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>
69 #include <net/tcp_memcontrol.h>
72 #include <asm/uaccess.h>
74 #include <trace/events/vmscan.h>
76 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
77 EXPORT_SYMBOL(memory_cgrp_subsys
);
79 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
81 #define MEM_CGROUP_RECLAIM_RETRIES 5
83 /* Socket memory accounting disabled? */
84 static bool cgroup_memory_nosocket
;
86 /* Whether the swap controller is active */
87 #ifdef CONFIG_MEMCG_SWAP
88 int do_swap_account __read_mostly
;
90 #define do_swap_account 0
93 /* Whether legacy memory+swap accounting is active */
94 static bool do_memsw_account(void)
96 return !cgroup_subsys_on_dfl(memory_cgrp_subsys
) && do_swap_account
;
99 static const char * const mem_cgroup_stat_names
[] = {
109 static const char * const mem_cgroup_events_names
[] = {
116 static const char * const mem_cgroup_lru_names
[] = {
124 #define THRESHOLDS_EVENTS_TARGET 128
125 #define SOFTLIMIT_EVENTS_TARGET 1024
126 #define NUMAINFO_EVENTS_TARGET 1024
129 * Cgroups above their limits are maintained in a RB-Tree, independent of
130 * their hierarchy representation
133 struct mem_cgroup_tree_per_zone
{
134 struct rb_root rb_root
;
138 struct mem_cgroup_tree_per_node
{
139 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
142 struct mem_cgroup_tree
{
143 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
146 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
149 struct mem_cgroup_eventfd_list
{
150 struct list_head list
;
151 struct eventfd_ctx
*eventfd
;
155 * cgroup_event represents events which userspace want to receive.
157 struct mem_cgroup_event
{
159 * memcg which the event belongs to.
161 struct mem_cgroup
*memcg
;
163 * eventfd to signal userspace about the event.
165 struct eventfd_ctx
*eventfd
;
167 * Each of these stored in a list by the cgroup.
169 struct list_head list
;
171 * register_event() callback will be used to add new userspace
172 * waiter for changes related to this event. Use eventfd_signal()
173 * on eventfd to send notification to userspace.
175 int (*register_event
)(struct mem_cgroup
*memcg
,
176 struct eventfd_ctx
*eventfd
, const char *args
);
178 * unregister_event() callback will be called when userspace closes
179 * the eventfd or on cgroup removing. This callback must be set,
180 * if you want provide notification functionality.
182 void (*unregister_event
)(struct mem_cgroup
*memcg
,
183 struct eventfd_ctx
*eventfd
);
185 * All fields below needed to unregister event when
186 * userspace closes eventfd.
189 wait_queue_head_t
*wqh
;
191 struct work_struct remove
;
194 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
195 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
197 /* Stuffs for move charges at task migration. */
199 * Types of charges to be moved.
201 #define MOVE_ANON 0x1U
202 #define MOVE_FILE 0x2U
203 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
205 /* "mc" and its members are protected by cgroup_mutex */
206 static struct move_charge_struct
{
207 spinlock_t lock
; /* for from, to */
208 struct mem_cgroup
*from
;
209 struct mem_cgroup
*to
;
211 unsigned long precharge
;
212 unsigned long moved_charge
;
213 unsigned long moved_swap
;
214 struct task_struct
*moving_task
; /* a task moving charges */
215 wait_queue_head_t waitq
; /* a waitq for other context */
217 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
218 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
222 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
223 * limit reclaim to prevent infinite loops, if they ever occur.
225 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
226 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
229 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
230 MEM_CGROUP_CHARGE_TYPE_ANON
,
231 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
232 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
236 /* for encoding cft->private value on file */
244 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
245 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
246 #define MEMFILE_ATTR(val) ((val) & 0xffff)
247 /* Used for OOM nofiier */
248 #define OOM_CONTROL (0)
251 * The memcg_create_mutex will be held whenever a new cgroup is created.
252 * As a consequence, any change that needs to protect against new child cgroups
253 * appearing has to hold it as well.
255 static DEFINE_MUTEX(memcg_create_mutex
);
257 /* Some nice accessors for the vmpressure. */
258 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
261 memcg
= root_mem_cgroup
;
262 return &memcg
->vmpressure
;
265 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
267 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
270 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
272 return (memcg
== root_mem_cgroup
);
276 * We restrict the id in the range of [1, 65535], so it can fit into
279 #define MEM_CGROUP_ID_MAX USHRT_MAX
281 static inline unsigned short mem_cgroup_id(struct mem_cgroup
*memcg
)
283 return memcg
->css
.id
;
287 * A helper function to get mem_cgroup from ID. must be called under
288 * rcu_read_lock(). The caller is responsible for calling
289 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
290 * refcnt from swap can be called against removed memcg.)
292 static inline struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
294 struct cgroup_subsys_state
*css
;
296 css
= css_from_id(id
, &memory_cgrp_subsys
);
297 return mem_cgroup_from_css(css
);
300 #ifdef CONFIG_MEMCG_KMEM
302 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
303 * The main reason for not using cgroup id for this:
304 * this works better in sparse environments, where we have a lot of memcgs,
305 * but only a few kmem-limited. Or also, if we have, for instance, 200
306 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
307 * 200 entry array for that.
309 * The current size of the caches array is stored in memcg_nr_cache_ids. It
310 * will double each time we have to increase it.
312 static DEFINE_IDA(memcg_cache_ida
);
313 int memcg_nr_cache_ids
;
315 /* Protects memcg_nr_cache_ids */
316 static DECLARE_RWSEM(memcg_cache_ids_sem
);
318 void memcg_get_cache_ids(void)
320 down_read(&memcg_cache_ids_sem
);
323 void memcg_put_cache_ids(void)
325 up_read(&memcg_cache_ids_sem
);
329 * MIN_SIZE is different than 1, because we would like to avoid going through
330 * the alloc/free process all the time. In a small machine, 4 kmem-limited
331 * cgroups is a reasonable guess. In the future, it could be a parameter or
332 * tunable, but that is strictly not necessary.
334 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
335 * this constant directly from cgroup, but it is understandable that this is
336 * better kept as an internal representation in cgroup.c. In any case, the
337 * cgrp_id space is not getting any smaller, and we don't have to necessarily
338 * increase ours as well if it increases.
340 #define MEMCG_CACHES_MIN_SIZE 4
341 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
344 * A lot of the calls to the cache allocation functions are expected to be
345 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
346 * conditional to this static branch, we'll have to allow modules that does
347 * kmem_cache_alloc and the such to see this symbol as well
349 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
350 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
352 #endif /* CONFIG_MEMCG_KMEM */
354 static struct mem_cgroup_per_zone
*
355 mem_cgroup_zone_zoneinfo(struct mem_cgroup
*memcg
, struct zone
*zone
)
357 int nid
= zone_to_nid(zone
);
358 int zid
= zone_idx(zone
);
360 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
364 * mem_cgroup_css_from_page - css of the memcg associated with a page
365 * @page: page of interest
367 * If memcg is bound to the default hierarchy, css of the memcg associated
368 * with @page is returned. The returned css remains associated with @page
369 * until it is released.
371 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
374 * XXX: The above description of behavior on the default hierarchy isn't
375 * strictly true yet as replace_page_cache_page() can modify the
376 * association before @page is released even on the default hierarchy;
377 * however, the current and planned usages don't mix the the two functions
378 * and replace_page_cache_page() will soon be updated to make the invariant
381 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
383 struct mem_cgroup
*memcg
;
387 memcg
= page
->mem_cgroup
;
389 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
390 memcg
= root_mem_cgroup
;
397 * page_cgroup_ino - return inode number of the memcg a page is charged to
400 * Look up the closest online ancestor of the memory cgroup @page is charged to
401 * and return its inode number or 0 if @page is not charged to any cgroup. It
402 * is safe to call this function without holding a reference to @page.
404 * Note, this function is inherently racy, because there is nothing to prevent
405 * the cgroup inode from getting torn down and potentially reallocated a moment
406 * after page_cgroup_ino() returns, so it only should be used by callers that
407 * do not care (such as procfs interfaces).
409 ino_t
page_cgroup_ino(struct page
*page
)
411 struct mem_cgroup
*memcg
;
412 unsigned long ino
= 0;
415 memcg
= READ_ONCE(page
->mem_cgroup
);
416 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
417 memcg
= parent_mem_cgroup(memcg
);
419 ino
= cgroup_ino(memcg
->css
.cgroup
);
424 static struct mem_cgroup_per_zone
*
425 mem_cgroup_page_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
427 int nid
= page_to_nid(page
);
428 int zid
= page_zonenum(page
);
430 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
433 static struct mem_cgroup_tree_per_zone
*
434 soft_limit_tree_node_zone(int nid
, int zid
)
436 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
439 static struct mem_cgroup_tree_per_zone
*
440 soft_limit_tree_from_page(struct page
*page
)
442 int nid
= page_to_nid(page
);
443 int zid
= page_zonenum(page
);
445 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
448 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone
*mz
,
449 struct mem_cgroup_tree_per_zone
*mctz
,
450 unsigned long new_usage_in_excess
)
452 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
453 struct rb_node
*parent
= NULL
;
454 struct mem_cgroup_per_zone
*mz_node
;
459 mz
->usage_in_excess
= new_usage_in_excess
;
460 if (!mz
->usage_in_excess
)
464 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
466 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
469 * We can't avoid mem cgroups that are over their soft
470 * limit by the same amount
472 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
475 rb_link_node(&mz
->tree_node
, parent
, p
);
476 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
480 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
481 struct mem_cgroup_tree_per_zone
*mctz
)
485 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
489 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
490 struct mem_cgroup_tree_per_zone
*mctz
)
494 spin_lock_irqsave(&mctz
->lock
, flags
);
495 __mem_cgroup_remove_exceeded(mz
, mctz
);
496 spin_unlock_irqrestore(&mctz
->lock
, flags
);
499 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
501 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
502 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
503 unsigned long excess
= 0;
505 if (nr_pages
> soft_limit
)
506 excess
= nr_pages
- soft_limit
;
511 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
513 unsigned long excess
;
514 struct mem_cgroup_per_zone
*mz
;
515 struct mem_cgroup_tree_per_zone
*mctz
;
517 mctz
= soft_limit_tree_from_page(page
);
519 * Necessary to update all ancestors when hierarchy is used.
520 * because their event counter is not touched.
522 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
523 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
524 excess
= soft_limit_excess(memcg
);
526 * We have to update the tree if mz is on RB-tree or
527 * mem is over its softlimit.
529 if (excess
|| mz
->on_tree
) {
532 spin_lock_irqsave(&mctz
->lock
, flags
);
533 /* if on-tree, remove it */
535 __mem_cgroup_remove_exceeded(mz
, mctz
);
537 * Insert again. mz->usage_in_excess will be updated.
538 * If excess is 0, no tree ops.
540 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
541 spin_unlock_irqrestore(&mctz
->lock
, flags
);
546 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
548 struct mem_cgroup_tree_per_zone
*mctz
;
549 struct mem_cgroup_per_zone
*mz
;
553 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
554 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
555 mctz
= soft_limit_tree_node_zone(nid
, zid
);
556 mem_cgroup_remove_exceeded(mz
, mctz
);
561 static struct mem_cgroup_per_zone
*
562 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
564 struct rb_node
*rightmost
= NULL
;
565 struct mem_cgroup_per_zone
*mz
;
569 rightmost
= rb_last(&mctz
->rb_root
);
571 goto done
; /* Nothing to reclaim from */
573 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
575 * Remove the node now but someone else can add it back,
576 * we will to add it back at the end of reclaim to its correct
577 * position in the tree.
579 __mem_cgroup_remove_exceeded(mz
, mctz
);
580 if (!soft_limit_excess(mz
->memcg
) ||
581 !css_tryget_online(&mz
->memcg
->css
))
587 static struct mem_cgroup_per_zone
*
588 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
590 struct mem_cgroup_per_zone
*mz
;
592 spin_lock_irq(&mctz
->lock
);
593 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
594 spin_unlock_irq(&mctz
->lock
);
599 * Return page count for single (non recursive) @memcg.
601 * Implementation Note: reading percpu statistics for memcg.
603 * Both of vmstat[] and percpu_counter has threshold and do periodic
604 * synchronization to implement "quick" read. There are trade-off between
605 * reading cost and precision of value. Then, we may have a chance to implement
606 * a periodic synchronization of counter in memcg's counter.
608 * But this _read() function is used for user interface now. The user accounts
609 * memory usage by memory cgroup and he _always_ requires exact value because
610 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
611 * have to visit all online cpus and make sum. So, for now, unnecessary
612 * synchronization is not implemented. (just implemented for cpu hotplug)
614 * If there are kernel internal actions which can make use of some not-exact
615 * value, and reading all cpu value can be performance bottleneck in some
616 * common workload, threshold and synchronization as vmstat[] should be
620 mem_cgroup_read_stat(struct mem_cgroup
*memcg
, enum mem_cgroup_stat_index idx
)
625 /* Per-cpu values can be negative, use a signed accumulator */
626 for_each_possible_cpu(cpu
)
627 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
629 * Summing races with updates, so val may be negative. Avoid exposing
630 * transient negative values.
637 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
638 enum mem_cgroup_events_index idx
)
640 unsigned long val
= 0;
643 for_each_possible_cpu(cpu
)
644 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
648 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
653 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
654 * counted as CACHE even if it's on ANON LRU.
657 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
660 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
663 if (PageTransHuge(page
))
664 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
667 /* pagein of a big page is an event. So, ignore page size */
669 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
671 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
672 nr_pages
= -nr_pages
; /* for event */
675 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
678 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
680 unsigned int lru_mask
)
682 unsigned long nr
= 0;
685 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
687 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
688 struct mem_cgroup_per_zone
*mz
;
692 if (!(BIT(lru
) & lru_mask
))
694 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
695 nr
+= mz
->lru_size
[lru
];
701 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
702 unsigned int lru_mask
)
704 unsigned long nr
= 0;
707 for_each_node_state(nid
, N_MEMORY
)
708 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
712 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
713 enum mem_cgroup_events_target target
)
715 unsigned long val
, next
;
717 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
718 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
719 /* from time_after() in jiffies.h */
720 if ((long)next
- (long)val
< 0) {
722 case MEM_CGROUP_TARGET_THRESH
:
723 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
725 case MEM_CGROUP_TARGET_SOFTLIMIT
:
726 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
728 case MEM_CGROUP_TARGET_NUMAINFO
:
729 next
= val
+ NUMAINFO_EVENTS_TARGET
;
734 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
741 * Check events in order.
744 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
746 /* threshold event is triggered in finer grain than soft limit */
747 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
748 MEM_CGROUP_TARGET_THRESH
))) {
750 bool do_numainfo __maybe_unused
;
752 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
753 MEM_CGROUP_TARGET_SOFTLIMIT
);
755 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
756 MEM_CGROUP_TARGET_NUMAINFO
);
758 mem_cgroup_threshold(memcg
);
759 if (unlikely(do_softlimit
))
760 mem_cgroup_update_tree(memcg
, page
);
762 if (unlikely(do_numainfo
))
763 atomic_inc(&memcg
->numainfo_events
);
768 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
771 * mm_update_next_owner() may clear mm->owner to NULL
772 * if it races with swapoff, page migration, etc.
773 * So this can be called with p == NULL.
778 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
780 EXPORT_SYMBOL(mem_cgroup_from_task
);
782 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
784 struct mem_cgroup
*memcg
= NULL
;
789 * Page cache insertions can happen withou an
790 * actual mm context, e.g. during disk probing
791 * on boot, loopback IO, acct() writes etc.
794 memcg
= root_mem_cgroup
;
796 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
797 if (unlikely(!memcg
))
798 memcg
= root_mem_cgroup
;
800 } while (!css_tryget_online(&memcg
->css
));
806 * mem_cgroup_iter - iterate over memory cgroup hierarchy
807 * @root: hierarchy root
808 * @prev: previously returned memcg, NULL on first invocation
809 * @reclaim: cookie for shared reclaim walks, NULL for full walks
811 * Returns references to children of the hierarchy below @root, or
812 * @root itself, or %NULL after a full round-trip.
814 * Caller must pass the return value in @prev on subsequent
815 * invocations for reference counting, or use mem_cgroup_iter_break()
816 * to cancel a hierarchy walk before the round-trip is complete.
818 * Reclaimers can specify a zone and a priority level in @reclaim to
819 * divide up the memcgs in the hierarchy among all concurrent
820 * reclaimers operating on the same zone and priority.
822 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
823 struct mem_cgroup
*prev
,
824 struct mem_cgroup_reclaim_cookie
*reclaim
)
826 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
827 struct cgroup_subsys_state
*css
= NULL
;
828 struct mem_cgroup
*memcg
= NULL
;
829 struct mem_cgroup
*pos
= NULL
;
831 if (mem_cgroup_disabled())
835 root
= root_mem_cgroup
;
837 if (prev
&& !reclaim
)
840 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
849 struct mem_cgroup_per_zone
*mz
;
851 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
852 iter
= &mz
->iter
[reclaim
->priority
];
854 if (prev
&& reclaim
->generation
!= iter
->generation
)
858 pos
= READ_ONCE(iter
->position
);
859 if (!pos
|| css_tryget(&pos
->css
))
862 * css reference reached zero, so iter->position will
863 * be cleared by ->css_released. However, we should not
864 * rely on this happening soon, because ->css_released
865 * is called from a work queue, and by busy-waiting we
866 * might block it. So we clear iter->position right
869 (void)cmpxchg(&iter
->position
, pos
, NULL
);
877 css
= css_next_descendant_pre(css
, &root
->css
);
880 * Reclaimers share the hierarchy walk, and a
881 * new one might jump in right at the end of
882 * the hierarchy - make sure they see at least
883 * one group and restart from the beginning.
891 * Verify the css and acquire a reference. The root
892 * is provided by the caller, so we know it's alive
893 * and kicking, and don't take an extra reference.
895 memcg
= mem_cgroup_from_css(css
);
897 if (css
== &root
->css
)
900 if (css_tryget(css
)) {
902 * Make sure the memcg is initialized:
903 * mem_cgroup_css_online() orders the the
904 * initialization against setting the flag.
906 if (smp_load_acquire(&memcg
->initialized
))
917 * The position could have already been updated by a competing
918 * thread, so check that the value hasn't changed since we read
919 * it to avoid reclaiming from the same cgroup twice.
921 (void)cmpxchg(&iter
->position
, pos
, memcg
);
929 reclaim
->generation
= iter
->generation
;
935 if (prev
&& prev
!= root
)
942 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
943 * @root: hierarchy root
944 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
946 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
947 struct mem_cgroup
*prev
)
950 root
= root_mem_cgroup
;
951 if (prev
&& prev
!= root
)
955 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
957 struct mem_cgroup
*memcg
= dead_memcg
;
958 struct mem_cgroup_reclaim_iter
*iter
;
959 struct mem_cgroup_per_zone
*mz
;
963 while ((memcg
= parent_mem_cgroup(memcg
))) {
965 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
966 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
967 for (i
= 0; i
<= DEF_PRIORITY
; i
++) {
969 cmpxchg(&iter
->position
,
978 * Iteration constructs for visiting all cgroups (under a tree). If
979 * loops are exited prematurely (break), mem_cgroup_iter_break() must
980 * be used for reference counting.
982 #define for_each_mem_cgroup_tree(iter, root) \
983 for (iter = mem_cgroup_iter(root, NULL, NULL); \
985 iter = mem_cgroup_iter(root, iter, NULL))
987 #define for_each_mem_cgroup(iter) \
988 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
990 iter = mem_cgroup_iter(NULL, iter, NULL))
993 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
994 * @zone: zone of the wanted lruvec
995 * @memcg: memcg of the wanted lruvec
997 * Returns the lru list vector holding pages for the given @zone and
998 * @mem. This can be the global zone lruvec, if the memory controller
1001 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1002 struct mem_cgroup
*memcg
)
1004 struct mem_cgroup_per_zone
*mz
;
1005 struct lruvec
*lruvec
;
1007 if (mem_cgroup_disabled()) {
1008 lruvec
= &zone
->lruvec
;
1012 mz
= mem_cgroup_zone_zoneinfo(memcg
, zone
);
1013 lruvec
= &mz
->lruvec
;
1016 * Since a node can be onlined after the mem_cgroup was created,
1017 * we have to be prepared to initialize lruvec->zone here;
1018 * and if offlined then reonlined, we need to reinitialize it.
1020 if (unlikely(lruvec
->zone
!= zone
))
1021 lruvec
->zone
= zone
;
1026 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1028 * @zone: zone of the page
1030 * This function is only safe when following the LRU page isolation
1031 * and putback protocol: the LRU lock must be held, and the page must
1032 * either be PageLRU() or the caller must have isolated/allocated it.
1034 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1036 struct mem_cgroup_per_zone
*mz
;
1037 struct mem_cgroup
*memcg
;
1038 struct lruvec
*lruvec
;
1040 if (mem_cgroup_disabled()) {
1041 lruvec
= &zone
->lruvec
;
1045 memcg
= page
->mem_cgroup
;
1047 * Swapcache readahead pages are added to the LRU - and
1048 * possibly migrated - before they are charged.
1051 memcg
= root_mem_cgroup
;
1053 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1054 lruvec
= &mz
->lruvec
;
1057 * Since a node can be onlined after the mem_cgroup was created,
1058 * we have to be prepared to initialize lruvec->zone here;
1059 * and if offlined then reonlined, we need to reinitialize it.
1061 if (unlikely(lruvec
->zone
!= zone
))
1062 lruvec
->zone
= zone
;
1067 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1068 * @lruvec: mem_cgroup per zone lru vector
1069 * @lru: index of lru list the page is sitting on
1070 * @nr_pages: positive when adding or negative when removing
1072 * This function must be called when a page is added to or removed from an
1075 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1078 struct mem_cgroup_per_zone
*mz
;
1079 unsigned long *lru_size
;
1081 if (mem_cgroup_disabled())
1084 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1085 lru_size
= mz
->lru_size
+ lru
;
1086 *lru_size
+= nr_pages
;
1087 VM_BUG_ON((long)(*lru_size
) < 0);
1090 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1092 struct mem_cgroup
*task_memcg
;
1093 struct task_struct
*p
;
1096 p
= find_lock_task_mm(task
);
1098 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1102 * All threads may have already detached their mm's, but the oom
1103 * killer still needs to detect if they have already been oom
1104 * killed to prevent needlessly killing additional tasks.
1107 task_memcg
= mem_cgroup_from_task(task
);
1108 css_get(&task_memcg
->css
);
1111 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1112 css_put(&task_memcg
->css
);
1117 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1118 * @memcg: the memory cgroup
1120 * Returns the maximum amount of memory @mem can be charged with, in
1123 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1125 unsigned long margin
= 0;
1126 unsigned long count
;
1127 unsigned long limit
;
1129 count
= page_counter_read(&memcg
->memory
);
1130 limit
= READ_ONCE(memcg
->memory
.limit
);
1132 margin
= limit
- count
;
1134 if (do_memsw_account()) {
1135 count
= page_counter_read(&memcg
->memsw
);
1136 limit
= READ_ONCE(memcg
->memsw
.limit
);
1138 margin
= min(margin
, limit
- count
);
1145 * A routine for checking "mem" is under move_account() or not.
1147 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1148 * moving cgroups. This is for waiting at high-memory pressure
1151 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1153 struct mem_cgroup
*from
;
1154 struct mem_cgroup
*to
;
1157 * Unlike task_move routines, we access mc.to, mc.from not under
1158 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1160 spin_lock(&mc
.lock
);
1166 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1167 mem_cgroup_is_descendant(to
, memcg
);
1169 spin_unlock(&mc
.lock
);
1173 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1175 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1176 if (mem_cgroup_under_move(memcg
)) {
1178 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1179 /* moving charge context might have finished. */
1182 finish_wait(&mc
.waitq
, &wait
);
1189 #define K(x) ((x) << (PAGE_SHIFT-10))
1191 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1192 * @memcg: The memory cgroup that went over limit
1193 * @p: Task that is going to be killed
1195 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1198 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1200 /* oom_info_lock ensures that parallel ooms do not interleave */
1201 static DEFINE_MUTEX(oom_info_lock
);
1202 struct mem_cgroup
*iter
;
1205 mutex_lock(&oom_info_lock
);
1209 pr_info("Task in ");
1210 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1211 pr_cont(" killed as a result of limit of ");
1213 pr_info("Memory limit reached of cgroup ");
1216 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1221 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1222 K((u64
)page_counter_read(&memcg
->memory
)),
1223 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1224 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1225 K((u64
)page_counter_read(&memcg
->memsw
)),
1226 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1227 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1228 K((u64
)page_counter_read(&memcg
->kmem
)),
1229 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1231 for_each_mem_cgroup_tree(iter
, memcg
) {
1232 pr_info("Memory cgroup stats for ");
1233 pr_cont_cgroup_path(iter
->css
.cgroup
);
1236 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1237 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
1239 pr_cont(" %s:%luKB", mem_cgroup_stat_names
[i
],
1240 K(mem_cgroup_read_stat(iter
, i
)));
1243 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1244 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1245 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1249 mutex_unlock(&oom_info_lock
);
1253 * This function returns the number of memcg under hierarchy tree. Returns
1254 * 1(self count) if no children.
1256 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1259 struct mem_cgroup
*iter
;
1261 for_each_mem_cgroup_tree(iter
, memcg
)
1267 * Return the memory (and swap, if configured) limit for a memcg.
1269 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1271 unsigned long limit
;
1273 limit
= memcg
->memory
.limit
;
1274 if (mem_cgroup_swappiness(memcg
)) {
1275 unsigned long memsw_limit
;
1277 memsw_limit
= memcg
->memsw
.limit
;
1278 limit
= min(limit
+ total_swap_pages
, memsw_limit
);
1283 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1286 struct oom_control oc
= {
1289 .gfp_mask
= gfp_mask
,
1292 struct mem_cgroup
*iter
;
1293 unsigned long chosen_points
= 0;
1294 unsigned long totalpages
;
1295 unsigned int points
= 0;
1296 struct task_struct
*chosen
= NULL
;
1298 mutex_lock(&oom_lock
);
1301 * If current has a pending SIGKILL or is exiting, then automatically
1302 * select it. The goal is to allow it to allocate so that it may
1303 * quickly exit and free its memory.
1305 if (fatal_signal_pending(current
) || task_will_free_mem(current
)) {
1306 mark_oom_victim(current
);
1310 check_panic_on_oom(&oc
, CONSTRAINT_MEMCG
, memcg
);
1311 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1312 for_each_mem_cgroup_tree(iter
, memcg
) {
1313 struct css_task_iter it
;
1314 struct task_struct
*task
;
1316 css_task_iter_start(&iter
->css
, &it
);
1317 while ((task
= css_task_iter_next(&it
))) {
1318 switch (oom_scan_process_thread(&oc
, task
, totalpages
)) {
1319 case OOM_SCAN_SELECT
:
1321 put_task_struct(chosen
);
1323 chosen_points
= ULONG_MAX
;
1324 get_task_struct(chosen
);
1326 case OOM_SCAN_CONTINUE
:
1328 case OOM_SCAN_ABORT
:
1329 css_task_iter_end(&it
);
1330 mem_cgroup_iter_break(memcg
, iter
);
1332 put_task_struct(chosen
);
1337 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1338 if (!points
|| points
< chosen_points
)
1340 /* Prefer thread group leaders for display purposes */
1341 if (points
== chosen_points
&&
1342 thread_group_leader(chosen
))
1346 put_task_struct(chosen
);
1348 chosen_points
= points
;
1349 get_task_struct(chosen
);
1351 css_task_iter_end(&it
);
1355 points
= chosen_points
* 1000 / totalpages
;
1356 oom_kill_process(&oc
, chosen
, points
, totalpages
, memcg
,
1357 "Memory cgroup out of memory");
1360 mutex_unlock(&oom_lock
);
1363 #if MAX_NUMNODES > 1
1366 * test_mem_cgroup_node_reclaimable
1367 * @memcg: the target memcg
1368 * @nid: the node ID to be checked.
1369 * @noswap : specify true here if the user wants flle only information.
1371 * This function returns whether the specified memcg contains any
1372 * reclaimable pages on a node. Returns true if there are any reclaimable
1373 * pages in the node.
1375 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1376 int nid
, bool noswap
)
1378 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1380 if (noswap
|| !total_swap_pages
)
1382 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1389 * Always updating the nodemask is not very good - even if we have an empty
1390 * list or the wrong list here, we can start from some node and traverse all
1391 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1394 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1398 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1399 * pagein/pageout changes since the last update.
1401 if (!atomic_read(&memcg
->numainfo_events
))
1403 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1406 /* make a nodemask where this memcg uses memory from */
1407 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1409 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1411 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1412 node_clear(nid
, memcg
->scan_nodes
);
1415 atomic_set(&memcg
->numainfo_events
, 0);
1416 atomic_set(&memcg
->numainfo_updating
, 0);
1420 * Selecting a node where we start reclaim from. Because what we need is just
1421 * reducing usage counter, start from anywhere is O,K. Considering
1422 * memory reclaim from current node, there are pros. and cons.
1424 * Freeing memory from current node means freeing memory from a node which
1425 * we'll use or we've used. So, it may make LRU bad. And if several threads
1426 * hit limits, it will see a contention on a node. But freeing from remote
1427 * node means more costs for memory reclaim because of memory latency.
1429 * Now, we use round-robin. Better algorithm is welcomed.
1431 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1435 mem_cgroup_may_update_nodemask(memcg
);
1436 node
= memcg
->last_scanned_node
;
1438 node
= next_node(node
, memcg
->scan_nodes
);
1439 if (node
== MAX_NUMNODES
)
1440 node
= first_node(memcg
->scan_nodes
);
1442 * We call this when we hit limit, not when pages are added to LRU.
1443 * No LRU may hold pages because all pages are UNEVICTABLE or
1444 * memcg is too small and all pages are not on LRU. In that case,
1445 * we use curret node.
1447 if (unlikely(node
== MAX_NUMNODES
))
1448 node
= numa_node_id();
1450 memcg
->last_scanned_node
= node
;
1454 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1460 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1463 unsigned long *total_scanned
)
1465 struct mem_cgroup
*victim
= NULL
;
1468 unsigned long excess
;
1469 unsigned long nr_scanned
;
1470 struct mem_cgroup_reclaim_cookie reclaim
= {
1475 excess
= soft_limit_excess(root_memcg
);
1478 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1483 * If we have not been able to reclaim
1484 * anything, it might because there are
1485 * no reclaimable pages under this hierarchy
1490 * We want to do more targeted reclaim.
1491 * excess >> 2 is not to excessive so as to
1492 * reclaim too much, nor too less that we keep
1493 * coming back to reclaim from this cgroup
1495 if (total
>= (excess
>> 2) ||
1496 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1501 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1503 *total_scanned
+= nr_scanned
;
1504 if (!soft_limit_excess(root_memcg
))
1507 mem_cgroup_iter_break(root_memcg
, victim
);
1511 #ifdef CONFIG_LOCKDEP
1512 static struct lockdep_map memcg_oom_lock_dep_map
= {
1513 .name
= "memcg_oom_lock",
1517 static DEFINE_SPINLOCK(memcg_oom_lock
);
1520 * Check OOM-Killer is already running under our hierarchy.
1521 * If someone is running, return false.
1523 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1525 struct mem_cgroup
*iter
, *failed
= NULL
;
1527 spin_lock(&memcg_oom_lock
);
1529 for_each_mem_cgroup_tree(iter
, memcg
) {
1530 if (iter
->oom_lock
) {
1532 * this subtree of our hierarchy is already locked
1533 * so we cannot give a lock.
1536 mem_cgroup_iter_break(memcg
, iter
);
1539 iter
->oom_lock
= true;
1544 * OK, we failed to lock the whole subtree so we have
1545 * to clean up what we set up to the failing subtree
1547 for_each_mem_cgroup_tree(iter
, memcg
) {
1548 if (iter
== failed
) {
1549 mem_cgroup_iter_break(memcg
, iter
);
1552 iter
->oom_lock
= false;
1555 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1557 spin_unlock(&memcg_oom_lock
);
1562 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1564 struct mem_cgroup
*iter
;
1566 spin_lock(&memcg_oom_lock
);
1567 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1568 for_each_mem_cgroup_tree(iter
, memcg
)
1569 iter
->oom_lock
= false;
1570 spin_unlock(&memcg_oom_lock
);
1573 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1575 struct mem_cgroup
*iter
;
1577 spin_lock(&memcg_oom_lock
);
1578 for_each_mem_cgroup_tree(iter
, memcg
)
1580 spin_unlock(&memcg_oom_lock
);
1583 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1585 struct mem_cgroup
*iter
;
1588 * When a new child is created while the hierarchy is under oom,
1589 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1591 spin_lock(&memcg_oom_lock
);
1592 for_each_mem_cgroup_tree(iter
, memcg
)
1593 if (iter
->under_oom
> 0)
1595 spin_unlock(&memcg_oom_lock
);
1598 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1600 struct oom_wait_info
{
1601 struct mem_cgroup
*memcg
;
1605 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1606 unsigned mode
, int sync
, void *arg
)
1608 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1609 struct mem_cgroup
*oom_wait_memcg
;
1610 struct oom_wait_info
*oom_wait_info
;
1612 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1613 oom_wait_memcg
= oom_wait_info
->memcg
;
1615 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1616 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1618 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1621 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1624 * For the following lockless ->under_oom test, the only required
1625 * guarantee is that it must see the state asserted by an OOM when
1626 * this function is called as a result of userland actions
1627 * triggered by the notification of the OOM. This is trivially
1628 * achieved by invoking mem_cgroup_mark_under_oom() before
1629 * triggering notification.
1631 if (memcg
&& memcg
->under_oom
)
1632 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1635 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1637 if (!current
->memcg_may_oom
)
1640 * We are in the middle of the charge context here, so we
1641 * don't want to block when potentially sitting on a callstack
1642 * that holds all kinds of filesystem and mm locks.
1644 * Also, the caller may handle a failed allocation gracefully
1645 * (like optional page cache readahead) and so an OOM killer
1646 * invocation might not even be necessary.
1648 * That's why we don't do anything here except remember the
1649 * OOM context and then deal with it at the end of the page
1650 * fault when the stack is unwound, the locks are released,
1651 * and when we know whether the fault was overall successful.
1653 css_get(&memcg
->css
);
1654 current
->memcg_in_oom
= memcg
;
1655 current
->memcg_oom_gfp_mask
= mask
;
1656 current
->memcg_oom_order
= order
;
1660 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1661 * @handle: actually kill/wait or just clean up the OOM state
1663 * This has to be called at the end of a page fault if the memcg OOM
1664 * handler was enabled.
1666 * Memcg supports userspace OOM handling where failed allocations must
1667 * sleep on a waitqueue until the userspace task resolves the
1668 * situation. Sleeping directly in the charge context with all kinds
1669 * of locks held is not a good idea, instead we remember an OOM state
1670 * in the task and mem_cgroup_oom_synchronize() has to be called at
1671 * the end of the page fault to complete the OOM handling.
1673 * Returns %true if an ongoing memcg OOM situation was detected and
1674 * completed, %false otherwise.
1676 bool mem_cgroup_oom_synchronize(bool handle
)
1678 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1679 struct oom_wait_info owait
;
1682 /* OOM is global, do not handle */
1686 if (!handle
|| oom_killer_disabled
)
1689 owait
.memcg
= memcg
;
1690 owait
.wait
.flags
= 0;
1691 owait
.wait
.func
= memcg_oom_wake_function
;
1692 owait
.wait
.private = current
;
1693 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1695 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1696 mem_cgroup_mark_under_oom(memcg
);
1698 locked
= mem_cgroup_oom_trylock(memcg
);
1701 mem_cgroup_oom_notify(memcg
);
1703 if (locked
&& !memcg
->oom_kill_disable
) {
1704 mem_cgroup_unmark_under_oom(memcg
);
1705 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1706 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1707 current
->memcg_oom_order
);
1710 mem_cgroup_unmark_under_oom(memcg
);
1711 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1715 mem_cgroup_oom_unlock(memcg
);
1717 * There is no guarantee that an OOM-lock contender
1718 * sees the wakeups triggered by the OOM kill
1719 * uncharges. Wake any sleepers explicitely.
1721 memcg_oom_recover(memcg
);
1724 current
->memcg_in_oom
= NULL
;
1725 css_put(&memcg
->css
);
1730 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1731 * @page: page that is going to change accounted state
1733 * This function must mark the beginning of an accounted page state
1734 * change to prevent double accounting when the page is concurrently
1735 * being moved to another memcg:
1737 * memcg = mem_cgroup_begin_page_stat(page);
1738 * if (TestClearPageState(page))
1739 * mem_cgroup_update_page_stat(memcg, state, -1);
1740 * mem_cgroup_end_page_stat(memcg);
1742 struct mem_cgroup
*mem_cgroup_begin_page_stat(struct page
*page
)
1744 struct mem_cgroup
*memcg
;
1745 unsigned long flags
;
1748 * The RCU lock is held throughout the transaction. The fast
1749 * path can get away without acquiring the memcg->move_lock
1750 * because page moving starts with an RCU grace period.
1752 * The RCU lock also protects the memcg from being freed when
1753 * the page state that is going to change is the only thing
1754 * preventing the page from being uncharged.
1755 * E.g. end-writeback clearing PageWriteback(), which allows
1756 * migration to go ahead and uncharge the page before the
1757 * account transaction might be complete.
1761 if (mem_cgroup_disabled())
1764 memcg
= page
->mem_cgroup
;
1765 if (unlikely(!memcg
))
1768 if (atomic_read(&memcg
->moving_account
) <= 0)
1771 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1772 if (memcg
!= page
->mem_cgroup
) {
1773 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1778 * When charge migration first begins, we can have locked and
1779 * unlocked page stat updates happening concurrently. Track
1780 * the task who has the lock for mem_cgroup_end_page_stat().
1782 memcg
->move_lock_task
= current
;
1783 memcg
->move_lock_flags
= flags
;
1787 EXPORT_SYMBOL(mem_cgroup_begin_page_stat
);
1790 * mem_cgroup_end_page_stat - finish a page state statistics transaction
1791 * @memcg: the memcg that was accounted against
1793 void mem_cgroup_end_page_stat(struct mem_cgroup
*memcg
)
1795 if (memcg
&& memcg
->move_lock_task
== current
) {
1796 unsigned long flags
= memcg
->move_lock_flags
;
1798 memcg
->move_lock_task
= NULL
;
1799 memcg
->move_lock_flags
= 0;
1801 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1806 EXPORT_SYMBOL(mem_cgroup_end_page_stat
);
1809 * size of first charge trial. "32" comes from vmscan.c's magic value.
1810 * TODO: maybe necessary to use big numbers in big irons.
1812 #define CHARGE_BATCH 32U
1813 struct memcg_stock_pcp
{
1814 struct mem_cgroup
*cached
; /* this never be root cgroup */
1815 unsigned int nr_pages
;
1816 struct work_struct work
;
1817 unsigned long flags
;
1818 #define FLUSHING_CACHED_CHARGE 0
1820 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1821 static DEFINE_MUTEX(percpu_charge_mutex
);
1824 * consume_stock: Try to consume stocked charge on this cpu.
1825 * @memcg: memcg to consume from.
1826 * @nr_pages: how many pages to charge.
1828 * The charges will only happen if @memcg matches the current cpu's memcg
1829 * stock, and at least @nr_pages are available in that stock. Failure to
1830 * service an allocation will refill the stock.
1832 * returns true if successful, false otherwise.
1834 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1836 struct memcg_stock_pcp
*stock
;
1839 if (nr_pages
> CHARGE_BATCH
)
1842 stock
= &get_cpu_var(memcg_stock
);
1843 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
1844 stock
->nr_pages
-= nr_pages
;
1847 put_cpu_var(memcg_stock
);
1852 * Returns stocks cached in percpu and reset cached information.
1854 static void drain_stock(struct memcg_stock_pcp
*stock
)
1856 struct mem_cgroup
*old
= stock
->cached
;
1858 if (stock
->nr_pages
) {
1859 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
1860 if (do_memsw_account())
1861 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
1862 css_put_many(&old
->css
, stock
->nr_pages
);
1863 stock
->nr_pages
= 0;
1865 stock
->cached
= NULL
;
1869 * This must be called under preempt disabled or must be called by
1870 * a thread which is pinned to local cpu.
1872 static void drain_local_stock(struct work_struct
*dummy
)
1874 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
1876 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
1880 * Cache charges(val) to local per_cpu area.
1881 * This will be consumed by consume_stock() function, later.
1883 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1885 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
1887 if (stock
->cached
!= memcg
) { /* reset if necessary */
1889 stock
->cached
= memcg
;
1891 stock
->nr_pages
+= nr_pages
;
1892 put_cpu_var(memcg_stock
);
1896 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1897 * of the hierarchy under it.
1899 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
1903 /* If someone's already draining, avoid adding running more workers. */
1904 if (!mutex_trylock(&percpu_charge_mutex
))
1906 /* Notify other cpus that system-wide "drain" is running */
1909 for_each_online_cpu(cpu
) {
1910 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1911 struct mem_cgroup
*memcg
;
1913 memcg
= stock
->cached
;
1914 if (!memcg
|| !stock
->nr_pages
)
1916 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
1918 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
1920 drain_local_stock(&stock
->work
);
1922 schedule_work_on(cpu
, &stock
->work
);
1927 mutex_unlock(&percpu_charge_mutex
);
1930 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
1931 unsigned long action
,
1934 int cpu
= (unsigned long)hcpu
;
1935 struct memcg_stock_pcp
*stock
;
1937 if (action
== CPU_ONLINE
)
1940 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
1943 stock
= &per_cpu(memcg_stock
, cpu
);
1948 static void reclaim_high(struct mem_cgroup
*memcg
,
1949 unsigned int nr_pages
,
1953 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
1955 mem_cgroup_events(memcg
, MEMCG_HIGH
, 1);
1956 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
1957 } while ((memcg
= parent_mem_cgroup(memcg
)));
1960 static void high_work_func(struct work_struct
*work
)
1962 struct mem_cgroup
*memcg
;
1964 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
1965 reclaim_high(memcg
, CHARGE_BATCH
, GFP_KERNEL
);
1969 * Scheduled by try_charge() to be executed from the userland return path
1970 * and reclaims memory over the high limit.
1972 void mem_cgroup_handle_over_high(void)
1974 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
1975 struct mem_cgroup
*memcg
;
1977 if (likely(!nr_pages
))
1980 memcg
= get_mem_cgroup_from_mm(current
->mm
);
1981 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
1982 css_put(&memcg
->css
);
1983 current
->memcg_nr_pages_over_high
= 0;
1986 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1987 unsigned int nr_pages
)
1989 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
1990 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1991 struct mem_cgroup
*mem_over_limit
;
1992 struct page_counter
*counter
;
1993 unsigned long nr_reclaimed
;
1994 bool may_swap
= true;
1995 bool drained
= false;
1997 if (mem_cgroup_is_root(memcg
))
2000 if (consume_stock(memcg
, nr_pages
))
2003 if (!do_memsw_account() ||
2004 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2005 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2007 if (do_memsw_account())
2008 page_counter_uncharge(&memcg
->memsw
, batch
);
2009 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2011 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2015 if (batch
> nr_pages
) {
2021 * Unlike in global OOM situations, memcg is not in a physical
2022 * memory shortage. Allow dying and OOM-killed tasks to
2023 * bypass the last charges so that they can exit quickly and
2024 * free their memory.
2026 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
2027 fatal_signal_pending(current
) ||
2028 current
->flags
& PF_EXITING
))
2031 if (unlikely(task_in_memcg_oom(current
)))
2034 if (!gfpflags_allow_blocking(gfp_mask
))
2037 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
2039 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2040 gfp_mask
, may_swap
);
2042 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2046 drain_all_stock(mem_over_limit
);
2051 if (gfp_mask
& __GFP_NORETRY
)
2054 * Even though the limit is exceeded at this point, reclaim
2055 * may have been able to free some pages. Retry the charge
2056 * before killing the task.
2058 * Only for regular pages, though: huge pages are rather
2059 * unlikely to succeed so close to the limit, and we fall back
2060 * to regular pages anyway in case of failure.
2062 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2065 * At task move, charge accounts can be doubly counted. So, it's
2066 * better to wait until the end of task_move if something is going on.
2068 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2074 if (gfp_mask
& __GFP_NOFAIL
)
2077 if (fatal_signal_pending(current
))
2080 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
2082 mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2083 get_order(nr_pages
* PAGE_SIZE
));
2085 if (!(gfp_mask
& __GFP_NOFAIL
))
2089 * The allocation either can't fail or will lead to more memory
2090 * being freed very soon. Allow memory usage go over the limit
2091 * temporarily by force charging it.
2093 page_counter_charge(&memcg
->memory
, nr_pages
);
2094 if (do_memsw_account())
2095 page_counter_charge(&memcg
->memsw
, nr_pages
);
2096 css_get_many(&memcg
->css
, nr_pages
);
2101 css_get_many(&memcg
->css
, batch
);
2102 if (batch
> nr_pages
)
2103 refill_stock(memcg
, batch
- nr_pages
);
2106 * If the hierarchy is above the normal consumption range, schedule
2107 * reclaim on returning to userland. We can perform reclaim here
2108 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2109 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2110 * not recorded as it most likely matches current's and won't
2111 * change in the meantime. As high limit is checked again before
2112 * reclaim, the cost of mismatch is negligible.
2115 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2116 /* Don't bother a random interrupted task */
2117 if (in_interrupt()) {
2118 schedule_work(&memcg
->high_work
);
2121 current
->memcg_nr_pages_over_high
+= batch
;
2122 set_notify_resume(current
);
2125 } while ((memcg
= parent_mem_cgroup(memcg
)));
2130 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2132 if (mem_cgroup_is_root(memcg
))
2135 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2136 if (do_memsw_account())
2137 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2139 css_put_many(&memcg
->css
, nr_pages
);
2142 static void lock_page_lru(struct page
*page
, int *isolated
)
2144 struct zone
*zone
= page_zone(page
);
2146 spin_lock_irq(&zone
->lru_lock
);
2147 if (PageLRU(page
)) {
2148 struct lruvec
*lruvec
;
2150 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2152 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2158 static void unlock_page_lru(struct page
*page
, int isolated
)
2160 struct zone
*zone
= page_zone(page
);
2163 struct lruvec
*lruvec
;
2165 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2166 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2168 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2170 spin_unlock_irq(&zone
->lru_lock
);
2173 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2178 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2181 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2182 * may already be on some other mem_cgroup's LRU. Take care of it.
2185 lock_page_lru(page
, &isolated
);
2188 * Nobody should be changing or seriously looking at
2189 * page->mem_cgroup at this point:
2191 * - the page is uncharged
2193 * - the page is off-LRU
2195 * - an anonymous fault has exclusive page access, except for
2196 * a locked page table
2198 * - a page cache insertion, a swapin fault, or a migration
2199 * have the page locked
2201 page
->mem_cgroup
= memcg
;
2204 unlock_page_lru(page
, isolated
);
2207 #ifdef CONFIG_MEMCG_KMEM
2208 static int memcg_alloc_cache_id(void)
2213 id
= ida_simple_get(&memcg_cache_ida
,
2214 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2218 if (id
< memcg_nr_cache_ids
)
2222 * There's no space for the new id in memcg_caches arrays,
2223 * so we have to grow them.
2225 down_write(&memcg_cache_ids_sem
);
2227 size
= 2 * (id
+ 1);
2228 if (size
< MEMCG_CACHES_MIN_SIZE
)
2229 size
= MEMCG_CACHES_MIN_SIZE
;
2230 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2231 size
= MEMCG_CACHES_MAX_SIZE
;
2233 err
= memcg_update_all_caches(size
);
2235 err
= memcg_update_all_list_lrus(size
);
2237 memcg_nr_cache_ids
= size
;
2239 up_write(&memcg_cache_ids_sem
);
2242 ida_simple_remove(&memcg_cache_ida
, id
);
2248 static void memcg_free_cache_id(int id
)
2250 ida_simple_remove(&memcg_cache_ida
, id
);
2253 struct memcg_kmem_cache_create_work
{
2254 struct mem_cgroup
*memcg
;
2255 struct kmem_cache
*cachep
;
2256 struct work_struct work
;
2259 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2261 struct memcg_kmem_cache_create_work
*cw
=
2262 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2263 struct mem_cgroup
*memcg
= cw
->memcg
;
2264 struct kmem_cache
*cachep
= cw
->cachep
;
2266 memcg_create_kmem_cache(memcg
, cachep
);
2268 css_put(&memcg
->css
);
2273 * Enqueue the creation of a per-memcg kmem_cache.
2275 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2276 struct kmem_cache
*cachep
)
2278 struct memcg_kmem_cache_create_work
*cw
;
2280 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2284 css_get(&memcg
->css
);
2287 cw
->cachep
= cachep
;
2288 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2290 schedule_work(&cw
->work
);
2293 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2294 struct kmem_cache
*cachep
)
2297 * We need to stop accounting when we kmalloc, because if the
2298 * corresponding kmalloc cache is not yet created, the first allocation
2299 * in __memcg_schedule_kmem_cache_create will recurse.
2301 * However, it is better to enclose the whole function. Depending on
2302 * the debugging options enabled, INIT_WORK(), for instance, can
2303 * trigger an allocation. This too, will make us recurse. Because at
2304 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2305 * the safest choice is to do it like this, wrapping the whole function.
2307 current
->memcg_kmem_skip_account
= 1;
2308 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2309 current
->memcg_kmem_skip_account
= 0;
2313 * Return the kmem_cache we're supposed to use for a slab allocation.
2314 * We try to use the current memcg's version of the cache.
2316 * If the cache does not exist yet, if we are the first user of it,
2317 * we either create it immediately, if possible, or create it asynchronously
2319 * In the latter case, we will let the current allocation go through with
2320 * the original cache.
2322 * Can't be called in interrupt context or from kernel threads.
2323 * This function needs to be called with rcu_read_lock() held.
2325 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2327 struct mem_cgroup
*memcg
;
2328 struct kmem_cache
*memcg_cachep
;
2331 VM_BUG_ON(!is_root_cache(cachep
));
2333 if (cachep
->flags
& SLAB_ACCOUNT
)
2334 gfp
|= __GFP_ACCOUNT
;
2336 if (!(gfp
& __GFP_ACCOUNT
))
2339 if (current
->memcg_kmem_skip_account
)
2342 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2343 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2347 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2348 if (likely(memcg_cachep
))
2349 return memcg_cachep
;
2352 * If we are in a safe context (can wait, and not in interrupt
2353 * context), we could be be predictable and return right away.
2354 * This would guarantee that the allocation being performed
2355 * already belongs in the new cache.
2357 * However, there are some clashes that can arrive from locking.
2358 * For instance, because we acquire the slab_mutex while doing
2359 * memcg_create_kmem_cache, this means no further allocation
2360 * could happen with the slab_mutex held. So it's better to
2363 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2365 css_put(&memcg
->css
);
2369 void __memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2371 if (!is_root_cache(cachep
))
2372 css_put(&cachep
->memcg_params
.memcg
->css
);
2375 int __memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2376 struct mem_cgroup
*memcg
)
2378 unsigned int nr_pages
= 1 << order
;
2379 struct page_counter
*counter
;
2382 if (!memcg_kmem_is_active(memcg
))
2385 if (!page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
))
2388 ret
= try_charge(memcg
, gfp
, nr_pages
);
2390 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2394 page
->mem_cgroup
= memcg
;
2399 int __memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2401 struct mem_cgroup
*memcg
;
2404 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2405 ret
= __memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2406 css_put(&memcg
->css
);
2410 void __memcg_kmem_uncharge(struct page
*page
, int order
)
2412 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2413 unsigned int nr_pages
= 1 << order
;
2418 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2420 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2421 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2422 if (do_memsw_account())
2423 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2425 page
->mem_cgroup
= NULL
;
2426 css_put_many(&memcg
->css
, nr_pages
);
2428 #endif /* CONFIG_MEMCG_KMEM */
2430 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2433 * Because tail pages are not marked as "used", set it. We're under
2434 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2435 * charge/uncharge will be never happen and move_account() is done under
2436 * compound_lock(), so we don't have to take care of races.
2438 void mem_cgroup_split_huge_fixup(struct page
*head
)
2442 if (mem_cgroup_disabled())
2445 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2446 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2448 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2451 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2453 #ifdef CONFIG_MEMCG_SWAP
2454 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2457 int val
= (charge
) ? 1 : -1;
2458 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2462 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2463 * @entry: swap entry to be moved
2464 * @from: mem_cgroup which the entry is moved from
2465 * @to: mem_cgroup which the entry is moved to
2467 * It succeeds only when the swap_cgroup's record for this entry is the same
2468 * as the mem_cgroup's id of @from.
2470 * Returns 0 on success, -EINVAL on failure.
2472 * The caller must have charged to @to, IOW, called page_counter_charge() about
2473 * both res and memsw, and called css_get().
2475 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2476 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2478 unsigned short old_id
, new_id
;
2480 old_id
= mem_cgroup_id(from
);
2481 new_id
= mem_cgroup_id(to
);
2483 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2484 mem_cgroup_swap_statistics(from
, false);
2485 mem_cgroup_swap_statistics(to
, true);
2491 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2492 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2498 static DEFINE_MUTEX(memcg_limit_mutex
);
2500 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2501 unsigned long limit
)
2503 unsigned long curusage
;
2504 unsigned long oldusage
;
2505 bool enlarge
= false;
2510 * For keeping hierarchical_reclaim simple, how long we should retry
2511 * is depends on callers. We set our retry-count to be function
2512 * of # of children which we should visit in this loop.
2514 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2515 mem_cgroup_count_children(memcg
);
2517 oldusage
= page_counter_read(&memcg
->memory
);
2520 if (signal_pending(current
)) {
2525 mutex_lock(&memcg_limit_mutex
);
2526 if (limit
> memcg
->memsw
.limit
) {
2527 mutex_unlock(&memcg_limit_mutex
);
2531 if (limit
> memcg
->memory
.limit
)
2533 ret
= page_counter_limit(&memcg
->memory
, limit
);
2534 mutex_unlock(&memcg_limit_mutex
);
2539 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2541 curusage
= page_counter_read(&memcg
->memory
);
2542 /* Usage is reduced ? */
2543 if (curusage
>= oldusage
)
2546 oldusage
= curusage
;
2547 } while (retry_count
);
2549 if (!ret
&& enlarge
)
2550 memcg_oom_recover(memcg
);
2555 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2556 unsigned long limit
)
2558 unsigned long curusage
;
2559 unsigned long oldusage
;
2560 bool enlarge
= false;
2564 /* see mem_cgroup_resize_res_limit */
2565 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2566 mem_cgroup_count_children(memcg
);
2568 oldusage
= page_counter_read(&memcg
->memsw
);
2571 if (signal_pending(current
)) {
2576 mutex_lock(&memcg_limit_mutex
);
2577 if (limit
< memcg
->memory
.limit
) {
2578 mutex_unlock(&memcg_limit_mutex
);
2582 if (limit
> memcg
->memsw
.limit
)
2584 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2585 mutex_unlock(&memcg_limit_mutex
);
2590 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2592 curusage
= page_counter_read(&memcg
->memsw
);
2593 /* Usage is reduced ? */
2594 if (curusage
>= oldusage
)
2597 oldusage
= curusage
;
2598 } while (retry_count
);
2600 if (!ret
&& enlarge
)
2601 memcg_oom_recover(memcg
);
2606 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
2608 unsigned long *total_scanned
)
2610 unsigned long nr_reclaimed
= 0;
2611 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
2612 unsigned long reclaimed
;
2614 struct mem_cgroup_tree_per_zone
*mctz
;
2615 unsigned long excess
;
2616 unsigned long nr_scanned
;
2621 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
2623 * This loop can run a while, specially if mem_cgroup's continuously
2624 * keep exceeding their soft limit and putting the system under
2631 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2636 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
2637 gfp_mask
, &nr_scanned
);
2638 nr_reclaimed
+= reclaimed
;
2639 *total_scanned
+= nr_scanned
;
2640 spin_lock_irq(&mctz
->lock
);
2641 __mem_cgroup_remove_exceeded(mz
, mctz
);
2644 * If we failed to reclaim anything from this memory cgroup
2645 * it is time to move on to the next cgroup
2649 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2651 excess
= soft_limit_excess(mz
->memcg
);
2653 * One school of thought says that we should not add
2654 * back the node to the tree if reclaim returns 0.
2655 * But our reclaim could return 0, simply because due
2656 * to priority we are exposing a smaller subset of
2657 * memory to reclaim from. Consider this as a longer
2660 /* If excess == 0, no tree ops */
2661 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2662 spin_unlock_irq(&mctz
->lock
);
2663 css_put(&mz
->memcg
->css
);
2666 * Could not reclaim anything and there are no more
2667 * mem cgroups to try or we seem to be looping without
2668 * reclaiming anything.
2670 if (!nr_reclaimed
&&
2672 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2674 } while (!nr_reclaimed
);
2676 css_put(&next_mz
->memcg
->css
);
2677 return nr_reclaimed
;
2681 * Test whether @memcg has children, dead or alive. Note that this
2682 * function doesn't care whether @memcg has use_hierarchy enabled and
2683 * returns %true if there are child csses according to the cgroup
2684 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2686 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2691 * The lock does not prevent addition or deletion of children, but
2692 * it prevents a new child from being initialized based on this
2693 * parent in css_online(), so it's enough to decide whether
2694 * hierarchically inherited attributes can still be changed or not.
2696 lockdep_assert_held(&memcg_create_mutex
);
2699 ret
= css_next_child(NULL
, &memcg
->css
);
2705 * Reclaims as many pages from the given memcg as possible and moves
2706 * the rest to the parent.
2708 * Caller is responsible for holding css reference for memcg.
2710 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2712 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2714 /* we call try-to-free pages for make this cgroup empty */
2715 lru_add_drain_all();
2716 /* try to free all pages in this cgroup */
2717 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2720 if (signal_pending(current
))
2723 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2727 /* maybe some writeback is necessary */
2728 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2736 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
2737 char *buf
, size_t nbytes
,
2740 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2742 if (mem_cgroup_is_root(memcg
))
2744 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
2747 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
2750 return mem_cgroup_from_css(css
)->use_hierarchy
;
2753 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
2754 struct cftype
*cft
, u64 val
)
2757 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2758 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
2760 mutex_lock(&memcg_create_mutex
);
2762 if (memcg
->use_hierarchy
== val
)
2766 * If parent's use_hierarchy is set, we can't make any modifications
2767 * in the child subtrees. If it is unset, then the change can
2768 * occur, provided the current cgroup has no children.
2770 * For the root cgroup, parent_mem is NULL, we allow value to be
2771 * set if there are no children.
2773 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
2774 (val
== 1 || val
== 0)) {
2775 if (!memcg_has_children(memcg
))
2776 memcg
->use_hierarchy
= val
;
2783 mutex_unlock(&memcg_create_mutex
);
2788 static unsigned long tree_stat(struct mem_cgroup
*memcg
,
2789 enum mem_cgroup_stat_index idx
)
2791 struct mem_cgroup
*iter
;
2792 unsigned long val
= 0;
2794 for_each_mem_cgroup_tree(iter
, memcg
)
2795 val
+= mem_cgroup_read_stat(iter
, idx
);
2800 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
2804 if (mem_cgroup_is_root(memcg
)) {
2805 val
= tree_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
2806 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_RSS
);
2808 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
2811 val
= page_counter_read(&memcg
->memory
);
2813 val
= page_counter_read(&memcg
->memsw
);
2826 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
2829 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2830 struct page_counter
*counter
;
2832 switch (MEMFILE_TYPE(cft
->private)) {
2834 counter
= &memcg
->memory
;
2837 counter
= &memcg
->memsw
;
2840 counter
= &memcg
->kmem
;
2846 switch (MEMFILE_ATTR(cft
->private)) {
2848 if (counter
== &memcg
->memory
)
2849 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
2850 if (counter
== &memcg
->memsw
)
2851 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
2852 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
2854 return (u64
)counter
->limit
* PAGE_SIZE
;
2856 return (u64
)counter
->watermark
* PAGE_SIZE
;
2858 return counter
->failcnt
;
2859 case RES_SOFT_LIMIT
:
2860 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
2866 #ifdef CONFIG_MEMCG_KMEM
2867 static int memcg_activate_kmem(struct mem_cgroup
*memcg
,
2868 unsigned long nr_pages
)
2873 BUG_ON(memcg
->kmemcg_id
>= 0);
2874 BUG_ON(memcg
->kmem_acct_activated
);
2875 BUG_ON(memcg
->kmem_acct_active
);
2878 * For simplicity, we won't allow this to be disabled. It also can't
2879 * be changed if the cgroup has children already, or if tasks had
2882 * If tasks join before we set the limit, a person looking at
2883 * kmem.usage_in_bytes will have no way to determine when it took
2884 * place, which makes the value quite meaningless.
2886 * After it first became limited, changes in the value of the limit are
2887 * of course permitted.
2889 mutex_lock(&memcg_create_mutex
);
2890 if (cgroup_is_populated(memcg
->css
.cgroup
) ||
2891 (memcg
->use_hierarchy
&& memcg_has_children(memcg
)))
2893 mutex_unlock(&memcg_create_mutex
);
2897 memcg_id
= memcg_alloc_cache_id();
2904 * We couldn't have accounted to this cgroup, because it hasn't got
2905 * activated yet, so this should succeed.
2907 err
= page_counter_limit(&memcg
->kmem
, nr_pages
);
2910 static_branch_inc(&memcg_kmem_enabled_key
);
2912 * A memory cgroup is considered kmem-active as soon as it gets
2913 * kmemcg_id. Setting the id after enabling static branching will
2914 * guarantee no one starts accounting before all call sites are
2917 memcg
->kmemcg_id
= memcg_id
;
2918 memcg
->kmem_acct_activated
= true;
2919 memcg
->kmem_acct_active
= true;
2924 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2925 unsigned long limit
)
2929 mutex_lock(&memcg_limit_mutex
);
2930 if (!memcg_kmem_is_active(memcg
))
2931 ret
= memcg_activate_kmem(memcg
, limit
);
2933 ret
= page_counter_limit(&memcg
->kmem
, limit
);
2934 mutex_unlock(&memcg_limit_mutex
);
2938 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
2941 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
2946 mutex_lock(&memcg_limit_mutex
);
2948 * If the parent cgroup is not kmem-active now, it cannot be activated
2949 * after this point, because it has at least one child already.
2951 if (memcg_kmem_is_active(parent
))
2952 ret
= memcg_activate_kmem(memcg
, PAGE_COUNTER_MAX
);
2953 mutex_unlock(&memcg_limit_mutex
);
2957 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2958 unsigned long limit
)
2962 #endif /* CONFIG_MEMCG_KMEM */
2965 * The user of this function is...
2968 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
2969 char *buf
, size_t nbytes
, loff_t off
)
2971 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2972 unsigned long nr_pages
;
2975 buf
= strstrip(buf
);
2976 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
2980 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
2982 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
2986 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
2988 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
2991 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
2994 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
2998 case RES_SOFT_LIMIT
:
2999 memcg
->soft_limit
= nr_pages
;
3003 return ret
?: nbytes
;
3006 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3007 size_t nbytes
, loff_t off
)
3009 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3010 struct page_counter
*counter
;
3012 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3014 counter
= &memcg
->memory
;
3017 counter
= &memcg
->memsw
;
3020 counter
= &memcg
->kmem
;
3026 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3028 page_counter_reset_watermark(counter
);
3031 counter
->failcnt
= 0;
3040 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3043 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3047 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3048 struct cftype
*cft
, u64 val
)
3050 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3052 if (val
& ~MOVE_MASK
)
3056 * No kind of locking is needed in here, because ->can_attach() will
3057 * check this value once in the beginning of the process, and then carry
3058 * on with stale data. This means that changes to this value will only
3059 * affect task migrations starting after the change.
3061 memcg
->move_charge_at_immigrate
= val
;
3065 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3066 struct cftype
*cft
, u64 val
)
3073 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3077 unsigned int lru_mask
;
3080 static const struct numa_stat stats
[] = {
3081 { "total", LRU_ALL
},
3082 { "file", LRU_ALL_FILE
},
3083 { "anon", LRU_ALL_ANON
},
3084 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3086 const struct numa_stat
*stat
;
3089 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3091 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3092 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3093 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3094 for_each_node_state(nid
, N_MEMORY
) {
3095 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3097 seq_printf(m
, " N%d=%lu", nid
, nr
);
3102 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3103 struct mem_cgroup
*iter
;
3106 for_each_mem_cgroup_tree(iter
, memcg
)
3107 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3108 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3109 for_each_node_state(nid
, N_MEMORY
) {
3111 for_each_mem_cgroup_tree(iter
, memcg
)
3112 nr
+= mem_cgroup_node_nr_lru_pages(
3113 iter
, nid
, stat
->lru_mask
);
3114 seq_printf(m
, " N%d=%lu", nid
, nr
);
3121 #endif /* CONFIG_NUMA */
3123 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3125 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3126 unsigned long memory
, memsw
;
3127 struct mem_cgroup
*mi
;
3130 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3131 MEM_CGROUP_STAT_NSTATS
);
3132 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3133 MEM_CGROUP_EVENTS_NSTATS
);
3134 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3136 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3137 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3139 seq_printf(m
, "%s %lu\n", mem_cgroup_stat_names
[i
],
3140 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3143 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3144 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3145 mem_cgroup_read_events(memcg
, i
));
3147 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3148 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3149 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3151 /* Hierarchical information */
3152 memory
= memsw
= PAGE_COUNTER_MAX
;
3153 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3154 memory
= min(memory
, mi
->memory
.limit
);
3155 memsw
= min(memsw
, mi
->memsw
.limit
);
3157 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3158 (u64
)memory
* PAGE_SIZE
);
3159 if (do_memsw_account())
3160 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3161 (u64
)memsw
* PAGE_SIZE
);
3163 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3164 unsigned long long val
= 0;
3166 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3168 for_each_mem_cgroup_tree(mi
, memcg
)
3169 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3170 seq_printf(m
, "total_%s %llu\n", mem_cgroup_stat_names
[i
], val
);
3173 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3174 unsigned long long val
= 0;
3176 for_each_mem_cgroup_tree(mi
, memcg
)
3177 val
+= mem_cgroup_read_events(mi
, i
);
3178 seq_printf(m
, "total_%s %llu\n",
3179 mem_cgroup_events_names
[i
], val
);
3182 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3183 unsigned long long val
= 0;
3185 for_each_mem_cgroup_tree(mi
, memcg
)
3186 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3187 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3190 #ifdef CONFIG_DEBUG_VM
3193 struct mem_cgroup_per_zone
*mz
;
3194 struct zone_reclaim_stat
*rstat
;
3195 unsigned long recent_rotated
[2] = {0, 0};
3196 unsigned long recent_scanned
[2] = {0, 0};
3198 for_each_online_node(nid
)
3199 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3200 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
3201 rstat
= &mz
->lruvec
.reclaim_stat
;
3203 recent_rotated
[0] += rstat
->recent_rotated
[0];
3204 recent_rotated
[1] += rstat
->recent_rotated
[1];
3205 recent_scanned
[0] += rstat
->recent_scanned
[0];
3206 recent_scanned
[1] += rstat
->recent_scanned
[1];
3208 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3209 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3210 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3211 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3218 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3221 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3223 return mem_cgroup_swappiness(memcg
);
3226 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3227 struct cftype
*cft
, u64 val
)
3229 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3235 memcg
->swappiness
= val
;
3237 vm_swappiness
= val
;
3242 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3244 struct mem_cgroup_threshold_ary
*t
;
3245 unsigned long usage
;
3250 t
= rcu_dereference(memcg
->thresholds
.primary
);
3252 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3257 usage
= mem_cgroup_usage(memcg
, swap
);
3260 * current_threshold points to threshold just below or equal to usage.
3261 * If it's not true, a threshold was crossed after last
3262 * call of __mem_cgroup_threshold().
3264 i
= t
->current_threshold
;
3267 * Iterate backward over array of thresholds starting from
3268 * current_threshold and check if a threshold is crossed.
3269 * If none of thresholds below usage is crossed, we read
3270 * only one element of the array here.
3272 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3273 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3275 /* i = current_threshold + 1 */
3279 * Iterate forward over array of thresholds starting from
3280 * current_threshold+1 and check if a threshold is crossed.
3281 * If none of thresholds above usage is crossed, we read
3282 * only one element of the array here.
3284 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3285 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3287 /* Update current_threshold */
3288 t
->current_threshold
= i
- 1;
3293 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3296 __mem_cgroup_threshold(memcg
, false);
3297 if (do_memsw_account())
3298 __mem_cgroup_threshold(memcg
, true);
3300 memcg
= parent_mem_cgroup(memcg
);
3304 static int compare_thresholds(const void *a
, const void *b
)
3306 const struct mem_cgroup_threshold
*_a
= a
;
3307 const struct mem_cgroup_threshold
*_b
= b
;
3309 if (_a
->threshold
> _b
->threshold
)
3312 if (_a
->threshold
< _b
->threshold
)
3318 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3320 struct mem_cgroup_eventfd_list
*ev
;
3322 spin_lock(&memcg_oom_lock
);
3324 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3325 eventfd_signal(ev
->eventfd
, 1);
3327 spin_unlock(&memcg_oom_lock
);
3331 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3333 struct mem_cgroup
*iter
;
3335 for_each_mem_cgroup_tree(iter
, memcg
)
3336 mem_cgroup_oom_notify_cb(iter
);
3339 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3340 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3342 struct mem_cgroup_thresholds
*thresholds
;
3343 struct mem_cgroup_threshold_ary
*new;
3344 unsigned long threshold
;
3345 unsigned long usage
;
3348 ret
= page_counter_memparse(args
, "-1", &threshold
);
3352 mutex_lock(&memcg
->thresholds_lock
);
3355 thresholds
= &memcg
->thresholds
;
3356 usage
= mem_cgroup_usage(memcg
, false);
3357 } else if (type
== _MEMSWAP
) {
3358 thresholds
= &memcg
->memsw_thresholds
;
3359 usage
= mem_cgroup_usage(memcg
, true);
3363 /* Check if a threshold crossed before adding a new one */
3364 if (thresholds
->primary
)
3365 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3367 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3369 /* Allocate memory for new array of thresholds */
3370 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3378 /* Copy thresholds (if any) to new array */
3379 if (thresholds
->primary
) {
3380 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3381 sizeof(struct mem_cgroup_threshold
));
3384 /* Add new threshold */
3385 new->entries
[size
- 1].eventfd
= eventfd
;
3386 new->entries
[size
- 1].threshold
= threshold
;
3388 /* Sort thresholds. Registering of new threshold isn't time-critical */
3389 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3390 compare_thresholds
, NULL
);
3392 /* Find current threshold */
3393 new->current_threshold
= -1;
3394 for (i
= 0; i
< size
; i
++) {
3395 if (new->entries
[i
].threshold
<= usage
) {
3397 * new->current_threshold will not be used until
3398 * rcu_assign_pointer(), so it's safe to increment
3401 ++new->current_threshold
;
3406 /* Free old spare buffer and save old primary buffer as spare */
3407 kfree(thresholds
->spare
);
3408 thresholds
->spare
= thresholds
->primary
;
3410 rcu_assign_pointer(thresholds
->primary
, new);
3412 /* To be sure that nobody uses thresholds */
3416 mutex_unlock(&memcg
->thresholds_lock
);
3421 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3422 struct eventfd_ctx
*eventfd
, const char *args
)
3424 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3427 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3428 struct eventfd_ctx
*eventfd
, const char *args
)
3430 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3433 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3434 struct eventfd_ctx
*eventfd
, enum res_type type
)
3436 struct mem_cgroup_thresholds
*thresholds
;
3437 struct mem_cgroup_threshold_ary
*new;
3438 unsigned long usage
;
3441 mutex_lock(&memcg
->thresholds_lock
);
3444 thresholds
= &memcg
->thresholds
;
3445 usage
= mem_cgroup_usage(memcg
, false);
3446 } else if (type
== _MEMSWAP
) {
3447 thresholds
= &memcg
->memsw_thresholds
;
3448 usage
= mem_cgroup_usage(memcg
, true);
3452 if (!thresholds
->primary
)
3455 /* Check if a threshold crossed before removing */
3456 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3458 /* Calculate new number of threshold */
3460 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3461 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3465 new = thresholds
->spare
;
3467 /* Set thresholds array to NULL if we don't have thresholds */
3476 /* Copy thresholds and find current threshold */
3477 new->current_threshold
= -1;
3478 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3479 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3482 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3483 if (new->entries
[j
].threshold
<= usage
) {
3485 * new->current_threshold will not be used
3486 * until rcu_assign_pointer(), so it's safe to increment
3489 ++new->current_threshold
;
3495 /* Swap primary and spare array */
3496 thresholds
->spare
= thresholds
->primary
;
3497 /* If all events are unregistered, free the spare array */
3499 kfree(thresholds
->spare
);
3500 thresholds
->spare
= NULL
;
3503 rcu_assign_pointer(thresholds
->primary
, new);
3505 /* To be sure that nobody uses thresholds */
3508 mutex_unlock(&memcg
->thresholds_lock
);
3511 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3512 struct eventfd_ctx
*eventfd
)
3514 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3517 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3518 struct eventfd_ctx
*eventfd
)
3520 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3523 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3524 struct eventfd_ctx
*eventfd
, const char *args
)
3526 struct mem_cgroup_eventfd_list
*event
;
3528 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3532 spin_lock(&memcg_oom_lock
);
3534 event
->eventfd
= eventfd
;
3535 list_add(&event
->list
, &memcg
->oom_notify
);
3537 /* already in OOM ? */
3538 if (memcg
->under_oom
)
3539 eventfd_signal(eventfd
, 1);
3540 spin_unlock(&memcg_oom_lock
);
3545 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3546 struct eventfd_ctx
*eventfd
)
3548 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3550 spin_lock(&memcg_oom_lock
);
3552 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3553 if (ev
->eventfd
== eventfd
) {
3554 list_del(&ev
->list
);
3559 spin_unlock(&memcg_oom_lock
);
3562 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3564 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3566 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3567 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3571 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3572 struct cftype
*cft
, u64 val
)
3574 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3576 /* cannot set to root cgroup and only 0 and 1 are allowed */
3577 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3580 memcg
->oom_kill_disable
= val
;
3582 memcg_oom_recover(memcg
);
3587 #ifdef CONFIG_MEMCG_KMEM
3588 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
3592 ret
= memcg_propagate_kmem(memcg
);
3596 return tcp_init_cgroup(memcg
, ss
);
3599 static void memcg_deactivate_kmem(struct mem_cgroup
*memcg
)
3601 struct cgroup_subsys_state
*css
;
3602 struct mem_cgroup
*parent
, *child
;
3605 if (!memcg
->kmem_acct_active
)
3609 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3610 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3611 * guarantees no cache will be created for this cgroup after we are
3612 * done (see memcg_create_kmem_cache()).
3614 memcg
->kmem_acct_active
= false;
3616 memcg_deactivate_kmem_caches(memcg
);
3618 kmemcg_id
= memcg
->kmemcg_id
;
3619 BUG_ON(kmemcg_id
< 0);
3621 parent
= parent_mem_cgroup(memcg
);
3623 parent
= root_mem_cgroup
;
3626 * Change kmemcg_id of this cgroup and all its descendants to the
3627 * parent's id, and then move all entries from this cgroup's list_lrus
3628 * to ones of the parent. After we have finished, all list_lrus
3629 * corresponding to this cgroup are guaranteed to remain empty. The
3630 * ordering is imposed by list_lru_node->lock taken by
3631 * memcg_drain_all_list_lrus().
3633 css_for_each_descendant_pre(css
, &memcg
->css
) {
3634 child
= mem_cgroup_from_css(css
);
3635 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
3636 child
->kmemcg_id
= parent
->kmemcg_id
;
3637 if (!memcg
->use_hierarchy
)
3640 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
3642 memcg_free_cache_id(kmemcg_id
);
3645 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
3647 if (memcg
->kmem_acct_activated
) {
3648 memcg_destroy_kmem_caches(memcg
);
3649 static_branch_dec(&memcg_kmem_enabled_key
);
3650 WARN_ON(page_counter_read(&memcg
->kmem
));
3652 tcp_destroy_cgroup(memcg
);
3655 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
3660 static void memcg_deactivate_kmem(struct mem_cgroup
*memcg
)
3664 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
3669 #ifdef CONFIG_CGROUP_WRITEBACK
3671 struct list_head
*mem_cgroup_cgwb_list(struct mem_cgroup
*memcg
)
3673 return &memcg
->cgwb_list
;
3676 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3678 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3681 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3683 wb_domain_exit(&memcg
->cgwb_domain
);
3686 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3688 wb_domain_size_changed(&memcg
->cgwb_domain
);
3691 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3693 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3695 if (!memcg
->css
.parent
)
3698 return &memcg
->cgwb_domain
;
3702 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3703 * @wb: bdi_writeback in question
3704 * @pfilepages: out parameter for number of file pages
3705 * @pheadroom: out parameter for number of allocatable pages according to memcg
3706 * @pdirty: out parameter for number of dirty pages
3707 * @pwriteback: out parameter for number of pages under writeback
3709 * Determine the numbers of file, headroom, dirty, and writeback pages in
3710 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3711 * is a bit more involved.
3713 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3714 * headroom is calculated as the lowest headroom of itself and the
3715 * ancestors. Note that this doesn't consider the actual amount of
3716 * available memory in the system. The caller should further cap
3717 * *@pheadroom accordingly.
3719 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
3720 unsigned long *pheadroom
, unsigned long *pdirty
,
3721 unsigned long *pwriteback
)
3723 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3724 struct mem_cgroup
*parent
;
3726 *pdirty
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
3728 /* this should eventually include NR_UNSTABLE_NFS */
3729 *pwriteback
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
3730 *pfilepages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
3731 (1 << LRU_ACTIVE_FILE
));
3732 *pheadroom
= PAGE_COUNTER_MAX
;
3734 while ((parent
= parent_mem_cgroup(memcg
))) {
3735 unsigned long ceiling
= min(memcg
->memory
.limit
, memcg
->high
);
3736 unsigned long used
= page_counter_read(&memcg
->memory
);
3738 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
3743 #else /* CONFIG_CGROUP_WRITEBACK */
3745 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3750 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3754 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3758 #endif /* CONFIG_CGROUP_WRITEBACK */
3761 * DO NOT USE IN NEW FILES.
3763 * "cgroup.event_control" implementation.
3765 * This is way over-engineered. It tries to support fully configurable
3766 * events for each user. Such level of flexibility is completely
3767 * unnecessary especially in the light of the planned unified hierarchy.
3769 * Please deprecate this and replace with something simpler if at all
3774 * Unregister event and free resources.
3776 * Gets called from workqueue.
3778 static void memcg_event_remove(struct work_struct
*work
)
3780 struct mem_cgroup_event
*event
=
3781 container_of(work
, struct mem_cgroup_event
, remove
);
3782 struct mem_cgroup
*memcg
= event
->memcg
;
3784 remove_wait_queue(event
->wqh
, &event
->wait
);
3786 event
->unregister_event(memcg
, event
->eventfd
);
3788 /* Notify userspace the event is going away. */
3789 eventfd_signal(event
->eventfd
, 1);
3791 eventfd_ctx_put(event
->eventfd
);
3793 css_put(&memcg
->css
);
3797 * Gets called on POLLHUP on eventfd when user closes it.
3799 * Called with wqh->lock held and interrupts disabled.
3801 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
3802 int sync
, void *key
)
3804 struct mem_cgroup_event
*event
=
3805 container_of(wait
, struct mem_cgroup_event
, wait
);
3806 struct mem_cgroup
*memcg
= event
->memcg
;
3807 unsigned long flags
= (unsigned long)key
;
3809 if (flags
& POLLHUP
) {
3811 * If the event has been detached at cgroup removal, we
3812 * can simply return knowing the other side will cleanup
3815 * We can't race against event freeing since the other
3816 * side will require wqh->lock via remove_wait_queue(),
3819 spin_lock(&memcg
->event_list_lock
);
3820 if (!list_empty(&event
->list
)) {
3821 list_del_init(&event
->list
);
3823 * We are in atomic context, but cgroup_event_remove()
3824 * may sleep, so we have to call it in workqueue.
3826 schedule_work(&event
->remove
);
3828 spin_unlock(&memcg
->event_list_lock
);
3834 static void memcg_event_ptable_queue_proc(struct file
*file
,
3835 wait_queue_head_t
*wqh
, poll_table
*pt
)
3837 struct mem_cgroup_event
*event
=
3838 container_of(pt
, struct mem_cgroup_event
, pt
);
3841 add_wait_queue(wqh
, &event
->wait
);
3845 * DO NOT USE IN NEW FILES.
3847 * Parse input and register new cgroup event handler.
3849 * Input must be in format '<event_fd> <control_fd> <args>'.
3850 * Interpretation of args is defined by control file implementation.
3852 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
3853 char *buf
, size_t nbytes
, loff_t off
)
3855 struct cgroup_subsys_state
*css
= of_css(of
);
3856 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3857 struct mem_cgroup_event
*event
;
3858 struct cgroup_subsys_state
*cfile_css
;
3859 unsigned int efd
, cfd
;
3866 buf
= strstrip(buf
);
3868 efd
= simple_strtoul(buf
, &endp
, 10);
3873 cfd
= simple_strtoul(buf
, &endp
, 10);
3874 if ((*endp
!= ' ') && (*endp
!= '\0'))
3878 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
3882 event
->memcg
= memcg
;
3883 INIT_LIST_HEAD(&event
->list
);
3884 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
3885 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
3886 INIT_WORK(&event
->remove
, memcg_event_remove
);
3894 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
3895 if (IS_ERR(event
->eventfd
)) {
3896 ret
= PTR_ERR(event
->eventfd
);
3903 goto out_put_eventfd
;
3906 /* the process need read permission on control file */
3907 /* AV: shouldn't we check that it's been opened for read instead? */
3908 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
3913 * Determine the event callbacks and set them in @event. This used
3914 * to be done via struct cftype but cgroup core no longer knows
3915 * about these events. The following is crude but the whole thing
3916 * is for compatibility anyway.
3918 * DO NOT ADD NEW FILES.
3920 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
3922 if (!strcmp(name
, "memory.usage_in_bytes")) {
3923 event
->register_event
= mem_cgroup_usage_register_event
;
3924 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
3925 } else if (!strcmp(name
, "memory.oom_control")) {
3926 event
->register_event
= mem_cgroup_oom_register_event
;
3927 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
3928 } else if (!strcmp(name
, "memory.pressure_level")) {
3929 event
->register_event
= vmpressure_register_event
;
3930 event
->unregister_event
= vmpressure_unregister_event
;
3931 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
3932 event
->register_event
= memsw_cgroup_usage_register_event
;
3933 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
3940 * Verify @cfile should belong to @css. Also, remaining events are
3941 * automatically removed on cgroup destruction but the removal is
3942 * asynchronous, so take an extra ref on @css.
3944 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
3945 &memory_cgrp_subsys
);
3947 if (IS_ERR(cfile_css
))
3949 if (cfile_css
!= css
) {
3954 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
3958 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
3960 spin_lock(&memcg
->event_list_lock
);
3961 list_add(&event
->list
, &memcg
->event_list
);
3962 spin_unlock(&memcg
->event_list_lock
);
3974 eventfd_ctx_put(event
->eventfd
);
3983 static struct cftype mem_cgroup_legacy_files
[] = {
3985 .name
= "usage_in_bytes",
3986 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
3987 .read_u64
= mem_cgroup_read_u64
,
3990 .name
= "max_usage_in_bytes",
3991 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
3992 .write
= mem_cgroup_reset
,
3993 .read_u64
= mem_cgroup_read_u64
,
3996 .name
= "limit_in_bytes",
3997 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
3998 .write
= mem_cgroup_write
,
3999 .read_u64
= mem_cgroup_read_u64
,
4002 .name
= "soft_limit_in_bytes",
4003 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4004 .write
= mem_cgroup_write
,
4005 .read_u64
= mem_cgroup_read_u64
,
4009 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4010 .write
= mem_cgroup_reset
,
4011 .read_u64
= mem_cgroup_read_u64
,
4015 .seq_show
= memcg_stat_show
,
4018 .name
= "force_empty",
4019 .write
= mem_cgroup_force_empty_write
,
4022 .name
= "use_hierarchy",
4023 .write_u64
= mem_cgroup_hierarchy_write
,
4024 .read_u64
= mem_cgroup_hierarchy_read
,
4027 .name
= "cgroup.event_control", /* XXX: for compat */
4028 .write
= memcg_write_event_control
,
4029 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
4032 .name
= "swappiness",
4033 .read_u64
= mem_cgroup_swappiness_read
,
4034 .write_u64
= mem_cgroup_swappiness_write
,
4037 .name
= "move_charge_at_immigrate",
4038 .read_u64
= mem_cgroup_move_charge_read
,
4039 .write_u64
= mem_cgroup_move_charge_write
,
4042 .name
= "oom_control",
4043 .seq_show
= mem_cgroup_oom_control_read
,
4044 .write_u64
= mem_cgroup_oom_control_write
,
4045 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4048 .name
= "pressure_level",
4052 .name
= "numa_stat",
4053 .seq_show
= memcg_numa_stat_show
,
4056 #ifdef CONFIG_MEMCG_KMEM
4058 .name
= "kmem.limit_in_bytes",
4059 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4060 .write
= mem_cgroup_write
,
4061 .read_u64
= mem_cgroup_read_u64
,
4064 .name
= "kmem.usage_in_bytes",
4065 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4066 .read_u64
= mem_cgroup_read_u64
,
4069 .name
= "kmem.failcnt",
4070 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4071 .write
= mem_cgroup_reset
,
4072 .read_u64
= mem_cgroup_read_u64
,
4075 .name
= "kmem.max_usage_in_bytes",
4076 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4077 .write
= mem_cgroup_reset
,
4078 .read_u64
= mem_cgroup_read_u64
,
4080 #ifdef CONFIG_SLABINFO
4082 .name
= "kmem.slabinfo",
4083 .seq_start
= slab_start
,
4084 .seq_next
= slab_next
,
4085 .seq_stop
= slab_stop
,
4086 .seq_show
= memcg_slab_show
,
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 struct mem_cgroup
*mem_cgroup_alloc(void)
4130 struct mem_cgroup
*memcg
;
4133 size
= sizeof(struct mem_cgroup
);
4134 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4136 memcg
= kzalloc(size
, GFP_KERNEL
);
4140 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4144 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4150 free_percpu(memcg
->stat
);
4157 * At destroying mem_cgroup, references from swap_cgroup can remain.
4158 * (scanning all at force_empty is too costly...)
4160 * Instead of clearing all references at force_empty, we remember
4161 * the number of reference from swap_cgroup and free mem_cgroup when
4162 * it goes down to 0.
4164 * Removal of cgroup itself succeeds regardless of refs from swap.
4167 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
4171 cancel_work_sync(&memcg
->high_work
);
4173 mem_cgroup_remove_from_trees(memcg
);
4176 free_mem_cgroup_per_zone_info(memcg
, node
);
4178 free_percpu(memcg
->stat
);
4179 memcg_wb_domain_exit(memcg
);
4183 static struct cgroup_subsys_state
* __ref
4184 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4186 struct mem_cgroup
*memcg
;
4187 long error
= -ENOMEM
;
4190 memcg
= mem_cgroup_alloc();
4192 return ERR_PTR(error
);
4195 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
4199 if (parent_css
== NULL
) {
4200 root_mem_cgroup
= memcg
;
4201 page_counter_init(&memcg
->memory
, NULL
);
4202 memcg
->high
= PAGE_COUNTER_MAX
;
4203 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4204 page_counter_init(&memcg
->memsw
, NULL
);
4205 page_counter_init(&memcg
->kmem
, NULL
);
4208 INIT_WORK(&memcg
->high_work
, high_work_func
);
4209 memcg
->last_scanned_node
= MAX_NUMNODES
;
4210 INIT_LIST_HEAD(&memcg
->oom_notify
);
4211 memcg
->move_charge_at_immigrate
= 0;
4212 mutex_init(&memcg
->thresholds_lock
);
4213 spin_lock_init(&memcg
->move_lock
);
4214 vmpressure_init(&memcg
->vmpressure
);
4215 INIT_LIST_HEAD(&memcg
->event_list
);
4216 spin_lock_init(&memcg
->event_list_lock
);
4217 #ifdef CONFIG_MEMCG_KMEM
4218 memcg
->kmemcg_id
= -1;
4220 #ifdef CONFIG_CGROUP_WRITEBACK
4221 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4224 memcg
->socket_pressure
= jiffies
;
4229 __mem_cgroup_free(memcg
);
4230 return ERR_PTR(error
);
4234 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4236 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4237 struct mem_cgroup
*parent
= mem_cgroup_from_css(css
->parent
);
4240 if (css
->id
> MEM_CGROUP_ID_MAX
)
4246 mutex_lock(&memcg_create_mutex
);
4248 memcg
->use_hierarchy
= parent
->use_hierarchy
;
4249 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4250 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4252 if (parent
->use_hierarchy
) {
4253 page_counter_init(&memcg
->memory
, &parent
->memory
);
4254 memcg
->high
= PAGE_COUNTER_MAX
;
4255 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4256 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4257 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4260 * No need to take a reference to the parent because cgroup
4261 * core guarantees its existence.
4264 page_counter_init(&memcg
->memory
, NULL
);
4265 memcg
->high
= PAGE_COUNTER_MAX
;
4266 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4267 page_counter_init(&memcg
->memsw
, NULL
);
4268 page_counter_init(&memcg
->kmem
, NULL
);
4270 * Deeper hierachy with use_hierarchy == false doesn't make
4271 * much sense so let cgroup subsystem know about this
4272 * unfortunate state in our controller.
4274 if (parent
!= root_mem_cgroup
)
4275 memory_cgrp_subsys
.broken_hierarchy
= true;
4277 mutex_unlock(&memcg_create_mutex
);
4279 ret
= memcg_init_kmem(memcg
, &memory_cgrp_subsys
);
4284 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4285 static_branch_inc(&memcg_sockets_enabled_key
);
4289 * Make sure the memcg is initialized: mem_cgroup_iter()
4290 * orders reading memcg->initialized against its callers
4291 * reading the memcg members.
4293 smp_store_release(&memcg
->initialized
, 1);
4298 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4300 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4301 struct mem_cgroup_event
*event
, *tmp
;
4304 * Unregister events and notify userspace.
4305 * Notify userspace about cgroup removing only after rmdir of cgroup
4306 * directory to avoid race between userspace and kernelspace.
4308 spin_lock(&memcg
->event_list_lock
);
4309 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4310 list_del_init(&event
->list
);
4311 schedule_work(&event
->remove
);
4313 spin_unlock(&memcg
->event_list_lock
);
4315 vmpressure_cleanup(&memcg
->vmpressure
);
4317 memcg_deactivate_kmem(memcg
);
4319 wb_memcg_offline(memcg
);
4322 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4324 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4326 invalidate_reclaim_iterators(memcg
);
4329 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4331 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4333 memcg_destroy_kmem(memcg
);
4335 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4336 static_branch_dec(&memcg_sockets_enabled_key
);
4338 __mem_cgroup_free(memcg
);
4342 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4343 * @css: the target css
4345 * Reset the states of the mem_cgroup associated with @css. This is
4346 * invoked when the userland requests disabling on the default hierarchy
4347 * but the memcg is pinned through dependency. The memcg should stop
4348 * applying policies and should revert to the vanilla state as it may be
4349 * made visible again.
4351 * The current implementation only resets the essential configurations.
4352 * This needs to be expanded to cover all the visible parts.
4354 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4356 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4358 mem_cgroup_resize_limit(memcg
, PAGE_COUNTER_MAX
);
4359 mem_cgroup_resize_memsw_limit(memcg
, PAGE_COUNTER_MAX
);
4360 memcg_update_kmem_limit(memcg
, PAGE_COUNTER_MAX
);
4362 memcg
->high
= PAGE_COUNTER_MAX
;
4363 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4364 memcg_wb_domain_size_changed(memcg
);
4368 /* Handlers for move charge at task migration. */
4369 static int mem_cgroup_do_precharge(unsigned long count
)
4373 /* Try a single bulk charge without reclaim first, kswapd may wake */
4374 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4376 mc
.precharge
+= count
;
4380 /* Try charges one by one with reclaim */
4382 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4392 * get_mctgt_type - get target type of moving charge
4393 * @vma: the vma the pte to be checked belongs
4394 * @addr: the address corresponding to the pte to be checked
4395 * @ptent: the pte to be checked
4396 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4399 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4400 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4401 * move charge. if @target is not NULL, the page is stored in target->page
4402 * with extra refcnt got(Callers should handle it).
4403 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4404 * target for charge migration. if @target is not NULL, the entry is stored
4407 * Called with pte lock held.
4414 enum mc_target_type
{
4420 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4421 unsigned long addr
, pte_t ptent
)
4423 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4425 if (!page
|| !page_mapped(page
))
4427 if (PageAnon(page
)) {
4428 if (!(mc
.flags
& MOVE_ANON
))
4431 if (!(mc
.flags
& MOVE_FILE
))
4434 if (!get_page_unless_zero(page
))
4441 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4442 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4444 struct page
*page
= NULL
;
4445 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4447 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4450 * Because lookup_swap_cache() updates some statistics counter,
4451 * we call find_get_page() with swapper_space directly.
4453 page
= find_get_page(swap_address_space(ent
), ent
.val
);
4454 if (do_memsw_account())
4455 entry
->val
= ent
.val
;
4460 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4461 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4467 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4468 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4470 struct page
*page
= NULL
;
4471 struct address_space
*mapping
;
4474 if (!vma
->vm_file
) /* anonymous vma */
4476 if (!(mc
.flags
& MOVE_FILE
))
4479 mapping
= vma
->vm_file
->f_mapping
;
4480 pgoff
= linear_page_index(vma
, addr
);
4482 /* page is moved even if it's not RSS of this task(page-faulted). */
4484 /* shmem/tmpfs may report page out on swap: account for that too. */
4485 if (shmem_mapping(mapping
)) {
4486 page
= find_get_entry(mapping
, pgoff
);
4487 if (radix_tree_exceptional_entry(page
)) {
4488 swp_entry_t swp
= radix_to_swp_entry(page
);
4489 if (do_memsw_account())
4491 page
= find_get_page(swap_address_space(swp
), swp
.val
);
4494 page
= find_get_page(mapping
, pgoff
);
4496 page
= find_get_page(mapping
, pgoff
);
4502 * mem_cgroup_move_account - move account of the page
4504 * @nr_pages: number of regular pages (>1 for huge pages)
4505 * @from: mem_cgroup which the page is moved from.
4506 * @to: mem_cgroup which the page is moved to. @from != @to.
4508 * The caller must confirm following.
4509 * - page is not on LRU (isolate_page() is useful.)
4510 * - compound_lock is held when nr_pages > 1
4512 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4515 static int mem_cgroup_move_account(struct page
*page
,
4516 unsigned int nr_pages
,
4517 struct mem_cgroup
*from
,
4518 struct mem_cgroup
*to
)
4520 unsigned long flags
;
4524 VM_BUG_ON(from
== to
);
4525 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4527 * The page is isolated from LRU. So, collapse function
4528 * will not handle this page. But page splitting can happen.
4529 * Do this check under compound_page_lock(). The caller should
4533 if (nr_pages
> 1 && !PageTransHuge(page
))
4537 * Prevent mem_cgroup_replace_page() from looking at
4538 * page->mem_cgroup of its source page while we change it.
4540 if (!trylock_page(page
))
4544 if (page
->mem_cgroup
!= from
)
4547 anon
= PageAnon(page
);
4549 spin_lock_irqsave(&from
->move_lock
, flags
);
4551 if (!anon
&& page_mapped(page
)) {
4552 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4554 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4559 * move_lock grabbed above and caller set from->moving_account, so
4560 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4561 * So mapping should be stable for dirty pages.
4563 if (!anon
&& PageDirty(page
)) {
4564 struct address_space
*mapping
= page_mapping(page
);
4566 if (mapping_cap_account_dirty(mapping
)) {
4567 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4569 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4574 if (PageWriteback(page
)) {
4575 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4577 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4582 * It is safe to change page->mem_cgroup here because the page
4583 * is referenced, charged, and isolated - we can't race with
4584 * uncharging, charging, migration, or LRU putback.
4587 /* caller should have done css_get */
4588 page
->mem_cgroup
= to
;
4589 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4593 local_irq_disable();
4594 mem_cgroup_charge_statistics(to
, page
, nr_pages
);
4595 memcg_check_events(to
, page
);
4596 mem_cgroup_charge_statistics(from
, page
, -nr_pages
);
4597 memcg_check_events(from
, page
);
4605 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4606 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4608 struct page
*page
= NULL
;
4609 enum mc_target_type ret
= MC_TARGET_NONE
;
4610 swp_entry_t ent
= { .val
= 0 };
4612 if (pte_present(ptent
))
4613 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4614 else if (is_swap_pte(ptent
))
4615 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
4616 else if (pte_none(ptent
))
4617 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4619 if (!page
&& !ent
.val
)
4623 * Do only loose check w/o serialization.
4624 * mem_cgroup_move_account() checks the page is valid or
4625 * not under LRU exclusion.
4627 if (page
->mem_cgroup
== mc
.from
) {
4628 ret
= MC_TARGET_PAGE
;
4630 target
->page
= page
;
4632 if (!ret
|| !target
)
4635 /* There is a swap entry and a page doesn't exist or isn't charged */
4636 if (ent
.val
&& !ret
&&
4637 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4638 ret
= MC_TARGET_SWAP
;
4645 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4647 * We don't consider swapping or file mapped pages because THP does not
4648 * support them for now.
4649 * Caller should make sure that pmd_trans_huge(pmd) is true.
4651 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4652 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4654 struct page
*page
= NULL
;
4655 enum mc_target_type ret
= MC_TARGET_NONE
;
4657 page
= pmd_page(pmd
);
4658 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4659 if (!(mc
.flags
& MOVE_ANON
))
4661 if (page
->mem_cgroup
== mc
.from
) {
4662 ret
= MC_TARGET_PAGE
;
4665 target
->page
= page
;
4671 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4672 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4674 return MC_TARGET_NONE
;
4678 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4679 unsigned long addr
, unsigned long end
,
4680 struct mm_walk
*walk
)
4682 struct vm_area_struct
*vma
= walk
->vma
;
4686 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
4687 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4688 mc
.precharge
+= HPAGE_PMD_NR
;
4693 if (pmd_trans_unstable(pmd
))
4695 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4696 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4697 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4698 mc
.precharge
++; /* increment precharge temporarily */
4699 pte_unmap_unlock(pte
- 1, ptl
);
4705 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4707 unsigned long precharge
;
4709 struct mm_walk mem_cgroup_count_precharge_walk
= {
4710 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4713 down_read(&mm
->mmap_sem
);
4714 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk
);
4715 up_read(&mm
->mmap_sem
);
4717 precharge
= mc
.precharge
;
4723 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4725 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4727 VM_BUG_ON(mc
.moving_task
);
4728 mc
.moving_task
= current
;
4729 return mem_cgroup_do_precharge(precharge
);
4732 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4733 static void __mem_cgroup_clear_mc(void)
4735 struct mem_cgroup
*from
= mc
.from
;
4736 struct mem_cgroup
*to
= mc
.to
;
4738 /* we must uncharge all the leftover precharges from mc.to */
4740 cancel_charge(mc
.to
, mc
.precharge
);
4744 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4745 * we must uncharge here.
4747 if (mc
.moved_charge
) {
4748 cancel_charge(mc
.from
, mc
.moved_charge
);
4749 mc
.moved_charge
= 0;
4751 /* we must fixup refcnts and charges */
4752 if (mc
.moved_swap
) {
4753 /* uncharge swap account from the old cgroup */
4754 if (!mem_cgroup_is_root(mc
.from
))
4755 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4758 * we charged both to->memory and to->memsw, so we
4759 * should uncharge to->memory.
4761 if (!mem_cgroup_is_root(mc
.to
))
4762 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4764 css_put_many(&mc
.from
->css
, mc
.moved_swap
);
4766 /* we've already done css_get(mc.to) */
4769 memcg_oom_recover(from
);
4770 memcg_oom_recover(to
);
4771 wake_up_all(&mc
.waitq
);
4774 static void mem_cgroup_clear_mc(void)
4777 * we must clear moving_task before waking up waiters at the end of
4780 mc
.moving_task
= NULL
;
4781 __mem_cgroup_clear_mc();
4782 spin_lock(&mc
.lock
);
4785 spin_unlock(&mc
.lock
);
4788 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4790 struct cgroup_subsys_state
*css
;
4791 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
4792 struct mem_cgroup
*from
;
4793 struct task_struct
*leader
, *p
;
4794 struct mm_struct
*mm
;
4795 unsigned long move_flags
;
4798 /* charge immigration isn't supported on the default hierarchy */
4799 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4803 * Multi-process migrations only happen on the default hierarchy
4804 * where charge immigration is not used. Perform charge
4805 * immigration if @tset contains a leader and whine if there are
4809 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
4812 memcg
= mem_cgroup_from_css(css
);
4818 * We are now commited to this value whatever it is. Changes in this
4819 * tunable will only affect upcoming migrations, not the current one.
4820 * So we need to save it, and keep it going.
4822 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
4826 from
= mem_cgroup_from_task(p
);
4828 VM_BUG_ON(from
== memcg
);
4830 mm
= get_task_mm(p
);
4833 /* We move charges only when we move a owner of the mm */
4834 if (mm
->owner
== p
) {
4837 VM_BUG_ON(mc
.precharge
);
4838 VM_BUG_ON(mc
.moved_charge
);
4839 VM_BUG_ON(mc
.moved_swap
);
4841 spin_lock(&mc
.lock
);
4844 mc
.flags
= move_flags
;
4845 spin_unlock(&mc
.lock
);
4846 /* We set mc.moving_task later */
4848 ret
= mem_cgroup_precharge_mc(mm
);
4850 mem_cgroup_clear_mc();
4856 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4859 mem_cgroup_clear_mc();
4862 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4863 unsigned long addr
, unsigned long end
,
4864 struct mm_walk
*walk
)
4867 struct vm_area_struct
*vma
= walk
->vma
;
4870 enum mc_target_type target_type
;
4871 union mc_target target
;
4875 * We don't take compound_lock() here but no race with splitting thp
4877 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
4878 * under splitting, which means there's no concurrent thp split,
4879 * - if another thread runs into split_huge_page() just after we
4880 * entered this if-block, the thread must wait for page table lock
4881 * to be unlocked in __split_huge_page_splitting(), where the main
4882 * part of thp split is not executed yet.
4884 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
4885 if (mc
.precharge
< HPAGE_PMD_NR
) {
4889 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
4890 if (target_type
== MC_TARGET_PAGE
) {
4892 if (!isolate_lru_page(page
)) {
4893 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
4895 mc
.precharge
-= HPAGE_PMD_NR
;
4896 mc
.moved_charge
+= HPAGE_PMD_NR
;
4898 putback_lru_page(page
);
4906 if (pmd_trans_unstable(pmd
))
4909 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4910 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4911 pte_t ptent
= *(pte
++);
4917 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
4918 case MC_TARGET_PAGE
:
4920 if (isolate_lru_page(page
))
4922 if (!mem_cgroup_move_account(page
, 1, mc
.from
, mc
.to
)) {
4924 /* we uncharge from mc.from later. */
4927 putback_lru_page(page
);
4928 put
: /* get_mctgt_type() gets the page */
4931 case MC_TARGET_SWAP
:
4933 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
4935 /* we fixup refcnts and charges later. */
4943 pte_unmap_unlock(pte
- 1, ptl
);
4948 * We have consumed all precharges we got in can_attach().
4949 * We try charge one by one, but don't do any additional
4950 * charges to mc.to if we have failed in charge once in attach()
4953 ret
= mem_cgroup_do_precharge(1);
4961 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
4963 struct mm_walk mem_cgroup_move_charge_walk
= {
4964 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4968 lru_add_drain_all();
4970 * Signal mem_cgroup_begin_page_stat() to take the memcg's
4971 * move_lock while we're moving its pages to another memcg.
4972 * Then wait for already started RCU-only updates to finish.
4974 atomic_inc(&mc
.from
->moving_account
);
4977 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
4979 * Someone who are holding the mmap_sem might be waiting in
4980 * waitq. So we cancel all extra charges, wake up all waiters,
4981 * and retry. Because we cancel precharges, we might not be able
4982 * to move enough charges, but moving charge is a best-effort
4983 * feature anyway, so it wouldn't be a big problem.
4985 __mem_cgroup_clear_mc();
4990 * When we have consumed all precharges and failed in doing
4991 * additional charge, the page walk just aborts.
4993 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk
);
4994 up_read(&mm
->mmap_sem
);
4995 atomic_dec(&mc
.from
->moving_account
);
4998 static void mem_cgroup_move_task(struct cgroup_taskset
*tset
)
5000 struct cgroup_subsys_state
*css
;
5001 struct task_struct
*p
= cgroup_taskset_first(tset
, &css
);
5002 struct mm_struct
*mm
= get_task_mm(p
);
5006 mem_cgroup_move_charge(mm
);
5010 mem_cgroup_clear_mc();
5012 #else /* !CONFIG_MMU */
5013 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5017 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5020 static void mem_cgroup_move_task(struct cgroup_taskset
*tset
)
5026 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5027 * to verify whether we're attached to the default hierarchy on each mount
5030 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5033 * use_hierarchy is forced on the default hierarchy. cgroup core
5034 * guarantees that @root doesn't have any children, so turning it
5035 * on for the root memcg is enough.
5037 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5038 root_mem_cgroup
->use_hierarchy
= true;
5040 root_mem_cgroup
->use_hierarchy
= false;
5043 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5046 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5048 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
5051 static int memory_low_show(struct seq_file
*m
, void *v
)
5053 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5054 unsigned long low
= READ_ONCE(memcg
->low
);
5056 if (low
== PAGE_COUNTER_MAX
)
5057 seq_puts(m
, "max\n");
5059 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
5064 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5065 char *buf
, size_t nbytes
, loff_t off
)
5067 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5071 buf
= strstrip(buf
);
5072 err
= page_counter_memparse(buf
, "max", &low
);
5081 static int memory_high_show(struct seq_file
*m
, void *v
)
5083 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5084 unsigned long high
= READ_ONCE(memcg
->high
);
5086 if (high
== PAGE_COUNTER_MAX
)
5087 seq_puts(m
, "max\n");
5089 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
5094 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5095 char *buf
, size_t nbytes
, loff_t off
)
5097 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5101 buf
= strstrip(buf
);
5102 err
= page_counter_memparse(buf
, "max", &high
);
5108 memcg_wb_domain_size_changed(memcg
);
5112 static int memory_max_show(struct seq_file
*m
, void *v
)
5114 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5115 unsigned long max
= READ_ONCE(memcg
->memory
.limit
);
5117 if (max
== PAGE_COUNTER_MAX
)
5118 seq_puts(m
, "max\n");
5120 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5125 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5126 char *buf
, size_t nbytes
, loff_t off
)
5128 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5132 buf
= strstrip(buf
);
5133 err
= page_counter_memparse(buf
, "max", &max
);
5137 err
= mem_cgroup_resize_limit(memcg
, max
);
5141 memcg_wb_domain_size_changed(memcg
);
5145 static int memory_events_show(struct seq_file
*m
, void *v
)
5147 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5149 seq_printf(m
, "low %lu\n", mem_cgroup_read_events(memcg
, MEMCG_LOW
));
5150 seq_printf(m
, "high %lu\n", mem_cgroup_read_events(memcg
, MEMCG_HIGH
));
5151 seq_printf(m
, "max %lu\n", mem_cgroup_read_events(memcg
, MEMCG_MAX
));
5152 seq_printf(m
, "oom %lu\n", mem_cgroup_read_events(memcg
, MEMCG_OOM
));
5157 static struct cftype memory_files
[] = {
5160 .flags
= CFTYPE_NOT_ON_ROOT
,
5161 .read_u64
= memory_current_read
,
5165 .flags
= CFTYPE_NOT_ON_ROOT
,
5166 .seq_show
= memory_low_show
,
5167 .write
= memory_low_write
,
5171 .flags
= CFTYPE_NOT_ON_ROOT
,
5172 .seq_show
= memory_high_show
,
5173 .write
= memory_high_write
,
5177 .flags
= CFTYPE_NOT_ON_ROOT
,
5178 .seq_show
= memory_max_show
,
5179 .write
= memory_max_write
,
5183 .flags
= CFTYPE_NOT_ON_ROOT
,
5184 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
5185 .seq_show
= memory_events_show
,
5190 struct cgroup_subsys memory_cgrp_subsys
= {
5191 .css_alloc
= mem_cgroup_css_alloc
,
5192 .css_online
= mem_cgroup_css_online
,
5193 .css_offline
= mem_cgroup_css_offline
,
5194 .css_released
= mem_cgroup_css_released
,
5195 .css_free
= mem_cgroup_css_free
,
5196 .css_reset
= mem_cgroup_css_reset
,
5197 .can_attach
= mem_cgroup_can_attach
,
5198 .cancel_attach
= mem_cgroup_cancel_attach
,
5199 .attach
= mem_cgroup_move_task
,
5200 .bind
= mem_cgroup_bind
,
5201 .dfl_cftypes
= memory_files
,
5202 .legacy_cftypes
= mem_cgroup_legacy_files
,
5207 * mem_cgroup_low - check if memory consumption is below the normal range
5208 * @root: the highest ancestor to consider
5209 * @memcg: the memory cgroup to check
5211 * Returns %true if memory consumption of @memcg, and that of all
5212 * configurable ancestors up to @root, is below the normal range.
5214 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5216 if (mem_cgroup_disabled())
5220 * The toplevel group doesn't have a configurable range, so
5221 * it's never low when looked at directly, and it is not
5222 * considered an ancestor when assessing the hierarchy.
5225 if (memcg
== root_mem_cgroup
)
5228 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5231 while (memcg
!= root
) {
5232 memcg
= parent_mem_cgroup(memcg
);
5234 if (memcg
== root_mem_cgroup
)
5237 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5244 * mem_cgroup_try_charge - try charging a page
5245 * @page: page to charge
5246 * @mm: mm context of the victim
5247 * @gfp_mask: reclaim mode
5248 * @memcgp: charged memcg return
5250 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5251 * pages according to @gfp_mask if necessary.
5253 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5254 * Otherwise, an error code is returned.
5256 * After page->mapping has been set up, the caller must finalize the
5257 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5258 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5260 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5261 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
5263 struct mem_cgroup
*memcg
= NULL
;
5264 unsigned int nr_pages
= 1;
5267 if (mem_cgroup_disabled())
5270 if (PageSwapCache(page
)) {
5272 * Every swap fault against a single page tries to charge the
5273 * page, bail as early as possible. shmem_unuse() encounters
5274 * already charged pages, too. The USED bit is protected by
5275 * the page lock, which serializes swap cache removal, which
5276 * in turn serializes uncharging.
5278 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5279 if (page
->mem_cgroup
)
5282 if (do_memsw_account()) {
5283 swp_entry_t ent
= { .val
= page_private(page
), };
5284 unsigned short id
= lookup_swap_cgroup_id(ent
);
5287 memcg
= mem_cgroup_from_id(id
);
5288 if (memcg
&& !css_tryget_online(&memcg
->css
))
5294 if (PageTransHuge(page
)) {
5295 nr_pages
<<= compound_order(page
);
5296 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5300 memcg
= get_mem_cgroup_from_mm(mm
);
5302 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5304 css_put(&memcg
->css
);
5311 * mem_cgroup_commit_charge - commit a page charge
5312 * @page: page to charge
5313 * @memcg: memcg to charge the page to
5314 * @lrucare: page might be on LRU already
5316 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5317 * after page->mapping has been set up. This must happen atomically
5318 * as part of the page instantiation, i.e. under the page table lock
5319 * for anonymous pages, under the page lock for page and swap cache.
5321 * In addition, the page must not be on the LRU during the commit, to
5322 * prevent racing with task migration. If it might be, use @lrucare.
5324 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5326 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5329 unsigned int nr_pages
= 1;
5331 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5332 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5334 if (mem_cgroup_disabled())
5337 * Swap faults will attempt to charge the same page multiple
5338 * times. But reuse_swap_page() might have removed the page
5339 * from swapcache already, so we can't check PageSwapCache().
5344 commit_charge(page
, memcg
, lrucare
);
5346 if (PageTransHuge(page
)) {
5347 nr_pages
<<= compound_order(page
);
5348 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5351 local_irq_disable();
5352 mem_cgroup_charge_statistics(memcg
, page
, nr_pages
);
5353 memcg_check_events(memcg
, page
);
5356 if (do_memsw_account() && PageSwapCache(page
)) {
5357 swp_entry_t entry
= { .val
= page_private(page
) };
5359 * The swap entry might not get freed for a long time,
5360 * let's not wait for it. The page already received a
5361 * memory+swap charge, drop the swap entry duplicate.
5363 mem_cgroup_uncharge_swap(entry
);
5368 * mem_cgroup_cancel_charge - cancel a page charge
5369 * @page: page to charge
5370 * @memcg: memcg to charge the page to
5372 * Cancel a charge transaction started by mem_cgroup_try_charge().
5374 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
)
5376 unsigned int nr_pages
= 1;
5378 if (mem_cgroup_disabled())
5381 * Swap faults will attempt to charge the same page multiple
5382 * times. But reuse_swap_page() might have removed the page
5383 * from swapcache already, so we can't check PageSwapCache().
5388 if (PageTransHuge(page
)) {
5389 nr_pages
<<= compound_order(page
);
5390 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5393 cancel_charge(memcg
, nr_pages
);
5396 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5397 unsigned long nr_anon
, unsigned long nr_file
,
5398 unsigned long nr_huge
, struct page
*dummy_page
)
5400 unsigned long nr_pages
= nr_anon
+ nr_file
;
5401 unsigned long flags
;
5403 if (!mem_cgroup_is_root(memcg
)) {
5404 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5405 if (do_memsw_account())
5406 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5407 memcg_oom_recover(memcg
);
5410 local_irq_save(flags
);
5411 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5412 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5413 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5414 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5415 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5416 memcg_check_events(memcg
, dummy_page
);
5417 local_irq_restore(flags
);
5419 if (!mem_cgroup_is_root(memcg
))
5420 css_put_many(&memcg
->css
, nr_pages
);
5423 static void uncharge_list(struct list_head
*page_list
)
5425 struct mem_cgroup
*memcg
= NULL
;
5426 unsigned long nr_anon
= 0;
5427 unsigned long nr_file
= 0;
5428 unsigned long nr_huge
= 0;
5429 unsigned long pgpgout
= 0;
5430 struct list_head
*next
;
5433 next
= page_list
->next
;
5435 unsigned int nr_pages
= 1;
5437 page
= list_entry(next
, struct page
, lru
);
5438 next
= page
->lru
.next
;
5440 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5441 VM_BUG_ON_PAGE(page_count(page
), page
);
5443 if (!page
->mem_cgroup
)
5447 * Nobody should be changing or seriously looking at
5448 * page->mem_cgroup at this point, we have fully
5449 * exclusive access to the page.
5452 if (memcg
!= page
->mem_cgroup
) {
5454 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5456 pgpgout
= nr_anon
= nr_file
= nr_huge
= 0;
5458 memcg
= page
->mem_cgroup
;
5461 if (PageTransHuge(page
)) {
5462 nr_pages
<<= compound_order(page
);
5463 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5464 nr_huge
+= nr_pages
;
5468 nr_anon
+= nr_pages
;
5470 nr_file
+= nr_pages
;
5472 page
->mem_cgroup
= NULL
;
5475 } while (next
!= page_list
);
5478 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5483 * mem_cgroup_uncharge - uncharge a page
5484 * @page: page to uncharge
5486 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5487 * mem_cgroup_commit_charge().
5489 void mem_cgroup_uncharge(struct page
*page
)
5491 if (mem_cgroup_disabled())
5494 /* Don't touch page->lru of any random page, pre-check: */
5495 if (!page
->mem_cgroup
)
5498 INIT_LIST_HEAD(&page
->lru
);
5499 uncharge_list(&page
->lru
);
5503 * mem_cgroup_uncharge_list - uncharge a list of page
5504 * @page_list: list of pages to uncharge
5506 * Uncharge a list of pages previously charged with
5507 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5509 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5511 if (mem_cgroup_disabled())
5514 if (!list_empty(page_list
))
5515 uncharge_list(page_list
);
5519 * mem_cgroup_replace_page - migrate a charge to another page
5520 * @oldpage: currently charged page
5521 * @newpage: page to transfer the charge to
5523 * Migrate the charge from @oldpage to @newpage.
5525 * Both pages must be locked, @newpage->mapping must be set up.
5526 * Either or both pages might be on the LRU already.
5528 void mem_cgroup_replace_page(struct page
*oldpage
, struct page
*newpage
)
5530 struct mem_cgroup
*memcg
;
5533 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5534 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5535 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5536 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5539 if (mem_cgroup_disabled())
5542 /* Page cache replacement: new page already charged? */
5543 if (newpage
->mem_cgroup
)
5546 /* Swapcache readahead pages can get replaced before being charged */
5547 memcg
= oldpage
->mem_cgroup
;
5551 lock_page_lru(oldpage
, &isolated
);
5552 oldpage
->mem_cgroup
= NULL
;
5553 unlock_page_lru(oldpage
, isolated
);
5555 commit_charge(newpage
, memcg
, true);
5560 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
5561 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
5563 void sock_update_memcg(struct sock
*sk
)
5565 struct mem_cgroup
*memcg
;
5567 /* Socket cloning can throw us here with sk_cgrp already
5568 * filled. It won't however, necessarily happen from
5569 * process context. So the test for root memcg given
5570 * the current task's memcg won't help us in this case.
5572 * Respecting the original socket's memcg is a better
5573 * decision in this case.
5576 BUG_ON(mem_cgroup_is_root(sk
->sk_memcg
));
5577 css_get(&sk
->sk_memcg
->css
);
5582 memcg
= mem_cgroup_from_task(current
);
5583 if (memcg
== root_mem_cgroup
)
5585 #ifdef CONFIG_MEMCG_KMEM
5586 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcp_mem
.active
)
5589 if (css_tryget_online(&memcg
->css
))
5590 sk
->sk_memcg
= memcg
;
5594 EXPORT_SYMBOL(sock_update_memcg
);
5596 void sock_release_memcg(struct sock
*sk
)
5598 WARN_ON(!sk
->sk_memcg
);
5599 css_put(&sk
->sk_memcg
->css
);
5603 * mem_cgroup_charge_skmem - charge socket memory
5604 * @memcg: memcg to charge
5605 * @nr_pages: number of pages to charge
5607 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5608 * @memcg's configured limit, %false if the charge had to be forced.
5610 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5612 gfp_t gfp_mask
= GFP_KERNEL
;
5614 #ifdef CONFIG_MEMCG_KMEM
5615 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5616 struct page_counter
*counter
;
5618 if (page_counter_try_charge(&memcg
->tcp_mem
.memory_allocated
,
5619 nr_pages
, &counter
)) {
5620 memcg
->tcp_mem
.memory_pressure
= 0;
5623 page_counter_charge(&memcg
->tcp_mem
.memory_allocated
, nr_pages
);
5624 memcg
->tcp_mem
.memory_pressure
= 1;
5628 /* Don't block in the packet receive path */
5630 gfp_mask
= GFP_NOWAIT
;
5632 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
5635 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
5640 * mem_cgroup_uncharge_skmem - uncharge socket memory
5641 * @memcg - memcg to uncharge
5642 * @nr_pages - number of pages to uncharge
5644 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5646 #ifdef CONFIG_MEMCG_KMEM
5647 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5648 page_counter_uncharge(&memcg
->tcp_mem
.memory_allocated
,
5653 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5654 css_put_many(&memcg
->css
, nr_pages
);
5657 #endif /* CONFIG_INET */
5659 static int __init
cgroup_memory(char *s
)
5663 while ((token
= strsep(&s
, ",")) != NULL
) {
5666 if (!strcmp(token
, "nosocket"))
5667 cgroup_memory_nosocket
= true;
5671 __setup("cgroup.memory=", cgroup_memory
);
5674 * subsys_initcall() for memory controller.
5676 * Some parts like hotcpu_notifier() have to be initialized from this context
5677 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5678 * everything that doesn't depend on a specific mem_cgroup structure should
5679 * be initialized from here.
5681 static int __init
mem_cgroup_init(void)
5685 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5687 for_each_possible_cpu(cpu
)
5688 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5691 for_each_node(node
) {
5692 struct mem_cgroup_tree_per_node
*rtpn
;
5695 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5696 node_online(node
) ? node
: NUMA_NO_NODE
);
5698 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5699 struct mem_cgroup_tree_per_zone
*rtpz
;
5701 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
5702 rtpz
->rb_root
= RB_ROOT
;
5703 spin_lock_init(&rtpz
->lock
);
5705 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5710 subsys_initcall(mem_cgroup_init
);
5712 #ifdef CONFIG_MEMCG_SWAP
5714 * mem_cgroup_swapout - transfer a memsw charge to swap
5715 * @page: page whose memsw charge to transfer
5716 * @entry: swap entry to move the charge to
5718 * Transfer the memsw charge of @page to @entry.
5720 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5722 struct mem_cgroup
*memcg
;
5723 unsigned short oldid
;
5725 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5726 VM_BUG_ON_PAGE(page_count(page
), page
);
5728 if (!do_memsw_account())
5731 memcg
= page
->mem_cgroup
;
5733 /* Readahead page, never charged */
5737 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5738 VM_BUG_ON_PAGE(oldid
, page
);
5739 mem_cgroup_swap_statistics(memcg
, true);
5741 page
->mem_cgroup
= NULL
;
5743 if (!mem_cgroup_is_root(memcg
))
5744 page_counter_uncharge(&memcg
->memory
, 1);
5747 * Interrupts should be disabled here because the caller holds the
5748 * mapping->tree_lock lock which is taken with interrupts-off. It is
5749 * important here to have the interrupts disabled because it is the
5750 * only synchronisation we have for udpating the per-CPU variables.
5752 VM_BUG_ON(!irqs_disabled());
5753 mem_cgroup_charge_statistics(memcg
, page
, -1);
5754 memcg_check_events(memcg
, page
);
5758 * mem_cgroup_uncharge_swap - uncharge a swap entry
5759 * @entry: swap entry to uncharge
5761 * Drop the memsw charge associated with @entry.
5763 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5765 struct mem_cgroup
*memcg
;
5768 if (!do_memsw_account())
5771 id
= swap_cgroup_record(entry
, 0);
5773 memcg
= mem_cgroup_from_id(id
);
5775 if (!mem_cgroup_is_root(memcg
))
5776 page_counter_uncharge(&memcg
->memsw
, 1);
5777 mem_cgroup_swap_statistics(memcg
, false);
5778 css_put(&memcg
->css
);
5783 /* for remember boot option*/
5784 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5785 static int really_do_swap_account __initdata
= 1;
5787 static int really_do_swap_account __initdata
;
5790 static int __init
enable_swap_account(char *s
)
5792 if (!strcmp(s
, "1"))
5793 really_do_swap_account
= 1;
5794 else if (!strcmp(s
, "0"))
5795 really_do_swap_account
= 0;
5798 __setup("swapaccount=", enable_swap_account
);
5800 static struct cftype memsw_cgroup_files
[] = {
5802 .name
= "memsw.usage_in_bytes",
5803 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5804 .read_u64
= mem_cgroup_read_u64
,
5807 .name
= "memsw.max_usage_in_bytes",
5808 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5809 .write
= mem_cgroup_reset
,
5810 .read_u64
= mem_cgroup_read_u64
,
5813 .name
= "memsw.limit_in_bytes",
5814 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5815 .write
= mem_cgroup_write
,
5816 .read_u64
= mem_cgroup_read_u64
,
5819 .name
= "memsw.failcnt",
5820 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5821 .write
= mem_cgroup_reset
,
5822 .read_u64
= mem_cgroup_read_u64
,
5824 { }, /* terminate */
5827 static int __init
mem_cgroup_swap_init(void)
5829 if (!mem_cgroup_disabled() && really_do_swap_account
) {
5830 do_swap_account
= 1;
5831 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
5832 memsw_cgroup_files
));
5836 subsys_initcall(mem_cgroup_swap_init
);
5838 #endif /* CONFIG_MEMCG_SWAP */