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)
253 /* Some nice accessors for the vmpressure. */
254 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
257 memcg
= root_mem_cgroup
;
258 return &memcg
->vmpressure
;
261 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
263 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
266 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
268 return (memcg
== root_mem_cgroup
);
272 * We restrict the id in the range of [1, 65535], so it can fit into
275 #define MEM_CGROUP_ID_MAX USHRT_MAX
277 static inline unsigned short mem_cgroup_id(struct mem_cgroup
*memcg
)
279 return memcg
->css
.id
;
283 * A helper function to get mem_cgroup from ID. must be called under
284 * rcu_read_lock(). The caller is responsible for calling
285 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
286 * refcnt from swap can be called against removed memcg.)
288 static inline struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
290 struct cgroup_subsys_state
*css
;
292 css
= css_from_id(id
, &memory_cgrp_subsys
);
293 return mem_cgroup_from_css(css
);
298 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
299 * The main reason for not using cgroup id for this:
300 * this works better in sparse environments, where we have a lot of memcgs,
301 * but only a few kmem-limited. Or also, if we have, for instance, 200
302 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
303 * 200 entry array for that.
305 * The current size of the caches array is stored in memcg_nr_cache_ids. It
306 * will double each time we have to increase it.
308 static DEFINE_IDA(memcg_cache_ida
);
309 int memcg_nr_cache_ids
;
311 /* Protects memcg_nr_cache_ids */
312 static DECLARE_RWSEM(memcg_cache_ids_sem
);
314 void memcg_get_cache_ids(void)
316 down_read(&memcg_cache_ids_sem
);
319 void memcg_put_cache_ids(void)
321 up_read(&memcg_cache_ids_sem
);
325 * MIN_SIZE is different than 1, because we would like to avoid going through
326 * the alloc/free process all the time. In a small machine, 4 kmem-limited
327 * cgroups is a reasonable guess. In the future, it could be a parameter or
328 * tunable, but that is strictly not necessary.
330 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
331 * this constant directly from cgroup, but it is understandable that this is
332 * better kept as an internal representation in cgroup.c. In any case, the
333 * cgrp_id space is not getting any smaller, and we don't have to necessarily
334 * increase ours as well if it increases.
336 #define MEMCG_CACHES_MIN_SIZE 4
337 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
340 * A lot of the calls to the cache allocation functions are expected to be
341 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
342 * conditional to this static branch, we'll have to allow modules that does
343 * kmem_cache_alloc and the such to see this symbol as well
345 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
346 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
348 #endif /* !CONFIG_SLOB */
350 static struct mem_cgroup_per_zone
*
351 mem_cgroup_zone_zoneinfo(struct mem_cgroup
*memcg
, struct zone
*zone
)
353 int nid
= zone_to_nid(zone
);
354 int zid
= zone_idx(zone
);
356 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
360 * mem_cgroup_css_from_page - css of the memcg associated with a page
361 * @page: page of interest
363 * If memcg is bound to the default hierarchy, css of the memcg associated
364 * with @page is returned. The returned css remains associated with @page
365 * until it is released.
367 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
370 * XXX: The above description of behavior on the default hierarchy isn't
371 * strictly true yet as replace_page_cache_page() can modify the
372 * association before @page is released even on the default hierarchy;
373 * however, the current and planned usages don't mix the the two functions
374 * and replace_page_cache_page() will soon be updated to make the invariant
377 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
379 struct mem_cgroup
*memcg
;
381 memcg
= page
->mem_cgroup
;
383 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
384 memcg
= root_mem_cgroup
;
390 * page_cgroup_ino - return inode number of the memcg a page is charged to
393 * Look up the closest online ancestor of the memory cgroup @page is charged to
394 * and return its inode number or 0 if @page is not charged to any cgroup. It
395 * is safe to call this function without holding a reference to @page.
397 * Note, this function is inherently racy, because there is nothing to prevent
398 * the cgroup inode from getting torn down and potentially reallocated a moment
399 * after page_cgroup_ino() returns, so it only should be used by callers that
400 * do not care (such as procfs interfaces).
402 ino_t
page_cgroup_ino(struct page
*page
)
404 struct mem_cgroup
*memcg
;
405 unsigned long ino
= 0;
408 memcg
= READ_ONCE(page
->mem_cgroup
);
409 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
410 memcg
= parent_mem_cgroup(memcg
);
412 ino
= cgroup_ino(memcg
->css
.cgroup
);
417 static struct mem_cgroup_per_zone
*
418 mem_cgroup_page_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
420 int nid
= page_to_nid(page
);
421 int zid
= page_zonenum(page
);
423 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
426 static struct mem_cgroup_tree_per_zone
*
427 soft_limit_tree_node_zone(int nid
, int zid
)
429 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
432 static struct mem_cgroup_tree_per_zone
*
433 soft_limit_tree_from_page(struct page
*page
)
435 int nid
= page_to_nid(page
);
436 int zid
= page_zonenum(page
);
438 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
441 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone
*mz
,
442 struct mem_cgroup_tree_per_zone
*mctz
,
443 unsigned long new_usage_in_excess
)
445 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
446 struct rb_node
*parent
= NULL
;
447 struct mem_cgroup_per_zone
*mz_node
;
452 mz
->usage_in_excess
= new_usage_in_excess
;
453 if (!mz
->usage_in_excess
)
457 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
459 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
462 * We can't avoid mem cgroups that are over their soft
463 * limit by the same amount
465 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
468 rb_link_node(&mz
->tree_node
, parent
, p
);
469 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
473 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
474 struct mem_cgroup_tree_per_zone
*mctz
)
478 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
482 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
483 struct mem_cgroup_tree_per_zone
*mctz
)
487 spin_lock_irqsave(&mctz
->lock
, flags
);
488 __mem_cgroup_remove_exceeded(mz
, mctz
);
489 spin_unlock_irqrestore(&mctz
->lock
, flags
);
492 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
494 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
495 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
496 unsigned long excess
= 0;
498 if (nr_pages
> soft_limit
)
499 excess
= nr_pages
- soft_limit
;
504 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
506 unsigned long excess
;
507 struct mem_cgroup_per_zone
*mz
;
508 struct mem_cgroup_tree_per_zone
*mctz
;
510 mctz
= soft_limit_tree_from_page(page
);
512 * Necessary to update all ancestors when hierarchy is used.
513 * because their event counter is not touched.
515 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
516 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
517 excess
= soft_limit_excess(memcg
);
519 * We have to update the tree if mz is on RB-tree or
520 * mem is over its softlimit.
522 if (excess
|| mz
->on_tree
) {
525 spin_lock_irqsave(&mctz
->lock
, flags
);
526 /* if on-tree, remove it */
528 __mem_cgroup_remove_exceeded(mz
, mctz
);
530 * Insert again. mz->usage_in_excess will be updated.
531 * If excess is 0, no tree ops.
533 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
534 spin_unlock_irqrestore(&mctz
->lock
, flags
);
539 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
541 struct mem_cgroup_tree_per_zone
*mctz
;
542 struct mem_cgroup_per_zone
*mz
;
546 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
547 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
548 mctz
= soft_limit_tree_node_zone(nid
, zid
);
549 mem_cgroup_remove_exceeded(mz
, mctz
);
554 static struct mem_cgroup_per_zone
*
555 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
557 struct rb_node
*rightmost
= NULL
;
558 struct mem_cgroup_per_zone
*mz
;
562 rightmost
= rb_last(&mctz
->rb_root
);
564 goto done
; /* Nothing to reclaim from */
566 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
568 * Remove the node now but someone else can add it back,
569 * we will to add it back at the end of reclaim to its correct
570 * position in the tree.
572 __mem_cgroup_remove_exceeded(mz
, mctz
);
573 if (!soft_limit_excess(mz
->memcg
) ||
574 !css_tryget_online(&mz
->memcg
->css
))
580 static struct mem_cgroup_per_zone
*
581 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
583 struct mem_cgroup_per_zone
*mz
;
585 spin_lock_irq(&mctz
->lock
);
586 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
587 spin_unlock_irq(&mctz
->lock
);
592 * Return page count for single (non recursive) @memcg.
594 * Implementation Note: reading percpu statistics for memcg.
596 * Both of vmstat[] and percpu_counter has threshold and do periodic
597 * synchronization to implement "quick" read. There are trade-off between
598 * reading cost and precision of value. Then, we may have a chance to implement
599 * a periodic synchronization of counter in memcg's counter.
601 * But this _read() function is used for user interface now. The user accounts
602 * memory usage by memory cgroup and he _always_ requires exact value because
603 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
604 * have to visit all online cpus and make sum. So, for now, unnecessary
605 * synchronization is not implemented. (just implemented for cpu hotplug)
607 * If there are kernel internal actions which can make use of some not-exact
608 * value, and reading all cpu value can be performance bottleneck in some
609 * common workload, threshold and synchronization as vmstat[] should be
613 mem_cgroup_read_stat(struct mem_cgroup
*memcg
, enum mem_cgroup_stat_index idx
)
618 /* Per-cpu values can be negative, use a signed accumulator */
619 for_each_possible_cpu(cpu
)
620 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
622 * Summing races with updates, so val may be negative. Avoid exposing
623 * transient negative values.
630 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
631 enum mem_cgroup_events_index idx
)
633 unsigned long val
= 0;
636 for_each_possible_cpu(cpu
)
637 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
641 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
643 bool compound
, int nr_pages
)
646 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
647 * counted as CACHE even if it's on ANON LRU.
650 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
653 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
657 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
658 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
662 /* pagein of a big page is an event. So, ignore page size */
664 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
666 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
667 nr_pages
= -nr_pages
; /* for event */
670 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
673 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
675 unsigned int lru_mask
)
677 unsigned long nr
= 0;
680 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
682 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
683 struct mem_cgroup_per_zone
*mz
;
687 if (!(BIT(lru
) & lru_mask
))
689 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
690 nr
+= mz
->lru_size
[lru
];
696 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
697 unsigned int lru_mask
)
699 unsigned long nr
= 0;
702 for_each_node_state(nid
, N_MEMORY
)
703 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
707 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
708 enum mem_cgroup_events_target target
)
710 unsigned long val
, next
;
712 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
713 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
714 /* from time_after() in jiffies.h */
715 if ((long)next
- (long)val
< 0) {
717 case MEM_CGROUP_TARGET_THRESH
:
718 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
720 case MEM_CGROUP_TARGET_SOFTLIMIT
:
721 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
723 case MEM_CGROUP_TARGET_NUMAINFO
:
724 next
= val
+ NUMAINFO_EVENTS_TARGET
;
729 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
736 * Check events in order.
739 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
741 /* threshold event is triggered in finer grain than soft limit */
742 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
743 MEM_CGROUP_TARGET_THRESH
))) {
745 bool do_numainfo __maybe_unused
;
747 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
748 MEM_CGROUP_TARGET_SOFTLIMIT
);
750 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
751 MEM_CGROUP_TARGET_NUMAINFO
);
753 mem_cgroup_threshold(memcg
);
754 if (unlikely(do_softlimit
))
755 mem_cgroup_update_tree(memcg
, page
);
757 if (unlikely(do_numainfo
))
758 atomic_inc(&memcg
->numainfo_events
);
763 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
766 * mm_update_next_owner() may clear mm->owner to NULL
767 * if it races with swapoff, page migration, etc.
768 * So this can be called with p == NULL.
773 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
775 EXPORT_SYMBOL(mem_cgroup_from_task
);
777 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
779 struct mem_cgroup
*memcg
= NULL
;
784 * Page cache insertions can happen withou an
785 * actual mm context, e.g. during disk probing
786 * on boot, loopback IO, acct() writes etc.
789 memcg
= root_mem_cgroup
;
791 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
792 if (unlikely(!memcg
))
793 memcg
= root_mem_cgroup
;
795 } while (!css_tryget_online(&memcg
->css
));
801 * mem_cgroup_iter - iterate over memory cgroup hierarchy
802 * @root: hierarchy root
803 * @prev: previously returned memcg, NULL on first invocation
804 * @reclaim: cookie for shared reclaim walks, NULL for full walks
806 * Returns references to children of the hierarchy below @root, or
807 * @root itself, or %NULL after a full round-trip.
809 * Caller must pass the return value in @prev on subsequent
810 * invocations for reference counting, or use mem_cgroup_iter_break()
811 * to cancel a hierarchy walk before the round-trip is complete.
813 * Reclaimers can specify a zone and a priority level in @reclaim to
814 * divide up the memcgs in the hierarchy among all concurrent
815 * reclaimers operating on the same zone and priority.
817 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
818 struct mem_cgroup
*prev
,
819 struct mem_cgroup_reclaim_cookie
*reclaim
)
821 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
822 struct cgroup_subsys_state
*css
= NULL
;
823 struct mem_cgroup
*memcg
= NULL
;
824 struct mem_cgroup
*pos
= NULL
;
826 if (mem_cgroup_disabled())
830 root
= root_mem_cgroup
;
832 if (prev
&& !reclaim
)
835 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
844 struct mem_cgroup_per_zone
*mz
;
846 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
847 iter
= &mz
->iter
[reclaim
->priority
];
849 if (prev
&& reclaim
->generation
!= iter
->generation
)
853 pos
= READ_ONCE(iter
->position
);
854 if (!pos
|| css_tryget(&pos
->css
))
857 * css reference reached zero, so iter->position will
858 * be cleared by ->css_released. However, we should not
859 * rely on this happening soon, because ->css_released
860 * is called from a work queue, and by busy-waiting we
861 * might block it. So we clear iter->position right
864 (void)cmpxchg(&iter
->position
, pos
, NULL
);
872 css
= css_next_descendant_pre(css
, &root
->css
);
875 * Reclaimers share the hierarchy walk, and a
876 * new one might jump in right at the end of
877 * the hierarchy - make sure they see at least
878 * one group and restart from the beginning.
886 * Verify the css and acquire a reference. The root
887 * is provided by the caller, so we know it's alive
888 * and kicking, and don't take an extra reference.
890 memcg
= mem_cgroup_from_css(css
);
892 if (css
== &root
->css
)
903 * The position could have already been updated by a competing
904 * thread, so check that the value hasn't changed since we read
905 * it to avoid reclaiming from the same cgroup twice.
907 (void)cmpxchg(&iter
->position
, pos
, memcg
);
915 reclaim
->generation
= iter
->generation
;
921 if (prev
&& prev
!= root
)
928 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
929 * @root: hierarchy root
930 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
932 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
933 struct mem_cgroup
*prev
)
936 root
= root_mem_cgroup
;
937 if (prev
&& prev
!= root
)
941 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
943 struct mem_cgroup
*memcg
= dead_memcg
;
944 struct mem_cgroup_reclaim_iter
*iter
;
945 struct mem_cgroup_per_zone
*mz
;
949 while ((memcg
= parent_mem_cgroup(memcg
))) {
951 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
952 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
953 for (i
= 0; i
<= DEF_PRIORITY
; i
++) {
955 cmpxchg(&iter
->position
,
964 * Iteration constructs for visiting all cgroups (under a tree). If
965 * loops are exited prematurely (break), mem_cgroup_iter_break() must
966 * be used for reference counting.
968 #define for_each_mem_cgroup_tree(iter, root) \
969 for (iter = mem_cgroup_iter(root, NULL, NULL); \
971 iter = mem_cgroup_iter(root, iter, NULL))
973 #define for_each_mem_cgroup(iter) \
974 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
976 iter = mem_cgroup_iter(NULL, iter, NULL))
979 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
980 * @zone: zone of the wanted lruvec
981 * @memcg: memcg of the wanted lruvec
983 * Returns the lru list vector holding pages for the given @zone and
984 * @mem. This can be the global zone lruvec, if the memory controller
987 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
988 struct mem_cgroup
*memcg
)
990 struct mem_cgroup_per_zone
*mz
;
991 struct lruvec
*lruvec
;
993 if (mem_cgroup_disabled()) {
994 lruvec
= &zone
->lruvec
;
998 mz
= mem_cgroup_zone_zoneinfo(memcg
, zone
);
999 lruvec
= &mz
->lruvec
;
1002 * Since a node can be onlined after the mem_cgroup was created,
1003 * we have to be prepared to initialize lruvec->zone here;
1004 * and if offlined then reonlined, we need to reinitialize it.
1006 if (unlikely(lruvec
->zone
!= zone
))
1007 lruvec
->zone
= zone
;
1012 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1014 * @zone: zone of the page
1016 * This function is only safe when following the LRU page isolation
1017 * and putback protocol: the LRU lock must be held, and the page must
1018 * either be PageLRU() or the caller must have isolated/allocated it.
1020 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1022 struct mem_cgroup_per_zone
*mz
;
1023 struct mem_cgroup
*memcg
;
1024 struct lruvec
*lruvec
;
1026 if (mem_cgroup_disabled()) {
1027 lruvec
= &zone
->lruvec
;
1031 memcg
= page
->mem_cgroup
;
1033 * Swapcache readahead pages are added to the LRU - and
1034 * possibly migrated - before they are charged.
1037 memcg
= root_mem_cgroup
;
1039 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1040 lruvec
= &mz
->lruvec
;
1043 * Since a node can be onlined after the mem_cgroup was created,
1044 * we have to be prepared to initialize lruvec->zone here;
1045 * and if offlined then reonlined, we need to reinitialize it.
1047 if (unlikely(lruvec
->zone
!= zone
))
1048 lruvec
->zone
= zone
;
1053 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1054 * @lruvec: mem_cgroup per zone lru vector
1055 * @lru: index of lru list the page is sitting on
1056 * @nr_pages: positive when adding or negative when removing
1058 * This function must be called when a page is added to or removed from an
1061 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1064 struct mem_cgroup_per_zone
*mz
;
1065 unsigned long *lru_size
;
1067 if (mem_cgroup_disabled())
1070 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1071 lru_size
= mz
->lru_size
+ lru
;
1072 *lru_size
+= nr_pages
;
1073 VM_BUG_ON((long)(*lru_size
) < 0);
1076 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1078 struct mem_cgroup
*task_memcg
;
1079 struct task_struct
*p
;
1082 p
= find_lock_task_mm(task
);
1084 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1088 * All threads may have already detached their mm's, but the oom
1089 * killer still needs to detect if they have already been oom
1090 * killed to prevent needlessly killing additional tasks.
1093 task_memcg
= mem_cgroup_from_task(task
);
1094 css_get(&task_memcg
->css
);
1097 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1098 css_put(&task_memcg
->css
);
1103 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1104 * @memcg: the memory cgroup
1106 * Returns the maximum amount of memory @mem can be charged with, in
1109 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1111 unsigned long margin
= 0;
1112 unsigned long count
;
1113 unsigned long limit
;
1115 count
= page_counter_read(&memcg
->memory
);
1116 limit
= READ_ONCE(memcg
->memory
.limit
);
1118 margin
= limit
- count
;
1120 if (do_memsw_account()) {
1121 count
= page_counter_read(&memcg
->memsw
);
1122 limit
= READ_ONCE(memcg
->memsw
.limit
);
1124 margin
= min(margin
, limit
- count
);
1131 * A routine for checking "mem" is under move_account() or not.
1133 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1134 * moving cgroups. This is for waiting at high-memory pressure
1137 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1139 struct mem_cgroup
*from
;
1140 struct mem_cgroup
*to
;
1143 * Unlike task_move routines, we access mc.to, mc.from not under
1144 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1146 spin_lock(&mc
.lock
);
1152 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1153 mem_cgroup_is_descendant(to
, memcg
);
1155 spin_unlock(&mc
.lock
);
1159 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1161 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1162 if (mem_cgroup_under_move(memcg
)) {
1164 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1165 /* moving charge context might have finished. */
1168 finish_wait(&mc
.waitq
, &wait
);
1175 #define K(x) ((x) << (PAGE_SHIFT-10))
1177 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1178 * @memcg: The memory cgroup that went over limit
1179 * @p: Task that is going to be killed
1181 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1184 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1186 /* oom_info_lock ensures that parallel ooms do not interleave */
1187 static DEFINE_MUTEX(oom_info_lock
);
1188 struct mem_cgroup
*iter
;
1191 mutex_lock(&oom_info_lock
);
1195 pr_info("Task in ");
1196 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1197 pr_cont(" killed as a result of limit of ");
1199 pr_info("Memory limit reached of cgroup ");
1202 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1207 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1208 K((u64
)page_counter_read(&memcg
->memory
)),
1209 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1210 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1211 K((u64
)page_counter_read(&memcg
->memsw
)),
1212 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1213 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1214 K((u64
)page_counter_read(&memcg
->kmem
)),
1215 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1217 for_each_mem_cgroup_tree(iter
, memcg
) {
1218 pr_info("Memory cgroup stats for ");
1219 pr_cont_cgroup_path(iter
->css
.cgroup
);
1222 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1223 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
1225 pr_cont(" %s:%luKB", mem_cgroup_stat_names
[i
],
1226 K(mem_cgroup_read_stat(iter
, i
)));
1229 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1230 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1231 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1235 mutex_unlock(&oom_info_lock
);
1239 * This function returns the number of memcg under hierarchy tree. Returns
1240 * 1(self count) if no children.
1242 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1245 struct mem_cgroup
*iter
;
1247 for_each_mem_cgroup_tree(iter
, memcg
)
1253 * Return the memory (and swap, if configured) limit for a memcg.
1255 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1257 unsigned long limit
;
1259 limit
= memcg
->memory
.limit
;
1260 if (mem_cgroup_swappiness(memcg
)) {
1261 unsigned long memsw_limit
;
1263 memsw_limit
= memcg
->memsw
.limit
;
1264 limit
= min(limit
+ total_swap_pages
, memsw_limit
);
1269 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1272 struct oom_control oc
= {
1275 .gfp_mask
= gfp_mask
,
1278 struct mem_cgroup
*iter
;
1279 unsigned long chosen_points
= 0;
1280 unsigned long totalpages
;
1281 unsigned int points
= 0;
1282 struct task_struct
*chosen
= NULL
;
1284 mutex_lock(&oom_lock
);
1287 * If current has a pending SIGKILL or is exiting, then automatically
1288 * select it. The goal is to allow it to allocate so that it may
1289 * quickly exit and free its memory.
1291 if (fatal_signal_pending(current
) || task_will_free_mem(current
)) {
1292 mark_oom_victim(current
);
1296 check_panic_on_oom(&oc
, CONSTRAINT_MEMCG
, memcg
);
1297 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1298 for_each_mem_cgroup_tree(iter
, memcg
) {
1299 struct css_task_iter it
;
1300 struct task_struct
*task
;
1302 css_task_iter_start(&iter
->css
, &it
);
1303 while ((task
= css_task_iter_next(&it
))) {
1304 switch (oom_scan_process_thread(&oc
, task
, totalpages
)) {
1305 case OOM_SCAN_SELECT
:
1307 put_task_struct(chosen
);
1309 chosen_points
= ULONG_MAX
;
1310 get_task_struct(chosen
);
1312 case OOM_SCAN_CONTINUE
:
1314 case OOM_SCAN_ABORT
:
1315 css_task_iter_end(&it
);
1316 mem_cgroup_iter_break(memcg
, iter
);
1318 put_task_struct(chosen
);
1323 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1324 if (!points
|| points
< chosen_points
)
1326 /* Prefer thread group leaders for display purposes */
1327 if (points
== chosen_points
&&
1328 thread_group_leader(chosen
))
1332 put_task_struct(chosen
);
1334 chosen_points
= points
;
1335 get_task_struct(chosen
);
1337 css_task_iter_end(&it
);
1341 points
= chosen_points
* 1000 / totalpages
;
1342 oom_kill_process(&oc
, chosen
, points
, totalpages
, memcg
,
1343 "Memory cgroup out of memory");
1346 mutex_unlock(&oom_lock
);
1349 #if MAX_NUMNODES > 1
1352 * test_mem_cgroup_node_reclaimable
1353 * @memcg: the target memcg
1354 * @nid: the node ID to be checked.
1355 * @noswap : specify true here if the user wants flle only information.
1357 * This function returns whether the specified memcg contains any
1358 * reclaimable pages on a node. Returns true if there are any reclaimable
1359 * pages in the node.
1361 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1362 int nid
, bool noswap
)
1364 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1366 if (noswap
|| !total_swap_pages
)
1368 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1375 * Always updating the nodemask is not very good - even if we have an empty
1376 * list or the wrong list here, we can start from some node and traverse all
1377 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1380 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1384 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1385 * pagein/pageout changes since the last update.
1387 if (!atomic_read(&memcg
->numainfo_events
))
1389 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1392 /* make a nodemask where this memcg uses memory from */
1393 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1395 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1397 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1398 node_clear(nid
, memcg
->scan_nodes
);
1401 atomic_set(&memcg
->numainfo_events
, 0);
1402 atomic_set(&memcg
->numainfo_updating
, 0);
1406 * Selecting a node where we start reclaim from. Because what we need is just
1407 * reducing usage counter, start from anywhere is O,K. Considering
1408 * memory reclaim from current node, there are pros. and cons.
1410 * Freeing memory from current node means freeing memory from a node which
1411 * we'll use or we've used. So, it may make LRU bad. And if several threads
1412 * hit limits, it will see a contention on a node. But freeing from remote
1413 * node means more costs for memory reclaim because of memory latency.
1415 * Now, we use round-robin. Better algorithm is welcomed.
1417 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1421 mem_cgroup_may_update_nodemask(memcg
);
1422 node
= memcg
->last_scanned_node
;
1424 node
= next_node(node
, memcg
->scan_nodes
);
1425 if (node
== MAX_NUMNODES
)
1426 node
= first_node(memcg
->scan_nodes
);
1428 * We call this when we hit limit, not when pages are added to LRU.
1429 * No LRU may hold pages because all pages are UNEVICTABLE or
1430 * memcg is too small and all pages are not on LRU. In that case,
1431 * we use curret node.
1433 if (unlikely(node
== MAX_NUMNODES
))
1434 node
= numa_node_id();
1436 memcg
->last_scanned_node
= node
;
1440 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1446 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1449 unsigned long *total_scanned
)
1451 struct mem_cgroup
*victim
= NULL
;
1454 unsigned long excess
;
1455 unsigned long nr_scanned
;
1456 struct mem_cgroup_reclaim_cookie reclaim
= {
1461 excess
= soft_limit_excess(root_memcg
);
1464 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1469 * If we have not been able to reclaim
1470 * anything, it might because there are
1471 * no reclaimable pages under this hierarchy
1476 * We want to do more targeted reclaim.
1477 * excess >> 2 is not to excessive so as to
1478 * reclaim too much, nor too less that we keep
1479 * coming back to reclaim from this cgroup
1481 if (total
>= (excess
>> 2) ||
1482 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1487 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1489 *total_scanned
+= nr_scanned
;
1490 if (!soft_limit_excess(root_memcg
))
1493 mem_cgroup_iter_break(root_memcg
, victim
);
1497 #ifdef CONFIG_LOCKDEP
1498 static struct lockdep_map memcg_oom_lock_dep_map
= {
1499 .name
= "memcg_oom_lock",
1503 static DEFINE_SPINLOCK(memcg_oom_lock
);
1506 * Check OOM-Killer is already running under our hierarchy.
1507 * If someone is running, return false.
1509 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1511 struct mem_cgroup
*iter
, *failed
= NULL
;
1513 spin_lock(&memcg_oom_lock
);
1515 for_each_mem_cgroup_tree(iter
, memcg
) {
1516 if (iter
->oom_lock
) {
1518 * this subtree of our hierarchy is already locked
1519 * so we cannot give a lock.
1522 mem_cgroup_iter_break(memcg
, iter
);
1525 iter
->oom_lock
= true;
1530 * OK, we failed to lock the whole subtree so we have
1531 * to clean up what we set up to the failing subtree
1533 for_each_mem_cgroup_tree(iter
, memcg
) {
1534 if (iter
== failed
) {
1535 mem_cgroup_iter_break(memcg
, iter
);
1538 iter
->oom_lock
= false;
1541 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1543 spin_unlock(&memcg_oom_lock
);
1548 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1550 struct mem_cgroup
*iter
;
1552 spin_lock(&memcg_oom_lock
);
1553 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1554 for_each_mem_cgroup_tree(iter
, memcg
)
1555 iter
->oom_lock
= false;
1556 spin_unlock(&memcg_oom_lock
);
1559 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1561 struct mem_cgroup
*iter
;
1563 spin_lock(&memcg_oom_lock
);
1564 for_each_mem_cgroup_tree(iter
, memcg
)
1566 spin_unlock(&memcg_oom_lock
);
1569 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1571 struct mem_cgroup
*iter
;
1574 * When a new child is created while the hierarchy is under oom,
1575 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1577 spin_lock(&memcg_oom_lock
);
1578 for_each_mem_cgroup_tree(iter
, memcg
)
1579 if (iter
->under_oom
> 0)
1581 spin_unlock(&memcg_oom_lock
);
1584 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1586 struct oom_wait_info
{
1587 struct mem_cgroup
*memcg
;
1591 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1592 unsigned mode
, int sync
, void *arg
)
1594 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1595 struct mem_cgroup
*oom_wait_memcg
;
1596 struct oom_wait_info
*oom_wait_info
;
1598 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1599 oom_wait_memcg
= oom_wait_info
->memcg
;
1601 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1602 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1604 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1607 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1610 * For the following lockless ->under_oom test, the only required
1611 * guarantee is that it must see the state asserted by an OOM when
1612 * this function is called as a result of userland actions
1613 * triggered by the notification of the OOM. This is trivially
1614 * achieved by invoking mem_cgroup_mark_under_oom() before
1615 * triggering notification.
1617 if (memcg
&& memcg
->under_oom
)
1618 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1621 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1623 if (!current
->memcg_may_oom
)
1626 * We are in the middle of the charge context here, so we
1627 * don't want to block when potentially sitting on a callstack
1628 * that holds all kinds of filesystem and mm locks.
1630 * Also, the caller may handle a failed allocation gracefully
1631 * (like optional page cache readahead) and so an OOM killer
1632 * invocation might not even be necessary.
1634 * That's why we don't do anything here except remember the
1635 * OOM context and then deal with it at the end of the page
1636 * fault when the stack is unwound, the locks are released,
1637 * and when we know whether the fault was overall successful.
1639 css_get(&memcg
->css
);
1640 current
->memcg_in_oom
= memcg
;
1641 current
->memcg_oom_gfp_mask
= mask
;
1642 current
->memcg_oom_order
= order
;
1646 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1647 * @handle: actually kill/wait or just clean up the OOM state
1649 * This has to be called at the end of a page fault if the memcg OOM
1650 * handler was enabled.
1652 * Memcg supports userspace OOM handling where failed allocations must
1653 * sleep on a waitqueue until the userspace task resolves the
1654 * situation. Sleeping directly in the charge context with all kinds
1655 * of locks held is not a good idea, instead we remember an OOM state
1656 * in the task and mem_cgroup_oom_synchronize() has to be called at
1657 * the end of the page fault to complete the OOM handling.
1659 * Returns %true if an ongoing memcg OOM situation was detected and
1660 * completed, %false otherwise.
1662 bool mem_cgroup_oom_synchronize(bool handle
)
1664 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1665 struct oom_wait_info owait
;
1668 /* OOM is global, do not handle */
1672 if (!handle
|| oom_killer_disabled
)
1675 owait
.memcg
= memcg
;
1676 owait
.wait
.flags
= 0;
1677 owait
.wait
.func
= memcg_oom_wake_function
;
1678 owait
.wait
.private = current
;
1679 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1681 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1682 mem_cgroup_mark_under_oom(memcg
);
1684 locked
= mem_cgroup_oom_trylock(memcg
);
1687 mem_cgroup_oom_notify(memcg
);
1689 if (locked
&& !memcg
->oom_kill_disable
) {
1690 mem_cgroup_unmark_under_oom(memcg
);
1691 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1692 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1693 current
->memcg_oom_order
);
1696 mem_cgroup_unmark_under_oom(memcg
);
1697 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1701 mem_cgroup_oom_unlock(memcg
);
1703 * There is no guarantee that an OOM-lock contender
1704 * sees the wakeups triggered by the OOM kill
1705 * uncharges. Wake any sleepers explicitely.
1707 memcg_oom_recover(memcg
);
1710 current
->memcg_in_oom
= NULL
;
1711 css_put(&memcg
->css
);
1716 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1717 * @page: page that is going to change accounted state
1719 * This function must mark the beginning of an accounted page state
1720 * change to prevent double accounting when the page is concurrently
1721 * being moved to another memcg:
1723 * memcg = mem_cgroup_begin_page_stat(page);
1724 * if (TestClearPageState(page))
1725 * mem_cgroup_update_page_stat(memcg, state, -1);
1726 * mem_cgroup_end_page_stat(memcg);
1728 struct mem_cgroup
*mem_cgroup_begin_page_stat(struct page
*page
)
1730 struct mem_cgroup
*memcg
;
1731 unsigned long flags
;
1734 * The RCU lock is held throughout the transaction. The fast
1735 * path can get away without acquiring the memcg->move_lock
1736 * because page moving starts with an RCU grace period.
1738 * The RCU lock also protects the memcg from being freed when
1739 * the page state that is going to change is the only thing
1740 * preventing the page from being uncharged.
1741 * E.g. end-writeback clearing PageWriteback(), which allows
1742 * migration to go ahead and uncharge the page before the
1743 * account transaction might be complete.
1747 if (mem_cgroup_disabled())
1750 memcg
= page
->mem_cgroup
;
1751 if (unlikely(!memcg
))
1754 if (atomic_read(&memcg
->moving_account
) <= 0)
1757 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1758 if (memcg
!= page
->mem_cgroup
) {
1759 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1764 * When charge migration first begins, we can have locked and
1765 * unlocked page stat updates happening concurrently. Track
1766 * the task who has the lock for mem_cgroup_end_page_stat().
1768 memcg
->move_lock_task
= current
;
1769 memcg
->move_lock_flags
= flags
;
1773 EXPORT_SYMBOL(mem_cgroup_begin_page_stat
);
1776 * mem_cgroup_end_page_stat - finish a page state statistics transaction
1777 * @memcg: the memcg that was accounted against
1779 void mem_cgroup_end_page_stat(struct mem_cgroup
*memcg
)
1781 if (memcg
&& memcg
->move_lock_task
== current
) {
1782 unsigned long flags
= memcg
->move_lock_flags
;
1784 memcg
->move_lock_task
= NULL
;
1785 memcg
->move_lock_flags
= 0;
1787 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1792 EXPORT_SYMBOL(mem_cgroup_end_page_stat
);
1795 * size of first charge trial. "32" comes from vmscan.c's magic value.
1796 * TODO: maybe necessary to use big numbers in big irons.
1798 #define CHARGE_BATCH 32U
1799 struct memcg_stock_pcp
{
1800 struct mem_cgroup
*cached
; /* this never be root cgroup */
1801 unsigned int nr_pages
;
1802 struct work_struct work
;
1803 unsigned long flags
;
1804 #define FLUSHING_CACHED_CHARGE 0
1806 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1807 static DEFINE_MUTEX(percpu_charge_mutex
);
1810 * consume_stock: Try to consume stocked charge on this cpu.
1811 * @memcg: memcg to consume from.
1812 * @nr_pages: how many pages to charge.
1814 * The charges will only happen if @memcg matches the current cpu's memcg
1815 * stock, and at least @nr_pages are available in that stock. Failure to
1816 * service an allocation will refill the stock.
1818 * returns true if successful, false otherwise.
1820 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1822 struct memcg_stock_pcp
*stock
;
1825 if (nr_pages
> CHARGE_BATCH
)
1828 stock
= &get_cpu_var(memcg_stock
);
1829 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
1830 stock
->nr_pages
-= nr_pages
;
1833 put_cpu_var(memcg_stock
);
1838 * Returns stocks cached in percpu and reset cached information.
1840 static void drain_stock(struct memcg_stock_pcp
*stock
)
1842 struct mem_cgroup
*old
= stock
->cached
;
1844 if (stock
->nr_pages
) {
1845 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
1846 if (do_memsw_account())
1847 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
1848 css_put_many(&old
->css
, stock
->nr_pages
);
1849 stock
->nr_pages
= 0;
1851 stock
->cached
= NULL
;
1855 * This must be called under preempt disabled or must be called by
1856 * a thread which is pinned to local cpu.
1858 static void drain_local_stock(struct work_struct
*dummy
)
1860 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
1862 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
1866 * Cache charges(val) to local per_cpu area.
1867 * This will be consumed by consume_stock() function, later.
1869 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1871 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
1873 if (stock
->cached
!= memcg
) { /* reset if necessary */
1875 stock
->cached
= memcg
;
1877 stock
->nr_pages
+= nr_pages
;
1878 put_cpu_var(memcg_stock
);
1882 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1883 * of the hierarchy under it.
1885 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
1889 /* If someone's already draining, avoid adding running more workers. */
1890 if (!mutex_trylock(&percpu_charge_mutex
))
1892 /* Notify other cpus that system-wide "drain" is running */
1895 for_each_online_cpu(cpu
) {
1896 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1897 struct mem_cgroup
*memcg
;
1899 memcg
= stock
->cached
;
1900 if (!memcg
|| !stock
->nr_pages
)
1902 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
1904 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
1906 drain_local_stock(&stock
->work
);
1908 schedule_work_on(cpu
, &stock
->work
);
1913 mutex_unlock(&percpu_charge_mutex
);
1916 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
1917 unsigned long action
,
1920 int cpu
= (unsigned long)hcpu
;
1921 struct memcg_stock_pcp
*stock
;
1923 if (action
== CPU_ONLINE
)
1926 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
1929 stock
= &per_cpu(memcg_stock
, cpu
);
1934 static void reclaim_high(struct mem_cgroup
*memcg
,
1935 unsigned int nr_pages
,
1939 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
1941 mem_cgroup_events(memcg
, MEMCG_HIGH
, 1);
1942 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
1943 } while ((memcg
= parent_mem_cgroup(memcg
)));
1946 static void high_work_func(struct work_struct
*work
)
1948 struct mem_cgroup
*memcg
;
1950 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
1951 reclaim_high(memcg
, CHARGE_BATCH
, GFP_KERNEL
);
1955 * Scheduled by try_charge() to be executed from the userland return path
1956 * and reclaims memory over the high limit.
1958 void mem_cgroup_handle_over_high(void)
1960 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
1961 struct mem_cgroup
*memcg
;
1963 if (likely(!nr_pages
))
1966 memcg
= get_mem_cgroup_from_mm(current
->mm
);
1967 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
1968 css_put(&memcg
->css
);
1969 current
->memcg_nr_pages_over_high
= 0;
1972 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1973 unsigned int nr_pages
)
1975 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
1976 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1977 struct mem_cgroup
*mem_over_limit
;
1978 struct page_counter
*counter
;
1979 unsigned long nr_reclaimed
;
1980 bool may_swap
= true;
1981 bool drained
= false;
1983 if (mem_cgroup_is_root(memcg
))
1986 if (consume_stock(memcg
, nr_pages
))
1989 if (!do_memsw_account() ||
1990 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
1991 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
1993 if (do_memsw_account())
1994 page_counter_uncharge(&memcg
->memsw
, batch
);
1995 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
1997 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2001 if (batch
> nr_pages
) {
2007 * Unlike in global OOM situations, memcg is not in a physical
2008 * memory shortage. Allow dying and OOM-killed tasks to
2009 * bypass the last charges so that they can exit quickly and
2010 * free their memory.
2012 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
2013 fatal_signal_pending(current
) ||
2014 current
->flags
& PF_EXITING
))
2017 if (unlikely(task_in_memcg_oom(current
)))
2020 if (!gfpflags_allow_blocking(gfp_mask
))
2023 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
2025 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2026 gfp_mask
, may_swap
);
2028 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2032 drain_all_stock(mem_over_limit
);
2037 if (gfp_mask
& __GFP_NORETRY
)
2040 * Even though the limit is exceeded at this point, reclaim
2041 * may have been able to free some pages. Retry the charge
2042 * before killing the task.
2044 * Only for regular pages, though: huge pages are rather
2045 * unlikely to succeed so close to the limit, and we fall back
2046 * to regular pages anyway in case of failure.
2048 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2051 * At task move, charge accounts can be doubly counted. So, it's
2052 * better to wait until the end of task_move if something is going on.
2054 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2060 if (gfp_mask
& __GFP_NOFAIL
)
2063 if (fatal_signal_pending(current
))
2066 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
2068 mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2069 get_order(nr_pages
* PAGE_SIZE
));
2071 if (!(gfp_mask
& __GFP_NOFAIL
))
2075 * The allocation either can't fail or will lead to more memory
2076 * being freed very soon. Allow memory usage go over the limit
2077 * temporarily by force charging it.
2079 page_counter_charge(&memcg
->memory
, nr_pages
);
2080 if (do_memsw_account())
2081 page_counter_charge(&memcg
->memsw
, nr_pages
);
2082 css_get_many(&memcg
->css
, nr_pages
);
2087 css_get_many(&memcg
->css
, batch
);
2088 if (batch
> nr_pages
)
2089 refill_stock(memcg
, batch
- nr_pages
);
2092 * If the hierarchy is above the normal consumption range, schedule
2093 * reclaim on returning to userland. We can perform reclaim here
2094 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2095 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2096 * not recorded as it most likely matches current's and won't
2097 * change in the meantime. As high limit is checked again before
2098 * reclaim, the cost of mismatch is negligible.
2101 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2102 /* Don't bother a random interrupted task */
2103 if (in_interrupt()) {
2104 schedule_work(&memcg
->high_work
);
2107 current
->memcg_nr_pages_over_high
+= batch
;
2108 set_notify_resume(current
);
2111 } while ((memcg
= parent_mem_cgroup(memcg
)));
2116 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2118 if (mem_cgroup_is_root(memcg
))
2121 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2122 if (do_memsw_account())
2123 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2125 css_put_many(&memcg
->css
, nr_pages
);
2128 static void lock_page_lru(struct page
*page
, int *isolated
)
2130 struct zone
*zone
= page_zone(page
);
2132 spin_lock_irq(&zone
->lru_lock
);
2133 if (PageLRU(page
)) {
2134 struct lruvec
*lruvec
;
2136 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2138 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2144 static void unlock_page_lru(struct page
*page
, int isolated
)
2146 struct zone
*zone
= page_zone(page
);
2149 struct lruvec
*lruvec
;
2151 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2152 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2154 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2156 spin_unlock_irq(&zone
->lru_lock
);
2159 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2164 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2167 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2168 * may already be on some other mem_cgroup's LRU. Take care of it.
2171 lock_page_lru(page
, &isolated
);
2174 * Nobody should be changing or seriously looking at
2175 * page->mem_cgroup at this point:
2177 * - the page is uncharged
2179 * - the page is off-LRU
2181 * - an anonymous fault has exclusive page access, except for
2182 * a locked page table
2184 * - a page cache insertion, a swapin fault, or a migration
2185 * have the page locked
2187 page
->mem_cgroup
= memcg
;
2190 unlock_page_lru(page
, isolated
);
2194 static int memcg_alloc_cache_id(void)
2199 id
= ida_simple_get(&memcg_cache_ida
,
2200 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2204 if (id
< memcg_nr_cache_ids
)
2208 * There's no space for the new id in memcg_caches arrays,
2209 * so we have to grow them.
2211 down_write(&memcg_cache_ids_sem
);
2213 size
= 2 * (id
+ 1);
2214 if (size
< MEMCG_CACHES_MIN_SIZE
)
2215 size
= MEMCG_CACHES_MIN_SIZE
;
2216 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2217 size
= MEMCG_CACHES_MAX_SIZE
;
2219 err
= memcg_update_all_caches(size
);
2221 err
= memcg_update_all_list_lrus(size
);
2223 memcg_nr_cache_ids
= size
;
2225 up_write(&memcg_cache_ids_sem
);
2228 ida_simple_remove(&memcg_cache_ida
, id
);
2234 static void memcg_free_cache_id(int id
)
2236 ida_simple_remove(&memcg_cache_ida
, id
);
2239 struct memcg_kmem_cache_create_work
{
2240 struct mem_cgroup
*memcg
;
2241 struct kmem_cache
*cachep
;
2242 struct work_struct work
;
2245 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2247 struct memcg_kmem_cache_create_work
*cw
=
2248 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2249 struct mem_cgroup
*memcg
= cw
->memcg
;
2250 struct kmem_cache
*cachep
= cw
->cachep
;
2252 memcg_create_kmem_cache(memcg
, cachep
);
2254 css_put(&memcg
->css
);
2259 * Enqueue the creation of a per-memcg kmem_cache.
2261 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2262 struct kmem_cache
*cachep
)
2264 struct memcg_kmem_cache_create_work
*cw
;
2266 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2270 css_get(&memcg
->css
);
2273 cw
->cachep
= cachep
;
2274 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2276 schedule_work(&cw
->work
);
2279 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2280 struct kmem_cache
*cachep
)
2283 * We need to stop accounting when we kmalloc, because if the
2284 * corresponding kmalloc cache is not yet created, the first allocation
2285 * in __memcg_schedule_kmem_cache_create will recurse.
2287 * However, it is better to enclose the whole function. Depending on
2288 * the debugging options enabled, INIT_WORK(), for instance, can
2289 * trigger an allocation. This too, will make us recurse. Because at
2290 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2291 * the safest choice is to do it like this, wrapping the whole function.
2293 current
->memcg_kmem_skip_account
= 1;
2294 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2295 current
->memcg_kmem_skip_account
= 0;
2299 * Return the kmem_cache we're supposed to use for a slab allocation.
2300 * We try to use the current memcg's version of the cache.
2302 * If the cache does not exist yet, if we are the first user of it,
2303 * we either create it immediately, if possible, or create it asynchronously
2305 * In the latter case, we will let the current allocation go through with
2306 * the original cache.
2308 * Can't be called in interrupt context or from kernel threads.
2309 * This function needs to be called with rcu_read_lock() held.
2311 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2313 struct mem_cgroup
*memcg
;
2314 struct kmem_cache
*memcg_cachep
;
2317 VM_BUG_ON(!is_root_cache(cachep
));
2319 if (cachep
->flags
& SLAB_ACCOUNT
)
2320 gfp
|= __GFP_ACCOUNT
;
2322 if (!(gfp
& __GFP_ACCOUNT
))
2325 if (current
->memcg_kmem_skip_account
)
2328 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2329 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2333 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2334 if (likely(memcg_cachep
))
2335 return memcg_cachep
;
2338 * If we are in a safe context (can wait, and not in interrupt
2339 * context), we could be be predictable and return right away.
2340 * This would guarantee that the allocation being performed
2341 * already belongs in the new cache.
2343 * However, there are some clashes that can arrive from locking.
2344 * For instance, because we acquire the slab_mutex while doing
2345 * memcg_create_kmem_cache, this means no further allocation
2346 * could happen with the slab_mutex held. So it's better to
2349 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2351 css_put(&memcg
->css
);
2355 void __memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2357 if (!is_root_cache(cachep
))
2358 css_put(&cachep
->memcg_params
.memcg
->css
);
2361 int __memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2362 struct mem_cgroup
*memcg
)
2364 unsigned int nr_pages
= 1 << order
;
2365 struct page_counter
*counter
;
2368 if (!memcg_kmem_online(memcg
))
2371 ret
= try_charge(memcg
, gfp
, nr_pages
);
2375 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2376 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2377 cancel_charge(memcg
, nr_pages
);
2381 page
->mem_cgroup
= memcg
;
2386 int __memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2388 struct mem_cgroup
*memcg
;
2391 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2392 ret
= __memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2393 css_put(&memcg
->css
);
2397 void __memcg_kmem_uncharge(struct page
*page
, int order
)
2399 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2400 unsigned int nr_pages
= 1 << order
;
2405 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2407 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2408 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2410 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2411 if (do_memsw_account())
2412 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2414 page
->mem_cgroup
= NULL
;
2415 css_put_many(&memcg
->css
, nr_pages
);
2417 #endif /* !CONFIG_SLOB */
2419 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2422 * Because tail pages are not marked as "used", set it. We're under
2423 * zone->lru_lock and migration entries setup in all page mappings.
2425 void mem_cgroup_split_huge_fixup(struct page
*head
)
2429 if (mem_cgroup_disabled())
2432 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2433 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2435 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2438 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2440 #ifdef CONFIG_MEMCG_SWAP
2441 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2444 int val
= (charge
) ? 1 : -1;
2445 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2449 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2450 * @entry: swap entry to be moved
2451 * @from: mem_cgroup which the entry is moved from
2452 * @to: mem_cgroup which the entry is moved to
2454 * It succeeds only when the swap_cgroup's record for this entry is the same
2455 * as the mem_cgroup's id of @from.
2457 * Returns 0 on success, -EINVAL on failure.
2459 * The caller must have charged to @to, IOW, called page_counter_charge() about
2460 * both res and memsw, and called css_get().
2462 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2463 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2465 unsigned short old_id
, new_id
;
2467 old_id
= mem_cgroup_id(from
);
2468 new_id
= mem_cgroup_id(to
);
2470 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2471 mem_cgroup_swap_statistics(from
, false);
2472 mem_cgroup_swap_statistics(to
, true);
2478 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2479 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2485 static DEFINE_MUTEX(memcg_limit_mutex
);
2487 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2488 unsigned long limit
)
2490 unsigned long curusage
;
2491 unsigned long oldusage
;
2492 bool enlarge
= false;
2497 * For keeping hierarchical_reclaim simple, how long we should retry
2498 * is depends on callers. We set our retry-count to be function
2499 * of # of children which we should visit in this loop.
2501 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2502 mem_cgroup_count_children(memcg
);
2504 oldusage
= page_counter_read(&memcg
->memory
);
2507 if (signal_pending(current
)) {
2512 mutex_lock(&memcg_limit_mutex
);
2513 if (limit
> memcg
->memsw
.limit
) {
2514 mutex_unlock(&memcg_limit_mutex
);
2518 if (limit
> memcg
->memory
.limit
)
2520 ret
= page_counter_limit(&memcg
->memory
, limit
);
2521 mutex_unlock(&memcg_limit_mutex
);
2526 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2528 curusage
= page_counter_read(&memcg
->memory
);
2529 /* Usage is reduced ? */
2530 if (curusage
>= oldusage
)
2533 oldusage
= curusage
;
2534 } while (retry_count
);
2536 if (!ret
&& enlarge
)
2537 memcg_oom_recover(memcg
);
2542 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2543 unsigned long limit
)
2545 unsigned long curusage
;
2546 unsigned long oldusage
;
2547 bool enlarge
= false;
2551 /* see mem_cgroup_resize_res_limit */
2552 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2553 mem_cgroup_count_children(memcg
);
2555 oldusage
= page_counter_read(&memcg
->memsw
);
2558 if (signal_pending(current
)) {
2563 mutex_lock(&memcg_limit_mutex
);
2564 if (limit
< memcg
->memory
.limit
) {
2565 mutex_unlock(&memcg_limit_mutex
);
2569 if (limit
> memcg
->memsw
.limit
)
2571 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2572 mutex_unlock(&memcg_limit_mutex
);
2577 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2579 curusage
= page_counter_read(&memcg
->memsw
);
2580 /* Usage is reduced ? */
2581 if (curusage
>= oldusage
)
2584 oldusage
= curusage
;
2585 } while (retry_count
);
2587 if (!ret
&& enlarge
)
2588 memcg_oom_recover(memcg
);
2593 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
2595 unsigned long *total_scanned
)
2597 unsigned long nr_reclaimed
= 0;
2598 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
2599 unsigned long reclaimed
;
2601 struct mem_cgroup_tree_per_zone
*mctz
;
2602 unsigned long excess
;
2603 unsigned long nr_scanned
;
2608 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
2610 * This loop can run a while, specially if mem_cgroup's continuously
2611 * keep exceeding their soft limit and putting the system under
2618 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2623 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
2624 gfp_mask
, &nr_scanned
);
2625 nr_reclaimed
+= reclaimed
;
2626 *total_scanned
+= nr_scanned
;
2627 spin_lock_irq(&mctz
->lock
);
2628 __mem_cgroup_remove_exceeded(mz
, mctz
);
2631 * If we failed to reclaim anything from this memory cgroup
2632 * it is time to move on to the next cgroup
2636 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2638 excess
= soft_limit_excess(mz
->memcg
);
2640 * One school of thought says that we should not add
2641 * back the node to the tree if reclaim returns 0.
2642 * But our reclaim could return 0, simply because due
2643 * to priority we are exposing a smaller subset of
2644 * memory to reclaim from. Consider this as a longer
2647 /* If excess == 0, no tree ops */
2648 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2649 spin_unlock_irq(&mctz
->lock
);
2650 css_put(&mz
->memcg
->css
);
2653 * Could not reclaim anything and there are no more
2654 * mem cgroups to try or we seem to be looping without
2655 * reclaiming anything.
2657 if (!nr_reclaimed
&&
2659 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2661 } while (!nr_reclaimed
);
2663 css_put(&next_mz
->memcg
->css
);
2664 return nr_reclaimed
;
2668 * Test whether @memcg has children, dead or alive. Note that this
2669 * function doesn't care whether @memcg has use_hierarchy enabled and
2670 * returns %true if there are child csses according to the cgroup
2671 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2673 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2678 ret
= css_next_child(NULL
, &memcg
->css
);
2684 * Reclaims as many pages from the given memcg as possible and moves
2685 * the rest to the parent.
2687 * Caller is responsible for holding css reference for memcg.
2689 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2691 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2693 /* we call try-to-free pages for make this cgroup empty */
2694 lru_add_drain_all();
2695 /* try to free all pages in this cgroup */
2696 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2699 if (signal_pending(current
))
2702 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2706 /* maybe some writeback is necessary */
2707 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2715 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
2716 char *buf
, size_t nbytes
,
2719 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2721 if (mem_cgroup_is_root(memcg
))
2723 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
2726 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
2729 return mem_cgroup_from_css(css
)->use_hierarchy
;
2732 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
2733 struct cftype
*cft
, u64 val
)
2736 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2737 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
2739 if (memcg
->use_hierarchy
== val
)
2743 * If parent's use_hierarchy is set, we can't make any modifications
2744 * in the child subtrees. If it is unset, then the change can
2745 * occur, provided the current cgroup has no children.
2747 * For the root cgroup, parent_mem is NULL, we allow value to be
2748 * set if there are no children.
2750 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
2751 (val
== 1 || val
== 0)) {
2752 if (!memcg_has_children(memcg
))
2753 memcg
->use_hierarchy
= val
;
2762 static unsigned long tree_stat(struct mem_cgroup
*memcg
,
2763 enum mem_cgroup_stat_index idx
)
2765 struct mem_cgroup
*iter
;
2766 unsigned long val
= 0;
2768 for_each_mem_cgroup_tree(iter
, memcg
)
2769 val
+= mem_cgroup_read_stat(iter
, idx
);
2774 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
2778 if (mem_cgroup_is_root(memcg
)) {
2779 val
= tree_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
2780 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_RSS
);
2782 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
2785 val
= page_counter_read(&memcg
->memory
);
2787 val
= page_counter_read(&memcg
->memsw
);
2800 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
2803 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2804 struct page_counter
*counter
;
2806 switch (MEMFILE_TYPE(cft
->private)) {
2808 counter
= &memcg
->memory
;
2811 counter
= &memcg
->memsw
;
2814 counter
= &memcg
->kmem
;
2817 counter
= &memcg
->tcpmem
;
2823 switch (MEMFILE_ATTR(cft
->private)) {
2825 if (counter
== &memcg
->memory
)
2826 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
2827 if (counter
== &memcg
->memsw
)
2828 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
2829 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
2831 return (u64
)counter
->limit
* PAGE_SIZE
;
2833 return (u64
)counter
->watermark
* PAGE_SIZE
;
2835 return counter
->failcnt
;
2836 case RES_SOFT_LIMIT
:
2837 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
2844 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2848 BUG_ON(memcg
->kmemcg_id
>= 0);
2849 BUG_ON(memcg
->kmem_state
);
2851 memcg_id
= memcg_alloc_cache_id();
2855 static_branch_inc(&memcg_kmem_enabled_key
);
2857 * A memory cgroup is considered kmem-online as soon as it gets
2858 * kmemcg_id. Setting the id after enabling static branching will
2859 * guarantee no one starts accounting before all call sites are
2862 memcg
->kmemcg_id
= memcg_id
;
2863 memcg
->kmem_state
= KMEM_ONLINE
;
2868 static int memcg_propagate_kmem(struct mem_cgroup
*parent
,
2869 struct mem_cgroup
*memcg
)
2873 mutex_lock(&memcg_limit_mutex
);
2875 * If the parent cgroup is not kmem-online now, it cannot be
2876 * onlined after this point, because it has at least one child
2879 if (memcg_kmem_online(parent
) ||
2880 (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nokmem
))
2881 ret
= memcg_online_kmem(memcg
);
2882 mutex_unlock(&memcg_limit_mutex
);
2886 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2888 struct cgroup_subsys_state
*css
;
2889 struct mem_cgroup
*parent
, *child
;
2892 if (memcg
->kmem_state
!= KMEM_ONLINE
)
2895 * Clear the online state before clearing memcg_caches array
2896 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2897 * guarantees that no cache will be created for this cgroup
2898 * after we are done (see memcg_create_kmem_cache()).
2900 memcg
->kmem_state
= KMEM_ALLOCATED
;
2902 memcg_deactivate_kmem_caches(memcg
);
2904 kmemcg_id
= memcg
->kmemcg_id
;
2905 BUG_ON(kmemcg_id
< 0);
2907 parent
= parent_mem_cgroup(memcg
);
2909 parent
= root_mem_cgroup
;
2912 * Change kmemcg_id of this cgroup and all its descendants to the
2913 * parent's id, and then move all entries from this cgroup's list_lrus
2914 * to ones of the parent. After we have finished, all list_lrus
2915 * corresponding to this cgroup are guaranteed to remain empty. The
2916 * ordering is imposed by list_lru_node->lock taken by
2917 * memcg_drain_all_list_lrus().
2919 css_for_each_descendant_pre(css
, &memcg
->css
) {
2920 child
= mem_cgroup_from_css(css
);
2921 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
2922 child
->kmemcg_id
= parent
->kmemcg_id
;
2923 if (!memcg
->use_hierarchy
)
2926 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
2928 memcg_free_cache_id(kmemcg_id
);
2931 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2933 /* css_alloc() failed, offlining didn't happen */
2934 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
2935 memcg_offline_kmem(memcg
);
2937 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
2938 memcg_destroy_kmem_caches(memcg
);
2939 static_branch_dec(&memcg_kmem_enabled_key
);
2940 WARN_ON(page_counter_read(&memcg
->kmem
));
2944 static int memcg_propagate_kmem(struct mem_cgroup
*parent
, struct mem_cgroup
*memcg
)
2948 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2952 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2955 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2958 #endif /* !CONFIG_SLOB */
2960 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2961 unsigned long limit
)
2965 mutex_lock(&memcg_limit_mutex
);
2966 /* Top-level cgroup doesn't propagate from root */
2967 if (!memcg_kmem_online(memcg
)) {
2968 if (cgroup_is_populated(memcg
->css
.cgroup
) ||
2969 (memcg
->use_hierarchy
&& memcg_has_children(memcg
)))
2973 ret
= memcg_online_kmem(memcg
);
2977 ret
= page_counter_limit(&memcg
->kmem
, limit
);
2979 mutex_unlock(&memcg_limit_mutex
);
2983 static int memcg_update_tcp_limit(struct mem_cgroup
*memcg
, unsigned long limit
)
2987 mutex_lock(&memcg_limit_mutex
);
2989 ret
= page_counter_limit(&memcg
->tcpmem
, limit
);
2993 if (!memcg
->tcpmem_active
) {
2995 * The active flag needs to be written after the static_key
2996 * update. This is what guarantees that the socket activation
2997 * function is the last one to run. See sock_update_memcg() for
2998 * details, and note that we don't mark any socket as belonging
2999 * to this memcg until that flag is up.
3001 * We need to do this, because static_keys will span multiple
3002 * sites, but we can't control their order. If we mark a socket
3003 * as accounted, but the accounting functions are not patched in
3004 * yet, we'll lose accounting.
3006 * We never race with the readers in sock_update_memcg(),
3007 * because when this value change, the code to process it is not
3010 static_branch_inc(&memcg_sockets_enabled_key
);
3011 memcg
->tcpmem_active
= true;
3014 mutex_unlock(&memcg_limit_mutex
);
3019 * The user of this function is...
3022 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3023 char *buf
, size_t nbytes
, loff_t off
)
3025 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3026 unsigned long nr_pages
;
3029 buf
= strstrip(buf
);
3030 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3034 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3036 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3040 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3042 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
3045 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
3048 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
3051 ret
= memcg_update_tcp_limit(memcg
, nr_pages
);
3055 case RES_SOFT_LIMIT
:
3056 memcg
->soft_limit
= nr_pages
;
3060 return ret
?: nbytes
;
3063 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3064 size_t nbytes
, loff_t off
)
3066 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3067 struct page_counter
*counter
;
3069 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3071 counter
= &memcg
->memory
;
3074 counter
= &memcg
->memsw
;
3077 counter
= &memcg
->kmem
;
3080 counter
= &memcg
->tcpmem
;
3086 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3088 page_counter_reset_watermark(counter
);
3091 counter
->failcnt
= 0;
3100 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3103 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3107 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3108 struct cftype
*cft
, u64 val
)
3110 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3112 if (val
& ~MOVE_MASK
)
3116 * No kind of locking is needed in here, because ->can_attach() will
3117 * check this value once in the beginning of the process, and then carry
3118 * on with stale data. This means that changes to this value will only
3119 * affect task migrations starting after the change.
3121 memcg
->move_charge_at_immigrate
= val
;
3125 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3126 struct cftype
*cft
, u64 val
)
3133 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3137 unsigned int lru_mask
;
3140 static const struct numa_stat stats
[] = {
3141 { "total", LRU_ALL
},
3142 { "file", LRU_ALL_FILE
},
3143 { "anon", LRU_ALL_ANON
},
3144 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3146 const struct numa_stat
*stat
;
3149 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3151 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3152 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3153 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3154 for_each_node_state(nid
, N_MEMORY
) {
3155 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3157 seq_printf(m
, " N%d=%lu", nid
, nr
);
3162 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3163 struct mem_cgroup
*iter
;
3166 for_each_mem_cgroup_tree(iter
, memcg
)
3167 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3168 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3169 for_each_node_state(nid
, N_MEMORY
) {
3171 for_each_mem_cgroup_tree(iter
, memcg
)
3172 nr
+= mem_cgroup_node_nr_lru_pages(
3173 iter
, nid
, stat
->lru_mask
);
3174 seq_printf(m
, " N%d=%lu", nid
, nr
);
3181 #endif /* CONFIG_NUMA */
3183 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3185 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3186 unsigned long memory
, memsw
;
3187 struct mem_cgroup
*mi
;
3190 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3191 MEM_CGROUP_STAT_NSTATS
);
3192 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3193 MEM_CGROUP_EVENTS_NSTATS
);
3194 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3196 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3197 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3199 seq_printf(m
, "%s %lu\n", mem_cgroup_stat_names
[i
],
3200 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3203 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3204 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3205 mem_cgroup_read_events(memcg
, i
));
3207 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3208 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3209 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3211 /* Hierarchical information */
3212 memory
= memsw
= PAGE_COUNTER_MAX
;
3213 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3214 memory
= min(memory
, mi
->memory
.limit
);
3215 memsw
= min(memsw
, mi
->memsw
.limit
);
3217 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3218 (u64
)memory
* PAGE_SIZE
);
3219 if (do_memsw_account())
3220 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3221 (u64
)memsw
* PAGE_SIZE
);
3223 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3224 unsigned long long val
= 0;
3226 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3228 for_each_mem_cgroup_tree(mi
, memcg
)
3229 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3230 seq_printf(m
, "total_%s %llu\n", mem_cgroup_stat_names
[i
], val
);
3233 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3234 unsigned long long val
= 0;
3236 for_each_mem_cgroup_tree(mi
, memcg
)
3237 val
+= mem_cgroup_read_events(mi
, i
);
3238 seq_printf(m
, "total_%s %llu\n",
3239 mem_cgroup_events_names
[i
], val
);
3242 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3243 unsigned long long val
= 0;
3245 for_each_mem_cgroup_tree(mi
, memcg
)
3246 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3247 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3250 #ifdef CONFIG_DEBUG_VM
3253 struct mem_cgroup_per_zone
*mz
;
3254 struct zone_reclaim_stat
*rstat
;
3255 unsigned long recent_rotated
[2] = {0, 0};
3256 unsigned long recent_scanned
[2] = {0, 0};
3258 for_each_online_node(nid
)
3259 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3260 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
3261 rstat
= &mz
->lruvec
.reclaim_stat
;
3263 recent_rotated
[0] += rstat
->recent_rotated
[0];
3264 recent_rotated
[1] += rstat
->recent_rotated
[1];
3265 recent_scanned
[0] += rstat
->recent_scanned
[0];
3266 recent_scanned
[1] += rstat
->recent_scanned
[1];
3268 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3269 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3270 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3271 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3278 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3281 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3283 return mem_cgroup_swappiness(memcg
);
3286 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3287 struct cftype
*cft
, u64 val
)
3289 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3295 memcg
->swappiness
= val
;
3297 vm_swappiness
= val
;
3302 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3304 struct mem_cgroup_threshold_ary
*t
;
3305 unsigned long usage
;
3310 t
= rcu_dereference(memcg
->thresholds
.primary
);
3312 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3317 usage
= mem_cgroup_usage(memcg
, swap
);
3320 * current_threshold points to threshold just below or equal to usage.
3321 * If it's not true, a threshold was crossed after last
3322 * call of __mem_cgroup_threshold().
3324 i
= t
->current_threshold
;
3327 * Iterate backward over array of thresholds starting from
3328 * current_threshold and check if a threshold is crossed.
3329 * If none of thresholds below usage is crossed, we read
3330 * only one element of the array here.
3332 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3333 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3335 /* i = current_threshold + 1 */
3339 * Iterate forward over array of thresholds starting from
3340 * current_threshold+1 and check if a threshold is crossed.
3341 * If none of thresholds above usage is crossed, we read
3342 * only one element of the array here.
3344 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3345 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3347 /* Update current_threshold */
3348 t
->current_threshold
= i
- 1;
3353 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3356 __mem_cgroup_threshold(memcg
, false);
3357 if (do_memsw_account())
3358 __mem_cgroup_threshold(memcg
, true);
3360 memcg
= parent_mem_cgroup(memcg
);
3364 static int compare_thresholds(const void *a
, const void *b
)
3366 const struct mem_cgroup_threshold
*_a
= a
;
3367 const struct mem_cgroup_threshold
*_b
= b
;
3369 if (_a
->threshold
> _b
->threshold
)
3372 if (_a
->threshold
< _b
->threshold
)
3378 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3380 struct mem_cgroup_eventfd_list
*ev
;
3382 spin_lock(&memcg_oom_lock
);
3384 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3385 eventfd_signal(ev
->eventfd
, 1);
3387 spin_unlock(&memcg_oom_lock
);
3391 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3393 struct mem_cgroup
*iter
;
3395 for_each_mem_cgroup_tree(iter
, memcg
)
3396 mem_cgroup_oom_notify_cb(iter
);
3399 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3400 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3402 struct mem_cgroup_thresholds
*thresholds
;
3403 struct mem_cgroup_threshold_ary
*new;
3404 unsigned long threshold
;
3405 unsigned long usage
;
3408 ret
= page_counter_memparse(args
, "-1", &threshold
);
3412 mutex_lock(&memcg
->thresholds_lock
);
3415 thresholds
= &memcg
->thresholds
;
3416 usage
= mem_cgroup_usage(memcg
, false);
3417 } else if (type
== _MEMSWAP
) {
3418 thresholds
= &memcg
->memsw_thresholds
;
3419 usage
= mem_cgroup_usage(memcg
, true);
3423 /* Check if a threshold crossed before adding a new one */
3424 if (thresholds
->primary
)
3425 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3427 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3429 /* Allocate memory for new array of thresholds */
3430 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3438 /* Copy thresholds (if any) to new array */
3439 if (thresholds
->primary
) {
3440 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3441 sizeof(struct mem_cgroup_threshold
));
3444 /* Add new threshold */
3445 new->entries
[size
- 1].eventfd
= eventfd
;
3446 new->entries
[size
- 1].threshold
= threshold
;
3448 /* Sort thresholds. Registering of new threshold isn't time-critical */
3449 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3450 compare_thresholds
, NULL
);
3452 /* Find current threshold */
3453 new->current_threshold
= -1;
3454 for (i
= 0; i
< size
; i
++) {
3455 if (new->entries
[i
].threshold
<= usage
) {
3457 * new->current_threshold will not be used until
3458 * rcu_assign_pointer(), so it's safe to increment
3461 ++new->current_threshold
;
3466 /* Free old spare buffer and save old primary buffer as spare */
3467 kfree(thresholds
->spare
);
3468 thresholds
->spare
= thresholds
->primary
;
3470 rcu_assign_pointer(thresholds
->primary
, new);
3472 /* To be sure that nobody uses thresholds */
3476 mutex_unlock(&memcg
->thresholds_lock
);
3481 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3482 struct eventfd_ctx
*eventfd
, const char *args
)
3484 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3487 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3488 struct eventfd_ctx
*eventfd
, const char *args
)
3490 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3493 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3494 struct eventfd_ctx
*eventfd
, enum res_type type
)
3496 struct mem_cgroup_thresholds
*thresholds
;
3497 struct mem_cgroup_threshold_ary
*new;
3498 unsigned long usage
;
3501 mutex_lock(&memcg
->thresholds_lock
);
3504 thresholds
= &memcg
->thresholds
;
3505 usage
= mem_cgroup_usage(memcg
, false);
3506 } else if (type
== _MEMSWAP
) {
3507 thresholds
= &memcg
->memsw_thresholds
;
3508 usage
= mem_cgroup_usage(memcg
, true);
3512 if (!thresholds
->primary
)
3515 /* Check if a threshold crossed before removing */
3516 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3518 /* Calculate new number of threshold */
3520 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3521 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3525 new = thresholds
->spare
;
3527 /* Set thresholds array to NULL if we don't have thresholds */
3536 /* Copy thresholds and find current threshold */
3537 new->current_threshold
= -1;
3538 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3539 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3542 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3543 if (new->entries
[j
].threshold
<= usage
) {
3545 * new->current_threshold will not be used
3546 * until rcu_assign_pointer(), so it's safe to increment
3549 ++new->current_threshold
;
3555 /* Swap primary and spare array */
3556 thresholds
->spare
= thresholds
->primary
;
3558 rcu_assign_pointer(thresholds
->primary
, new);
3560 /* To be sure that nobody uses thresholds */
3563 /* If all events are unregistered, free the spare array */
3565 kfree(thresholds
->spare
);
3566 thresholds
->spare
= NULL
;
3569 mutex_unlock(&memcg
->thresholds_lock
);
3572 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3573 struct eventfd_ctx
*eventfd
)
3575 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3578 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3579 struct eventfd_ctx
*eventfd
)
3581 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3584 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3585 struct eventfd_ctx
*eventfd
, const char *args
)
3587 struct mem_cgroup_eventfd_list
*event
;
3589 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3593 spin_lock(&memcg_oom_lock
);
3595 event
->eventfd
= eventfd
;
3596 list_add(&event
->list
, &memcg
->oom_notify
);
3598 /* already in OOM ? */
3599 if (memcg
->under_oom
)
3600 eventfd_signal(eventfd
, 1);
3601 spin_unlock(&memcg_oom_lock
);
3606 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3607 struct eventfd_ctx
*eventfd
)
3609 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3611 spin_lock(&memcg_oom_lock
);
3613 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3614 if (ev
->eventfd
== eventfd
) {
3615 list_del(&ev
->list
);
3620 spin_unlock(&memcg_oom_lock
);
3623 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3625 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3627 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3628 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3632 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3633 struct cftype
*cft
, u64 val
)
3635 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3637 /* cannot set to root cgroup and only 0 and 1 are allowed */
3638 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3641 memcg
->oom_kill_disable
= val
;
3643 memcg_oom_recover(memcg
);
3648 #ifdef CONFIG_CGROUP_WRITEBACK
3650 struct list_head
*mem_cgroup_cgwb_list(struct mem_cgroup
*memcg
)
3652 return &memcg
->cgwb_list
;
3655 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3657 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3660 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3662 wb_domain_exit(&memcg
->cgwb_domain
);
3665 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3667 wb_domain_size_changed(&memcg
->cgwb_domain
);
3670 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3672 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3674 if (!memcg
->css
.parent
)
3677 return &memcg
->cgwb_domain
;
3681 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3682 * @wb: bdi_writeback in question
3683 * @pfilepages: out parameter for number of file pages
3684 * @pheadroom: out parameter for number of allocatable pages according to memcg
3685 * @pdirty: out parameter for number of dirty pages
3686 * @pwriteback: out parameter for number of pages under writeback
3688 * Determine the numbers of file, headroom, dirty, and writeback pages in
3689 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3690 * is a bit more involved.
3692 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3693 * headroom is calculated as the lowest headroom of itself and the
3694 * ancestors. Note that this doesn't consider the actual amount of
3695 * available memory in the system. The caller should further cap
3696 * *@pheadroom accordingly.
3698 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
3699 unsigned long *pheadroom
, unsigned long *pdirty
,
3700 unsigned long *pwriteback
)
3702 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3703 struct mem_cgroup
*parent
;
3705 *pdirty
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
3707 /* this should eventually include NR_UNSTABLE_NFS */
3708 *pwriteback
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
3709 *pfilepages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
3710 (1 << LRU_ACTIVE_FILE
));
3711 *pheadroom
= PAGE_COUNTER_MAX
;
3713 while ((parent
= parent_mem_cgroup(memcg
))) {
3714 unsigned long ceiling
= min(memcg
->memory
.limit
, memcg
->high
);
3715 unsigned long used
= page_counter_read(&memcg
->memory
);
3717 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
3722 #else /* CONFIG_CGROUP_WRITEBACK */
3724 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3729 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3733 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3737 #endif /* CONFIG_CGROUP_WRITEBACK */
3740 * DO NOT USE IN NEW FILES.
3742 * "cgroup.event_control" implementation.
3744 * This is way over-engineered. It tries to support fully configurable
3745 * events for each user. Such level of flexibility is completely
3746 * unnecessary especially in the light of the planned unified hierarchy.
3748 * Please deprecate this and replace with something simpler if at all
3753 * Unregister event and free resources.
3755 * Gets called from workqueue.
3757 static void memcg_event_remove(struct work_struct
*work
)
3759 struct mem_cgroup_event
*event
=
3760 container_of(work
, struct mem_cgroup_event
, remove
);
3761 struct mem_cgroup
*memcg
= event
->memcg
;
3763 remove_wait_queue(event
->wqh
, &event
->wait
);
3765 event
->unregister_event(memcg
, event
->eventfd
);
3767 /* Notify userspace the event is going away. */
3768 eventfd_signal(event
->eventfd
, 1);
3770 eventfd_ctx_put(event
->eventfd
);
3772 css_put(&memcg
->css
);
3776 * Gets called on POLLHUP on eventfd when user closes it.
3778 * Called with wqh->lock held and interrupts disabled.
3780 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
3781 int sync
, void *key
)
3783 struct mem_cgroup_event
*event
=
3784 container_of(wait
, struct mem_cgroup_event
, wait
);
3785 struct mem_cgroup
*memcg
= event
->memcg
;
3786 unsigned long flags
= (unsigned long)key
;
3788 if (flags
& POLLHUP
) {
3790 * If the event has been detached at cgroup removal, we
3791 * can simply return knowing the other side will cleanup
3794 * We can't race against event freeing since the other
3795 * side will require wqh->lock via remove_wait_queue(),
3798 spin_lock(&memcg
->event_list_lock
);
3799 if (!list_empty(&event
->list
)) {
3800 list_del_init(&event
->list
);
3802 * We are in atomic context, but cgroup_event_remove()
3803 * may sleep, so we have to call it in workqueue.
3805 schedule_work(&event
->remove
);
3807 spin_unlock(&memcg
->event_list_lock
);
3813 static void memcg_event_ptable_queue_proc(struct file
*file
,
3814 wait_queue_head_t
*wqh
, poll_table
*pt
)
3816 struct mem_cgroup_event
*event
=
3817 container_of(pt
, struct mem_cgroup_event
, pt
);
3820 add_wait_queue(wqh
, &event
->wait
);
3824 * DO NOT USE IN NEW FILES.
3826 * Parse input and register new cgroup event handler.
3828 * Input must be in format '<event_fd> <control_fd> <args>'.
3829 * Interpretation of args is defined by control file implementation.
3831 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
3832 char *buf
, size_t nbytes
, loff_t off
)
3834 struct cgroup_subsys_state
*css
= of_css(of
);
3835 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3836 struct mem_cgroup_event
*event
;
3837 struct cgroup_subsys_state
*cfile_css
;
3838 unsigned int efd
, cfd
;
3845 buf
= strstrip(buf
);
3847 efd
= simple_strtoul(buf
, &endp
, 10);
3852 cfd
= simple_strtoul(buf
, &endp
, 10);
3853 if ((*endp
!= ' ') && (*endp
!= '\0'))
3857 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
3861 event
->memcg
= memcg
;
3862 INIT_LIST_HEAD(&event
->list
);
3863 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
3864 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
3865 INIT_WORK(&event
->remove
, memcg_event_remove
);
3873 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
3874 if (IS_ERR(event
->eventfd
)) {
3875 ret
= PTR_ERR(event
->eventfd
);
3882 goto out_put_eventfd
;
3885 /* the process need read permission on control file */
3886 /* AV: shouldn't we check that it's been opened for read instead? */
3887 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
3892 * Determine the event callbacks and set them in @event. This used
3893 * to be done via struct cftype but cgroup core no longer knows
3894 * about these events. The following is crude but the whole thing
3895 * is for compatibility anyway.
3897 * DO NOT ADD NEW FILES.
3899 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
3901 if (!strcmp(name
, "memory.usage_in_bytes")) {
3902 event
->register_event
= mem_cgroup_usage_register_event
;
3903 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
3904 } else if (!strcmp(name
, "memory.oom_control")) {
3905 event
->register_event
= mem_cgroup_oom_register_event
;
3906 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
3907 } else if (!strcmp(name
, "memory.pressure_level")) {
3908 event
->register_event
= vmpressure_register_event
;
3909 event
->unregister_event
= vmpressure_unregister_event
;
3910 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
3911 event
->register_event
= memsw_cgroup_usage_register_event
;
3912 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
3919 * Verify @cfile should belong to @css. Also, remaining events are
3920 * automatically removed on cgroup destruction but the removal is
3921 * asynchronous, so take an extra ref on @css.
3923 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
3924 &memory_cgrp_subsys
);
3926 if (IS_ERR(cfile_css
))
3928 if (cfile_css
!= css
) {
3933 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
3937 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
3939 spin_lock(&memcg
->event_list_lock
);
3940 list_add(&event
->list
, &memcg
->event_list
);
3941 spin_unlock(&memcg
->event_list_lock
);
3953 eventfd_ctx_put(event
->eventfd
);
3962 static struct cftype mem_cgroup_legacy_files
[] = {
3964 .name
= "usage_in_bytes",
3965 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
3966 .read_u64
= mem_cgroup_read_u64
,
3969 .name
= "max_usage_in_bytes",
3970 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
3971 .write
= mem_cgroup_reset
,
3972 .read_u64
= mem_cgroup_read_u64
,
3975 .name
= "limit_in_bytes",
3976 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
3977 .write
= mem_cgroup_write
,
3978 .read_u64
= mem_cgroup_read_u64
,
3981 .name
= "soft_limit_in_bytes",
3982 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
3983 .write
= mem_cgroup_write
,
3984 .read_u64
= mem_cgroup_read_u64
,
3988 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
3989 .write
= mem_cgroup_reset
,
3990 .read_u64
= mem_cgroup_read_u64
,
3994 .seq_show
= memcg_stat_show
,
3997 .name
= "force_empty",
3998 .write
= mem_cgroup_force_empty_write
,
4001 .name
= "use_hierarchy",
4002 .write_u64
= mem_cgroup_hierarchy_write
,
4003 .read_u64
= mem_cgroup_hierarchy_read
,
4006 .name
= "cgroup.event_control", /* XXX: for compat */
4007 .write
= memcg_write_event_control
,
4008 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
4011 .name
= "swappiness",
4012 .read_u64
= mem_cgroup_swappiness_read
,
4013 .write_u64
= mem_cgroup_swappiness_write
,
4016 .name
= "move_charge_at_immigrate",
4017 .read_u64
= mem_cgroup_move_charge_read
,
4018 .write_u64
= mem_cgroup_move_charge_write
,
4021 .name
= "oom_control",
4022 .seq_show
= mem_cgroup_oom_control_read
,
4023 .write_u64
= mem_cgroup_oom_control_write
,
4024 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4027 .name
= "pressure_level",
4031 .name
= "numa_stat",
4032 .seq_show
= memcg_numa_stat_show
,
4036 .name
= "kmem.limit_in_bytes",
4037 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4038 .write
= mem_cgroup_write
,
4039 .read_u64
= mem_cgroup_read_u64
,
4042 .name
= "kmem.usage_in_bytes",
4043 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4044 .read_u64
= mem_cgroup_read_u64
,
4047 .name
= "kmem.failcnt",
4048 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4049 .write
= mem_cgroup_reset
,
4050 .read_u64
= mem_cgroup_read_u64
,
4053 .name
= "kmem.max_usage_in_bytes",
4054 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4055 .write
= mem_cgroup_reset
,
4056 .read_u64
= mem_cgroup_read_u64
,
4058 #ifdef CONFIG_SLABINFO
4060 .name
= "kmem.slabinfo",
4061 .seq_start
= slab_start
,
4062 .seq_next
= slab_next
,
4063 .seq_stop
= slab_stop
,
4064 .seq_show
= memcg_slab_show
,
4068 .name
= "kmem.tcp.limit_in_bytes",
4069 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4070 .write
= mem_cgroup_write
,
4071 .read_u64
= mem_cgroup_read_u64
,
4074 .name
= "kmem.tcp.usage_in_bytes",
4075 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4076 .read_u64
= mem_cgroup_read_u64
,
4079 .name
= "kmem.tcp.failcnt",
4080 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4081 .write
= mem_cgroup_reset
,
4082 .read_u64
= mem_cgroup_read_u64
,
4085 .name
= "kmem.tcp.max_usage_in_bytes",
4086 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4087 .write
= mem_cgroup_reset
,
4088 .read_u64
= mem_cgroup_read_u64
,
4090 { }, /* terminate */
4093 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4095 struct mem_cgroup_per_node
*pn
;
4096 struct mem_cgroup_per_zone
*mz
;
4097 int zone
, tmp
= node
;
4099 * This routine is called against possible nodes.
4100 * But it's BUG to call kmalloc() against offline node.
4102 * TODO: this routine can waste much memory for nodes which will
4103 * never be onlined. It's better to use memory hotplug callback
4106 if (!node_state(node
, N_NORMAL_MEMORY
))
4108 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4112 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4113 mz
= &pn
->zoneinfo
[zone
];
4114 lruvec_init(&mz
->lruvec
);
4115 mz
->usage_in_excess
= 0;
4116 mz
->on_tree
= false;
4119 memcg
->nodeinfo
[node
] = pn
;
4123 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4125 kfree(memcg
->nodeinfo
[node
]);
4128 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4132 memcg_wb_domain_exit(memcg
);
4134 free_mem_cgroup_per_zone_info(memcg
, node
);
4135 free_percpu(memcg
->stat
);
4139 static struct mem_cgroup
*mem_cgroup_alloc(void)
4141 struct mem_cgroup
*memcg
;
4145 size
= sizeof(struct mem_cgroup
);
4146 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4148 memcg
= kzalloc(size
, GFP_KERNEL
);
4152 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4157 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
4160 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4163 INIT_WORK(&memcg
->high_work
, high_work_func
);
4164 memcg
->last_scanned_node
= MAX_NUMNODES
;
4165 INIT_LIST_HEAD(&memcg
->oom_notify
);
4166 mutex_init(&memcg
->thresholds_lock
);
4167 spin_lock_init(&memcg
->move_lock
);
4168 vmpressure_init(&memcg
->vmpressure
);
4169 INIT_LIST_HEAD(&memcg
->event_list
);
4170 spin_lock_init(&memcg
->event_list_lock
);
4171 memcg
->socket_pressure
= jiffies
;
4173 memcg
->kmemcg_id
= -1;
4175 #ifdef CONFIG_CGROUP_WRITEBACK
4176 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4180 mem_cgroup_free(memcg
);
4184 static struct cgroup_subsys_state
* __ref
4185 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4187 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
4188 struct mem_cgroup
*memcg
;
4189 long error
= -ENOMEM
;
4191 memcg
= mem_cgroup_alloc();
4193 return ERR_PTR(error
);
4195 memcg
->high
= PAGE_COUNTER_MAX
;
4196 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4198 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4199 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4201 if (parent
&& parent
->use_hierarchy
) {
4202 memcg
->use_hierarchy
= true;
4203 page_counter_init(&memcg
->memory
, &parent
->memory
);
4204 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4205 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4206 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
4208 page_counter_init(&memcg
->memory
, NULL
);
4209 page_counter_init(&memcg
->memsw
, NULL
);
4210 page_counter_init(&memcg
->kmem
, NULL
);
4211 page_counter_init(&memcg
->tcpmem
, NULL
);
4213 * Deeper hierachy with use_hierarchy == false doesn't make
4214 * much sense so let cgroup subsystem know about this
4215 * unfortunate state in our controller.
4217 if (parent
!= root_mem_cgroup
)
4218 memory_cgrp_subsys
.broken_hierarchy
= true;
4221 /* The following stuff does not apply to the root */
4223 root_mem_cgroup
= memcg
;
4227 error
= memcg_propagate_kmem(parent
, memcg
);
4231 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4232 static_branch_inc(&memcg_sockets_enabled_key
);
4236 mem_cgroup_free(memcg
);
4241 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4243 if (css
->id
> MEM_CGROUP_ID_MAX
)
4249 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4251 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4252 struct mem_cgroup_event
*event
, *tmp
;
4255 * Unregister events and notify userspace.
4256 * Notify userspace about cgroup removing only after rmdir of cgroup
4257 * directory to avoid race between userspace and kernelspace.
4259 spin_lock(&memcg
->event_list_lock
);
4260 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4261 list_del_init(&event
->list
);
4262 schedule_work(&event
->remove
);
4264 spin_unlock(&memcg
->event_list_lock
);
4266 memcg_offline_kmem(memcg
);
4267 wb_memcg_offline(memcg
);
4270 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4272 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4274 invalidate_reclaim_iterators(memcg
);
4277 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4279 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4281 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4282 static_branch_dec(&memcg_sockets_enabled_key
);
4284 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
4285 static_branch_dec(&memcg_sockets_enabled_key
);
4287 vmpressure_cleanup(&memcg
->vmpressure
);
4288 cancel_work_sync(&memcg
->high_work
);
4289 mem_cgroup_remove_from_trees(memcg
);
4290 memcg_free_kmem(memcg
);
4291 mem_cgroup_free(memcg
);
4295 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4296 * @css: the target css
4298 * Reset the states of the mem_cgroup associated with @css. This is
4299 * invoked when the userland requests disabling on the default hierarchy
4300 * but the memcg is pinned through dependency. The memcg should stop
4301 * applying policies and should revert to the vanilla state as it may be
4302 * made visible again.
4304 * The current implementation only resets the essential configurations.
4305 * This needs to be expanded to cover all the visible parts.
4307 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4309 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4311 mem_cgroup_resize_limit(memcg
, PAGE_COUNTER_MAX
);
4312 mem_cgroup_resize_memsw_limit(memcg
, PAGE_COUNTER_MAX
);
4313 memcg_update_kmem_limit(memcg
, PAGE_COUNTER_MAX
);
4315 memcg
->high
= PAGE_COUNTER_MAX
;
4316 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4317 memcg_wb_domain_size_changed(memcg
);
4321 /* Handlers for move charge at task migration. */
4322 static int mem_cgroup_do_precharge(unsigned long count
)
4326 /* Try a single bulk charge without reclaim first, kswapd may wake */
4327 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4329 mc
.precharge
+= count
;
4333 /* Try charges one by one with reclaim */
4335 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4345 * get_mctgt_type - get target type of moving charge
4346 * @vma: the vma the pte to be checked belongs
4347 * @addr: the address corresponding to the pte to be checked
4348 * @ptent: the pte to be checked
4349 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4352 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4353 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4354 * move charge. if @target is not NULL, the page is stored in target->page
4355 * with extra refcnt got(Callers should handle it).
4356 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4357 * target for charge migration. if @target is not NULL, the entry is stored
4360 * Called with pte lock held.
4367 enum mc_target_type
{
4373 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4374 unsigned long addr
, pte_t ptent
)
4376 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4378 if (!page
|| !page_mapped(page
))
4380 if (PageAnon(page
)) {
4381 if (!(mc
.flags
& MOVE_ANON
))
4384 if (!(mc
.flags
& MOVE_FILE
))
4387 if (!get_page_unless_zero(page
))
4394 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4395 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4397 struct page
*page
= NULL
;
4398 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4400 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4403 * Because lookup_swap_cache() updates some statistics counter,
4404 * we call find_get_page() with swapper_space directly.
4406 page
= find_get_page(swap_address_space(ent
), ent
.val
);
4407 if (do_memsw_account())
4408 entry
->val
= ent
.val
;
4413 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4414 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4420 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4421 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4423 struct page
*page
= NULL
;
4424 struct address_space
*mapping
;
4427 if (!vma
->vm_file
) /* anonymous vma */
4429 if (!(mc
.flags
& MOVE_FILE
))
4432 mapping
= vma
->vm_file
->f_mapping
;
4433 pgoff
= linear_page_index(vma
, addr
);
4435 /* page is moved even if it's not RSS of this task(page-faulted). */
4437 /* shmem/tmpfs may report page out on swap: account for that too. */
4438 if (shmem_mapping(mapping
)) {
4439 page
= find_get_entry(mapping
, pgoff
);
4440 if (radix_tree_exceptional_entry(page
)) {
4441 swp_entry_t swp
= radix_to_swp_entry(page
);
4442 if (do_memsw_account())
4444 page
= find_get_page(swap_address_space(swp
), swp
.val
);
4447 page
= find_get_page(mapping
, pgoff
);
4449 page
= find_get_page(mapping
, pgoff
);
4455 * mem_cgroup_move_account - move account of the page
4457 * @nr_pages: number of regular pages (>1 for huge pages)
4458 * @from: mem_cgroup which the page is moved from.
4459 * @to: mem_cgroup which the page is moved to. @from != @to.
4461 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4463 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4466 static int mem_cgroup_move_account(struct page
*page
,
4468 struct mem_cgroup
*from
,
4469 struct mem_cgroup
*to
)
4471 unsigned long flags
;
4472 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
4476 VM_BUG_ON(from
== to
);
4477 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4478 VM_BUG_ON(compound
&& !PageTransHuge(page
));
4481 * Prevent mem_cgroup_replace_page() from looking at
4482 * page->mem_cgroup of its source page while we change it.
4485 if (!trylock_page(page
))
4489 if (page
->mem_cgroup
!= from
)
4492 anon
= PageAnon(page
);
4494 spin_lock_irqsave(&from
->move_lock
, flags
);
4496 if (!anon
&& page_mapped(page
)) {
4497 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4499 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4504 * move_lock grabbed above and caller set from->moving_account, so
4505 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4506 * So mapping should be stable for dirty pages.
4508 if (!anon
&& PageDirty(page
)) {
4509 struct address_space
*mapping
= page_mapping(page
);
4511 if (mapping_cap_account_dirty(mapping
)) {
4512 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4514 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4519 if (PageWriteback(page
)) {
4520 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4522 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4527 * It is safe to change page->mem_cgroup here because the page
4528 * is referenced, charged, and isolated - we can't race with
4529 * uncharging, charging, migration, or LRU putback.
4532 /* caller should have done css_get */
4533 page
->mem_cgroup
= to
;
4534 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4538 local_irq_disable();
4539 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
4540 memcg_check_events(to
, page
);
4541 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
4542 memcg_check_events(from
, page
);
4550 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4551 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4553 struct page
*page
= NULL
;
4554 enum mc_target_type ret
= MC_TARGET_NONE
;
4555 swp_entry_t ent
= { .val
= 0 };
4557 if (pte_present(ptent
))
4558 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4559 else if (is_swap_pte(ptent
))
4560 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
4561 else if (pte_none(ptent
))
4562 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4564 if (!page
&& !ent
.val
)
4568 * Do only loose check w/o serialization.
4569 * mem_cgroup_move_account() checks the page is valid or
4570 * not under LRU exclusion.
4572 if (page
->mem_cgroup
== mc
.from
) {
4573 ret
= MC_TARGET_PAGE
;
4575 target
->page
= page
;
4577 if (!ret
|| !target
)
4580 /* There is a swap entry and a page doesn't exist or isn't charged */
4581 if (ent
.val
&& !ret
&&
4582 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4583 ret
= MC_TARGET_SWAP
;
4590 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4592 * We don't consider swapping or file mapped pages because THP does not
4593 * support them for now.
4594 * Caller should make sure that pmd_trans_huge(pmd) is true.
4596 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4597 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4599 struct page
*page
= NULL
;
4600 enum mc_target_type ret
= MC_TARGET_NONE
;
4602 page
= pmd_page(pmd
);
4603 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4604 if (!(mc
.flags
& MOVE_ANON
))
4606 if (page
->mem_cgroup
== mc
.from
) {
4607 ret
= MC_TARGET_PAGE
;
4610 target
->page
= page
;
4616 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4617 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4619 return MC_TARGET_NONE
;
4623 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4624 unsigned long addr
, unsigned long end
,
4625 struct mm_walk
*walk
)
4627 struct vm_area_struct
*vma
= walk
->vma
;
4631 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
)) {
4632 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4633 mc
.precharge
+= HPAGE_PMD_NR
;
4638 if (pmd_trans_unstable(pmd
))
4640 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4641 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4642 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4643 mc
.precharge
++; /* increment precharge temporarily */
4644 pte_unmap_unlock(pte
- 1, ptl
);
4650 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4652 unsigned long precharge
;
4654 struct mm_walk mem_cgroup_count_precharge_walk
= {
4655 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4658 down_read(&mm
->mmap_sem
);
4659 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk
);
4660 up_read(&mm
->mmap_sem
);
4662 precharge
= mc
.precharge
;
4668 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4670 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4672 VM_BUG_ON(mc
.moving_task
);
4673 mc
.moving_task
= current
;
4674 return mem_cgroup_do_precharge(precharge
);
4677 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4678 static void __mem_cgroup_clear_mc(void)
4680 struct mem_cgroup
*from
= mc
.from
;
4681 struct mem_cgroup
*to
= mc
.to
;
4683 /* we must uncharge all the leftover precharges from mc.to */
4685 cancel_charge(mc
.to
, mc
.precharge
);
4689 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4690 * we must uncharge here.
4692 if (mc
.moved_charge
) {
4693 cancel_charge(mc
.from
, mc
.moved_charge
);
4694 mc
.moved_charge
= 0;
4696 /* we must fixup refcnts and charges */
4697 if (mc
.moved_swap
) {
4698 /* uncharge swap account from the old cgroup */
4699 if (!mem_cgroup_is_root(mc
.from
))
4700 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4703 * we charged both to->memory and to->memsw, so we
4704 * should uncharge to->memory.
4706 if (!mem_cgroup_is_root(mc
.to
))
4707 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4709 css_put_many(&mc
.from
->css
, mc
.moved_swap
);
4711 /* we've already done css_get(mc.to) */
4714 memcg_oom_recover(from
);
4715 memcg_oom_recover(to
);
4716 wake_up_all(&mc
.waitq
);
4719 static void mem_cgroup_clear_mc(void)
4722 * we must clear moving_task before waking up waiters at the end of
4725 mc
.moving_task
= NULL
;
4726 __mem_cgroup_clear_mc();
4727 spin_lock(&mc
.lock
);
4730 spin_unlock(&mc
.lock
);
4733 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4735 struct cgroup_subsys_state
*css
;
4736 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
4737 struct mem_cgroup
*from
;
4738 struct task_struct
*leader
, *p
;
4739 struct mm_struct
*mm
;
4740 unsigned long move_flags
;
4743 /* charge immigration isn't supported on the default hierarchy */
4744 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4748 * Multi-process migrations only happen on the default hierarchy
4749 * where charge immigration is not used. Perform charge
4750 * immigration if @tset contains a leader and whine if there are
4754 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
4757 memcg
= mem_cgroup_from_css(css
);
4763 * We are now commited to this value whatever it is. Changes in this
4764 * tunable will only affect upcoming migrations, not the current one.
4765 * So we need to save it, and keep it going.
4767 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
4771 from
= mem_cgroup_from_task(p
);
4773 VM_BUG_ON(from
== memcg
);
4775 mm
= get_task_mm(p
);
4778 /* We move charges only when we move a owner of the mm */
4779 if (mm
->owner
== p
) {
4782 VM_BUG_ON(mc
.precharge
);
4783 VM_BUG_ON(mc
.moved_charge
);
4784 VM_BUG_ON(mc
.moved_swap
);
4786 spin_lock(&mc
.lock
);
4789 mc
.flags
= move_flags
;
4790 spin_unlock(&mc
.lock
);
4791 /* We set mc.moving_task later */
4793 ret
= mem_cgroup_precharge_mc(mm
);
4795 mem_cgroup_clear_mc();
4801 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4804 mem_cgroup_clear_mc();
4807 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4808 unsigned long addr
, unsigned long end
,
4809 struct mm_walk
*walk
)
4812 struct vm_area_struct
*vma
= walk
->vma
;
4815 enum mc_target_type target_type
;
4816 union mc_target target
;
4819 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
)) {
4820 if (mc
.precharge
< HPAGE_PMD_NR
) {
4824 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
4825 if (target_type
== MC_TARGET_PAGE
) {
4827 if (!isolate_lru_page(page
)) {
4828 if (!mem_cgroup_move_account(page
, true,
4830 mc
.precharge
-= HPAGE_PMD_NR
;
4831 mc
.moved_charge
+= HPAGE_PMD_NR
;
4833 putback_lru_page(page
);
4841 if (pmd_trans_unstable(pmd
))
4844 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4845 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4846 pte_t ptent
= *(pte
++);
4852 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
4853 case MC_TARGET_PAGE
:
4856 * We can have a part of the split pmd here. Moving it
4857 * can be done but it would be too convoluted so simply
4858 * ignore such a partial THP and keep it in original
4859 * memcg. There should be somebody mapping the head.
4861 if (PageTransCompound(page
))
4863 if (isolate_lru_page(page
))
4865 if (!mem_cgroup_move_account(page
, false,
4868 /* we uncharge from mc.from later. */
4871 putback_lru_page(page
);
4872 put
: /* get_mctgt_type() gets the page */
4875 case MC_TARGET_SWAP
:
4877 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
4879 /* we fixup refcnts and charges later. */
4887 pte_unmap_unlock(pte
- 1, ptl
);
4892 * We have consumed all precharges we got in can_attach().
4893 * We try charge one by one, but don't do any additional
4894 * charges to mc.to if we have failed in charge once in attach()
4897 ret
= mem_cgroup_do_precharge(1);
4905 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
4907 struct mm_walk mem_cgroup_move_charge_walk
= {
4908 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4912 lru_add_drain_all();
4914 * Signal mem_cgroup_begin_page_stat() to take the memcg's
4915 * move_lock while we're moving its pages to another memcg.
4916 * Then wait for already started RCU-only updates to finish.
4918 atomic_inc(&mc
.from
->moving_account
);
4921 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
4923 * Someone who are holding the mmap_sem might be waiting in
4924 * waitq. So we cancel all extra charges, wake up all waiters,
4925 * and retry. Because we cancel precharges, we might not be able
4926 * to move enough charges, but moving charge is a best-effort
4927 * feature anyway, so it wouldn't be a big problem.
4929 __mem_cgroup_clear_mc();
4934 * When we have consumed all precharges and failed in doing
4935 * additional charge, the page walk just aborts.
4937 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk
);
4938 up_read(&mm
->mmap_sem
);
4939 atomic_dec(&mc
.from
->moving_account
);
4942 static void mem_cgroup_move_task(struct cgroup_taskset
*tset
)
4944 struct cgroup_subsys_state
*css
;
4945 struct task_struct
*p
= cgroup_taskset_first(tset
, &css
);
4946 struct mm_struct
*mm
= get_task_mm(p
);
4950 mem_cgroup_move_charge(mm
);
4954 mem_cgroup_clear_mc();
4956 #else /* !CONFIG_MMU */
4957 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4961 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4964 static void mem_cgroup_move_task(struct cgroup_taskset
*tset
)
4970 * Cgroup retains root cgroups across [un]mount cycles making it necessary
4971 * to verify whether we're attached to the default hierarchy on each mount
4974 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
4977 * use_hierarchy is forced on the default hierarchy. cgroup core
4978 * guarantees that @root doesn't have any children, so turning it
4979 * on for the root memcg is enough.
4981 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4982 root_mem_cgroup
->use_hierarchy
= true;
4984 root_mem_cgroup
->use_hierarchy
= false;
4987 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
4990 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4992 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
4995 static int memory_low_show(struct seq_file
*m
, void *v
)
4997 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
4998 unsigned long low
= READ_ONCE(memcg
->low
);
5000 if (low
== PAGE_COUNTER_MAX
)
5001 seq_puts(m
, "max\n");
5003 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
5008 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5009 char *buf
, size_t nbytes
, loff_t off
)
5011 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5015 buf
= strstrip(buf
);
5016 err
= page_counter_memparse(buf
, "max", &low
);
5025 static int memory_high_show(struct seq_file
*m
, void *v
)
5027 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5028 unsigned long high
= READ_ONCE(memcg
->high
);
5030 if (high
== PAGE_COUNTER_MAX
)
5031 seq_puts(m
, "max\n");
5033 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
5038 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5039 char *buf
, size_t nbytes
, loff_t off
)
5041 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5045 buf
= strstrip(buf
);
5046 err
= page_counter_memparse(buf
, "max", &high
);
5052 memcg_wb_domain_size_changed(memcg
);
5056 static int memory_max_show(struct seq_file
*m
, void *v
)
5058 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5059 unsigned long max
= READ_ONCE(memcg
->memory
.limit
);
5061 if (max
== PAGE_COUNTER_MAX
)
5062 seq_puts(m
, "max\n");
5064 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5069 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5070 char *buf
, size_t nbytes
, loff_t off
)
5072 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5076 buf
= strstrip(buf
);
5077 err
= page_counter_memparse(buf
, "max", &max
);
5081 err
= mem_cgroup_resize_limit(memcg
, max
);
5085 memcg_wb_domain_size_changed(memcg
);
5089 static int memory_events_show(struct seq_file
*m
, void *v
)
5091 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5093 seq_printf(m
, "low %lu\n", mem_cgroup_read_events(memcg
, MEMCG_LOW
));
5094 seq_printf(m
, "high %lu\n", mem_cgroup_read_events(memcg
, MEMCG_HIGH
));
5095 seq_printf(m
, "max %lu\n", mem_cgroup_read_events(memcg
, MEMCG_MAX
));
5096 seq_printf(m
, "oom %lu\n", mem_cgroup_read_events(memcg
, MEMCG_OOM
));
5101 static struct cftype memory_files
[] = {
5104 .flags
= CFTYPE_NOT_ON_ROOT
,
5105 .read_u64
= memory_current_read
,
5109 .flags
= CFTYPE_NOT_ON_ROOT
,
5110 .seq_show
= memory_low_show
,
5111 .write
= memory_low_write
,
5115 .flags
= CFTYPE_NOT_ON_ROOT
,
5116 .seq_show
= memory_high_show
,
5117 .write
= memory_high_write
,
5121 .flags
= CFTYPE_NOT_ON_ROOT
,
5122 .seq_show
= memory_max_show
,
5123 .write
= memory_max_write
,
5127 .flags
= CFTYPE_NOT_ON_ROOT
,
5128 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
5129 .seq_show
= memory_events_show
,
5134 struct cgroup_subsys memory_cgrp_subsys
= {
5135 .css_alloc
= mem_cgroup_css_alloc
,
5136 .css_online
= mem_cgroup_css_online
,
5137 .css_offline
= mem_cgroup_css_offline
,
5138 .css_released
= mem_cgroup_css_released
,
5139 .css_free
= mem_cgroup_css_free
,
5140 .css_reset
= mem_cgroup_css_reset
,
5141 .can_attach
= mem_cgroup_can_attach
,
5142 .cancel_attach
= mem_cgroup_cancel_attach
,
5143 .attach
= mem_cgroup_move_task
,
5144 .bind
= mem_cgroup_bind
,
5145 .dfl_cftypes
= memory_files
,
5146 .legacy_cftypes
= mem_cgroup_legacy_files
,
5151 * mem_cgroup_low - check if memory consumption is below the normal range
5152 * @root: the highest ancestor to consider
5153 * @memcg: the memory cgroup to check
5155 * Returns %true if memory consumption of @memcg, and that of all
5156 * configurable ancestors up to @root, is below the normal range.
5158 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5160 if (mem_cgroup_disabled())
5164 * The toplevel group doesn't have a configurable range, so
5165 * it's never low when looked at directly, and it is not
5166 * considered an ancestor when assessing the hierarchy.
5169 if (memcg
== root_mem_cgroup
)
5172 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5175 while (memcg
!= root
) {
5176 memcg
= parent_mem_cgroup(memcg
);
5178 if (memcg
== root_mem_cgroup
)
5181 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5188 * mem_cgroup_try_charge - try charging a page
5189 * @page: page to charge
5190 * @mm: mm context of the victim
5191 * @gfp_mask: reclaim mode
5192 * @memcgp: charged memcg return
5194 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5195 * pages according to @gfp_mask if necessary.
5197 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5198 * Otherwise, an error code is returned.
5200 * After page->mapping has been set up, the caller must finalize the
5201 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5202 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5204 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5205 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
5208 struct mem_cgroup
*memcg
= NULL
;
5209 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5212 if (mem_cgroup_disabled())
5215 if (PageSwapCache(page
)) {
5217 * Every swap fault against a single page tries to charge the
5218 * page, bail as early as possible. shmem_unuse() encounters
5219 * already charged pages, too. The USED bit is protected by
5220 * the page lock, which serializes swap cache removal, which
5221 * in turn serializes uncharging.
5223 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5224 if (page
->mem_cgroup
)
5227 if (do_memsw_account()) {
5228 swp_entry_t ent
= { .val
= page_private(page
), };
5229 unsigned short id
= lookup_swap_cgroup_id(ent
);
5232 memcg
= mem_cgroup_from_id(id
);
5233 if (memcg
&& !css_tryget_online(&memcg
->css
))
5240 memcg
= get_mem_cgroup_from_mm(mm
);
5242 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5244 css_put(&memcg
->css
);
5251 * mem_cgroup_commit_charge - commit a page charge
5252 * @page: page to charge
5253 * @memcg: memcg to charge the page to
5254 * @lrucare: page might be on LRU already
5256 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5257 * after page->mapping has been set up. This must happen atomically
5258 * as part of the page instantiation, i.e. under the page table lock
5259 * for anonymous pages, under the page lock for page and swap cache.
5261 * In addition, the page must not be on the LRU during the commit, to
5262 * prevent racing with task migration. If it might be, use @lrucare.
5264 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5266 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5267 bool lrucare
, bool compound
)
5269 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5271 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5272 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5274 if (mem_cgroup_disabled())
5277 * Swap faults will attempt to charge the same page multiple
5278 * times. But reuse_swap_page() might have removed the page
5279 * from swapcache already, so we can't check PageSwapCache().
5284 commit_charge(page
, memcg
, lrucare
);
5286 local_irq_disable();
5287 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
5288 memcg_check_events(memcg
, page
);
5291 if (do_memsw_account() && PageSwapCache(page
)) {
5292 swp_entry_t entry
= { .val
= page_private(page
) };
5294 * The swap entry might not get freed for a long time,
5295 * let's not wait for it. The page already received a
5296 * memory+swap charge, drop the swap entry duplicate.
5298 mem_cgroup_uncharge_swap(entry
);
5303 * mem_cgroup_cancel_charge - cancel a page charge
5304 * @page: page to charge
5305 * @memcg: memcg to charge the page to
5307 * Cancel a charge transaction started by mem_cgroup_try_charge().
5309 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5312 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5314 if (mem_cgroup_disabled())
5317 * Swap faults will attempt to charge the same page multiple
5318 * times. But reuse_swap_page() might have removed the page
5319 * from swapcache already, so we can't check PageSwapCache().
5324 cancel_charge(memcg
, nr_pages
);
5327 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5328 unsigned long nr_anon
, unsigned long nr_file
,
5329 unsigned long nr_huge
, struct page
*dummy_page
)
5331 unsigned long nr_pages
= nr_anon
+ nr_file
;
5332 unsigned long flags
;
5334 if (!mem_cgroup_is_root(memcg
)) {
5335 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5336 if (do_memsw_account())
5337 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5338 memcg_oom_recover(memcg
);
5341 local_irq_save(flags
);
5342 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5343 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5344 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5345 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5346 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5347 memcg_check_events(memcg
, dummy_page
);
5348 local_irq_restore(flags
);
5350 if (!mem_cgroup_is_root(memcg
))
5351 css_put_many(&memcg
->css
, nr_pages
);
5354 static void uncharge_list(struct list_head
*page_list
)
5356 struct mem_cgroup
*memcg
= NULL
;
5357 unsigned long nr_anon
= 0;
5358 unsigned long nr_file
= 0;
5359 unsigned long nr_huge
= 0;
5360 unsigned long pgpgout
= 0;
5361 struct list_head
*next
;
5364 next
= page_list
->next
;
5366 unsigned int nr_pages
= 1;
5368 page
= list_entry(next
, struct page
, lru
);
5369 next
= page
->lru
.next
;
5371 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5372 VM_BUG_ON_PAGE(page_count(page
), page
);
5374 if (!page
->mem_cgroup
)
5378 * Nobody should be changing or seriously looking at
5379 * page->mem_cgroup at this point, we have fully
5380 * exclusive access to the page.
5383 if (memcg
!= page
->mem_cgroup
) {
5385 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5387 pgpgout
= nr_anon
= nr_file
= nr_huge
= 0;
5389 memcg
= page
->mem_cgroup
;
5392 if (PageTransHuge(page
)) {
5393 nr_pages
<<= compound_order(page
);
5394 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5395 nr_huge
+= nr_pages
;
5399 nr_anon
+= nr_pages
;
5401 nr_file
+= nr_pages
;
5403 page
->mem_cgroup
= NULL
;
5406 } while (next
!= page_list
);
5409 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5414 * mem_cgroup_uncharge - uncharge a page
5415 * @page: page to uncharge
5417 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5418 * mem_cgroup_commit_charge().
5420 void mem_cgroup_uncharge(struct page
*page
)
5422 if (mem_cgroup_disabled())
5425 /* Don't touch page->lru of any random page, pre-check: */
5426 if (!page
->mem_cgroup
)
5429 INIT_LIST_HEAD(&page
->lru
);
5430 uncharge_list(&page
->lru
);
5434 * mem_cgroup_uncharge_list - uncharge a list of page
5435 * @page_list: list of pages to uncharge
5437 * Uncharge a list of pages previously charged with
5438 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5440 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5442 if (mem_cgroup_disabled())
5445 if (!list_empty(page_list
))
5446 uncharge_list(page_list
);
5450 * mem_cgroup_replace_page - migrate a charge to another page
5451 * @oldpage: currently charged page
5452 * @newpage: page to transfer the charge to
5454 * Migrate the charge from @oldpage to @newpage.
5456 * Both pages must be locked, @newpage->mapping must be set up.
5457 * Either or both pages might be on the LRU already.
5459 void mem_cgroup_replace_page(struct page
*oldpage
, struct page
*newpage
)
5461 struct mem_cgroup
*memcg
;
5464 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5465 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5466 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5467 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5470 if (mem_cgroup_disabled())
5473 /* Page cache replacement: new page already charged? */
5474 if (newpage
->mem_cgroup
)
5477 /* Swapcache readahead pages can get replaced before being charged */
5478 memcg
= oldpage
->mem_cgroup
;
5482 lock_page_lru(oldpage
, &isolated
);
5483 oldpage
->mem_cgroup
= NULL
;
5484 unlock_page_lru(oldpage
, isolated
);
5486 commit_charge(newpage
, memcg
, true);
5489 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
5490 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
5492 void sock_update_memcg(struct sock
*sk
)
5494 struct mem_cgroup
*memcg
;
5496 /* Socket cloning can throw us here with sk_cgrp already
5497 * filled. It won't however, necessarily happen from
5498 * process context. So the test for root memcg given
5499 * the current task's memcg won't help us in this case.
5501 * Respecting the original socket's memcg is a better
5502 * decision in this case.
5505 BUG_ON(mem_cgroup_is_root(sk
->sk_memcg
));
5506 css_get(&sk
->sk_memcg
->css
);
5511 memcg
= mem_cgroup_from_task(current
);
5512 if (memcg
== root_mem_cgroup
)
5514 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
5516 if (css_tryget_online(&memcg
->css
))
5517 sk
->sk_memcg
= memcg
;
5521 EXPORT_SYMBOL(sock_update_memcg
);
5523 void sock_release_memcg(struct sock
*sk
)
5525 WARN_ON(!sk
->sk_memcg
);
5526 css_put(&sk
->sk_memcg
->css
);
5530 * mem_cgroup_charge_skmem - charge socket memory
5531 * @memcg: memcg to charge
5532 * @nr_pages: number of pages to charge
5534 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5535 * @memcg's configured limit, %false if the charge had to be forced.
5537 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5539 gfp_t gfp_mask
= GFP_KERNEL
;
5541 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5542 struct page_counter
*fail
;
5544 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
5545 memcg
->tcpmem_pressure
= 0;
5548 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
5549 memcg
->tcpmem_pressure
= 1;
5553 /* Don't block in the packet receive path */
5555 gfp_mask
= GFP_NOWAIT
;
5557 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
5560 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
5565 * mem_cgroup_uncharge_skmem - uncharge socket memory
5566 * @memcg - memcg to uncharge
5567 * @nr_pages - number of pages to uncharge
5569 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5571 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5572 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
5576 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5577 css_put_many(&memcg
->css
, nr_pages
);
5580 static int __init
cgroup_memory(char *s
)
5584 while ((token
= strsep(&s
, ",")) != NULL
) {
5587 if (!strcmp(token
, "nosocket"))
5588 cgroup_memory_nosocket
= true;
5589 if (!strcmp(token
, "nokmem"))
5590 cgroup_memory_nokmem
= true;
5594 __setup("cgroup.memory=", cgroup_memory
);
5597 * subsys_initcall() for memory controller.
5599 * Some parts like hotcpu_notifier() have to be initialized from this context
5600 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5601 * everything that doesn't depend on a specific mem_cgroup structure should
5602 * be initialized from here.
5604 static int __init
mem_cgroup_init(void)
5608 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5610 for_each_possible_cpu(cpu
)
5611 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5614 for_each_node(node
) {
5615 struct mem_cgroup_tree_per_node
*rtpn
;
5618 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5619 node_online(node
) ? node
: NUMA_NO_NODE
);
5621 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5622 struct mem_cgroup_tree_per_zone
*rtpz
;
5624 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
5625 rtpz
->rb_root
= RB_ROOT
;
5626 spin_lock_init(&rtpz
->lock
);
5628 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5633 subsys_initcall(mem_cgroup_init
);
5635 #ifdef CONFIG_MEMCG_SWAP
5637 * mem_cgroup_swapout - transfer a memsw charge to swap
5638 * @page: page whose memsw charge to transfer
5639 * @entry: swap entry to move the charge to
5641 * Transfer the memsw charge of @page to @entry.
5643 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5645 struct mem_cgroup
*memcg
;
5646 unsigned short oldid
;
5648 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5649 VM_BUG_ON_PAGE(page_count(page
), page
);
5651 if (!do_memsw_account())
5654 memcg
= page
->mem_cgroup
;
5656 /* Readahead page, never charged */
5660 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5661 VM_BUG_ON_PAGE(oldid
, page
);
5662 mem_cgroup_swap_statistics(memcg
, true);
5664 page
->mem_cgroup
= NULL
;
5666 if (!mem_cgroup_is_root(memcg
))
5667 page_counter_uncharge(&memcg
->memory
, 1);
5670 * Interrupts should be disabled here because the caller holds the
5671 * mapping->tree_lock lock which is taken with interrupts-off. It is
5672 * important here to have the interrupts disabled because it is the
5673 * only synchronisation we have for udpating the per-CPU variables.
5675 VM_BUG_ON(!irqs_disabled());
5676 mem_cgroup_charge_statistics(memcg
, page
, false, -1);
5677 memcg_check_events(memcg
, page
);
5681 * mem_cgroup_uncharge_swap - uncharge a swap entry
5682 * @entry: swap entry to uncharge
5684 * Drop the memsw charge associated with @entry.
5686 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5688 struct mem_cgroup
*memcg
;
5691 if (!do_memsw_account())
5694 id
= swap_cgroup_record(entry
, 0);
5696 memcg
= mem_cgroup_from_id(id
);
5698 if (!mem_cgroup_is_root(memcg
))
5699 page_counter_uncharge(&memcg
->memsw
, 1);
5700 mem_cgroup_swap_statistics(memcg
, false);
5701 css_put(&memcg
->css
);
5706 /* for remember boot option*/
5707 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5708 static int really_do_swap_account __initdata
= 1;
5710 static int really_do_swap_account __initdata
;
5713 static int __init
enable_swap_account(char *s
)
5715 if (!strcmp(s
, "1"))
5716 really_do_swap_account
= 1;
5717 else if (!strcmp(s
, "0"))
5718 really_do_swap_account
= 0;
5721 __setup("swapaccount=", enable_swap_account
);
5723 static struct cftype memsw_cgroup_files
[] = {
5725 .name
= "memsw.usage_in_bytes",
5726 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5727 .read_u64
= mem_cgroup_read_u64
,
5730 .name
= "memsw.max_usage_in_bytes",
5731 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5732 .write
= mem_cgroup_reset
,
5733 .read_u64
= mem_cgroup_read_u64
,
5736 .name
= "memsw.limit_in_bytes",
5737 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5738 .write
= mem_cgroup_write
,
5739 .read_u64
= mem_cgroup_read_u64
,
5742 .name
= "memsw.failcnt",
5743 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5744 .write
= mem_cgroup_reset
,
5745 .read_u64
= mem_cgroup_read_u64
,
5747 { }, /* terminate */
5750 static int __init
mem_cgroup_swap_init(void)
5752 if (!mem_cgroup_disabled() && really_do_swap_account
) {
5753 do_swap_account
= 1;
5754 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
5755 memsw_cgroup_files
));
5759 subsys_initcall(mem_cgroup_swap_init
);
5761 #endif /* CONFIG_MEMCG_SWAP */