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 #define MEM_CGROUP_RECLAIM_RETRIES 5
80 static struct mem_cgroup
*root_mem_cgroup __read_mostly
;
81 struct cgroup_subsys_state
*mem_cgroup_root_css __read_mostly
;
83 /* Whether the swap controller is active */
84 #ifdef CONFIG_MEMCG_SWAP
85 int do_swap_account __read_mostly
;
87 #define do_swap_account 0
90 static const char * const mem_cgroup_stat_names
[] = {
100 static const char * const mem_cgroup_events_names
[] = {
107 static const char * const mem_cgroup_lru_names
[] = {
115 #define THRESHOLDS_EVENTS_TARGET 128
116 #define SOFTLIMIT_EVENTS_TARGET 1024
117 #define NUMAINFO_EVENTS_TARGET 1024
120 * Cgroups above their limits are maintained in a RB-Tree, independent of
121 * their hierarchy representation
124 struct mem_cgroup_tree_per_zone
{
125 struct rb_root rb_root
;
129 struct mem_cgroup_tree_per_node
{
130 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
133 struct mem_cgroup_tree
{
134 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
137 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
140 struct mem_cgroup_eventfd_list
{
141 struct list_head list
;
142 struct eventfd_ctx
*eventfd
;
146 * cgroup_event represents events which userspace want to receive.
148 struct mem_cgroup_event
{
150 * memcg which the event belongs to.
152 struct mem_cgroup
*memcg
;
154 * eventfd to signal userspace about the event.
156 struct eventfd_ctx
*eventfd
;
158 * Each of these stored in a list by the cgroup.
160 struct list_head list
;
162 * register_event() callback will be used to add new userspace
163 * waiter for changes related to this event. Use eventfd_signal()
164 * on eventfd to send notification to userspace.
166 int (*register_event
)(struct mem_cgroup
*memcg
,
167 struct eventfd_ctx
*eventfd
, const char *args
);
169 * unregister_event() callback will be called when userspace closes
170 * the eventfd or on cgroup removing. This callback must be set,
171 * if you want provide notification functionality.
173 void (*unregister_event
)(struct mem_cgroup
*memcg
,
174 struct eventfd_ctx
*eventfd
);
176 * All fields below needed to unregister event when
177 * userspace closes eventfd.
180 wait_queue_head_t
*wqh
;
182 struct work_struct remove
;
185 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
186 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
188 /* Stuffs for move charges at task migration. */
190 * Types of charges to be moved.
192 #define MOVE_ANON 0x1U
193 #define MOVE_FILE 0x2U
194 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
196 /* "mc" and its members are protected by cgroup_mutex */
197 static struct move_charge_struct
{
198 spinlock_t lock
; /* for from, to */
199 struct mem_cgroup
*from
;
200 struct mem_cgroup
*to
;
202 unsigned long precharge
;
203 unsigned long moved_charge
;
204 unsigned long moved_swap
;
205 struct task_struct
*moving_task
; /* a task moving charges */
206 wait_queue_head_t waitq
; /* a waitq for other context */
208 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
209 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
213 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
214 * limit reclaim to prevent infinite loops, if they ever occur.
216 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
217 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
220 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
221 MEM_CGROUP_CHARGE_TYPE_ANON
,
222 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
223 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
227 /* for encoding cft->private value on file */
235 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
236 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
237 #define MEMFILE_ATTR(val) ((val) & 0xffff)
238 /* Used for OOM nofiier */
239 #define OOM_CONTROL (0)
242 * The memcg_create_mutex will be held whenever a new cgroup is created.
243 * As a consequence, any change that needs to protect against new child cgroups
244 * appearing has to hold it as well.
246 static DEFINE_MUTEX(memcg_create_mutex
);
248 /* Some nice accessors for the vmpressure. */
249 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
252 memcg
= root_mem_cgroup
;
253 return &memcg
->vmpressure
;
256 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
258 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
261 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
263 return (memcg
== root_mem_cgroup
);
267 * We restrict the id in the range of [1, 65535], so it can fit into
270 #define MEM_CGROUP_ID_MAX USHRT_MAX
272 static inline unsigned short mem_cgroup_id(struct mem_cgroup
*memcg
)
274 return memcg
->css
.id
;
278 * A helper function to get mem_cgroup from ID. must be called under
279 * rcu_read_lock(). The caller is responsible for calling
280 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
281 * refcnt from swap can be called against removed memcg.)
283 static inline struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
285 struct cgroup_subsys_state
*css
;
287 css
= css_from_id(id
, &memory_cgrp_subsys
);
288 return mem_cgroup_from_css(css
);
291 /* Writing them here to avoid exposing memcg's inner layout */
292 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
294 void sock_update_memcg(struct sock
*sk
)
296 if (mem_cgroup_sockets_enabled
) {
297 struct mem_cgroup
*memcg
;
298 struct cg_proto
*cg_proto
;
300 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
302 /* Socket cloning can throw us here with sk_cgrp already
303 * filled. It won't however, necessarily happen from
304 * process context. So the test for root memcg given
305 * the current task's memcg won't help us in this case.
307 * Respecting the original socket's memcg is a better
308 * decision in this case.
311 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
312 css_get(&sk
->sk_cgrp
->memcg
->css
);
317 memcg
= mem_cgroup_from_task(current
);
318 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
319 if (cg_proto
&& cg_proto
->active
&&
320 css_tryget_online(&memcg
->css
)) {
321 sk
->sk_cgrp
= cg_proto
;
326 EXPORT_SYMBOL(sock_update_memcg
);
328 void sock_release_memcg(struct sock
*sk
)
330 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
331 struct mem_cgroup
*memcg
;
332 WARN_ON(!sk
->sk_cgrp
->memcg
);
333 memcg
= sk
->sk_cgrp
->memcg
;
334 css_put(&sk
->sk_cgrp
->memcg
->css
);
338 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
340 if (!memcg
|| mem_cgroup_is_root(memcg
))
343 return &memcg
->tcp_mem
;
345 EXPORT_SYMBOL(tcp_proto_cgroup
);
349 #ifdef CONFIG_MEMCG_KMEM
351 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
352 * The main reason for not using cgroup id for this:
353 * this works better in sparse environments, where we have a lot of memcgs,
354 * but only a few kmem-limited. Or also, if we have, for instance, 200
355 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
356 * 200 entry array for that.
358 * The current size of the caches array is stored in memcg_nr_cache_ids. It
359 * will double each time we have to increase it.
361 static DEFINE_IDA(memcg_cache_ida
);
362 int memcg_nr_cache_ids
;
364 /* Protects memcg_nr_cache_ids */
365 static DECLARE_RWSEM(memcg_cache_ids_sem
);
367 void memcg_get_cache_ids(void)
369 down_read(&memcg_cache_ids_sem
);
372 void memcg_put_cache_ids(void)
374 up_read(&memcg_cache_ids_sem
);
378 * MIN_SIZE is different than 1, because we would like to avoid going through
379 * the alloc/free process all the time. In a small machine, 4 kmem-limited
380 * cgroups is a reasonable guess. In the future, it could be a parameter or
381 * tunable, but that is strictly not necessary.
383 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
384 * this constant directly from cgroup, but it is understandable that this is
385 * better kept as an internal representation in cgroup.c. In any case, the
386 * cgrp_id space is not getting any smaller, and we don't have to necessarily
387 * increase ours as well if it increases.
389 #define MEMCG_CACHES_MIN_SIZE 4
390 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
393 * A lot of the calls to the cache allocation functions are expected to be
394 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
395 * conditional to this static branch, we'll have to allow modules that does
396 * kmem_cache_alloc and the such to see this symbol as well
398 struct static_key memcg_kmem_enabled_key
;
399 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
401 #endif /* CONFIG_MEMCG_KMEM */
403 static struct mem_cgroup_per_zone
*
404 mem_cgroup_zone_zoneinfo(struct mem_cgroup
*memcg
, struct zone
*zone
)
406 int nid
= zone_to_nid(zone
);
407 int zid
= zone_idx(zone
);
409 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
413 * mem_cgroup_css_from_page - css of the memcg associated with a page
414 * @page: page of interest
416 * If memcg is bound to the default hierarchy, css of the memcg associated
417 * with @page is returned. The returned css remains associated with @page
418 * until it is released.
420 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
423 * XXX: The above description of behavior on the default hierarchy isn't
424 * strictly true yet as replace_page_cache_page() can modify the
425 * association before @page is released even on the default hierarchy;
426 * however, the current and planned usages don't mix the the two functions
427 * and replace_page_cache_page() will soon be updated to make the invariant
430 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
432 struct mem_cgroup
*memcg
;
436 memcg
= page
->mem_cgroup
;
438 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
439 memcg
= root_mem_cgroup
;
446 * page_cgroup_ino - return inode number of the memcg a page is charged to
449 * Look up the closest online ancestor of the memory cgroup @page is charged to
450 * and return its inode number or 0 if @page is not charged to any cgroup. It
451 * is safe to call this function without holding a reference to @page.
453 * Note, this function is inherently racy, because there is nothing to prevent
454 * the cgroup inode from getting torn down and potentially reallocated a moment
455 * after page_cgroup_ino() returns, so it only should be used by callers that
456 * do not care (such as procfs interfaces).
458 ino_t
page_cgroup_ino(struct page
*page
)
460 struct mem_cgroup
*memcg
;
461 unsigned long ino
= 0;
464 memcg
= READ_ONCE(page
->mem_cgroup
);
465 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
466 memcg
= parent_mem_cgroup(memcg
);
468 ino
= cgroup_ino(memcg
->css
.cgroup
);
473 static struct mem_cgroup_per_zone
*
474 mem_cgroup_page_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
476 int nid
= page_to_nid(page
);
477 int zid
= page_zonenum(page
);
479 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
482 static struct mem_cgroup_tree_per_zone
*
483 soft_limit_tree_node_zone(int nid
, int zid
)
485 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
488 static struct mem_cgroup_tree_per_zone
*
489 soft_limit_tree_from_page(struct page
*page
)
491 int nid
= page_to_nid(page
);
492 int zid
= page_zonenum(page
);
494 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
497 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone
*mz
,
498 struct mem_cgroup_tree_per_zone
*mctz
,
499 unsigned long new_usage_in_excess
)
501 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
502 struct rb_node
*parent
= NULL
;
503 struct mem_cgroup_per_zone
*mz_node
;
508 mz
->usage_in_excess
= new_usage_in_excess
;
509 if (!mz
->usage_in_excess
)
513 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
515 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
518 * We can't avoid mem cgroups that are over their soft
519 * limit by the same amount
521 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
524 rb_link_node(&mz
->tree_node
, parent
, p
);
525 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
529 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
530 struct mem_cgroup_tree_per_zone
*mctz
)
534 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
538 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
539 struct mem_cgroup_tree_per_zone
*mctz
)
543 spin_lock_irqsave(&mctz
->lock
, flags
);
544 __mem_cgroup_remove_exceeded(mz
, mctz
);
545 spin_unlock_irqrestore(&mctz
->lock
, flags
);
548 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
550 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
551 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
552 unsigned long excess
= 0;
554 if (nr_pages
> soft_limit
)
555 excess
= nr_pages
- soft_limit
;
560 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
562 unsigned long excess
;
563 struct mem_cgroup_per_zone
*mz
;
564 struct mem_cgroup_tree_per_zone
*mctz
;
566 mctz
= soft_limit_tree_from_page(page
);
568 * Necessary to update all ancestors when hierarchy is used.
569 * because their event counter is not touched.
571 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
572 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
573 excess
= soft_limit_excess(memcg
);
575 * We have to update the tree if mz is on RB-tree or
576 * mem is over its softlimit.
578 if (excess
|| mz
->on_tree
) {
581 spin_lock_irqsave(&mctz
->lock
, flags
);
582 /* if on-tree, remove it */
584 __mem_cgroup_remove_exceeded(mz
, mctz
);
586 * Insert again. mz->usage_in_excess will be updated.
587 * If excess is 0, no tree ops.
589 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
590 spin_unlock_irqrestore(&mctz
->lock
, flags
);
595 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
597 struct mem_cgroup_tree_per_zone
*mctz
;
598 struct mem_cgroup_per_zone
*mz
;
602 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
603 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
604 mctz
= soft_limit_tree_node_zone(nid
, zid
);
605 mem_cgroup_remove_exceeded(mz
, mctz
);
610 static struct mem_cgroup_per_zone
*
611 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
613 struct rb_node
*rightmost
= NULL
;
614 struct mem_cgroup_per_zone
*mz
;
618 rightmost
= rb_last(&mctz
->rb_root
);
620 goto done
; /* Nothing to reclaim from */
622 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
624 * Remove the node now but someone else can add it back,
625 * we will to add it back at the end of reclaim to its correct
626 * position in the tree.
628 __mem_cgroup_remove_exceeded(mz
, mctz
);
629 if (!soft_limit_excess(mz
->memcg
) ||
630 !css_tryget_online(&mz
->memcg
->css
))
636 static struct mem_cgroup_per_zone
*
637 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
639 struct mem_cgroup_per_zone
*mz
;
641 spin_lock_irq(&mctz
->lock
);
642 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
643 spin_unlock_irq(&mctz
->lock
);
648 * Return page count for single (non recursive) @memcg.
650 * Implementation Note: reading percpu statistics for memcg.
652 * Both of vmstat[] and percpu_counter has threshold and do periodic
653 * synchronization to implement "quick" read. There are trade-off between
654 * reading cost and precision of value. Then, we may have a chance to implement
655 * a periodic synchronization of counter in memcg's counter.
657 * But this _read() function is used for user interface now. The user accounts
658 * memory usage by memory cgroup and he _always_ requires exact value because
659 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
660 * have to visit all online cpus and make sum. So, for now, unnecessary
661 * synchronization is not implemented. (just implemented for cpu hotplug)
663 * If there are kernel internal actions which can make use of some not-exact
664 * value, and reading all cpu value can be performance bottleneck in some
665 * common workload, threshold and synchronization as vmstat[] should be
669 mem_cgroup_read_stat(struct mem_cgroup
*memcg
, enum mem_cgroup_stat_index idx
)
674 /* Per-cpu values can be negative, use a signed accumulator */
675 for_each_possible_cpu(cpu
)
676 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
678 * Summing races with updates, so val may be negative. Avoid exposing
679 * transient negative values.
686 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
687 enum mem_cgroup_events_index idx
)
689 unsigned long val
= 0;
692 for_each_possible_cpu(cpu
)
693 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
697 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
702 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
703 * counted as CACHE even if it's on ANON LRU.
706 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
709 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
712 if (PageTransHuge(page
))
713 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
716 /* pagein of a big page is an event. So, ignore page size */
718 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
720 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
721 nr_pages
= -nr_pages
; /* for event */
724 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
727 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
729 unsigned int lru_mask
)
731 unsigned long nr
= 0;
734 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
736 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
737 struct mem_cgroup_per_zone
*mz
;
741 if (!(BIT(lru
) & lru_mask
))
743 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
744 nr
+= mz
->lru_size
[lru
];
750 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
751 unsigned int lru_mask
)
753 unsigned long nr
= 0;
756 for_each_node_state(nid
, N_MEMORY
)
757 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
761 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
762 enum mem_cgroup_events_target target
)
764 unsigned long val
, next
;
766 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
767 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
768 /* from time_after() in jiffies.h */
769 if ((long)next
- (long)val
< 0) {
771 case MEM_CGROUP_TARGET_THRESH
:
772 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
774 case MEM_CGROUP_TARGET_SOFTLIMIT
:
775 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
777 case MEM_CGROUP_TARGET_NUMAINFO
:
778 next
= val
+ NUMAINFO_EVENTS_TARGET
;
783 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
790 * Check events in order.
793 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
795 /* threshold event is triggered in finer grain than soft limit */
796 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
797 MEM_CGROUP_TARGET_THRESH
))) {
799 bool do_numainfo __maybe_unused
;
801 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
802 MEM_CGROUP_TARGET_SOFTLIMIT
);
804 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
805 MEM_CGROUP_TARGET_NUMAINFO
);
807 mem_cgroup_threshold(memcg
);
808 if (unlikely(do_softlimit
))
809 mem_cgroup_update_tree(memcg
, page
);
811 if (unlikely(do_numainfo
))
812 atomic_inc(&memcg
->numainfo_events
);
817 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
820 * mm_update_next_owner() may clear mm->owner to NULL
821 * if it races with swapoff, page migration, etc.
822 * So this can be called with p == NULL.
827 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
829 EXPORT_SYMBOL(mem_cgroup_from_task
);
831 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
833 struct mem_cgroup
*memcg
= NULL
;
838 * Page cache insertions can happen withou an
839 * actual mm context, e.g. during disk probing
840 * on boot, loopback IO, acct() writes etc.
843 memcg
= root_mem_cgroup
;
845 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
846 if (unlikely(!memcg
))
847 memcg
= root_mem_cgroup
;
849 } while (!css_tryget_online(&memcg
->css
));
855 * mem_cgroup_iter - iterate over memory cgroup hierarchy
856 * @root: hierarchy root
857 * @prev: previously returned memcg, NULL on first invocation
858 * @reclaim: cookie for shared reclaim walks, NULL for full walks
860 * Returns references to children of the hierarchy below @root, or
861 * @root itself, or %NULL after a full round-trip.
863 * Caller must pass the return value in @prev on subsequent
864 * invocations for reference counting, or use mem_cgroup_iter_break()
865 * to cancel a hierarchy walk before the round-trip is complete.
867 * Reclaimers can specify a zone and a priority level in @reclaim to
868 * divide up the memcgs in the hierarchy among all concurrent
869 * reclaimers operating on the same zone and priority.
871 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
872 struct mem_cgroup
*prev
,
873 struct mem_cgroup_reclaim_cookie
*reclaim
)
875 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
876 struct cgroup_subsys_state
*css
= NULL
;
877 struct mem_cgroup
*memcg
= NULL
;
878 struct mem_cgroup
*pos
= NULL
;
880 if (mem_cgroup_disabled())
884 root
= root_mem_cgroup
;
886 if (prev
&& !reclaim
)
889 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
898 struct mem_cgroup_per_zone
*mz
;
900 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
901 iter
= &mz
->iter
[reclaim
->priority
];
903 if (prev
&& reclaim
->generation
!= iter
->generation
)
907 pos
= READ_ONCE(iter
->position
);
908 if (!pos
|| css_tryget(&pos
->css
))
911 * css reference reached zero, so iter->position will
912 * be cleared by ->css_released. However, we should not
913 * rely on this happening soon, because ->css_released
914 * is called from a work queue, and by busy-waiting we
915 * might block it. So we clear iter->position right
918 (void)cmpxchg(&iter
->position
, pos
, NULL
);
926 css
= css_next_descendant_pre(css
, &root
->css
);
929 * Reclaimers share the hierarchy walk, and a
930 * new one might jump in right at the end of
931 * the hierarchy - make sure they see at least
932 * one group and restart from the beginning.
940 * Verify the css and acquire a reference. The root
941 * is provided by the caller, so we know it's alive
942 * and kicking, and don't take an extra reference.
944 memcg
= mem_cgroup_from_css(css
);
946 if (css
== &root
->css
)
949 if (css_tryget(css
)) {
951 * Make sure the memcg is initialized:
952 * mem_cgroup_css_online() orders the the
953 * initialization against setting the flag.
955 if (smp_load_acquire(&memcg
->initialized
))
966 * The position could have already been updated by a competing
967 * thread, so check that the value hasn't changed since we read
968 * it to avoid reclaiming from the same cgroup twice.
970 (void)cmpxchg(&iter
->position
, pos
, memcg
);
978 reclaim
->generation
= iter
->generation
;
984 if (prev
&& prev
!= root
)
991 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
992 * @root: hierarchy root
993 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
995 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
996 struct mem_cgroup
*prev
)
999 root
= root_mem_cgroup
;
1000 if (prev
&& prev
!= root
)
1001 css_put(&prev
->css
);
1004 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
1006 struct mem_cgroup
*memcg
= dead_memcg
;
1007 struct mem_cgroup_reclaim_iter
*iter
;
1008 struct mem_cgroup_per_zone
*mz
;
1012 while ((memcg
= parent_mem_cgroup(memcg
))) {
1013 for_each_node(nid
) {
1014 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1015 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
1016 for (i
= 0; i
<= DEF_PRIORITY
; i
++) {
1017 iter
= &mz
->iter
[i
];
1018 cmpxchg(&iter
->position
,
1027 * Iteration constructs for visiting all cgroups (under a tree). If
1028 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1029 * be used for reference counting.
1031 #define for_each_mem_cgroup_tree(iter, root) \
1032 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1034 iter = mem_cgroup_iter(root, iter, NULL))
1036 #define for_each_mem_cgroup(iter) \
1037 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1039 iter = mem_cgroup_iter(NULL, iter, NULL))
1042 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1043 * @zone: zone of the wanted lruvec
1044 * @memcg: memcg of the wanted lruvec
1046 * Returns the lru list vector holding pages for the given @zone and
1047 * @mem. This can be the global zone lruvec, if the memory controller
1050 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1051 struct mem_cgroup
*memcg
)
1053 struct mem_cgroup_per_zone
*mz
;
1054 struct lruvec
*lruvec
;
1056 if (mem_cgroup_disabled()) {
1057 lruvec
= &zone
->lruvec
;
1061 mz
= mem_cgroup_zone_zoneinfo(memcg
, zone
);
1062 lruvec
= &mz
->lruvec
;
1065 * Since a node can be onlined after the mem_cgroup was created,
1066 * we have to be prepared to initialize lruvec->zone here;
1067 * and if offlined then reonlined, we need to reinitialize it.
1069 if (unlikely(lruvec
->zone
!= zone
))
1070 lruvec
->zone
= zone
;
1075 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1077 * @zone: zone of the page
1079 * This function is only safe when following the LRU page isolation
1080 * and putback protocol: the LRU lock must be held, and the page must
1081 * either be PageLRU() or the caller must have isolated/allocated it.
1083 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1085 struct mem_cgroup_per_zone
*mz
;
1086 struct mem_cgroup
*memcg
;
1087 struct lruvec
*lruvec
;
1089 if (mem_cgroup_disabled()) {
1090 lruvec
= &zone
->lruvec
;
1094 memcg
= page
->mem_cgroup
;
1096 * Swapcache readahead pages are added to the LRU - and
1097 * possibly migrated - before they are charged.
1100 memcg
= root_mem_cgroup
;
1102 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1103 lruvec
= &mz
->lruvec
;
1106 * Since a node can be onlined after the mem_cgroup was created,
1107 * we have to be prepared to initialize lruvec->zone here;
1108 * and if offlined then reonlined, we need to reinitialize it.
1110 if (unlikely(lruvec
->zone
!= zone
))
1111 lruvec
->zone
= zone
;
1116 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1117 * @lruvec: mem_cgroup per zone lru vector
1118 * @lru: index of lru list the page is sitting on
1119 * @nr_pages: positive when adding or negative when removing
1121 * This function must be called when a page is added to or removed from an
1124 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1127 struct mem_cgroup_per_zone
*mz
;
1128 unsigned long *lru_size
;
1130 if (mem_cgroup_disabled())
1133 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1134 lru_size
= mz
->lru_size
+ lru
;
1135 *lru_size
+= nr_pages
;
1136 VM_BUG_ON((long)(*lru_size
) < 0);
1139 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1141 struct mem_cgroup
*task_memcg
;
1142 struct task_struct
*p
;
1145 p
= find_lock_task_mm(task
);
1147 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1151 * All threads may have already detached their mm's, but the oom
1152 * killer still needs to detect if they have already been oom
1153 * killed to prevent needlessly killing additional tasks.
1156 task_memcg
= mem_cgroup_from_task(task
);
1157 css_get(&task_memcg
->css
);
1160 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1161 css_put(&task_memcg
->css
);
1165 #define mem_cgroup_from_counter(counter, member) \
1166 container_of(counter, struct mem_cgroup, member)
1169 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1170 * @memcg: the memory cgroup
1172 * Returns the maximum amount of memory @mem can be charged with, in
1175 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1177 unsigned long margin
= 0;
1178 unsigned long count
;
1179 unsigned long limit
;
1181 count
= page_counter_read(&memcg
->memory
);
1182 limit
= READ_ONCE(memcg
->memory
.limit
);
1184 margin
= limit
- count
;
1186 if (do_swap_account
) {
1187 count
= page_counter_read(&memcg
->memsw
);
1188 limit
= READ_ONCE(memcg
->memsw
.limit
);
1190 margin
= min(margin
, limit
- count
);
1197 * A routine for checking "mem" is under move_account() or not.
1199 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1200 * moving cgroups. This is for waiting at high-memory pressure
1203 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1205 struct mem_cgroup
*from
;
1206 struct mem_cgroup
*to
;
1209 * Unlike task_move routines, we access mc.to, mc.from not under
1210 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1212 spin_lock(&mc
.lock
);
1218 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1219 mem_cgroup_is_descendant(to
, memcg
);
1221 spin_unlock(&mc
.lock
);
1225 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1227 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1228 if (mem_cgroup_under_move(memcg
)) {
1230 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1231 /* moving charge context might have finished. */
1234 finish_wait(&mc
.waitq
, &wait
);
1241 #define K(x) ((x) << (PAGE_SHIFT-10))
1243 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1244 * @memcg: The memory cgroup that went over limit
1245 * @p: Task that is going to be killed
1247 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1250 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1252 /* oom_info_lock ensures that parallel ooms do not interleave */
1253 static DEFINE_MUTEX(oom_info_lock
);
1254 struct mem_cgroup
*iter
;
1257 mutex_lock(&oom_info_lock
);
1261 pr_info("Task in ");
1262 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1263 pr_cont(" killed as a result of limit of ");
1265 pr_info("Memory limit reached of cgroup ");
1268 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1273 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1274 K((u64
)page_counter_read(&memcg
->memory
)),
1275 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1276 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1277 K((u64
)page_counter_read(&memcg
->memsw
)),
1278 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1279 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1280 K((u64
)page_counter_read(&memcg
->kmem
)),
1281 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1283 for_each_mem_cgroup_tree(iter
, memcg
) {
1284 pr_info("Memory cgroup stats for ");
1285 pr_cont_cgroup_path(iter
->css
.cgroup
);
1288 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1289 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1291 pr_cont(" %s:%luKB", mem_cgroup_stat_names
[i
],
1292 K(mem_cgroup_read_stat(iter
, i
)));
1295 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1296 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1297 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1301 mutex_unlock(&oom_info_lock
);
1305 * This function returns the number of memcg under hierarchy tree. Returns
1306 * 1(self count) if no children.
1308 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1311 struct mem_cgroup
*iter
;
1313 for_each_mem_cgroup_tree(iter
, memcg
)
1319 * Return the memory (and swap, if configured) limit for a memcg.
1321 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1323 unsigned long limit
;
1325 limit
= memcg
->memory
.limit
;
1326 if (mem_cgroup_swappiness(memcg
)) {
1327 unsigned long memsw_limit
;
1329 memsw_limit
= memcg
->memsw
.limit
;
1330 limit
= min(limit
+ total_swap_pages
, memsw_limit
);
1335 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1338 struct oom_control oc
= {
1341 .gfp_mask
= gfp_mask
,
1344 struct mem_cgroup
*iter
;
1345 unsigned long chosen_points
= 0;
1346 unsigned long totalpages
;
1347 unsigned int points
= 0;
1348 struct task_struct
*chosen
= NULL
;
1350 mutex_lock(&oom_lock
);
1353 * If current has a pending SIGKILL or is exiting, then automatically
1354 * select it. The goal is to allow it to allocate so that it may
1355 * quickly exit and free its memory.
1357 if (fatal_signal_pending(current
) || task_will_free_mem(current
)) {
1358 mark_oom_victim(current
);
1362 check_panic_on_oom(&oc
, CONSTRAINT_MEMCG
, memcg
);
1363 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1364 for_each_mem_cgroup_tree(iter
, memcg
) {
1365 struct css_task_iter it
;
1366 struct task_struct
*task
;
1368 css_task_iter_start(&iter
->css
, &it
);
1369 while ((task
= css_task_iter_next(&it
))) {
1370 switch (oom_scan_process_thread(&oc
, task
, totalpages
)) {
1371 case OOM_SCAN_SELECT
:
1373 put_task_struct(chosen
);
1375 chosen_points
= ULONG_MAX
;
1376 get_task_struct(chosen
);
1378 case OOM_SCAN_CONTINUE
:
1380 case OOM_SCAN_ABORT
:
1381 css_task_iter_end(&it
);
1382 mem_cgroup_iter_break(memcg
, iter
);
1384 put_task_struct(chosen
);
1389 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1390 if (!points
|| points
< chosen_points
)
1392 /* Prefer thread group leaders for display purposes */
1393 if (points
== chosen_points
&&
1394 thread_group_leader(chosen
))
1398 put_task_struct(chosen
);
1400 chosen_points
= points
;
1401 get_task_struct(chosen
);
1403 css_task_iter_end(&it
);
1407 points
= chosen_points
* 1000 / totalpages
;
1408 oom_kill_process(&oc
, chosen
, points
, totalpages
, memcg
,
1409 "Memory cgroup out of memory");
1412 mutex_unlock(&oom_lock
);
1415 #if MAX_NUMNODES > 1
1418 * test_mem_cgroup_node_reclaimable
1419 * @memcg: the target memcg
1420 * @nid: the node ID to be checked.
1421 * @noswap : specify true here if the user wants flle only information.
1423 * This function returns whether the specified memcg contains any
1424 * reclaimable pages on a node. Returns true if there are any reclaimable
1425 * pages in the node.
1427 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1428 int nid
, bool noswap
)
1430 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1432 if (noswap
|| !total_swap_pages
)
1434 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1441 * Always updating the nodemask is not very good - even if we have an empty
1442 * list or the wrong list here, we can start from some node and traverse all
1443 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1446 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1450 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1451 * pagein/pageout changes since the last update.
1453 if (!atomic_read(&memcg
->numainfo_events
))
1455 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1458 /* make a nodemask where this memcg uses memory from */
1459 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1461 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1463 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1464 node_clear(nid
, memcg
->scan_nodes
);
1467 atomic_set(&memcg
->numainfo_events
, 0);
1468 atomic_set(&memcg
->numainfo_updating
, 0);
1472 * Selecting a node where we start reclaim from. Because what we need is just
1473 * reducing usage counter, start from anywhere is O,K. Considering
1474 * memory reclaim from current node, there are pros. and cons.
1476 * Freeing memory from current node means freeing memory from a node which
1477 * we'll use or we've used. So, it may make LRU bad. And if several threads
1478 * hit limits, it will see a contention on a node. But freeing from remote
1479 * node means more costs for memory reclaim because of memory latency.
1481 * Now, we use round-robin. Better algorithm is welcomed.
1483 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1487 mem_cgroup_may_update_nodemask(memcg
);
1488 node
= memcg
->last_scanned_node
;
1490 node
= next_node(node
, memcg
->scan_nodes
);
1491 if (node
== MAX_NUMNODES
)
1492 node
= first_node(memcg
->scan_nodes
);
1494 * We call this when we hit limit, not when pages are added to LRU.
1495 * No LRU may hold pages because all pages are UNEVICTABLE or
1496 * memcg is too small and all pages are not on LRU. In that case,
1497 * we use curret node.
1499 if (unlikely(node
== MAX_NUMNODES
))
1500 node
= numa_node_id();
1502 memcg
->last_scanned_node
= node
;
1506 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1512 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1515 unsigned long *total_scanned
)
1517 struct mem_cgroup
*victim
= NULL
;
1520 unsigned long excess
;
1521 unsigned long nr_scanned
;
1522 struct mem_cgroup_reclaim_cookie reclaim
= {
1527 excess
= soft_limit_excess(root_memcg
);
1530 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1535 * If we have not been able to reclaim
1536 * anything, it might because there are
1537 * no reclaimable pages under this hierarchy
1542 * We want to do more targeted reclaim.
1543 * excess >> 2 is not to excessive so as to
1544 * reclaim too much, nor too less that we keep
1545 * coming back to reclaim from this cgroup
1547 if (total
>= (excess
>> 2) ||
1548 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1553 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1555 *total_scanned
+= nr_scanned
;
1556 if (!soft_limit_excess(root_memcg
))
1559 mem_cgroup_iter_break(root_memcg
, victim
);
1563 #ifdef CONFIG_LOCKDEP
1564 static struct lockdep_map memcg_oom_lock_dep_map
= {
1565 .name
= "memcg_oom_lock",
1569 static DEFINE_SPINLOCK(memcg_oom_lock
);
1572 * Check OOM-Killer is already running under our hierarchy.
1573 * If someone is running, return false.
1575 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1577 struct mem_cgroup
*iter
, *failed
= NULL
;
1579 spin_lock(&memcg_oom_lock
);
1581 for_each_mem_cgroup_tree(iter
, memcg
) {
1582 if (iter
->oom_lock
) {
1584 * this subtree of our hierarchy is already locked
1585 * so we cannot give a lock.
1588 mem_cgroup_iter_break(memcg
, iter
);
1591 iter
->oom_lock
= true;
1596 * OK, we failed to lock the whole subtree so we have
1597 * to clean up what we set up to the failing subtree
1599 for_each_mem_cgroup_tree(iter
, memcg
) {
1600 if (iter
== failed
) {
1601 mem_cgroup_iter_break(memcg
, iter
);
1604 iter
->oom_lock
= false;
1607 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1609 spin_unlock(&memcg_oom_lock
);
1614 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1616 struct mem_cgroup
*iter
;
1618 spin_lock(&memcg_oom_lock
);
1619 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1620 for_each_mem_cgroup_tree(iter
, memcg
)
1621 iter
->oom_lock
= false;
1622 spin_unlock(&memcg_oom_lock
);
1625 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1627 struct mem_cgroup
*iter
;
1629 spin_lock(&memcg_oom_lock
);
1630 for_each_mem_cgroup_tree(iter
, memcg
)
1632 spin_unlock(&memcg_oom_lock
);
1635 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1637 struct mem_cgroup
*iter
;
1640 * When a new child is created while the hierarchy is under oom,
1641 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1643 spin_lock(&memcg_oom_lock
);
1644 for_each_mem_cgroup_tree(iter
, memcg
)
1645 if (iter
->under_oom
> 0)
1647 spin_unlock(&memcg_oom_lock
);
1650 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1652 struct oom_wait_info
{
1653 struct mem_cgroup
*memcg
;
1657 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1658 unsigned mode
, int sync
, void *arg
)
1660 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1661 struct mem_cgroup
*oom_wait_memcg
;
1662 struct oom_wait_info
*oom_wait_info
;
1664 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1665 oom_wait_memcg
= oom_wait_info
->memcg
;
1667 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1668 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1670 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1673 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1676 * For the following lockless ->under_oom test, the only required
1677 * guarantee is that it must see the state asserted by an OOM when
1678 * this function is called as a result of userland actions
1679 * triggered by the notification of the OOM. This is trivially
1680 * achieved by invoking mem_cgroup_mark_under_oom() before
1681 * triggering notification.
1683 if (memcg
&& memcg
->under_oom
)
1684 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1687 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1689 if (!current
->memcg_may_oom
)
1692 * We are in the middle of the charge context here, so we
1693 * don't want to block when potentially sitting on a callstack
1694 * that holds all kinds of filesystem and mm locks.
1696 * Also, the caller may handle a failed allocation gracefully
1697 * (like optional page cache readahead) and so an OOM killer
1698 * invocation might not even be necessary.
1700 * That's why we don't do anything here except remember the
1701 * OOM context and then deal with it at the end of the page
1702 * fault when the stack is unwound, the locks are released,
1703 * and when we know whether the fault was overall successful.
1705 css_get(&memcg
->css
);
1706 current
->memcg_in_oom
= memcg
;
1707 current
->memcg_oom_gfp_mask
= mask
;
1708 current
->memcg_oom_order
= order
;
1712 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1713 * @handle: actually kill/wait or just clean up the OOM state
1715 * This has to be called at the end of a page fault if the memcg OOM
1716 * handler was enabled.
1718 * Memcg supports userspace OOM handling where failed allocations must
1719 * sleep on a waitqueue until the userspace task resolves the
1720 * situation. Sleeping directly in the charge context with all kinds
1721 * of locks held is not a good idea, instead we remember an OOM state
1722 * in the task and mem_cgroup_oom_synchronize() has to be called at
1723 * the end of the page fault to complete the OOM handling.
1725 * Returns %true if an ongoing memcg OOM situation was detected and
1726 * completed, %false otherwise.
1728 bool mem_cgroup_oom_synchronize(bool handle
)
1730 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1731 struct oom_wait_info owait
;
1734 /* OOM is global, do not handle */
1738 if (!handle
|| oom_killer_disabled
)
1741 owait
.memcg
= memcg
;
1742 owait
.wait
.flags
= 0;
1743 owait
.wait
.func
= memcg_oom_wake_function
;
1744 owait
.wait
.private = current
;
1745 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1747 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1748 mem_cgroup_mark_under_oom(memcg
);
1750 locked
= mem_cgroup_oom_trylock(memcg
);
1753 mem_cgroup_oom_notify(memcg
);
1755 if (locked
&& !memcg
->oom_kill_disable
) {
1756 mem_cgroup_unmark_under_oom(memcg
);
1757 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1758 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1759 current
->memcg_oom_order
);
1762 mem_cgroup_unmark_under_oom(memcg
);
1763 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1767 mem_cgroup_oom_unlock(memcg
);
1769 * There is no guarantee that an OOM-lock contender
1770 * sees the wakeups triggered by the OOM kill
1771 * uncharges. Wake any sleepers explicitely.
1773 memcg_oom_recover(memcg
);
1776 current
->memcg_in_oom
= NULL
;
1777 css_put(&memcg
->css
);
1782 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1783 * @page: page that is going to change accounted state
1785 * This function must mark the beginning of an accounted page state
1786 * change to prevent double accounting when the page is concurrently
1787 * being moved to another memcg:
1789 * memcg = mem_cgroup_begin_page_stat(page);
1790 * if (TestClearPageState(page))
1791 * mem_cgroup_update_page_stat(memcg, state, -1);
1792 * mem_cgroup_end_page_stat(memcg);
1794 struct mem_cgroup
*mem_cgroup_begin_page_stat(struct page
*page
)
1796 struct mem_cgroup
*memcg
;
1797 unsigned long flags
;
1800 * The RCU lock is held throughout the transaction. The fast
1801 * path can get away without acquiring the memcg->move_lock
1802 * because page moving starts with an RCU grace period.
1804 * The RCU lock also protects the memcg from being freed when
1805 * the page state that is going to change is the only thing
1806 * preventing the page from being uncharged.
1807 * E.g. end-writeback clearing PageWriteback(), which allows
1808 * migration to go ahead and uncharge the page before the
1809 * account transaction might be complete.
1813 if (mem_cgroup_disabled())
1816 memcg
= page
->mem_cgroup
;
1817 if (unlikely(!memcg
))
1820 if (atomic_read(&memcg
->moving_account
) <= 0)
1823 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1824 if (memcg
!= page
->mem_cgroup
) {
1825 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1830 * When charge migration first begins, we can have locked and
1831 * unlocked page stat updates happening concurrently. Track
1832 * the task who has the lock for mem_cgroup_end_page_stat().
1834 memcg
->move_lock_task
= current
;
1835 memcg
->move_lock_flags
= flags
;
1839 EXPORT_SYMBOL(mem_cgroup_begin_page_stat
);
1842 * mem_cgroup_end_page_stat - finish a page state statistics transaction
1843 * @memcg: the memcg that was accounted against
1845 void mem_cgroup_end_page_stat(struct mem_cgroup
*memcg
)
1847 if (memcg
&& memcg
->move_lock_task
== current
) {
1848 unsigned long flags
= memcg
->move_lock_flags
;
1850 memcg
->move_lock_task
= NULL
;
1851 memcg
->move_lock_flags
= 0;
1853 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1858 EXPORT_SYMBOL(mem_cgroup_end_page_stat
);
1861 * size of first charge trial. "32" comes from vmscan.c's magic value.
1862 * TODO: maybe necessary to use big numbers in big irons.
1864 #define CHARGE_BATCH 32U
1865 struct memcg_stock_pcp
{
1866 struct mem_cgroup
*cached
; /* this never be root cgroup */
1867 unsigned int nr_pages
;
1868 struct work_struct work
;
1869 unsigned long flags
;
1870 #define FLUSHING_CACHED_CHARGE 0
1872 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1873 static DEFINE_MUTEX(percpu_charge_mutex
);
1876 * consume_stock: Try to consume stocked charge on this cpu.
1877 * @memcg: memcg to consume from.
1878 * @nr_pages: how many pages to charge.
1880 * The charges will only happen if @memcg matches the current cpu's memcg
1881 * stock, and at least @nr_pages are available in that stock. Failure to
1882 * service an allocation will refill the stock.
1884 * returns true if successful, false otherwise.
1886 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1888 struct memcg_stock_pcp
*stock
;
1891 if (nr_pages
> CHARGE_BATCH
)
1894 stock
= &get_cpu_var(memcg_stock
);
1895 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
1896 stock
->nr_pages
-= nr_pages
;
1899 put_cpu_var(memcg_stock
);
1904 * Returns stocks cached in percpu and reset cached information.
1906 static void drain_stock(struct memcg_stock_pcp
*stock
)
1908 struct mem_cgroup
*old
= stock
->cached
;
1910 if (stock
->nr_pages
) {
1911 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
1912 if (do_swap_account
)
1913 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
1914 css_put_many(&old
->css
, stock
->nr_pages
);
1915 stock
->nr_pages
= 0;
1917 stock
->cached
= NULL
;
1921 * This must be called under preempt disabled or must be called by
1922 * a thread which is pinned to local cpu.
1924 static void drain_local_stock(struct work_struct
*dummy
)
1926 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
1928 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
1932 * Cache charges(val) to local per_cpu area.
1933 * This will be consumed by consume_stock() function, later.
1935 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1937 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
1939 if (stock
->cached
!= memcg
) { /* reset if necessary */
1941 stock
->cached
= memcg
;
1943 stock
->nr_pages
+= nr_pages
;
1944 put_cpu_var(memcg_stock
);
1948 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1949 * of the hierarchy under it.
1951 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
1955 /* If someone's already draining, avoid adding running more workers. */
1956 if (!mutex_trylock(&percpu_charge_mutex
))
1958 /* Notify other cpus that system-wide "drain" is running */
1961 for_each_online_cpu(cpu
) {
1962 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1963 struct mem_cgroup
*memcg
;
1965 memcg
= stock
->cached
;
1966 if (!memcg
|| !stock
->nr_pages
)
1968 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
1970 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
1972 drain_local_stock(&stock
->work
);
1974 schedule_work_on(cpu
, &stock
->work
);
1979 mutex_unlock(&percpu_charge_mutex
);
1982 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
1983 unsigned long action
,
1986 int cpu
= (unsigned long)hcpu
;
1987 struct memcg_stock_pcp
*stock
;
1989 if (action
== CPU_ONLINE
)
1992 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
1995 stock
= &per_cpu(memcg_stock
, cpu
);
2001 * Scheduled by try_charge() to be executed from the userland return path
2002 * and reclaims memory over the high limit.
2004 void mem_cgroup_handle_over_high(void)
2006 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
2007 struct mem_cgroup
*memcg
, *pos
;
2009 if (likely(!nr_pages
))
2012 pos
= memcg
= get_mem_cgroup_from_mm(current
->mm
);
2015 if (page_counter_read(&pos
->memory
) <= pos
->high
)
2017 mem_cgroup_events(pos
, MEMCG_HIGH
, 1);
2018 try_to_free_mem_cgroup_pages(pos
, nr_pages
, GFP_KERNEL
, true);
2019 } while ((pos
= parent_mem_cgroup(pos
)));
2021 css_put(&memcg
->css
);
2022 current
->memcg_nr_pages_over_high
= 0;
2025 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2026 unsigned int nr_pages
)
2028 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2029 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2030 struct mem_cgroup
*mem_over_limit
;
2031 struct page_counter
*counter
;
2032 unsigned long nr_reclaimed
;
2033 bool may_swap
= true;
2034 bool drained
= false;
2036 if (mem_cgroup_is_root(memcg
))
2039 if (consume_stock(memcg
, nr_pages
))
2042 if (!do_swap_account
||
2043 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2044 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2046 if (do_swap_account
)
2047 page_counter_uncharge(&memcg
->memsw
, batch
);
2048 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2050 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2054 if (batch
> nr_pages
) {
2060 * Unlike in global OOM situations, memcg is not in a physical
2061 * memory shortage. Allow dying and OOM-killed tasks to
2062 * bypass the last charges so that they can exit quickly and
2063 * free their memory.
2065 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
2066 fatal_signal_pending(current
) ||
2067 current
->flags
& PF_EXITING
))
2070 if (unlikely(task_in_memcg_oom(current
)))
2073 if (!gfpflags_allow_blocking(gfp_mask
))
2076 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
2078 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2079 gfp_mask
, may_swap
);
2081 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2085 drain_all_stock(mem_over_limit
);
2090 if (gfp_mask
& __GFP_NORETRY
)
2093 * Even though the limit is exceeded at this point, reclaim
2094 * may have been able to free some pages. Retry the charge
2095 * before killing the task.
2097 * Only for regular pages, though: huge pages are rather
2098 * unlikely to succeed so close to the limit, and we fall back
2099 * to regular pages anyway in case of failure.
2101 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2104 * At task move, charge accounts can be doubly counted. So, it's
2105 * better to wait until the end of task_move if something is going on.
2107 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2113 if (gfp_mask
& __GFP_NOFAIL
)
2116 if (fatal_signal_pending(current
))
2119 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
2121 mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2122 get_order(nr_pages
* PAGE_SIZE
));
2124 if (!(gfp_mask
& __GFP_NOFAIL
))
2128 * The allocation either can't fail or will lead to more memory
2129 * being freed very soon. Allow memory usage go over the limit
2130 * temporarily by force charging it.
2132 page_counter_charge(&memcg
->memory
, nr_pages
);
2133 if (do_swap_account
)
2134 page_counter_charge(&memcg
->memsw
, nr_pages
);
2135 css_get_many(&memcg
->css
, nr_pages
);
2140 css_get_many(&memcg
->css
, batch
);
2141 if (batch
> nr_pages
)
2142 refill_stock(memcg
, batch
- nr_pages
);
2145 * If the hierarchy is above the normal consumption range, schedule
2146 * reclaim on returning to userland. We can perform reclaim here
2147 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2148 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2149 * not recorded as it most likely matches current's and won't
2150 * change in the meantime. As high limit is checked again before
2151 * reclaim, the cost of mismatch is negligible.
2154 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2155 current
->memcg_nr_pages_over_high
+= batch
;
2156 set_notify_resume(current
);
2159 } while ((memcg
= parent_mem_cgroup(memcg
)));
2164 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2166 if (mem_cgroup_is_root(memcg
))
2169 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2170 if (do_swap_account
)
2171 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2173 css_put_many(&memcg
->css
, nr_pages
);
2176 static void lock_page_lru(struct page
*page
, int *isolated
)
2178 struct zone
*zone
= page_zone(page
);
2180 spin_lock_irq(&zone
->lru_lock
);
2181 if (PageLRU(page
)) {
2182 struct lruvec
*lruvec
;
2184 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2186 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2192 static void unlock_page_lru(struct page
*page
, int isolated
)
2194 struct zone
*zone
= page_zone(page
);
2197 struct lruvec
*lruvec
;
2199 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2200 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2202 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2204 spin_unlock_irq(&zone
->lru_lock
);
2207 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2212 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2215 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2216 * may already be on some other mem_cgroup's LRU. Take care of it.
2219 lock_page_lru(page
, &isolated
);
2222 * Nobody should be changing or seriously looking at
2223 * page->mem_cgroup at this point:
2225 * - the page is uncharged
2227 * - the page is off-LRU
2229 * - an anonymous fault has exclusive page access, except for
2230 * a locked page table
2232 * - a page cache insertion, a swapin fault, or a migration
2233 * have the page locked
2235 page
->mem_cgroup
= memcg
;
2238 unlock_page_lru(page
, isolated
);
2241 #ifdef CONFIG_MEMCG_KMEM
2242 static int memcg_alloc_cache_id(void)
2247 id
= ida_simple_get(&memcg_cache_ida
,
2248 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2252 if (id
< memcg_nr_cache_ids
)
2256 * There's no space for the new id in memcg_caches arrays,
2257 * so we have to grow them.
2259 down_write(&memcg_cache_ids_sem
);
2261 size
= 2 * (id
+ 1);
2262 if (size
< MEMCG_CACHES_MIN_SIZE
)
2263 size
= MEMCG_CACHES_MIN_SIZE
;
2264 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2265 size
= MEMCG_CACHES_MAX_SIZE
;
2267 err
= memcg_update_all_caches(size
);
2269 err
= memcg_update_all_list_lrus(size
);
2271 memcg_nr_cache_ids
= size
;
2273 up_write(&memcg_cache_ids_sem
);
2276 ida_simple_remove(&memcg_cache_ida
, id
);
2282 static void memcg_free_cache_id(int id
)
2284 ida_simple_remove(&memcg_cache_ida
, id
);
2287 struct memcg_kmem_cache_create_work
{
2288 struct mem_cgroup
*memcg
;
2289 struct kmem_cache
*cachep
;
2290 struct work_struct work
;
2293 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2295 struct memcg_kmem_cache_create_work
*cw
=
2296 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2297 struct mem_cgroup
*memcg
= cw
->memcg
;
2298 struct kmem_cache
*cachep
= cw
->cachep
;
2300 memcg_create_kmem_cache(memcg
, cachep
);
2302 css_put(&memcg
->css
);
2307 * Enqueue the creation of a per-memcg kmem_cache.
2309 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2310 struct kmem_cache
*cachep
)
2312 struct memcg_kmem_cache_create_work
*cw
;
2314 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2318 css_get(&memcg
->css
);
2321 cw
->cachep
= cachep
;
2322 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2324 schedule_work(&cw
->work
);
2327 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2328 struct kmem_cache
*cachep
)
2331 * We need to stop accounting when we kmalloc, because if the
2332 * corresponding kmalloc cache is not yet created, the first allocation
2333 * in __memcg_schedule_kmem_cache_create will recurse.
2335 * However, it is better to enclose the whole function. Depending on
2336 * the debugging options enabled, INIT_WORK(), for instance, can
2337 * trigger an allocation. This too, will make us recurse. Because at
2338 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2339 * the safest choice is to do it like this, wrapping the whole function.
2341 current
->memcg_kmem_skip_account
= 1;
2342 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2343 current
->memcg_kmem_skip_account
= 0;
2347 * Return the kmem_cache we're supposed to use for a slab allocation.
2348 * We try to use the current memcg's version of the cache.
2350 * If the cache does not exist yet, if we are the first user of it,
2351 * we either create it immediately, if possible, or create it asynchronously
2353 * In the latter case, we will let the current allocation go through with
2354 * the original cache.
2356 * Can't be called in interrupt context or from kernel threads.
2357 * This function needs to be called with rcu_read_lock() held.
2359 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2361 struct mem_cgroup
*memcg
;
2362 struct kmem_cache
*memcg_cachep
;
2365 VM_BUG_ON(!is_root_cache(cachep
));
2367 if (cachep
->flags
& SLAB_ACCOUNT
)
2368 gfp
|= __GFP_ACCOUNT
;
2370 if (!(gfp
& __GFP_ACCOUNT
))
2373 if (current
->memcg_kmem_skip_account
)
2376 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2377 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2381 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2382 if (likely(memcg_cachep
))
2383 return memcg_cachep
;
2386 * If we are in a safe context (can wait, and not in interrupt
2387 * context), we could be be predictable and return right away.
2388 * This would guarantee that the allocation being performed
2389 * already belongs in the new cache.
2391 * However, there are some clashes that can arrive from locking.
2392 * For instance, because we acquire the slab_mutex while doing
2393 * memcg_create_kmem_cache, this means no further allocation
2394 * could happen with the slab_mutex held. So it's better to
2397 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2399 css_put(&memcg
->css
);
2403 void __memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2405 if (!is_root_cache(cachep
))
2406 css_put(&cachep
->memcg_params
.memcg
->css
);
2409 int __memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2410 struct mem_cgroup
*memcg
)
2412 unsigned int nr_pages
= 1 << order
;
2413 struct page_counter
*counter
;
2416 if (!memcg_kmem_is_active(memcg
))
2419 if (!page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
))
2422 ret
= try_charge(memcg
, gfp
, nr_pages
);
2424 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2428 page
->mem_cgroup
= memcg
;
2433 int __memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2435 struct mem_cgroup
*memcg
;
2438 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2439 ret
= __memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2440 css_put(&memcg
->css
);
2444 void __memcg_kmem_uncharge(struct page
*page
, int order
)
2446 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2447 unsigned int nr_pages
= 1 << order
;
2452 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2454 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2455 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2456 if (do_swap_account
)
2457 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2459 page
->mem_cgroup
= NULL
;
2460 css_put_many(&memcg
->css
, nr_pages
);
2462 #endif /* CONFIG_MEMCG_KMEM */
2464 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2467 * Because tail pages are not marked as "used", set it. We're under
2468 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2469 * charge/uncharge will be never happen and move_account() is done under
2470 * compound_lock(), so we don't have to take care of races.
2472 void mem_cgroup_split_huge_fixup(struct page
*head
)
2476 if (mem_cgroup_disabled())
2479 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2480 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2482 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2485 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2487 #ifdef CONFIG_MEMCG_SWAP
2488 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2491 int val
= (charge
) ? 1 : -1;
2492 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2496 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2497 * @entry: swap entry to be moved
2498 * @from: mem_cgroup which the entry is moved from
2499 * @to: mem_cgroup which the entry is moved to
2501 * It succeeds only when the swap_cgroup's record for this entry is the same
2502 * as the mem_cgroup's id of @from.
2504 * Returns 0 on success, -EINVAL on failure.
2506 * The caller must have charged to @to, IOW, called page_counter_charge() about
2507 * both res and memsw, and called css_get().
2509 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2510 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2512 unsigned short old_id
, new_id
;
2514 old_id
= mem_cgroup_id(from
);
2515 new_id
= mem_cgroup_id(to
);
2517 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2518 mem_cgroup_swap_statistics(from
, false);
2519 mem_cgroup_swap_statistics(to
, true);
2525 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2526 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2532 static DEFINE_MUTEX(memcg_limit_mutex
);
2534 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2535 unsigned long limit
)
2537 unsigned long curusage
;
2538 unsigned long oldusage
;
2539 bool enlarge
= false;
2544 * For keeping hierarchical_reclaim simple, how long we should retry
2545 * is depends on callers. We set our retry-count to be function
2546 * of # of children which we should visit in this loop.
2548 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2549 mem_cgroup_count_children(memcg
);
2551 oldusage
= page_counter_read(&memcg
->memory
);
2554 if (signal_pending(current
)) {
2559 mutex_lock(&memcg_limit_mutex
);
2560 if (limit
> memcg
->memsw
.limit
) {
2561 mutex_unlock(&memcg_limit_mutex
);
2565 if (limit
> memcg
->memory
.limit
)
2567 ret
= page_counter_limit(&memcg
->memory
, limit
);
2568 mutex_unlock(&memcg_limit_mutex
);
2573 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2575 curusage
= page_counter_read(&memcg
->memory
);
2576 /* Usage is reduced ? */
2577 if (curusage
>= oldusage
)
2580 oldusage
= curusage
;
2581 } while (retry_count
);
2583 if (!ret
&& enlarge
)
2584 memcg_oom_recover(memcg
);
2589 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2590 unsigned long limit
)
2592 unsigned long curusage
;
2593 unsigned long oldusage
;
2594 bool enlarge
= false;
2598 /* see mem_cgroup_resize_res_limit */
2599 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2600 mem_cgroup_count_children(memcg
);
2602 oldusage
= page_counter_read(&memcg
->memsw
);
2605 if (signal_pending(current
)) {
2610 mutex_lock(&memcg_limit_mutex
);
2611 if (limit
< memcg
->memory
.limit
) {
2612 mutex_unlock(&memcg_limit_mutex
);
2616 if (limit
> memcg
->memsw
.limit
)
2618 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2619 mutex_unlock(&memcg_limit_mutex
);
2624 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2626 curusage
= page_counter_read(&memcg
->memsw
);
2627 /* Usage is reduced ? */
2628 if (curusage
>= oldusage
)
2631 oldusage
= curusage
;
2632 } while (retry_count
);
2634 if (!ret
&& enlarge
)
2635 memcg_oom_recover(memcg
);
2640 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
2642 unsigned long *total_scanned
)
2644 unsigned long nr_reclaimed
= 0;
2645 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
2646 unsigned long reclaimed
;
2648 struct mem_cgroup_tree_per_zone
*mctz
;
2649 unsigned long excess
;
2650 unsigned long nr_scanned
;
2655 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
2657 * This loop can run a while, specially if mem_cgroup's continuously
2658 * keep exceeding their soft limit and putting the system under
2665 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2670 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
2671 gfp_mask
, &nr_scanned
);
2672 nr_reclaimed
+= reclaimed
;
2673 *total_scanned
+= nr_scanned
;
2674 spin_lock_irq(&mctz
->lock
);
2675 __mem_cgroup_remove_exceeded(mz
, mctz
);
2678 * If we failed to reclaim anything from this memory cgroup
2679 * it is time to move on to the next cgroup
2683 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2685 excess
= soft_limit_excess(mz
->memcg
);
2687 * One school of thought says that we should not add
2688 * back the node to the tree if reclaim returns 0.
2689 * But our reclaim could return 0, simply because due
2690 * to priority we are exposing a smaller subset of
2691 * memory to reclaim from. Consider this as a longer
2694 /* If excess == 0, no tree ops */
2695 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2696 spin_unlock_irq(&mctz
->lock
);
2697 css_put(&mz
->memcg
->css
);
2700 * Could not reclaim anything and there are no more
2701 * mem cgroups to try or we seem to be looping without
2702 * reclaiming anything.
2704 if (!nr_reclaimed
&&
2706 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2708 } while (!nr_reclaimed
);
2710 css_put(&next_mz
->memcg
->css
);
2711 return nr_reclaimed
;
2715 * Test whether @memcg has children, dead or alive. Note that this
2716 * function doesn't care whether @memcg has use_hierarchy enabled and
2717 * returns %true if there are child csses according to the cgroup
2718 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2720 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2725 * The lock does not prevent addition or deletion of children, but
2726 * it prevents a new child from being initialized based on this
2727 * parent in css_online(), so it's enough to decide whether
2728 * hierarchically inherited attributes can still be changed or not.
2730 lockdep_assert_held(&memcg_create_mutex
);
2733 ret
= css_next_child(NULL
, &memcg
->css
);
2739 * Reclaims as many pages from the given memcg as possible and moves
2740 * the rest to the parent.
2742 * Caller is responsible for holding css reference for memcg.
2744 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2746 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2748 /* we call try-to-free pages for make this cgroup empty */
2749 lru_add_drain_all();
2750 /* try to free all pages in this cgroup */
2751 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2754 if (signal_pending(current
))
2757 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2761 /* maybe some writeback is necessary */
2762 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2770 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
2771 char *buf
, size_t nbytes
,
2774 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2776 if (mem_cgroup_is_root(memcg
))
2778 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
2781 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
2784 return mem_cgroup_from_css(css
)->use_hierarchy
;
2787 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
2788 struct cftype
*cft
, u64 val
)
2791 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2792 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
2794 mutex_lock(&memcg_create_mutex
);
2796 if (memcg
->use_hierarchy
== val
)
2800 * If parent's use_hierarchy is set, we can't make any modifications
2801 * in the child subtrees. If it is unset, then the change can
2802 * occur, provided the current cgroup has no children.
2804 * For the root cgroup, parent_mem is NULL, we allow value to be
2805 * set if there are no children.
2807 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
2808 (val
== 1 || val
== 0)) {
2809 if (!memcg_has_children(memcg
))
2810 memcg
->use_hierarchy
= val
;
2817 mutex_unlock(&memcg_create_mutex
);
2822 static unsigned long tree_stat(struct mem_cgroup
*memcg
,
2823 enum mem_cgroup_stat_index idx
)
2825 struct mem_cgroup
*iter
;
2826 unsigned long val
= 0;
2828 for_each_mem_cgroup_tree(iter
, memcg
)
2829 val
+= mem_cgroup_read_stat(iter
, idx
);
2834 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
2838 if (mem_cgroup_is_root(memcg
)) {
2839 val
= tree_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
2840 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_RSS
);
2842 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
2845 val
= page_counter_read(&memcg
->memory
);
2847 val
= page_counter_read(&memcg
->memsw
);
2860 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
2863 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2864 struct page_counter
*counter
;
2866 switch (MEMFILE_TYPE(cft
->private)) {
2868 counter
= &memcg
->memory
;
2871 counter
= &memcg
->memsw
;
2874 counter
= &memcg
->kmem
;
2880 switch (MEMFILE_ATTR(cft
->private)) {
2882 if (counter
== &memcg
->memory
)
2883 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
2884 if (counter
== &memcg
->memsw
)
2885 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
2886 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
2888 return (u64
)counter
->limit
* PAGE_SIZE
;
2890 return (u64
)counter
->watermark
* PAGE_SIZE
;
2892 return counter
->failcnt
;
2893 case RES_SOFT_LIMIT
:
2894 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
2900 #ifdef CONFIG_MEMCG_KMEM
2901 static int memcg_activate_kmem(struct mem_cgroup
*memcg
,
2902 unsigned long nr_pages
)
2907 BUG_ON(memcg
->kmemcg_id
>= 0);
2908 BUG_ON(memcg
->kmem_acct_activated
);
2909 BUG_ON(memcg
->kmem_acct_active
);
2912 * For simplicity, we won't allow this to be disabled. It also can't
2913 * be changed if the cgroup has children already, or if tasks had
2916 * If tasks join before we set the limit, a person looking at
2917 * kmem.usage_in_bytes will have no way to determine when it took
2918 * place, which makes the value quite meaningless.
2920 * After it first became limited, changes in the value of the limit are
2921 * of course permitted.
2923 mutex_lock(&memcg_create_mutex
);
2924 if (cgroup_is_populated(memcg
->css
.cgroup
) ||
2925 (memcg
->use_hierarchy
&& memcg_has_children(memcg
)))
2927 mutex_unlock(&memcg_create_mutex
);
2931 memcg_id
= memcg_alloc_cache_id();
2938 * We couldn't have accounted to this cgroup, because it hasn't got
2939 * activated yet, so this should succeed.
2941 err
= page_counter_limit(&memcg
->kmem
, nr_pages
);
2944 static_key_slow_inc(&memcg_kmem_enabled_key
);
2946 * A memory cgroup is considered kmem-active as soon as it gets
2947 * kmemcg_id. Setting the id after enabling static branching will
2948 * guarantee no one starts accounting before all call sites are
2951 memcg
->kmemcg_id
= memcg_id
;
2952 memcg
->kmem_acct_activated
= true;
2953 memcg
->kmem_acct_active
= true;
2958 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2959 unsigned long limit
)
2963 mutex_lock(&memcg_limit_mutex
);
2964 if (!memcg_kmem_is_active(memcg
))
2965 ret
= memcg_activate_kmem(memcg
, limit
);
2967 ret
= page_counter_limit(&memcg
->kmem
, limit
);
2968 mutex_unlock(&memcg_limit_mutex
);
2972 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
2975 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
2980 mutex_lock(&memcg_limit_mutex
);
2982 * If the parent cgroup is not kmem-active now, it cannot be activated
2983 * after this point, because it has at least one child already.
2985 if (memcg_kmem_is_active(parent
))
2986 ret
= memcg_activate_kmem(memcg
, PAGE_COUNTER_MAX
);
2987 mutex_unlock(&memcg_limit_mutex
);
2991 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2992 unsigned long limit
)
2996 #endif /* CONFIG_MEMCG_KMEM */
2999 * The user of this function is...
3002 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3003 char *buf
, size_t nbytes
, loff_t off
)
3005 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3006 unsigned long nr_pages
;
3009 buf
= strstrip(buf
);
3010 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3014 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3016 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3020 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3022 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
3025 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
3028 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
3032 case RES_SOFT_LIMIT
:
3033 memcg
->soft_limit
= nr_pages
;
3037 return ret
?: nbytes
;
3040 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3041 size_t nbytes
, loff_t off
)
3043 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3044 struct page_counter
*counter
;
3046 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3048 counter
= &memcg
->memory
;
3051 counter
= &memcg
->memsw
;
3054 counter
= &memcg
->kmem
;
3060 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3062 page_counter_reset_watermark(counter
);
3065 counter
->failcnt
= 0;
3074 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3077 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3081 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3082 struct cftype
*cft
, u64 val
)
3084 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3086 if (val
& ~MOVE_MASK
)
3090 * No kind of locking is needed in here, because ->can_attach() will
3091 * check this value once in the beginning of the process, and then carry
3092 * on with stale data. This means that changes to this value will only
3093 * affect task migrations starting after the change.
3095 memcg
->move_charge_at_immigrate
= val
;
3099 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3100 struct cftype
*cft
, u64 val
)
3107 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3111 unsigned int lru_mask
;
3114 static const struct numa_stat stats
[] = {
3115 { "total", LRU_ALL
},
3116 { "file", LRU_ALL_FILE
},
3117 { "anon", LRU_ALL_ANON
},
3118 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3120 const struct numa_stat
*stat
;
3123 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3125 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3126 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3127 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3128 for_each_node_state(nid
, N_MEMORY
) {
3129 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3131 seq_printf(m
, " N%d=%lu", nid
, nr
);
3136 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3137 struct mem_cgroup
*iter
;
3140 for_each_mem_cgroup_tree(iter
, memcg
)
3141 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3142 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3143 for_each_node_state(nid
, N_MEMORY
) {
3145 for_each_mem_cgroup_tree(iter
, memcg
)
3146 nr
+= mem_cgroup_node_nr_lru_pages(
3147 iter
, nid
, stat
->lru_mask
);
3148 seq_printf(m
, " N%d=%lu", nid
, nr
);
3155 #endif /* CONFIG_NUMA */
3157 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3159 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3160 unsigned long memory
, memsw
;
3161 struct mem_cgroup
*mi
;
3164 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3165 MEM_CGROUP_STAT_NSTATS
);
3166 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3167 MEM_CGROUP_EVENTS_NSTATS
);
3168 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3170 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3171 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
3173 seq_printf(m
, "%s %lu\n", mem_cgroup_stat_names
[i
],
3174 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3177 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3178 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3179 mem_cgroup_read_events(memcg
, i
));
3181 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3182 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3183 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3185 /* Hierarchical information */
3186 memory
= memsw
= PAGE_COUNTER_MAX
;
3187 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3188 memory
= min(memory
, mi
->memory
.limit
);
3189 memsw
= min(memsw
, mi
->memsw
.limit
);
3191 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3192 (u64
)memory
* PAGE_SIZE
);
3193 if (do_swap_account
)
3194 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3195 (u64
)memsw
* PAGE_SIZE
);
3197 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3198 unsigned long long val
= 0;
3200 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
3202 for_each_mem_cgroup_tree(mi
, memcg
)
3203 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3204 seq_printf(m
, "total_%s %llu\n", mem_cgroup_stat_names
[i
], val
);
3207 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3208 unsigned long long val
= 0;
3210 for_each_mem_cgroup_tree(mi
, memcg
)
3211 val
+= mem_cgroup_read_events(mi
, i
);
3212 seq_printf(m
, "total_%s %llu\n",
3213 mem_cgroup_events_names
[i
], val
);
3216 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3217 unsigned long long val
= 0;
3219 for_each_mem_cgroup_tree(mi
, memcg
)
3220 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3221 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3224 #ifdef CONFIG_DEBUG_VM
3227 struct mem_cgroup_per_zone
*mz
;
3228 struct zone_reclaim_stat
*rstat
;
3229 unsigned long recent_rotated
[2] = {0, 0};
3230 unsigned long recent_scanned
[2] = {0, 0};
3232 for_each_online_node(nid
)
3233 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3234 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
3235 rstat
= &mz
->lruvec
.reclaim_stat
;
3237 recent_rotated
[0] += rstat
->recent_rotated
[0];
3238 recent_rotated
[1] += rstat
->recent_rotated
[1];
3239 recent_scanned
[0] += rstat
->recent_scanned
[0];
3240 recent_scanned
[1] += rstat
->recent_scanned
[1];
3242 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3243 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3244 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3245 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3252 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3255 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3257 return mem_cgroup_swappiness(memcg
);
3260 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3261 struct cftype
*cft
, u64 val
)
3263 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3269 memcg
->swappiness
= val
;
3271 vm_swappiness
= val
;
3276 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3278 struct mem_cgroup_threshold_ary
*t
;
3279 unsigned long usage
;
3284 t
= rcu_dereference(memcg
->thresholds
.primary
);
3286 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3291 usage
= mem_cgroup_usage(memcg
, swap
);
3294 * current_threshold points to threshold just below or equal to usage.
3295 * If it's not true, a threshold was crossed after last
3296 * call of __mem_cgroup_threshold().
3298 i
= t
->current_threshold
;
3301 * Iterate backward over array of thresholds starting from
3302 * current_threshold and check if a threshold is crossed.
3303 * If none of thresholds below usage is crossed, we read
3304 * only one element of the array here.
3306 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3307 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3309 /* i = current_threshold + 1 */
3313 * Iterate forward over array of thresholds starting from
3314 * current_threshold+1 and check if a threshold is crossed.
3315 * If none of thresholds above usage is crossed, we read
3316 * only one element of the array here.
3318 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3319 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3321 /* Update current_threshold */
3322 t
->current_threshold
= i
- 1;
3327 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3330 __mem_cgroup_threshold(memcg
, false);
3331 if (do_swap_account
)
3332 __mem_cgroup_threshold(memcg
, true);
3334 memcg
= parent_mem_cgroup(memcg
);
3338 static int compare_thresholds(const void *a
, const void *b
)
3340 const struct mem_cgroup_threshold
*_a
= a
;
3341 const struct mem_cgroup_threshold
*_b
= b
;
3343 if (_a
->threshold
> _b
->threshold
)
3346 if (_a
->threshold
< _b
->threshold
)
3352 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3354 struct mem_cgroup_eventfd_list
*ev
;
3356 spin_lock(&memcg_oom_lock
);
3358 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3359 eventfd_signal(ev
->eventfd
, 1);
3361 spin_unlock(&memcg_oom_lock
);
3365 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3367 struct mem_cgroup
*iter
;
3369 for_each_mem_cgroup_tree(iter
, memcg
)
3370 mem_cgroup_oom_notify_cb(iter
);
3373 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3374 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3376 struct mem_cgroup_thresholds
*thresholds
;
3377 struct mem_cgroup_threshold_ary
*new;
3378 unsigned long threshold
;
3379 unsigned long usage
;
3382 ret
= page_counter_memparse(args
, "-1", &threshold
);
3386 mutex_lock(&memcg
->thresholds_lock
);
3389 thresholds
= &memcg
->thresholds
;
3390 usage
= mem_cgroup_usage(memcg
, false);
3391 } else if (type
== _MEMSWAP
) {
3392 thresholds
= &memcg
->memsw_thresholds
;
3393 usage
= mem_cgroup_usage(memcg
, true);
3397 /* Check if a threshold crossed before adding a new one */
3398 if (thresholds
->primary
)
3399 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3401 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3403 /* Allocate memory for new array of thresholds */
3404 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3412 /* Copy thresholds (if any) to new array */
3413 if (thresholds
->primary
) {
3414 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3415 sizeof(struct mem_cgroup_threshold
));
3418 /* Add new threshold */
3419 new->entries
[size
- 1].eventfd
= eventfd
;
3420 new->entries
[size
- 1].threshold
= threshold
;
3422 /* Sort thresholds. Registering of new threshold isn't time-critical */
3423 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3424 compare_thresholds
, NULL
);
3426 /* Find current threshold */
3427 new->current_threshold
= -1;
3428 for (i
= 0; i
< size
; i
++) {
3429 if (new->entries
[i
].threshold
<= usage
) {
3431 * new->current_threshold will not be used until
3432 * rcu_assign_pointer(), so it's safe to increment
3435 ++new->current_threshold
;
3440 /* Free old spare buffer and save old primary buffer as spare */
3441 kfree(thresholds
->spare
);
3442 thresholds
->spare
= thresholds
->primary
;
3444 rcu_assign_pointer(thresholds
->primary
, new);
3446 /* To be sure that nobody uses thresholds */
3450 mutex_unlock(&memcg
->thresholds_lock
);
3455 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3456 struct eventfd_ctx
*eventfd
, const char *args
)
3458 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3461 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3462 struct eventfd_ctx
*eventfd
, const char *args
)
3464 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3467 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3468 struct eventfd_ctx
*eventfd
, enum res_type type
)
3470 struct mem_cgroup_thresholds
*thresholds
;
3471 struct mem_cgroup_threshold_ary
*new;
3472 unsigned long usage
;
3475 mutex_lock(&memcg
->thresholds_lock
);
3478 thresholds
= &memcg
->thresholds
;
3479 usage
= mem_cgroup_usage(memcg
, false);
3480 } else if (type
== _MEMSWAP
) {
3481 thresholds
= &memcg
->memsw_thresholds
;
3482 usage
= mem_cgroup_usage(memcg
, true);
3486 if (!thresholds
->primary
)
3489 /* Check if a threshold crossed before removing */
3490 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3492 /* Calculate new number of threshold */
3494 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3495 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3499 new = thresholds
->spare
;
3501 /* Set thresholds array to NULL if we don't have thresholds */
3510 /* Copy thresholds and find current threshold */
3511 new->current_threshold
= -1;
3512 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3513 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3516 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3517 if (new->entries
[j
].threshold
<= usage
) {
3519 * new->current_threshold will not be used
3520 * until rcu_assign_pointer(), so it's safe to increment
3523 ++new->current_threshold
;
3529 /* Swap primary and spare array */
3530 thresholds
->spare
= thresholds
->primary
;
3531 /* If all events are unregistered, free the spare array */
3533 kfree(thresholds
->spare
);
3534 thresholds
->spare
= NULL
;
3537 rcu_assign_pointer(thresholds
->primary
, new);
3539 /* To be sure that nobody uses thresholds */
3542 mutex_unlock(&memcg
->thresholds_lock
);
3545 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3546 struct eventfd_ctx
*eventfd
)
3548 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3551 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3552 struct eventfd_ctx
*eventfd
)
3554 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3557 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3558 struct eventfd_ctx
*eventfd
, const char *args
)
3560 struct mem_cgroup_eventfd_list
*event
;
3562 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3566 spin_lock(&memcg_oom_lock
);
3568 event
->eventfd
= eventfd
;
3569 list_add(&event
->list
, &memcg
->oom_notify
);
3571 /* already in OOM ? */
3572 if (memcg
->under_oom
)
3573 eventfd_signal(eventfd
, 1);
3574 spin_unlock(&memcg_oom_lock
);
3579 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3580 struct eventfd_ctx
*eventfd
)
3582 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3584 spin_lock(&memcg_oom_lock
);
3586 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3587 if (ev
->eventfd
== eventfd
) {
3588 list_del(&ev
->list
);
3593 spin_unlock(&memcg_oom_lock
);
3596 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3598 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3600 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3601 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3605 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3606 struct cftype
*cft
, u64 val
)
3608 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3610 /* cannot set to root cgroup and only 0 and 1 are allowed */
3611 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3614 memcg
->oom_kill_disable
= val
;
3616 memcg_oom_recover(memcg
);
3621 #ifdef CONFIG_MEMCG_KMEM
3622 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
3626 ret
= memcg_propagate_kmem(memcg
);
3630 return mem_cgroup_sockets_init(memcg
, ss
);
3633 static void memcg_deactivate_kmem(struct mem_cgroup
*memcg
)
3635 struct cgroup_subsys_state
*css
;
3636 struct mem_cgroup
*parent
, *child
;
3639 if (!memcg
->kmem_acct_active
)
3643 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3644 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3645 * guarantees no cache will be created for this cgroup after we are
3646 * done (see memcg_create_kmem_cache()).
3648 memcg
->kmem_acct_active
= false;
3650 memcg_deactivate_kmem_caches(memcg
);
3652 kmemcg_id
= memcg
->kmemcg_id
;
3653 BUG_ON(kmemcg_id
< 0);
3655 parent
= parent_mem_cgroup(memcg
);
3657 parent
= root_mem_cgroup
;
3660 * Change kmemcg_id of this cgroup and all its descendants to the
3661 * parent's id, and then move all entries from this cgroup's list_lrus
3662 * to ones of the parent. After we have finished, all list_lrus
3663 * corresponding to this cgroup are guaranteed to remain empty. The
3664 * ordering is imposed by list_lru_node->lock taken by
3665 * memcg_drain_all_list_lrus().
3667 css_for_each_descendant_pre(css
, &memcg
->css
) {
3668 child
= mem_cgroup_from_css(css
);
3669 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
3670 child
->kmemcg_id
= parent
->kmemcg_id
;
3671 if (!memcg
->use_hierarchy
)
3674 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
3676 memcg_free_cache_id(kmemcg_id
);
3679 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
3681 if (memcg
->kmem_acct_activated
) {
3682 memcg_destroy_kmem_caches(memcg
);
3683 static_key_slow_dec(&memcg_kmem_enabled_key
);
3684 WARN_ON(page_counter_read(&memcg
->kmem
));
3686 mem_cgroup_sockets_destroy(memcg
);
3689 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
3694 static void memcg_deactivate_kmem(struct mem_cgroup
*memcg
)
3698 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
3703 #ifdef CONFIG_CGROUP_WRITEBACK
3705 struct list_head
*mem_cgroup_cgwb_list(struct mem_cgroup
*memcg
)
3707 return &memcg
->cgwb_list
;
3710 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3712 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3715 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3717 wb_domain_exit(&memcg
->cgwb_domain
);
3720 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3722 wb_domain_size_changed(&memcg
->cgwb_domain
);
3725 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3727 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3729 if (!memcg
->css
.parent
)
3732 return &memcg
->cgwb_domain
;
3736 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3737 * @wb: bdi_writeback in question
3738 * @pfilepages: out parameter for number of file pages
3739 * @pheadroom: out parameter for number of allocatable pages according to memcg
3740 * @pdirty: out parameter for number of dirty pages
3741 * @pwriteback: out parameter for number of pages under writeback
3743 * Determine the numbers of file, headroom, dirty, and writeback pages in
3744 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3745 * is a bit more involved.
3747 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3748 * headroom is calculated as the lowest headroom of itself and the
3749 * ancestors. Note that this doesn't consider the actual amount of
3750 * available memory in the system. The caller should further cap
3751 * *@pheadroom accordingly.
3753 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
3754 unsigned long *pheadroom
, unsigned long *pdirty
,
3755 unsigned long *pwriteback
)
3757 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3758 struct mem_cgroup
*parent
;
3760 *pdirty
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
3762 /* this should eventually include NR_UNSTABLE_NFS */
3763 *pwriteback
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
3764 *pfilepages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
3765 (1 << LRU_ACTIVE_FILE
));
3766 *pheadroom
= PAGE_COUNTER_MAX
;
3768 while ((parent
= parent_mem_cgroup(memcg
))) {
3769 unsigned long ceiling
= min(memcg
->memory
.limit
, memcg
->high
);
3770 unsigned long used
= page_counter_read(&memcg
->memory
);
3772 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
3777 #else /* CONFIG_CGROUP_WRITEBACK */
3779 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3784 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3788 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3792 #endif /* CONFIG_CGROUP_WRITEBACK */
3795 * DO NOT USE IN NEW FILES.
3797 * "cgroup.event_control" implementation.
3799 * This is way over-engineered. It tries to support fully configurable
3800 * events for each user. Such level of flexibility is completely
3801 * unnecessary especially in the light of the planned unified hierarchy.
3803 * Please deprecate this and replace with something simpler if at all
3808 * Unregister event and free resources.
3810 * Gets called from workqueue.
3812 static void memcg_event_remove(struct work_struct
*work
)
3814 struct mem_cgroup_event
*event
=
3815 container_of(work
, struct mem_cgroup_event
, remove
);
3816 struct mem_cgroup
*memcg
= event
->memcg
;
3818 remove_wait_queue(event
->wqh
, &event
->wait
);
3820 event
->unregister_event(memcg
, event
->eventfd
);
3822 /* Notify userspace the event is going away. */
3823 eventfd_signal(event
->eventfd
, 1);
3825 eventfd_ctx_put(event
->eventfd
);
3827 css_put(&memcg
->css
);
3831 * Gets called on POLLHUP on eventfd when user closes it.
3833 * Called with wqh->lock held and interrupts disabled.
3835 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
3836 int sync
, void *key
)
3838 struct mem_cgroup_event
*event
=
3839 container_of(wait
, struct mem_cgroup_event
, wait
);
3840 struct mem_cgroup
*memcg
= event
->memcg
;
3841 unsigned long flags
= (unsigned long)key
;
3843 if (flags
& POLLHUP
) {
3845 * If the event has been detached at cgroup removal, we
3846 * can simply return knowing the other side will cleanup
3849 * We can't race against event freeing since the other
3850 * side will require wqh->lock via remove_wait_queue(),
3853 spin_lock(&memcg
->event_list_lock
);
3854 if (!list_empty(&event
->list
)) {
3855 list_del_init(&event
->list
);
3857 * We are in atomic context, but cgroup_event_remove()
3858 * may sleep, so we have to call it in workqueue.
3860 schedule_work(&event
->remove
);
3862 spin_unlock(&memcg
->event_list_lock
);
3868 static void memcg_event_ptable_queue_proc(struct file
*file
,
3869 wait_queue_head_t
*wqh
, poll_table
*pt
)
3871 struct mem_cgroup_event
*event
=
3872 container_of(pt
, struct mem_cgroup_event
, pt
);
3875 add_wait_queue(wqh
, &event
->wait
);
3879 * DO NOT USE IN NEW FILES.
3881 * Parse input and register new cgroup event handler.
3883 * Input must be in format '<event_fd> <control_fd> <args>'.
3884 * Interpretation of args is defined by control file implementation.
3886 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
3887 char *buf
, size_t nbytes
, loff_t off
)
3889 struct cgroup_subsys_state
*css
= of_css(of
);
3890 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3891 struct mem_cgroup_event
*event
;
3892 struct cgroup_subsys_state
*cfile_css
;
3893 unsigned int efd
, cfd
;
3900 buf
= strstrip(buf
);
3902 efd
= simple_strtoul(buf
, &endp
, 10);
3907 cfd
= simple_strtoul(buf
, &endp
, 10);
3908 if ((*endp
!= ' ') && (*endp
!= '\0'))
3912 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
3916 event
->memcg
= memcg
;
3917 INIT_LIST_HEAD(&event
->list
);
3918 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
3919 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
3920 INIT_WORK(&event
->remove
, memcg_event_remove
);
3928 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
3929 if (IS_ERR(event
->eventfd
)) {
3930 ret
= PTR_ERR(event
->eventfd
);
3937 goto out_put_eventfd
;
3940 /* the process need read permission on control file */
3941 /* AV: shouldn't we check that it's been opened for read instead? */
3942 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
3947 * Determine the event callbacks and set them in @event. This used
3948 * to be done via struct cftype but cgroup core no longer knows
3949 * about these events. The following is crude but the whole thing
3950 * is for compatibility anyway.
3952 * DO NOT ADD NEW FILES.
3954 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
3956 if (!strcmp(name
, "memory.usage_in_bytes")) {
3957 event
->register_event
= mem_cgroup_usage_register_event
;
3958 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
3959 } else if (!strcmp(name
, "memory.oom_control")) {
3960 event
->register_event
= mem_cgroup_oom_register_event
;
3961 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
3962 } else if (!strcmp(name
, "memory.pressure_level")) {
3963 event
->register_event
= vmpressure_register_event
;
3964 event
->unregister_event
= vmpressure_unregister_event
;
3965 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
3966 event
->register_event
= memsw_cgroup_usage_register_event
;
3967 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
3974 * Verify @cfile should belong to @css. Also, remaining events are
3975 * automatically removed on cgroup destruction but the removal is
3976 * asynchronous, so take an extra ref on @css.
3978 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
3979 &memory_cgrp_subsys
);
3981 if (IS_ERR(cfile_css
))
3983 if (cfile_css
!= css
) {
3988 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
3992 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
3994 spin_lock(&memcg
->event_list_lock
);
3995 list_add(&event
->list
, &memcg
->event_list
);
3996 spin_unlock(&memcg
->event_list_lock
);
4008 eventfd_ctx_put(event
->eventfd
);
4017 static struct cftype mem_cgroup_legacy_files
[] = {
4019 .name
= "usage_in_bytes",
4020 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4021 .read_u64
= mem_cgroup_read_u64
,
4024 .name
= "max_usage_in_bytes",
4025 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4026 .write
= mem_cgroup_reset
,
4027 .read_u64
= mem_cgroup_read_u64
,
4030 .name
= "limit_in_bytes",
4031 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4032 .write
= mem_cgroup_write
,
4033 .read_u64
= mem_cgroup_read_u64
,
4036 .name
= "soft_limit_in_bytes",
4037 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4038 .write
= mem_cgroup_write
,
4039 .read_u64
= mem_cgroup_read_u64
,
4043 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4044 .write
= mem_cgroup_reset
,
4045 .read_u64
= mem_cgroup_read_u64
,
4049 .seq_show
= memcg_stat_show
,
4052 .name
= "force_empty",
4053 .write
= mem_cgroup_force_empty_write
,
4056 .name
= "use_hierarchy",
4057 .write_u64
= mem_cgroup_hierarchy_write
,
4058 .read_u64
= mem_cgroup_hierarchy_read
,
4061 .name
= "cgroup.event_control", /* XXX: for compat */
4062 .write
= memcg_write_event_control
,
4063 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
4066 .name
= "swappiness",
4067 .read_u64
= mem_cgroup_swappiness_read
,
4068 .write_u64
= mem_cgroup_swappiness_write
,
4071 .name
= "move_charge_at_immigrate",
4072 .read_u64
= mem_cgroup_move_charge_read
,
4073 .write_u64
= mem_cgroup_move_charge_write
,
4076 .name
= "oom_control",
4077 .seq_show
= mem_cgroup_oom_control_read
,
4078 .write_u64
= mem_cgroup_oom_control_write
,
4079 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4082 .name
= "pressure_level",
4086 .name
= "numa_stat",
4087 .seq_show
= memcg_numa_stat_show
,
4090 #ifdef CONFIG_MEMCG_KMEM
4092 .name
= "kmem.limit_in_bytes",
4093 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4094 .write
= mem_cgroup_write
,
4095 .read_u64
= mem_cgroup_read_u64
,
4098 .name
= "kmem.usage_in_bytes",
4099 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4100 .read_u64
= mem_cgroup_read_u64
,
4103 .name
= "kmem.failcnt",
4104 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4105 .write
= mem_cgroup_reset
,
4106 .read_u64
= mem_cgroup_read_u64
,
4109 .name
= "kmem.max_usage_in_bytes",
4110 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4111 .write
= mem_cgroup_reset
,
4112 .read_u64
= mem_cgroup_read_u64
,
4114 #ifdef CONFIG_SLABINFO
4116 .name
= "kmem.slabinfo",
4117 .seq_start
= slab_start
,
4118 .seq_next
= slab_next
,
4119 .seq_stop
= slab_stop
,
4120 .seq_show
= memcg_slab_show
,
4124 { }, /* terminate */
4127 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4129 struct mem_cgroup_per_node
*pn
;
4130 struct mem_cgroup_per_zone
*mz
;
4131 int zone
, tmp
= node
;
4133 * This routine is called against possible nodes.
4134 * But it's BUG to call kmalloc() against offline node.
4136 * TODO: this routine can waste much memory for nodes which will
4137 * never be onlined. It's better to use memory hotplug callback
4140 if (!node_state(node
, N_NORMAL_MEMORY
))
4142 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4146 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4147 mz
= &pn
->zoneinfo
[zone
];
4148 lruvec_init(&mz
->lruvec
);
4149 mz
->usage_in_excess
= 0;
4150 mz
->on_tree
= false;
4153 memcg
->nodeinfo
[node
] = pn
;
4157 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4159 kfree(memcg
->nodeinfo
[node
]);
4162 static struct mem_cgroup
*mem_cgroup_alloc(void)
4164 struct mem_cgroup
*memcg
;
4167 size
= sizeof(struct mem_cgroup
);
4168 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4170 memcg
= kzalloc(size
, GFP_KERNEL
);
4174 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4178 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4184 free_percpu(memcg
->stat
);
4191 * At destroying mem_cgroup, references from swap_cgroup can remain.
4192 * (scanning all at force_empty is too costly...)
4194 * Instead of clearing all references at force_empty, we remember
4195 * the number of reference from swap_cgroup and free mem_cgroup when
4196 * it goes down to 0.
4198 * Removal of cgroup itself succeeds regardless of refs from swap.
4201 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
4205 mem_cgroup_remove_from_trees(memcg
);
4208 free_mem_cgroup_per_zone_info(memcg
, node
);
4210 free_percpu(memcg
->stat
);
4211 memcg_wb_domain_exit(memcg
);
4216 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4218 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
4220 if (!memcg
->memory
.parent
)
4222 return mem_cgroup_from_counter(memcg
->memory
.parent
, memory
);
4224 EXPORT_SYMBOL(parent_mem_cgroup
);
4226 static struct cgroup_subsys_state
* __ref
4227 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4229 struct mem_cgroup
*memcg
;
4230 long error
= -ENOMEM
;
4233 memcg
= mem_cgroup_alloc();
4235 return ERR_PTR(error
);
4238 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
4242 if (parent_css
== NULL
) {
4243 root_mem_cgroup
= memcg
;
4244 mem_cgroup_root_css
= &memcg
->css
;
4245 page_counter_init(&memcg
->memory
, NULL
);
4246 memcg
->high
= PAGE_COUNTER_MAX
;
4247 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4248 page_counter_init(&memcg
->memsw
, NULL
);
4249 page_counter_init(&memcg
->kmem
, NULL
);
4252 memcg
->last_scanned_node
= MAX_NUMNODES
;
4253 INIT_LIST_HEAD(&memcg
->oom_notify
);
4254 memcg
->move_charge_at_immigrate
= 0;
4255 mutex_init(&memcg
->thresholds_lock
);
4256 spin_lock_init(&memcg
->move_lock
);
4257 vmpressure_init(&memcg
->vmpressure
);
4258 INIT_LIST_HEAD(&memcg
->event_list
);
4259 spin_lock_init(&memcg
->event_list_lock
);
4260 #ifdef CONFIG_MEMCG_KMEM
4261 memcg
->kmemcg_id
= -1;
4263 #ifdef CONFIG_CGROUP_WRITEBACK
4264 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4269 __mem_cgroup_free(memcg
);
4270 return ERR_PTR(error
);
4274 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4276 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4277 struct mem_cgroup
*parent
= mem_cgroup_from_css(css
->parent
);
4280 if (css
->id
> MEM_CGROUP_ID_MAX
)
4286 mutex_lock(&memcg_create_mutex
);
4288 memcg
->use_hierarchy
= parent
->use_hierarchy
;
4289 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4290 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4292 if (parent
->use_hierarchy
) {
4293 page_counter_init(&memcg
->memory
, &parent
->memory
);
4294 memcg
->high
= PAGE_COUNTER_MAX
;
4295 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4296 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4297 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4300 * No need to take a reference to the parent because cgroup
4301 * core guarantees its existence.
4304 page_counter_init(&memcg
->memory
, NULL
);
4305 memcg
->high
= PAGE_COUNTER_MAX
;
4306 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4307 page_counter_init(&memcg
->memsw
, NULL
);
4308 page_counter_init(&memcg
->kmem
, NULL
);
4310 * Deeper hierachy with use_hierarchy == false doesn't make
4311 * much sense so let cgroup subsystem know about this
4312 * unfortunate state in our controller.
4314 if (parent
!= root_mem_cgroup
)
4315 memory_cgrp_subsys
.broken_hierarchy
= true;
4317 mutex_unlock(&memcg_create_mutex
);
4319 ret
= memcg_init_kmem(memcg
, &memory_cgrp_subsys
);
4324 * Make sure the memcg is initialized: mem_cgroup_iter()
4325 * orders reading memcg->initialized against its callers
4326 * reading the memcg members.
4328 smp_store_release(&memcg
->initialized
, 1);
4333 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4335 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4336 struct mem_cgroup_event
*event
, *tmp
;
4339 * Unregister events and notify userspace.
4340 * Notify userspace about cgroup removing only after rmdir of cgroup
4341 * directory to avoid race between userspace and kernelspace.
4343 spin_lock(&memcg
->event_list_lock
);
4344 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4345 list_del_init(&event
->list
);
4346 schedule_work(&event
->remove
);
4348 spin_unlock(&memcg
->event_list_lock
);
4350 vmpressure_cleanup(&memcg
->vmpressure
);
4352 memcg_deactivate_kmem(memcg
);
4354 wb_memcg_offline(memcg
);
4357 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4359 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4361 invalidate_reclaim_iterators(memcg
);
4364 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4366 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4368 memcg_destroy_kmem(memcg
);
4369 __mem_cgroup_free(memcg
);
4373 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4374 * @css: the target css
4376 * Reset the states of the mem_cgroup associated with @css. This is
4377 * invoked when the userland requests disabling on the default hierarchy
4378 * but the memcg is pinned through dependency. The memcg should stop
4379 * applying policies and should revert to the vanilla state as it may be
4380 * made visible again.
4382 * The current implementation only resets the essential configurations.
4383 * This needs to be expanded to cover all the visible parts.
4385 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4387 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4389 mem_cgroup_resize_limit(memcg
, PAGE_COUNTER_MAX
);
4390 mem_cgroup_resize_memsw_limit(memcg
, PAGE_COUNTER_MAX
);
4391 memcg_update_kmem_limit(memcg
, PAGE_COUNTER_MAX
);
4393 memcg
->high
= PAGE_COUNTER_MAX
;
4394 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4395 memcg_wb_domain_size_changed(memcg
);
4399 /* Handlers for move charge at task migration. */
4400 static int mem_cgroup_do_precharge(unsigned long count
)
4404 /* Try a single bulk charge without reclaim first, kswapd may wake */
4405 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4407 mc
.precharge
+= count
;
4411 /* Try charges one by one with reclaim */
4413 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4423 * get_mctgt_type - get target type of moving charge
4424 * @vma: the vma the pte to be checked belongs
4425 * @addr: the address corresponding to the pte to be checked
4426 * @ptent: the pte to be checked
4427 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4430 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4431 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4432 * move charge. if @target is not NULL, the page is stored in target->page
4433 * with extra refcnt got(Callers should handle it).
4434 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4435 * target for charge migration. if @target is not NULL, the entry is stored
4438 * Called with pte lock held.
4445 enum mc_target_type
{
4451 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4452 unsigned long addr
, pte_t ptent
)
4454 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4456 if (!page
|| !page_mapped(page
))
4458 if (PageAnon(page
)) {
4459 if (!(mc
.flags
& MOVE_ANON
))
4462 if (!(mc
.flags
& MOVE_FILE
))
4465 if (!get_page_unless_zero(page
))
4472 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4473 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4475 struct page
*page
= NULL
;
4476 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4478 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4481 * Because lookup_swap_cache() updates some statistics counter,
4482 * we call find_get_page() with swapper_space directly.
4484 page
= find_get_page(swap_address_space(ent
), ent
.val
);
4485 if (do_swap_account
)
4486 entry
->val
= ent
.val
;
4491 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4492 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4498 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4499 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4501 struct page
*page
= NULL
;
4502 struct address_space
*mapping
;
4505 if (!vma
->vm_file
) /* anonymous vma */
4507 if (!(mc
.flags
& MOVE_FILE
))
4510 mapping
= vma
->vm_file
->f_mapping
;
4511 pgoff
= linear_page_index(vma
, addr
);
4513 /* page is moved even if it's not RSS of this task(page-faulted). */
4515 /* shmem/tmpfs may report page out on swap: account for that too. */
4516 if (shmem_mapping(mapping
)) {
4517 page
= find_get_entry(mapping
, pgoff
);
4518 if (radix_tree_exceptional_entry(page
)) {
4519 swp_entry_t swp
= radix_to_swp_entry(page
);
4520 if (do_swap_account
)
4522 page
= find_get_page(swap_address_space(swp
), swp
.val
);
4525 page
= find_get_page(mapping
, pgoff
);
4527 page
= find_get_page(mapping
, pgoff
);
4533 * mem_cgroup_move_account - move account of the page
4535 * @nr_pages: number of regular pages (>1 for huge pages)
4536 * @from: mem_cgroup which the page is moved from.
4537 * @to: mem_cgroup which the page is moved to. @from != @to.
4539 * The caller must confirm following.
4540 * - page is not on LRU (isolate_page() is useful.)
4541 * - compound_lock is held when nr_pages > 1
4543 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4546 static int mem_cgroup_move_account(struct page
*page
,
4547 unsigned int nr_pages
,
4548 struct mem_cgroup
*from
,
4549 struct mem_cgroup
*to
)
4551 unsigned long flags
;
4555 VM_BUG_ON(from
== to
);
4556 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4558 * The page is isolated from LRU. So, collapse function
4559 * will not handle this page. But page splitting can happen.
4560 * Do this check under compound_page_lock(). The caller should
4564 if (nr_pages
> 1 && !PageTransHuge(page
))
4568 * Prevent mem_cgroup_replace_page() from looking at
4569 * page->mem_cgroup of its source page while we change it.
4571 if (!trylock_page(page
))
4575 if (page
->mem_cgroup
!= from
)
4578 anon
= PageAnon(page
);
4580 spin_lock_irqsave(&from
->move_lock
, flags
);
4582 if (!anon
&& page_mapped(page
)) {
4583 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4585 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4590 * move_lock grabbed above and caller set from->moving_account, so
4591 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4592 * So mapping should be stable for dirty pages.
4594 if (!anon
&& PageDirty(page
)) {
4595 struct address_space
*mapping
= page_mapping(page
);
4597 if (mapping_cap_account_dirty(mapping
)) {
4598 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4600 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4605 if (PageWriteback(page
)) {
4606 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4608 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4613 * It is safe to change page->mem_cgroup here because the page
4614 * is referenced, charged, and isolated - we can't race with
4615 * uncharging, charging, migration, or LRU putback.
4618 /* caller should have done css_get */
4619 page
->mem_cgroup
= to
;
4620 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4624 local_irq_disable();
4625 mem_cgroup_charge_statistics(to
, page
, nr_pages
);
4626 memcg_check_events(to
, page
);
4627 mem_cgroup_charge_statistics(from
, page
, -nr_pages
);
4628 memcg_check_events(from
, page
);
4636 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4637 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4639 struct page
*page
= NULL
;
4640 enum mc_target_type ret
= MC_TARGET_NONE
;
4641 swp_entry_t ent
= { .val
= 0 };
4643 if (pte_present(ptent
))
4644 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4645 else if (is_swap_pte(ptent
))
4646 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
4647 else if (pte_none(ptent
))
4648 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4650 if (!page
&& !ent
.val
)
4654 * Do only loose check w/o serialization.
4655 * mem_cgroup_move_account() checks the page is valid or
4656 * not under LRU exclusion.
4658 if (page
->mem_cgroup
== mc
.from
) {
4659 ret
= MC_TARGET_PAGE
;
4661 target
->page
= page
;
4663 if (!ret
|| !target
)
4666 /* There is a swap entry and a page doesn't exist or isn't charged */
4667 if (ent
.val
&& !ret
&&
4668 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4669 ret
= MC_TARGET_SWAP
;
4676 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4678 * We don't consider swapping or file mapped pages because THP does not
4679 * support them for now.
4680 * Caller should make sure that pmd_trans_huge(pmd) is true.
4682 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4683 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4685 struct page
*page
= NULL
;
4686 enum mc_target_type ret
= MC_TARGET_NONE
;
4688 page
= pmd_page(pmd
);
4689 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4690 if (!(mc
.flags
& MOVE_ANON
))
4692 if (page
->mem_cgroup
== mc
.from
) {
4693 ret
= MC_TARGET_PAGE
;
4696 target
->page
= page
;
4702 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4703 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4705 return MC_TARGET_NONE
;
4709 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4710 unsigned long addr
, unsigned long end
,
4711 struct mm_walk
*walk
)
4713 struct vm_area_struct
*vma
= walk
->vma
;
4717 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
4718 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4719 mc
.precharge
+= HPAGE_PMD_NR
;
4724 if (pmd_trans_unstable(pmd
))
4726 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4727 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4728 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4729 mc
.precharge
++; /* increment precharge temporarily */
4730 pte_unmap_unlock(pte
- 1, ptl
);
4736 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4738 unsigned long precharge
;
4740 struct mm_walk mem_cgroup_count_precharge_walk
= {
4741 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4744 down_read(&mm
->mmap_sem
);
4745 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk
);
4746 up_read(&mm
->mmap_sem
);
4748 precharge
= mc
.precharge
;
4754 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4756 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4758 VM_BUG_ON(mc
.moving_task
);
4759 mc
.moving_task
= current
;
4760 return mem_cgroup_do_precharge(precharge
);
4763 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4764 static void __mem_cgroup_clear_mc(void)
4766 struct mem_cgroup
*from
= mc
.from
;
4767 struct mem_cgroup
*to
= mc
.to
;
4769 /* we must uncharge all the leftover precharges from mc.to */
4771 cancel_charge(mc
.to
, mc
.precharge
);
4775 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4776 * we must uncharge here.
4778 if (mc
.moved_charge
) {
4779 cancel_charge(mc
.from
, mc
.moved_charge
);
4780 mc
.moved_charge
= 0;
4782 /* we must fixup refcnts and charges */
4783 if (mc
.moved_swap
) {
4784 /* uncharge swap account from the old cgroup */
4785 if (!mem_cgroup_is_root(mc
.from
))
4786 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4789 * we charged both to->memory and to->memsw, so we
4790 * should uncharge to->memory.
4792 if (!mem_cgroup_is_root(mc
.to
))
4793 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4795 css_put_many(&mc
.from
->css
, mc
.moved_swap
);
4797 /* we've already done css_get(mc.to) */
4800 memcg_oom_recover(from
);
4801 memcg_oom_recover(to
);
4802 wake_up_all(&mc
.waitq
);
4805 static void mem_cgroup_clear_mc(void)
4808 * we must clear moving_task before waking up waiters at the end of
4811 mc
.moving_task
= NULL
;
4812 __mem_cgroup_clear_mc();
4813 spin_lock(&mc
.lock
);
4816 spin_unlock(&mc
.lock
);
4819 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4821 struct cgroup_subsys_state
*css
;
4822 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
4823 struct mem_cgroup
*from
;
4824 struct task_struct
*leader
, *p
;
4825 struct mm_struct
*mm
;
4826 unsigned long move_flags
;
4829 /* charge immigration isn't supported on the default hierarchy */
4830 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4834 * Multi-process migrations only happen on the default hierarchy
4835 * where charge immigration is not used. Perform charge
4836 * immigration if @tset contains a leader and whine if there are
4840 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
4843 memcg
= mem_cgroup_from_css(css
);
4849 * We are now commited to this value whatever it is. Changes in this
4850 * tunable will only affect upcoming migrations, not the current one.
4851 * So we need to save it, and keep it going.
4853 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
4857 from
= mem_cgroup_from_task(p
);
4859 VM_BUG_ON(from
== memcg
);
4861 mm
= get_task_mm(p
);
4864 /* We move charges only when we move a owner of the mm */
4865 if (mm
->owner
== p
) {
4868 VM_BUG_ON(mc
.precharge
);
4869 VM_BUG_ON(mc
.moved_charge
);
4870 VM_BUG_ON(mc
.moved_swap
);
4872 spin_lock(&mc
.lock
);
4875 mc
.flags
= move_flags
;
4876 spin_unlock(&mc
.lock
);
4877 /* We set mc.moving_task later */
4879 ret
= mem_cgroup_precharge_mc(mm
);
4881 mem_cgroup_clear_mc();
4887 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4890 mem_cgroup_clear_mc();
4893 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4894 unsigned long addr
, unsigned long end
,
4895 struct mm_walk
*walk
)
4898 struct vm_area_struct
*vma
= walk
->vma
;
4901 enum mc_target_type target_type
;
4902 union mc_target target
;
4906 * We don't take compound_lock() here but no race with splitting thp
4908 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
4909 * under splitting, which means there's no concurrent thp split,
4910 * - if another thread runs into split_huge_page() just after we
4911 * entered this if-block, the thread must wait for page table lock
4912 * to be unlocked in __split_huge_page_splitting(), where the main
4913 * part of thp split is not executed yet.
4915 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
4916 if (mc
.precharge
< HPAGE_PMD_NR
) {
4920 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
4921 if (target_type
== MC_TARGET_PAGE
) {
4923 if (!isolate_lru_page(page
)) {
4924 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
4926 mc
.precharge
-= HPAGE_PMD_NR
;
4927 mc
.moved_charge
+= HPAGE_PMD_NR
;
4929 putback_lru_page(page
);
4937 if (pmd_trans_unstable(pmd
))
4940 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4941 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4942 pte_t ptent
= *(pte
++);
4948 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
4949 case MC_TARGET_PAGE
:
4951 if (isolate_lru_page(page
))
4953 if (!mem_cgroup_move_account(page
, 1, mc
.from
, mc
.to
)) {
4955 /* we uncharge from mc.from later. */
4958 putback_lru_page(page
);
4959 put
: /* get_mctgt_type() gets the page */
4962 case MC_TARGET_SWAP
:
4964 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
4966 /* we fixup refcnts and charges later. */
4974 pte_unmap_unlock(pte
- 1, ptl
);
4979 * We have consumed all precharges we got in can_attach().
4980 * We try charge one by one, but don't do any additional
4981 * charges to mc.to if we have failed in charge once in attach()
4984 ret
= mem_cgroup_do_precharge(1);
4992 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
4994 struct mm_walk mem_cgroup_move_charge_walk
= {
4995 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4999 lru_add_drain_all();
5001 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5002 * move_lock while we're moving its pages to another memcg.
5003 * Then wait for already started RCU-only updates to finish.
5005 atomic_inc(&mc
.from
->moving_account
);
5008 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
5010 * Someone who are holding the mmap_sem might be waiting in
5011 * waitq. So we cancel all extra charges, wake up all waiters,
5012 * and retry. Because we cancel precharges, we might not be able
5013 * to move enough charges, but moving charge is a best-effort
5014 * feature anyway, so it wouldn't be a big problem.
5016 __mem_cgroup_clear_mc();
5021 * When we have consumed all precharges and failed in doing
5022 * additional charge, the page walk just aborts.
5024 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk
);
5025 up_read(&mm
->mmap_sem
);
5026 atomic_dec(&mc
.from
->moving_account
);
5029 static void mem_cgroup_move_task(struct cgroup_taskset
*tset
)
5031 struct cgroup_subsys_state
*css
;
5032 struct task_struct
*p
= cgroup_taskset_first(tset
, &css
);
5033 struct mm_struct
*mm
= get_task_mm(p
);
5037 mem_cgroup_move_charge(mm
);
5041 mem_cgroup_clear_mc();
5043 #else /* !CONFIG_MMU */
5044 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5048 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5051 static void mem_cgroup_move_task(struct cgroup_taskset
*tset
)
5057 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5058 * to verify whether we're attached to the default hierarchy on each mount
5061 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5064 * use_hierarchy is forced on the default hierarchy. cgroup core
5065 * guarantees that @root doesn't have any children, so turning it
5066 * on for the root memcg is enough.
5068 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5069 root_mem_cgroup
->use_hierarchy
= true;
5071 root_mem_cgroup
->use_hierarchy
= false;
5074 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5077 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5079 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
5082 static int memory_low_show(struct seq_file
*m
, void *v
)
5084 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5085 unsigned long low
= READ_ONCE(memcg
->low
);
5087 if (low
== PAGE_COUNTER_MAX
)
5088 seq_puts(m
, "max\n");
5090 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
5095 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5096 char *buf
, size_t nbytes
, loff_t off
)
5098 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5102 buf
= strstrip(buf
);
5103 err
= page_counter_memparse(buf
, "max", &low
);
5112 static int memory_high_show(struct seq_file
*m
, void *v
)
5114 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5115 unsigned long high
= READ_ONCE(memcg
->high
);
5117 if (high
== PAGE_COUNTER_MAX
)
5118 seq_puts(m
, "max\n");
5120 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
5125 static ssize_t
memory_high_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", &high
);
5139 memcg_wb_domain_size_changed(memcg
);
5143 static int memory_max_show(struct seq_file
*m
, void *v
)
5145 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5146 unsigned long max
= READ_ONCE(memcg
->memory
.limit
);
5148 if (max
== PAGE_COUNTER_MAX
)
5149 seq_puts(m
, "max\n");
5151 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5156 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5157 char *buf
, size_t nbytes
, loff_t off
)
5159 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5163 buf
= strstrip(buf
);
5164 err
= page_counter_memparse(buf
, "max", &max
);
5168 err
= mem_cgroup_resize_limit(memcg
, max
);
5172 memcg_wb_domain_size_changed(memcg
);
5176 static int memory_events_show(struct seq_file
*m
, void *v
)
5178 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5180 seq_printf(m
, "low %lu\n", mem_cgroup_read_events(memcg
, MEMCG_LOW
));
5181 seq_printf(m
, "high %lu\n", mem_cgroup_read_events(memcg
, MEMCG_HIGH
));
5182 seq_printf(m
, "max %lu\n", mem_cgroup_read_events(memcg
, MEMCG_MAX
));
5183 seq_printf(m
, "oom %lu\n", mem_cgroup_read_events(memcg
, MEMCG_OOM
));
5188 static struct cftype memory_files
[] = {
5191 .flags
= CFTYPE_NOT_ON_ROOT
,
5192 .read_u64
= memory_current_read
,
5196 .flags
= CFTYPE_NOT_ON_ROOT
,
5197 .seq_show
= memory_low_show
,
5198 .write
= memory_low_write
,
5202 .flags
= CFTYPE_NOT_ON_ROOT
,
5203 .seq_show
= memory_high_show
,
5204 .write
= memory_high_write
,
5208 .flags
= CFTYPE_NOT_ON_ROOT
,
5209 .seq_show
= memory_max_show
,
5210 .write
= memory_max_write
,
5214 .flags
= CFTYPE_NOT_ON_ROOT
,
5215 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
5216 .seq_show
= memory_events_show
,
5221 struct cgroup_subsys memory_cgrp_subsys
= {
5222 .css_alloc
= mem_cgroup_css_alloc
,
5223 .css_online
= mem_cgroup_css_online
,
5224 .css_offline
= mem_cgroup_css_offline
,
5225 .css_released
= mem_cgroup_css_released
,
5226 .css_free
= mem_cgroup_css_free
,
5227 .css_reset
= mem_cgroup_css_reset
,
5228 .can_attach
= mem_cgroup_can_attach
,
5229 .cancel_attach
= mem_cgroup_cancel_attach
,
5230 .attach
= mem_cgroup_move_task
,
5231 .bind
= mem_cgroup_bind
,
5232 .dfl_cftypes
= memory_files
,
5233 .legacy_cftypes
= mem_cgroup_legacy_files
,
5238 * mem_cgroup_low - check if memory consumption is below the normal range
5239 * @root: the highest ancestor to consider
5240 * @memcg: the memory cgroup to check
5242 * Returns %true if memory consumption of @memcg, and that of all
5243 * configurable ancestors up to @root, is below the normal range.
5245 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5247 if (mem_cgroup_disabled())
5251 * The toplevel group doesn't have a configurable range, so
5252 * it's never low when looked at directly, and it is not
5253 * considered an ancestor when assessing the hierarchy.
5256 if (memcg
== root_mem_cgroup
)
5259 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5262 while (memcg
!= root
) {
5263 memcg
= parent_mem_cgroup(memcg
);
5265 if (memcg
== root_mem_cgroup
)
5268 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5275 * mem_cgroup_try_charge - try charging a page
5276 * @page: page to charge
5277 * @mm: mm context of the victim
5278 * @gfp_mask: reclaim mode
5279 * @memcgp: charged memcg return
5281 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5282 * pages according to @gfp_mask if necessary.
5284 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5285 * Otherwise, an error code is returned.
5287 * After page->mapping has been set up, the caller must finalize the
5288 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5289 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5291 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5292 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
5294 struct mem_cgroup
*memcg
= NULL
;
5295 unsigned int nr_pages
= 1;
5298 if (mem_cgroup_disabled())
5301 if (PageSwapCache(page
)) {
5303 * Every swap fault against a single page tries to charge the
5304 * page, bail as early as possible. shmem_unuse() encounters
5305 * already charged pages, too. The USED bit is protected by
5306 * the page lock, which serializes swap cache removal, which
5307 * in turn serializes uncharging.
5309 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5310 if (page
->mem_cgroup
)
5313 if (do_swap_account
) {
5314 swp_entry_t ent
= { .val
= page_private(page
), };
5315 unsigned short id
= lookup_swap_cgroup_id(ent
);
5318 memcg
= mem_cgroup_from_id(id
);
5319 if (memcg
&& !css_tryget_online(&memcg
->css
))
5325 if (PageTransHuge(page
)) {
5326 nr_pages
<<= compound_order(page
);
5327 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5331 memcg
= get_mem_cgroup_from_mm(mm
);
5333 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5335 css_put(&memcg
->css
);
5342 * mem_cgroup_commit_charge - commit a page charge
5343 * @page: page to charge
5344 * @memcg: memcg to charge the page to
5345 * @lrucare: page might be on LRU already
5347 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5348 * after page->mapping has been set up. This must happen atomically
5349 * as part of the page instantiation, i.e. under the page table lock
5350 * for anonymous pages, under the page lock for page and swap cache.
5352 * In addition, the page must not be on the LRU during the commit, to
5353 * prevent racing with task migration. If it might be, use @lrucare.
5355 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5357 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5360 unsigned int nr_pages
= 1;
5362 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5363 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5365 if (mem_cgroup_disabled())
5368 * Swap faults will attempt to charge the same page multiple
5369 * times. But reuse_swap_page() might have removed the page
5370 * from swapcache already, so we can't check PageSwapCache().
5375 commit_charge(page
, memcg
, lrucare
);
5377 if (PageTransHuge(page
)) {
5378 nr_pages
<<= compound_order(page
);
5379 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5382 local_irq_disable();
5383 mem_cgroup_charge_statistics(memcg
, page
, nr_pages
);
5384 memcg_check_events(memcg
, page
);
5387 if (do_swap_account
&& PageSwapCache(page
)) {
5388 swp_entry_t entry
= { .val
= page_private(page
) };
5390 * The swap entry might not get freed for a long time,
5391 * let's not wait for it. The page already received a
5392 * memory+swap charge, drop the swap entry duplicate.
5394 mem_cgroup_uncharge_swap(entry
);
5399 * mem_cgroup_cancel_charge - cancel a page charge
5400 * @page: page to charge
5401 * @memcg: memcg to charge the page to
5403 * Cancel a charge transaction started by mem_cgroup_try_charge().
5405 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
)
5407 unsigned int nr_pages
= 1;
5409 if (mem_cgroup_disabled())
5412 * Swap faults will attempt to charge the same page multiple
5413 * times. But reuse_swap_page() might have removed the page
5414 * from swapcache already, so we can't check PageSwapCache().
5419 if (PageTransHuge(page
)) {
5420 nr_pages
<<= compound_order(page
);
5421 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5424 cancel_charge(memcg
, nr_pages
);
5427 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5428 unsigned long nr_anon
, unsigned long nr_file
,
5429 unsigned long nr_huge
, struct page
*dummy_page
)
5431 unsigned long nr_pages
= nr_anon
+ nr_file
;
5432 unsigned long flags
;
5434 if (!mem_cgroup_is_root(memcg
)) {
5435 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5436 if (do_swap_account
)
5437 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5438 memcg_oom_recover(memcg
);
5441 local_irq_save(flags
);
5442 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5443 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5444 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5445 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5446 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5447 memcg_check_events(memcg
, dummy_page
);
5448 local_irq_restore(flags
);
5450 if (!mem_cgroup_is_root(memcg
))
5451 css_put_many(&memcg
->css
, nr_pages
);
5454 static void uncharge_list(struct list_head
*page_list
)
5456 struct mem_cgroup
*memcg
= NULL
;
5457 unsigned long nr_anon
= 0;
5458 unsigned long nr_file
= 0;
5459 unsigned long nr_huge
= 0;
5460 unsigned long pgpgout
= 0;
5461 struct list_head
*next
;
5464 next
= page_list
->next
;
5466 unsigned int nr_pages
= 1;
5468 page
= list_entry(next
, struct page
, lru
);
5469 next
= page
->lru
.next
;
5471 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5472 VM_BUG_ON_PAGE(page_count(page
), page
);
5474 if (!page
->mem_cgroup
)
5478 * Nobody should be changing or seriously looking at
5479 * page->mem_cgroup at this point, we have fully
5480 * exclusive access to the page.
5483 if (memcg
!= page
->mem_cgroup
) {
5485 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5487 pgpgout
= nr_anon
= nr_file
= nr_huge
= 0;
5489 memcg
= page
->mem_cgroup
;
5492 if (PageTransHuge(page
)) {
5493 nr_pages
<<= compound_order(page
);
5494 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5495 nr_huge
+= nr_pages
;
5499 nr_anon
+= nr_pages
;
5501 nr_file
+= nr_pages
;
5503 page
->mem_cgroup
= NULL
;
5506 } while (next
!= page_list
);
5509 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5514 * mem_cgroup_uncharge - uncharge a page
5515 * @page: page to uncharge
5517 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5518 * mem_cgroup_commit_charge().
5520 void mem_cgroup_uncharge(struct page
*page
)
5522 if (mem_cgroup_disabled())
5525 /* Don't touch page->lru of any random page, pre-check: */
5526 if (!page
->mem_cgroup
)
5529 INIT_LIST_HEAD(&page
->lru
);
5530 uncharge_list(&page
->lru
);
5534 * mem_cgroup_uncharge_list - uncharge a list of page
5535 * @page_list: list of pages to uncharge
5537 * Uncharge a list of pages previously charged with
5538 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5540 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5542 if (mem_cgroup_disabled())
5545 if (!list_empty(page_list
))
5546 uncharge_list(page_list
);
5550 * mem_cgroup_replace_page - migrate a charge to another page
5551 * @oldpage: currently charged page
5552 * @newpage: page to transfer the charge to
5554 * Migrate the charge from @oldpage to @newpage.
5556 * Both pages must be locked, @newpage->mapping must be set up.
5557 * Either or both pages might be on the LRU already.
5559 void mem_cgroup_replace_page(struct page
*oldpage
, struct page
*newpage
)
5561 struct mem_cgroup
*memcg
;
5564 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5565 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5566 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5567 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5570 if (mem_cgroup_disabled())
5573 /* Page cache replacement: new page already charged? */
5574 if (newpage
->mem_cgroup
)
5577 /* Swapcache readahead pages can get replaced before being charged */
5578 memcg
= oldpage
->mem_cgroup
;
5582 lock_page_lru(oldpage
, &isolated
);
5583 oldpage
->mem_cgroup
= NULL
;
5584 unlock_page_lru(oldpage
, isolated
);
5586 commit_charge(newpage
, memcg
, true);
5590 * subsys_initcall() for memory controller.
5592 * Some parts like hotcpu_notifier() have to be initialized from this context
5593 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5594 * everything that doesn't depend on a specific mem_cgroup structure should
5595 * be initialized from here.
5597 static int __init
mem_cgroup_init(void)
5601 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5603 for_each_possible_cpu(cpu
)
5604 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5607 for_each_node(node
) {
5608 struct mem_cgroup_tree_per_node
*rtpn
;
5611 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5612 node_online(node
) ? node
: NUMA_NO_NODE
);
5614 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5615 struct mem_cgroup_tree_per_zone
*rtpz
;
5617 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
5618 rtpz
->rb_root
= RB_ROOT
;
5619 spin_lock_init(&rtpz
->lock
);
5621 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5626 subsys_initcall(mem_cgroup_init
);
5628 #ifdef CONFIG_MEMCG_SWAP
5630 * mem_cgroup_swapout - transfer a memsw charge to swap
5631 * @page: page whose memsw charge to transfer
5632 * @entry: swap entry to move the charge to
5634 * Transfer the memsw charge of @page to @entry.
5636 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5638 struct mem_cgroup
*memcg
;
5639 unsigned short oldid
;
5641 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5642 VM_BUG_ON_PAGE(page_count(page
), page
);
5644 if (!do_swap_account
)
5647 memcg
= page
->mem_cgroup
;
5649 /* Readahead page, never charged */
5653 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5654 VM_BUG_ON_PAGE(oldid
, page
);
5655 mem_cgroup_swap_statistics(memcg
, true);
5657 page
->mem_cgroup
= NULL
;
5659 if (!mem_cgroup_is_root(memcg
))
5660 page_counter_uncharge(&memcg
->memory
, 1);
5663 * Interrupts should be disabled here because the caller holds the
5664 * mapping->tree_lock lock which is taken with interrupts-off. It is
5665 * important here to have the interrupts disabled because it is the
5666 * only synchronisation we have for udpating the per-CPU variables.
5668 VM_BUG_ON(!irqs_disabled());
5669 mem_cgroup_charge_statistics(memcg
, page
, -1);
5670 memcg_check_events(memcg
, page
);
5674 * mem_cgroup_uncharge_swap - uncharge a swap entry
5675 * @entry: swap entry to uncharge
5677 * Drop the memsw charge associated with @entry.
5679 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5681 struct mem_cgroup
*memcg
;
5684 if (!do_swap_account
)
5687 id
= swap_cgroup_record(entry
, 0);
5689 memcg
= mem_cgroup_from_id(id
);
5691 if (!mem_cgroup_is_root(memcg
))
5692 page_counter_uncharge(&memcg
->memsw
, 1);
5693 mem_cgroup_swap_statistics(memcg
, false);
5694 css_put(&memcg
->css
);
5699 /* for remember boot option*/
5700 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5701 static int really_do_swap_account __initdata
= 1;
5703 static int really_do_swap_account __initdata
;
5706 static int __init
enable_swap_account(char *s
)
5708 if (!strcmp(s
, "1"))
5709 really_do_swap_account
= 1;
5710 else if (!strcmp(s
, "0"))
5711 really_do_swap_account
= 0;
5714 __setup("swapaccount=", enable_swap_account
);
5716 static struct cftype memsw_cgroup_files
[] = {
5718 .name
= "memsw.usage_in_bytes",
5719 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5720 .read_u64
= mem_cgroup_read_u64
,
5723 .name
= "memsw.max_usage_in_bytes",
5724 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5725 .write
= mem_cgroup_reset
,
5726 .read_u64
= mem_cgroup_read_u64
,
5729 .name
= "memsw.limit_in_bytes",
5730 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5731 .write
= mem_cgroup_write
,
5732 .read_u64
= mem_cgroup_read_u64
,
5735 .name
= "memsw.failcnt",
5736 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5737 .write
= mem_cgroup_reset
,
5738 .read_u64
= mem_cgroup_read_u64
,
5740 { }, /* terminate */
5743 static int __init
mem_cgroup_swap_init(void)
5745 if (!mem_cgroup_disabled() && really_do_swap_account
) {
5746 do_swap_account
= 1;
5747 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
5748 memsw_cgroup_files
));
5752 subsys_initcall(mem_cgroup_swap_init
);
5754 #endif /* CONFIG_MEMCG_SWAP */