
);
4638 if (mapping_cap_account_dirty(mapping
)) {
4639 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4641 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4646 if (PageWriteback(page
)) {
4647 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4649 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4654 * It is safe to change page->mem_cgroup here because the page
4655 * is referenced, charged, and isolated - we can't race with
4656 * uncharging, charging, migration, or LRU putback.
4659 /* caller should have done css_get */
4660 page
->mem_cgroup
= to
;
4661 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4665 local_irq_disable();
4666 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
4667 memcg_check_events(to
, page
);
4668 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
4669 memcg_check_events(from
, page
);
4677 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4678 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4680 struct page
*page
= NULL
;
4681 enum mc_target_type ret
= MC_TARGET_NONE
;
4682 swp_entry_t ent
= { .val
= 0 };
4684 if (pte_present(ptent
))
4685 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4686 else if (is_swap_pte(ptent
))
4687 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
4688 else if (pte_none(ptent
))
4689 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4691 if (!page
&& !ent
.val
)
4695 * Do only loose check w/o serialization.
4696 * mem_cgroup_move_account() checks the page is valid or
4697 * not under LRU exclusion.
4699 if (page
->mem_cgroup
== mc
.from
) {
4700 ret
= MC_TARGET_PAGE
;
4702 target
->page
= page
;
4704 if (!ret
|| !target
)
4707 /* There is a swap entry and a page doesn't exist or isn't charged */
4708 if (ent
.val
&& !ret
&&
4709 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4710 ret
= MC_TARGET_SWAP
;
4717 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4719 * We don't consider swapping or file mapped pages because THP does not
4720 * support them for now.
4721 * Caller should make sure that pmd_trans_huge(pmd) is true.
4723 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4724 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4726 struct page
*page
= NULL
;
4727 enum mc_target_type ret
= MC_TARGET_NONE
;
4729 page
= pmd_page(pmd
);
4730 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4731 if (!(mc
.flags
& MOVE_ANON
))
4733 if (page
->mem_cgroup
== mc
.from
) {
4734 ret
= MC_TARGET_PAGE
;
4737 target
->page
= page
;
4743 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4744 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4746 return MC_TARGET_NONE
;
4750 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4751 unsigned long addr
, unsigned long end
,
4752 struct mm_walk
*walk
)
4754 struct vm_area_struct
*vma
= walk
->vma
;
4758 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
)) {
4759 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4760 mc
.precharge
+= HPAGE_PMD_NR
;
4765 if (pmd_trans_unstable(pmd
))
4767 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4768 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4769 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4770 mc
.precharge
++; /* increment precharge temporarily */
4771 pte_unmap_unlock(pte
- 1, ptl
);
4777 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4779 unsigned long precharge
;
4781 struct mm_walk mem_cgroup_count_precharge_walk
= {
4782 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4785 down_read(&mm
->mmap_sem
);
4786 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk
);
4787 up_read(&mm
->mmap_sem
);
4789 precharge
= mc
.precharge
;
4795 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4797 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4799 VM_BUG_ON(mc
.moving_task
);
4800 mc
.moving_task
= current
;
4801 return mem_cgroup_do_precharge(precharge
);
4804 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4805 static void __mem_cgroup_clear_mc(void)
4807 struct mem_cgroup
*from
= mc
.from
;
4808 struct mem_cgroup
*to
= mc
.to
;
4810 /* we must uncharge all the leftover precharges from mc.to */
4812 cancel_charge(mc
.to
, mc
.precharge
);
4816 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4817 * we must uncharge here.
4819 if (mc
.moved_charge
) {
4820 cancel_charge(mc
.from
, mc
.moved_charge
);
4821 mc
.moved_charge
= 0;
4823 /* we must fixup refcnts and charges */
4824 if (mc
.moved_swap
) {
4825 /* uncharge swap account from the old cgroup */
4826 if (!mem_cgroup_is_root(mc
.from
))
4827 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4830 * we charged both to->memory and to->memsw, so we
4831 * should uncharge to->memory.
4833 if (!mem_cgroup_is_root(mc
.to
))
4834 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4836 css_put_many(&mc
.from
->css
, mc
.moved_swap
);
4838 /* we've already done css_get(mc.to) */
4841 memcg_oom_recover(from
);
4842 memcg_oom_recover(to
);
4843 wake_up_all(&mc
.waitq
);
4846 static void mem_cgroup_clear_mc(void)
4849 * we must clear moving_task before waking up waiters at the end of
4852 mc
.moving_task
= NULL
;
4853 __mem_cgroup_clear_mc();
4854 spin_lock(&mc
.lock
);
4857 spin_unlock(&mc
.lock
);
4860 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4862 struct cgroup_subsys_state
*css
;
4863 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
4864 struct mem_cgroup
*from
;
4865 struct task_struct
*leader
, *p
;
4866 struct mm_struct
*mm
;
4867 unsigned long move_flags
;
4870 /* charge immigration isn't supported on the default hierarchy */
4871 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4875 * Multi-process migrations only happen on the default hierarchy
4876 * where charge immigration is not used. Perform charge
4877 * immigration if @tset contains a leader and whine if there are
4881 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
4884 memcg
= mem_cgroup_from_css(css
);
4890 * We are now commited to this value whatever it is. Changes in this
4891 * tunable will only affect upcoming migrations, not the current one.
4892 * So we need to save it, and keep it going.
4894 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
4898 from
= mem_cgroup_from_task(p
);
4900 VM_BUG_ON(from
== memcg
);
4902 mm
= get_task_mm(p
);
4905 /* We move charges only when we move a owner of the mm */
4906 if (mm
->owner
== p
) {
4909 VM_BUG_ON(mc
.precharge
);
4910 VM_BUG_ON(mc
.moved_charge
);
4911 VM_BUG_ON(mc
.moved_swap
);
4913 spin_lock(&mc
.lock
);
4916 mc
.flags
= move_flags
;
4917 spin_unlock(&mc
.lock
);
4918 /* We set mc.moving_task later */
4920 ret
= mem_cgroup_precharge_mc(mm
);
4922 mem_cgroup_clear_mc();
4928 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4931 mem_cgroup_clear_mc();
4934 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4935 unsigned long addr
, unsigned long end
,
4936 struct mm_walk
*walk
)
4939 struct vm_area_struct
*vma
= walk
->vma
;
4942 enum mc_target_type target_type
;
4943 union mc_target target
;
4946 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
)) {
4947 if (mc
.precharge
< HPAGE_PMD_NR
) {
4951 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
4952 if (target_type
== MC_TARGET_PAGE
) {
4954 if (!isolate_lru_page(page
)) {
4955 if (!mem_cgroup_move_account(page
, true,
4957 mc
.precharge
-= HPAGE_PMD_NR
;
4958 mc
.moved_charge
+= HPAGE_PMD_NR
;
4960 putback_lru_page(page
);
4968 if (pmd_trans_unstable(pmd
))
4971 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4972 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4973 pte_t ptent
= *(pte
++);
4979 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
4980 case MC_TARGET_PAGE
:
4983 * We can have a part of the split pmd here. Moving it
4984 * can be done but it would be too convoluted so simply
4985 * ignore such a partial THP and keep it in original
4986 * memcg. There should be somebody mapping the head.
4988 if (PageTransCompound(page
))
4990 if (isolate_lru_page(page
))
4992 if (!mem_cgroup_move_account(page
, false,
4995 /* we uncharge from mc.from later. */
4998 putback_lru_page(page
);
4999 put
: /* get_mctgt_type() gets the page */
5002 case MC_TARGET_SWAP
:
5004 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
5006 /* we fixup refcnts and charges later. */
5014 pte_unmap_unlock(pte
- 1, ptl
);
5019 * We have consumed all precharges we got in can_attach().
5020 * We try charge one by one, but don't do any additional
5021 * charges to mc.to if we have failed in charge once in attach()
5024 ret
= mem_cgroup_do_precharge(1);
5032 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
5034 struct mm_walk mem_cgroup_move_charge_walk
= {
5035 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
5039 lru_add_drain_all();
5041 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5042 * move_lock while we're moving its pages to another memcg.
5043 * Then wait for already started RCU-only updates to finish.
5045 atomic_inc(&mc
.from
->moving_account
);
5048 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
5050 * Someone who are holding the mmap_sem might be waiting in
5051 * waitq. So we cancel all extra charges, wake up all waiters,
5052 * and retry. Because we cancel precharges, we might not be able
5053 * to move enough charges, but moving charge is a best-effort
5054 * feature anyway, so it wouldn't be a big problem.
5056 __mem_cgroup_clear_mc();
5061 * When we have consumed all precharges and failed in doing
5062 * additional charge, the page walk just aborts.
5064 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk
);
5065 up_read(&mm
->mmap_sem
);
5066 atomic_dec(&mc
.from
->moving_account
);
5069 static void mem_cgroup_move_task(struct cgroup_taskset
*tset
)
5071 struct cgroup_subsys_state
*css
;
5072 struct task_struct
*p
= cgroup_taskset_first(tset
, &css
);
5073 struct mm_struct
*mm
= get_task_mm(p
);
5077 mem_cgroup_move_charge(mm
);
5081 mem_cgroup_clear_mc();
5083 #else /* !CONFIG_MMU */
5084 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5088 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5091 static void mem_cgroup_move_task(struct cgroup_taskset
*tset
)
5097 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5098 * to verify whether we're attached to the default hierarchy on each mount
5101 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5104 * use_hierarchy is forced on the default hierarchy. cgroup core
5105 * guarantees that @root doesn't have any children, so turning it
5106 * on for the root memcg is enough.
5108 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5109 root_mem_cgroup
->use_hierarchy
= true;
5111 root_mem_cgroup
->use_hierarchy
= false;
5114 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5117 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5119 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
5122 static int memory_low_show(struct seq_file
*m
, void *v
)
5124 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5125 unsigned long low
= READ_ONCE(memcg
->low
);
5127 if (low
== PAGE_COUNTER_MAX
)
5128 seq_puts(m
, "max\n");
5130 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
5135 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5136 char *buf
, size_t nbytes
, loff_t off
)
5138 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5142 buf
= strstrip(buf
);
5143 err
= page_counter_memparse(buf
, "max", &low
);
5152 static int memory_high_show(struct seq_file
*m
, void *v
)
5154 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5155 unsigned long high
= READ_ONCE(memcg
->high
);
5157 if (high
== PAGE_COUNTER_MAX
)
5158 seq_puts(m
, "max\n");
5160 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
5165 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5166 char *buf
, size_t nbytes
, loff_t off
)
5168 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5172 buf
= strstrip(buf
);
5173 err
= page_counter_memparse(buf
, "max", &high
);
5179 memcg_wb_domain_size_changed(memcg
);
5183 static int memory_max_show(struct seq_file
*m
, void *v
)
5185 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5186 unsigned long max
= READ_ONCE(memcg
->memory
.limit
);
5188 if (max
== PAGE_COUNTER_MAX
)
5189 seq_puts(m
, "max\n");
5191 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5196 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5197 char *buf
, size_t nbytes
, loff_t off
)
5199 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5203 buf
= strstrip(buf
);
5204 err
= page_counter_memparse(buf
, "max", &max
);
5208 err
= mem_cgroup_resize_limit(memcg
, max
);
5212 memcg_wb_domain_size_changed(memcg
);
5216 static int memory_events_show(struct seq_file
*m
, void *v
)
5218 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5220 seq_printf(m
, "low %lu\n", mem_cgroup_read_events(memcg
, MEMCG_LOW
));
5221 seq_printf(m
, "high %lu\n", mem_cgroup_read_events(memcg
, MEMCG_HIGH
));
5222 seq_printf(m
, "max %lu\n", mem_cgroup_read_events(memcg
, MEMCG_MAX
));
5223 seq_printf(m
, "oom %lu\n", mem_cgroup_read_events(memcg
, MEMCG_OOM
));
5228 static struct cftype memory_files
[] = {
5231 .flags
= CFTYPE_NOT_ON_ROOT
,
5232 .read_u64
= memory_current_read
,
5236 .flags
= CFTYPE_NOT_ON_ROOT
,
5237 .seq_show
= memory_low_show
,
5238 .write
= memory_low_write
,
5242 .flags
= CFTYPE_NOT_ON_ROOT
,
5243 .seq_show
= memory_high_show
,
5244 .write
= memory_high_write
,
5248 .flags
= CFTYPE_NOT_ON_ROOT
,
5249 .seq_show
= memory_max_show
,
5250 .write
= memory_max_write
,
5254 .flags
= CFTYPE_NOT_ON_ROOT
,
5255 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
5256 .seq_show
= memory_events_show
,
5261 struct cgroup_subsys memory_cgrp_subsys
= {
5262 .css_alloc
= mem_cgroup_css_alloc
,
5263 .css_online
= mem_cgroup_css_online
,
5264 .css_offline
= mem_cgroup_css_offline
,
5265 .css_released
= mem_cgroup_css_released
,
5266 .css_free
= mem_cgroup_css_free
,
5267 .css_reset
= mem_cgroup_css_reset
,
5268 .can_attach
= mem_cgroup_can_attach
,
5269 .cancel_attach
= mem_cgroup_cancel_attach
,
5270 .attach
= mem_cgroup_move_task
,
5271 .bind
= mem_cgroup_bind
,
5272 .dfl_cftypes
= memory_files
,
5273 .legacy_cftypes
= mem_cgroup_legacy_files
,
5278 * mem_cgroup_low - check if memory consumption is below the normal range
5279 * @root: the highest ancestor to consider
5280 * @memcg: the memory cgroup to check
5282 * Returns %true if memory consumption of @memcg, and that of all
5283 * configurable ancestors up to @root, is below the normal range.
5285 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5287 if (mem_cgroup_disabled())
5291 * The toplevel group doesn't have a configurable range, so
5292 * it's never low when looked at directly, and it is not
5293 * considered an ancestor when assessing the hierarchy.
5296 if (memcg
== root_mem_cgroup
)
5299 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5302 while (memcg
!= root
) {
5303 memcg
= parent_mem_cgroup(memcg
);
5305 if (memcg
== root_mem_cgroup
)
5308 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5315 * mem_cgroup_try_charge - try charging a page
5316 * @page: page to charge
5317 * @mm: mm context of the victim
5318 * @gfp_mask: reclaim mode
5319 * @memcgp: charged memcg return
5321 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5322 * pages according to @gfp_mask if necessary.
5324 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5325 * Otherwise, an error code is returned.
5327 * After page->mapping has been set up, the caller must finalize the
5328 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5329 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5331 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5332 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
5335 struct mem_cgroup
*memcg
= NULL
;
5336 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5339 if (mem_cgroup_disabled())
5342 if (PageSwapCache(page
)) {
5344 * Every swap fault against a single page tries to charge the
5345 * page, bail as early as possible. shmem_unuse() encounters
5346 * already charged pages, too. The USED bit is protected by
5347 * the page lock, which serializes swap cache removal, which
5348 * in turn serializes uncharging.
5350 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5351 if (page
->mem_cgroup
)
5354 if (do_memsw_account()) {
5355 swp_entry_t ent
= { .val
= page_private(page
), };
5356 unsigned short id
= lookup_swap_cgroup_id(ent
);
5359 memcg
= mem_cgroup_from_id(id
);
5360 if (memcg
&& !css_tryget_online(&memcg
->css
))
5367 memcg
= get_mem_cgroup_from_mm(mm
);
5369 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5371 css_put(&memcg
->css
);
5378 * mem_cgroup_commit_charge - commit a page charge
5379 * @page: page to charge
5380 * @memcg: memcg to charge the page to
5381 * @lrucare: page might be on LRU already
5383 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5384 * after page->mapping has been set up. This must happen atomically
5385 * as part of the page instantiation, i.e. under the page table lock
5386 * for anonymous pages, under the page lock for page and swap cache.
5388 * In addition, the page must not be on the LRU during the commit, to
5389 * prevent racing with task migration. If it might be, use @lrucare.
5391 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5393 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5394 bool lrucare
, bool compound
)
5396 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5398 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5399 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5401 if (mem_cgroup_disabled())
5404 * Swap faults will attempt to charge the same page multiple
5405 * times. But reuse_swap_page() might have removed the page
5406 * from swapcache already, so we can't check PageSwapCache().
5411 commit_charge(page
, memcg
, lrucare
);
5413 local_irq_disable();
5414 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
5415 memcg_check_events(memcg
, page
);
5418 if (do_memsw_account() && PageSwapCache(page
)) {
5419 swp_entry_t entry
= { .val
= page_private(page
) };
5421 * The swap entry might not get freed for a long time,
5422 * let's not wait for it. The page already received a
5423 * memory+swap charge, drop the swap entry duplicate.
5425 mem_cgroup_uncharge_swap(entry
);
5430 * mem_cgroup_cancel_charge - cancel a page charge
5431 * @page: page to charge
5432 * @memcg: memcg to charge the page to
5434 * Cancel a charge transaction started by mem_cgroup_try_charge().
5436 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5439 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5441 if (mem_cgroup_disabled())
5444 * Swap faults will attempt to charge the same page multiple
5445 * times. But reuse_swap_page() might have removed the page
5446 * from swapcache already, so we can't check PageSwapCache().
5451 cancel_charge(memcg
, nr_pages
);
5454 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5455 unsigned long nr_anon
, unsigned long nr_file
,
5456 unsigned long nr_huge
, struct page
*dummy_page
)
5458 unsigned long nr_pages
= nr_anon
+ nr_file
;
5459 unsigned long flags
;
5461 if (!mem_cgroup_is_root(memcg
)) {
5462 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5463 if (do_memsw_account())
5464 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5465 memcg_oom_recover(memcg
);
5468 local_irq_save(flags
);
5469 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5470 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5471 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5472 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5473 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5474 memcg_check_events(memcg
, dummy_page
);
5475 local_irq_restore(flags
);
5477 if (!mem_cgroup_is_root(memcg
))
5478 css_put_many(&memcg
->css
, nr_pages
);
5481 static void uncharge_list(struct list_head
*page_list
)
5483 struct mem_cgroup
*memcg
= NULL
;
5484 unsigned long nr_anon
= 0;
5485 unsigned long nr_file
= 0;
5486 unsigned long nr_huge
= 0;
5487 unsigned long pgpgout
= 0;
5488 struct list_head
*next
;
5491 next
= page_list
->next
;
5493 unsigned int nr_pages
= 1;
5495 page
= list_entry(next
, struct page
, lru
);
5496 next
= page
->lru
.next
;
5498 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5499 VM_BUG_ON_PAGE(page_count(page
), page
);
5501 if (!page
->mem_cgroup
)
5505 * Nobody should be changing or seriously looking at
5506 * page->mem_cgroup at this point, we have fully
5507 * exclusive access to the page.
5510 if (memcg
!= page
->mem_cgroup
) {
5512 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5514 pgpgout
= nr_anon
= nr_file
= nr_huge
= 0;
5516 memcg
= page
->mem_cgroup
;
5519 if (PageTransHuge(page
)) {
5520 nr_pages
<<= compound_order(page
);
5521 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5522 nr_huge
+= nr_pages
;
5526 nr_anon
+= nr_pages
;
5528 nr_file
+= nr_pages
;
5530 page
->mem_cgroup
= NULL
;
5533 } while (next
!= page_list
);
5536 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5541 * mem_cgroup_uncharge - uncharge a page
5542 * @page: page to uncharge
5544 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5545 * mem_cgroup_commit_charge().
5547 void mem_cgroup_uncharge(struct page
*page
)
5549 if (mem_cgroup_disabled())
5552 /* Don't touch page->lru of any random page, pre-check: */
5553 if (!page
->mem_cgroup
)
5556 INIT_LIST_HEAD(&page
->lru
);
5557 uncharge_list(&page
->lru
);
5561 * mem_cgroup_uncharge_list - uncharge a list of page
5562 * @page_list: list of pages to uncharge
5564 * Uncharge a list of pages previously charged with
5565 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5567 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5569 if (mem_cgroup_disabled())
5572 if (!list_empty(page_list
))
5573 uncharge_list(page_list
);
5577 * mem_cgroup_replace_page - migrate a charge to another page
5578 * @oldpage: currently charged page
5579 * @newpage: page to transfer the charge to
5581 * Migrate the charge from @oldpage to @newpage.
5583 * Both pages must be locked, @newpage->mapping must be set up.
5584 * Either or both pages might be on the LRU already.
5586 void mem_cgroup_replace_page(struct page
*oldpage
, struct page
*newpage
)
5588 struct mem_cgroup
*memcg
;
5591 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5592 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5593 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5594 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5597 if (mem_cgroup_disabled())
5600 /* Page cache replacement: new page already charged? */
5601 if (newpage
->mem_cgroup
)
5604 /* Swapcache readahead pages can get replaced before being charged */
5605 memcg
= oldpage
->mem_cgroup
;
5609 lock_page_lru(oldpage
, &isolated
);
5610 oldpage
->mem_cgroup
= NULL
;
5611 unlock_page_lru(oldpage
, isolated
);
5613 commit_charge(newpage
, memcg
, true);
5618 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
5619 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
5621 void sock_update_memcg(struct sock
*sk
)
5623 struct mem_cgroup
*memcg
;
5625 /* Socket cloning can throw us here with sk_cgrp already
5626 * filled. It won't however, necessarily happen from
5627 * process context. So the test for root memcg given
5628 * the current task's memcg won't help us in this case.
5630 * Respecting the original socket's memcg is a better
5631 * decision in this case.
5634 BUG_ON(mem_cgroup_is_root(sk
->sk_memcg
));
5635 css_get(&sk
->sk_memcg
->css
);
5640 memcg
= mem_cgroup_from_task(current
);
5641 if (memcg
== root_mem_cgroup
)
5643 #ifdef CONFIG_MEMCG_LEGACY_KMEM
5644 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcp_mem
.active
)
5647 if (css_tryget_online(&memcg
->css
))
5648 sk
->sk_memcg
= memcg
;
5652 EXPORT_SYMBOL(sock_update_memcg
);
5654 void sock_release_memcg(struct sock
*sk
)
5656 WARN_ON(!sk
->sk_memcg
);
5657 css_put(&sk
->sk_memcg
->css
);
5661 * mem_cgroup_charge_skmem - charge socket memory
5662 * @memcg: memcg to charge
5663 * @nr_pages: number of pages to charge
5665 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5666 * @memcg's configured limit, %false if the charge had to be forced.
5668 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5670 gfp_t gfp_mask
= GFP_KERNEL
;
5672 #ifdef CONFIG_MEMCG_LEGACY_KMEM
5673 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5674 struct page_counter
*counter
;
5676 if (page_counter_try_charge(&memcg
->tcp_mem
.memory_allocated
,
5677 nr_pages
, &counter
)) {
5678 memcg
->tcp_mem
.memory_pressure
= 0;
5681 page_counter_charge(&memcg
->tcp_mem
.memory_allocated
, nr_pages
);
5682 memcg
->tcp_mem
.memory_pressure
= 1;
5686 /* Don't block in the packet receive path */
5688 gfp_mask
= GFP_NOWAIT
;
5690 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
5693 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
5698 * mem_cgroup_uncharge_skmem - uncharge socket memory
5699 * @memcg - memcg to uncharge
5700 * @nr_pages - number of pages to uncharge
5702 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5704 #ifdef CONFIG_MEMCG_LEGACY_KMEM
5705 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5706 page_counter_uncharge(&memcg
->tcp_mem
.memory_allocated
,
5711 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5712 css_put_many(&memcg
->css
, nr_pages
);
5715 #endif /* CONFIG_INET */
5717 static int __init
cgroup_memory(char *s
)
5721 while ((token
= strsep(&s
, ",")) != NULL
) {
5724 if (!strcmp(token
, "nosocket"))
5725 cgroup_memory_nosocket
= true;
5726 if (!strcmp(token
, "nokmem"))
5727 cgroup_memory_nokmem
= true;
5731 __setup("cgroup.memory=", cgroup_memory
);
5734 * subsys_initcall() for memory controller.
5736 * Some parts like hotcpu_notifier() have to be initialized from this context
5737 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5738 * everything that doesn't depend on a specific mem_cgroup structure should
5739 * be initialized from here.
5741 static int __init
mem_cgroup_init(void)
5745 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5747 for_each_possible_cpu(cpu
)
5748 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5751 for_each_node(node
) {
5752 struct mem_cgroup_tree_per_node
*rtpn
;
5755 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5756 node_online(node
) ? node
: NUMA_NO_NODE
);
5758 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5759 struct mem_cgroup_tree_per_zone
*rtpz
;
5761 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
5762 rtpz
->rb_root
= RB_ROOT
;
5763 spin_lock_init(&rtpz
->lock
);
5765 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5770 subsys_initcall(mem_cgroup_init
);
5772 #ifdef CONFIG_MEMCG_SWAP
5774 * mem_cgroup_swapout - transfer a memsw charge to swap
5775 * @page: page whose memsw charge to transfer
5776 * @entry: swap entry to move the charge to
5778 * Transfer the memsw charge of @page to @entry.
5780 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5782 struct mem_cgroup
*memcg
;
5783 unsigned short oldid
;
5785 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5786 VM_BUG_ON_PAGE(page_count(page
), page
);
5788 if (!do_memsw_account())
5791 memcg
= page
->mem_cgroup
;
5793 /* Readahead page, never charged */
5797 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5798 VM_BUG_ON_PAGE(oldid
, page
);
5799 mem_cgroup_swap_statistics(memcg
, true);
5801 page
->mem_cgroup
= NULL
;
5803 if (!mem_cgroup_is_root(memcg
))
5804 page_counter_uncharge(&memcg
->memory
, 1);
5807 * Interrupts should be disabled here because the caller holds the
5808 * mapping->tree_lock lock which is taken with interrupts-off. It is
5809 * important here to have the interrupts disabled because it is the
5810 * only synchronisation we have for udpating the per-CPU variables.
5812 VM_BUG_ON(!irqs_disabled());
5813 mem_cgroup_charge_statistics(memcg
, page
, false, -1);
5814 memcg_check_events(memcg
, page
);
5818 * mem_cgroup_uncharge_swap - uncharge a swap entry
5819 * @entry: swap entry to uncharge
5821 * Drop the memsw charge associated with @entry.
5823 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5825 struct mem_cgroup
*memcg
;
5828 if (!do_memsw_account())
5831 id
= swap_cgroup_record(entry
, 0);
5833 memcg
= mem_cgroup_from_id(id
);
5835 if (!mem_cgroup_is_root(memcg
))
5836 page_counter_uncharge(&memcg
->memsw
, 1);
5837 mem_cgroup_swap_statistics(memcg
, false);
5838 css_put(&memcg
->css
);
5843 /* for remember boot option*/
5844 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5845 static int really_do_swap_account __initdata
= 1;
5847 static int really_do_swap_account __initdata
;
5850 static int __init
enable_swap_account(char *s
)
5852 if (!strcmp(s
, "1"))
5853 really_do_swap_account
= 1;
5854 else if (!strcmp(s
, "0"))
5855 really_do_swap_account
= 0;
5858 __setup("swapaccount=", enable_swap_account
);
5860 static struct cftype memsw_cgroup_files
[] = {
5862 .name
= "memsw.usage_in_bytes",
5863 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5864 .read_u64
= mem_cgroup_read_u64
,
5867 .name
= "memsw.max_usage_in_bytes",
5868 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5869 .write
= mem_cgroup_reset
,
5870 .read_u64
= mem_cgroup_read_u64
,
5873 .name
= "memsw.limit_in_bytes",
5874 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5875 .write
= mem_cgroup_write
,
5876 .read_u64
= mem_cgroup_read_u64
,
5879 .name
= "memsw.failcnt",
5880 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5881 .write
= mem_cgroup_reset
,
5882 .read_u64
= mem_cgroup_read_u64
,
5884 { }, /* terminate */
5887 static int __init
mem_cgroup_swap_init(void)
5889 if (!mem_cgroup_disabled() && really_do_swap_account
) {
5890 do_swap_account
= 1;
5891 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
5892 memsw_cgroup_files
));
5896 subsys_initcall(mem_cgroup_swap_init
);
5898 #endif /* CONFIG_MEMCG_SWAP */