1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
65 #include <linux/tracehook.h>
71 #include <asm/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
76 EXPORT_SYMBOL(memory_cgrp_subsys
);
78 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
80 #define MEM_CGROUP_RECLAIM_RETRIES 5
82 /* Socket memory accounting disabled? */
83 static bool cgroup_memory_nosocket
;
85 /* Kernel memory accounting disabled? */
86 static bool cgroup_memory_nokmem
;
88 /* Whether the swap controller is active */
89 #ifdef CONFIG_MEMCG_SWAP
90 int do_swap_account __read_mostly
;
92 #define do_swap_account 0
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys
) && do_swap_account
;
101 static const char * const mem_cgroup_stat_names
[] = {
111 static const char * const mem_cgroup_events_names
[] = {
118 static const char * const mem_cgroup_lru_names
[] = {
126 #define THRESHOLDS_EVENTS_TARGET 128
127 #define SOFTLIMIT_EVENTS_TARGET 1024
128 #define NUMAINFO_EVENTS_TARGET 1024
131 * Cgroups above their limits are maintained in a RB-Tree, independent of
132 * their hierarchy representation
135 struct mem_cgroup_tree_per_zone
{
136 struct rb_root rb_root
;
140 struct mem_cgroup_tree_per_node
{
141 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
144 struct mem_cgroup_tree
{
145 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
148 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
151 struct mem_cgroup_eventfd_list
{
152 struct list_head list
;
153 struct eventfd_ctx
*eventfd
;
157 * cgroup_event represents events which userspace want to receive.
159 struct mem_cgroup_event
{
161 * memcg which the event belongs to.
163 struct mem_cgroup
*memcg
;
165 * eventfd to signal userspace about the event.
167 struct eventfd_ctx
*eventfd
;
169 * Each of these stored in a list by the cgroup.
171 struct list_head list
;
173 * register_event() callback will be used to add new userspace
174 * waiter for changes related to this event. Use eventfd_signal()
175 * on eventfd to send notification to userspace.
177 int (*register_event
)(struct mem_cgroup
*memcg
,
178 struct eventfd_ctx
*eventfd
, const char *args
);
180 * unregister_event() callback will be called when userspace closes
181 * the eventfd or on cgroup removing. This callback must be set,
182 * if you want provide notification functionality.
184 void (*unregister_event
)(struct mem_cgroup
*memcg
,
185 struct eventfd_ctx
*eventfd
);
187 * All fields below needed to unregister event when
188 * userspace closes eventfd.
191 wait_queue_head_t
*wqh
;
193 struct work_struct remove
;
196 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
197 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
199 /* Stuffs for move charges at task migration. */
201 * Types of charges to be moved.
203 #define MOVE_ANON 0x1U
204 #define MOVE_FILE 0x2U
205 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
207 /* "mc" and its members are protected by cgroup_mutex */
208 static struct move_charge_struct
{
209 spinlock_t lock
; /* for from, to */
210 struct mm_struct
*mm
;
211 struct mem_cgroup
*from
;
212 struct mem_cgroup
*to
;
214 unsigned long precharge
;
215 unsigned long moved_charge
;
216 unsigned long moved_swap
;
217 struct task_struct
*moving_task
; /* a task moving charges */
218 wait_queue_head_t waitq
; /* a waitq for other context */
220 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
221 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
225 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
226 * limit reclaim to prevent infinite loops, if they ever occur.
228 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
229 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
232 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
233 MEM_CGROUP_CHARGE_TYPE_ANON
,
234 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
235 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
239 /* for encoding cft->private value on file */
248 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
249 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
250 #define MEMFILE_ATTR(val) ((val) & 0xffff)
251 /* Used for OOM nofiier */
252 #define OOM_CONTROL (0)
254 /* Some nice accessors for the vmpressure. */
255 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
258 memcg
= root_mem_cgroup
;
259 return &memcg
->vmpressure
;
262 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
264 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
267 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
269 return (memcg
== root_mem_cgroup
);
274 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
275 * The main reason for not using cgroup id for this:
276 * this works better in sparse environments, where we have a lot of memcgs,
277 * but only a few kmem-limited. Or also, if we have, for instance, 200
278 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
279 * 200 entry array for that.
281 * The current size of the caches array is stored in memcg_nr_cache_ids. It
282 * will double each time we have to increase it.
284 static DEFINE_IDA(memcg_cache_ida
);
285 int memcg_nr_cache_ids
;
287 /* Protects memcg_nr_cache_ids */
288 static DECLARE_RWSEM(memcg_cache_ids_sem
);
290 void memcg_get_cache_ids(void)
292 down_read(&memcg_cache_ids_sem
);
295 void memcg_put_cache_ids(void)
297 up_read(&memcg_cache_ids_sem
);
301 * MIN_SIZE is different than 1, because we would like to avoid going through
302 * the alloc/free process all the time. In a small machine, 4 kmem-limited
303 * cgroups is a reasonable guess. In the future, it could be a parameter or
304 * tunable, but that is strictly not necessary.
306 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
307 * this constant directly from cgroup, but it is understandable that this is
308 * better kept as an internal representation in cgroup.c. In any case, the
309 * cgrp_id space is not getting any smaller, and we don't have to necessarily
310 * increase ours as well if it increases.
312 #define MEMCG_CACHES_MIN_SIZE 4
313 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
316 * A lot of the calls to the cache allocation functions are expected to be
317 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
318 * conditional to this static branch, we'll have to allow modules that does
319 * kmem_cache_alloc and the such to see this symbol as well
321 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
322 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
324 #endif /* !CONFIG_SLOB */
326 static struct mem_cgroup_per_zone
*
327 mem_cgroup_zone_zoneinfo(struct mem_cgroup
*memcg
, struct zone
*zone
)
329 int nid
= zone_to_nid(zone
);
330 int zid
= zone_idx(zone
);
332 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
336 * mem_cgroup_css_from_page - css of the memcg associated with a page
337 * @page: page of interest
339 * If memcg is bound to the default hierarchy, css of the memcg associated
340 * with @page is returned. The returned css remains associated with @page
341 * until it is released.
343 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
346 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
348 struct mem_cgroup
*memcg
;
350 memcg
= page
->mem_cgroup
;
352 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
353 memcg
= root_mem_cgroup
;
359 * page_cgroup_ino - return inode number of the memcg a page is charged to
362 * Look up the closest online ancestor of the memory cgroup @page is charged to
363 * and return its inode number or 0 if @page is not charged to any cgroup. It
364 * is safe to call this function without holding a reference to @page.
366 * Note, this function is inherently racy, because there is nothing to prevent
367 * the cgroup inode from getting torn down and potentially reallocated a moment
368 * after page_cgroup_ino() returns, so it only should be used by callers that
369 * do not care (such as procfs interfaces).
371 ino_t
page_cgroup_ino(struct page
*page
)
373 struct mem_cgroup
*memcg
;
374 unsigned long ino
= 0;
377 memcg
= READ_ONCE(page
->mem_cgroup
);
378 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
379 memcg
= parent_mem_cgroup(memcg
);
381 ino
= cgroup_ino(memcg
->css
.cgroup
);
386 static struct mem_cgroup_per_zone
*
387 mem_cgroup_page_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
389 int nid
= page_to_nid(page
);
390 int zid
= page_zonenum(page
);
392 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
395 static struct mem_cgroup_tree_per_zone
*
396 soft_limit_tree_node_zone(int nid
, int zid
)
398 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
401 static struct mem_cgroup_tree_per_zone
*
402 soft_limit_tree_from_page(struct page
*page
)
404 int nid
= page_to_nid(page
);
405 int zid
= page_zonenum(page
);
407 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
410 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone
*mz
,
411 struct mem_cgroup_tree_per_zone
*mctz
,
412 unsigned long new_usage_in_excess
)
414 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
415 struct rb_node
*parent
= NULL
;
416 struct mem_cgroup_per_zone
*mz_node
;
421 mz
->usage_in_excess
= new_usage_in_excess
;
422 if (!mz
->usage_in_excess
)
426 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
428 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
431 * We can't avoid mem cgroups that are over their soft
432 * limit by the same amount
434 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
437 rb_link_node(&mz
->tree_node
, parent
, p
);
438 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
442 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
443 struct mem_cgroup_tree_per_zone
*mctz
)
447 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
451 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
452 struct mem_cgroup_tree_per_zone
*mctz
)
456 spin_lock_irqsave(&mctz
->lock
, flags
);
457 __mem_cgroup_remove_exceeded(mz
, mctz
);
458 spin_unlock_irqrestore(&mctz
->lock
, flags
);
461 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
463 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
464 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
465 unsigned long excess
= 0;
467 if (nr_pages
> soft_limit
)
468 excess
= nr_pages
- soft_limit
;
473 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
475 unsigned long excess
;
476 struct mem_cgroup_per_zone
*mz
;
477 struct mem_cgroup_tree_per_zone
*mctz
;
479 mctz
= soft_limit_tree_from_page(page
);
481 * Necessary to update all ancestors when hierarchy is used.
482 * because their event counter is not touched.
484 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
485 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
486 excess
= soft_limit_excess(memcg
);
488 * We have to update the tree if mz is on RB-tree or
489 * mem is over its softlimit.
491 if (excess
|| mz
->on_tree
) {
494 spin_lock_irqsave(&mctz
->lock
, flags
);
495 /* if on-tree, remove it */
497 __mem_cgroup_remove_exceeded(mz
, mctz
);
499 * Insert again. mz->usage_in_excess will be updated.
500 * If excess is 0, no tree ops.
502 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
503 spin_unlock_irqrestore(&mctz
->lock
, flags
);
508 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
510 struct mem_cgroup_tree_per_zone
*mctz
;
511 struct mem_cgroup_per_zone
*mz
;
515 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
516 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
517 mctz
= soft_limit_tree_node_zone(nid
, zid
);
518 mem_cgroup_remove_exceeded(mz
, mctz
);
523 static struct mem_cgroup_per_zone
*
524 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
526 struct rb_node
*rightmost
= NULL
;
527 struct mem_cgroup_per_zone
*mz
;
531 rightmost
= rb_last(&mctz
->rb_root
);
533 goto done
; /* Nothing to reclaim from */
535 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
537 * Remove the node now but someone else can add it back,
538 * we will to add it back at the end of reclaim to its correct
539 * position in the tree.
541 __mem_cgroup_remove_exceeded(mz
, mctz
);
542 if (!soft_limit_excess(mz
->memcg
) ||
543 !css_tryget_online(&mz
->memcg
->css
))
549 static struct mem_cgroup_per_zone
*
550 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
552 struct mem_cgroup_per_zone
*mz
;
554 spin_lock_irq(&mctz
->lock
);
555 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
556 spin_unlock_irq(&mctz
->lock
);
561 * Return page count for single (non recursive) @memcg.
563 * Implementation Note: reading percpu statistics for memcg.
565 * Both of vmstat[] and percpu_counter has threshold and do periodic
566 * synchronization to implement "quick" read. There are trade-off between
567 * reading cost and precision of value. Then, we may have a chance to implement
568 * a periodic synchronization of counter in memcg's counter.
570 * But this _read() function is used for user interface now. The user accounts
571 * memory usage by memory cgroup and he _always_ requires exact value because
572 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
573 * have to visit all online cpus and make sum. So, for now, unnecessary
574 * synchronization is not implemented. (just implemented for cpu hotplug)
576 * If there are kernel internal actions which can make use of some not-exact
577 * value, and reading all cpu value can be performance bottleneck in some
578 * common workload, threshold and synchronization as vmstat[] should be
582 mem_cgroup_read_stat(struct mem_cgroup
*memcg
, enum mem_cgroup_stat_index idx
)
587 /* Per-cpu values can be negative, use a signed accumulator */
588 for_each_possible_cpu(cpu
)
589 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
591 * Summing races with updates, so val may be negative. Avoid exposing
592 * transient negative values.
599 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
600 enum mem_cgroup_events_index idx
)
602 unsigned long val
= 0;
605 for_each_possible_cpu(cpu
)
606 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
610 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
612 bool compound
, int nr_pages
)
615 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
616 * counted as CACHE even if it's on ANON LRU.
619 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
622 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
626 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
627 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
631 /* pagein of a big page is an event. So, ignore page size */
633 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
635 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
636 nr_pages
= -nr_pages
; /* for event */
639 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
642 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
643 int nid
, unsigned int lru_mask
)
645 unsigned long nr
= 0;
648 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
650 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
651 struct mem_cgroup_per_zone
*mz
;
655 if (!(BIT(lru
) & lru_mask
))
657 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
658 nr
+= mz
->lru_size
[lru
];
664 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
665 unsigned int lru_mask
)
667 unsigned long nr
= 0;
670 for_each_node_state(nid
, N_MEMORY
)
671 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
675 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
676 enum mem_cgroup_events_target target
)
678 unsigned long val
, next
;
680 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
681 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
682 /* from time_after() in jiffies.h */
683 if ((long)next
- (long)val
< 0) {
685 case MEM_CGROUP_TARGET_THRESH
:
686 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
688 case MEM_CGROUP_TARGET_SOFTLIMIT
:
689 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
691 case MEM_CGROUP_TARGET_NUMAINFO
:
692 next
= val
+ NUMAINFO_EVENTS_TARGET
;
697 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
704 * Check events in order.
707 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
709 /* threshold event is triggered in finer grain than soft limit */
710 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
711 MEM_CGROUP_TARGET_THRESH
))) {
713 bool do_numainfo __maybe_unused
;
715 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
716 MEM_CGROUP_TARGET_SOFTLIMIT
);
718 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
719 MEM_CGROUP_TARGET_NUMAINFO
);
721 mem_cgroup_threshold(memcg
);
722 if (unlikely(do_softlimit
))
723 mem_cgroup_update_tree(memcg
, page
);
725 if (unlikely(do_numainfo
))
726 atomic_inc(&memcg
->numainfo_events
);
731 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
734 * mm_update_next_owner() may clear mm->owner to NULL
735 * if it races with swapoff, page migration, etc.
736 * So this can be called with p == NULL.
741 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
743 EXPORT_SYMBOL(mem_cgroup_from_task
);
745 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
747 struct mem_cgroup
*memcg
= NULL
;
752 * Page cache insertions can happen withou an
753 * actual mm context, e.g. during disk probing
754 * on boot, loopback IO, acct() writes etc.
757 memcg
= root_mem_cgroup
;
759 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
760 if (unlikely(!memcg
))
761 memcg
= root_mem_cgroup
;
763 } while (!css_tryget_online(&memcg
->css
));
769 * mem_cgroup_iter - iterate over memory cgroup hierarchy
770 * @root: hierarchy root
771 * @prev: previously returned memcg, NULL on first invocation
772 * @reclaim: cookie for shared reclaim walks, NULL for full walks
774 * Returns references to children of the hierarchy below @root, or
775 * @root itself, or %NULL after a full round-trip.
777 * Caller must pass the return value in @prev on subsequent
778 * invocations for reference counting, or use mem_cgroup_iter_break()
779 * to cancel a hierarchy walk before the round-trip is complete.
781 * Reclaimers can specify a zone and a priority level in @reclaim to
782 * divide up the memcgs in the hierarchy among all concurrent
783 * reclaimers operating on the same zone and priority.
785 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
786 struct mem_cgroup
*prev
,
787 struct mem_cgroup_reclaim_cookie
*reclaim
)
789 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
790 struct cgroup_subsys_state
*css
= NULL
;
791 struct mem_cgroup
*memcg
= NULL
;
792 struct mem_cgroup
*pos
= NULL
;
794 if (mem_cgroup_disabled())
798 root
= root_mem_cgroup
;
800 if (prev
&& !reclaim
)
803 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
812 struct mem_cgroup_per_zone
*mz
;
814 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
815 iter
= &mz
->iter
[reclaim
->priority
];
817 if (prev
&& reclaim
->generation
!= iter
->generation
)
821 pos
= READ_ONCE(iter
->position
);
822 if (!pos
|| css_tryget(&pos
->css
))
825 * css reference reached zero, so iter->position will
826 * be cleared by ->css_released. However, we should not
827 * rely on this happening soon, because ->css_released
828 * is called from a work queue, and by busy-waiting we
829 * might block it. So we clear iter->position right
832 (void)cmpxchg(&iter
->position
, pos
, NULL
);
840 css
= css_next_descendant_pre(css
, &root
->css
);
843 * Reclaimers share the hierarchy walk, and a
844 * new one might jump in right at the end of
845 * the hierarchy - make sure they see at least
846 * one group and restart from the beginning.
854 * Verify the css and acquire a reference. The root
855 * is provided by the caller, so we know it's alive
856 * and kicking, and don't take an extra reference.
858 memcg
= mem_cgroup_from_css(css
);
860 if (css
== &root
->css
)
871 * The position could have already been updated by a competing
872 * thread, so check that the value hasn't changed since we read
873 * it to avoid reclaiming from the same cgroup twice.
875 (void)cmpxchg(&iter
->position
, pos
, memcg
);
883 reclaim
->generation
= iter
->generation
;
889 if (prev
&& prev
!= root
)
896 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
897 * @root: hierarchy root
898 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
900 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
901 struct mem_cgroup
*prev
)
904 root
= root_mem_cgroup
;
905 if (prev
&& prev
!= root
)
909 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
911 struct mem_cgroup
*memcg
= dead_memcg
;
912 struct mem_cgroup_reclaim_iter
*iter
;
913 struct mem_cgroup_per_zone
*mz
;
917 while ((memcg
= parent_mem_cgroup(memcg
))) {
919 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
920 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
921 for (i
= 0; i
<= DEF_PRIORITY
; i
++) {
923 cmpxchg(&iter
->position
,
932 * Iteration constructs for visiting all cgroups (under a tree). If
933 * loops are exited prematurely (break), mem_cgroup_iter_break() must
934 * be used for reference counting.
936 #define for_each_mem_cgroup_tree(iter, root) \
937 for (iter = mem_cgroup_iter(root, NULL, NULL); \
939 iter = mem_cgroup_iter(root, iter, NULL))
941 #define for_each_mem_cgroup(iter) \
942 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
944 iter = mem_cgroup_iter(NULL, iter, NULL))
947 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
948 * @zone: zone of the wanted lruvec
949 * @memcg: memcg of the wanted lruvec
951 * Returns the lru list vector holding pages for the given @zone and
952 * @mem. This can be the global zone lruvec, if the memory controller
955 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
956 struct mem_cgroup
*memcg
)
958 struct mem_cgroup_per_zone
*mz
;
959 struct lruvec
*lruvec
;
961 if (mem_cgroup_disabled()) {
962 lruvec
= &zone
->lruvec
;
966 mz
= mem_cgroup_zone_zoneinfo(memcg
, zone
);
967 lruvec
= &mz
->lruvec
;
970 * Since a node can be onlined after the mem_cgroup was created,
971 * we have to be prepared to initialize lruvec->zone here;
972 * and if offlined then reonlined, we need to reinitialize it.
974 if (unlikely(lruvec
->zone
!= zone
))
980 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
982 * @zone: zone of the page
984 * This function is only safe when following the LRU page isolation
985 * and putback protocol: the LRU lock must be held, and the page must
986 * either be PageLRU() or the caller must have isolated/allocated it.
988 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
990 struct mem_cgroup_per_zone
*mz
;
991 struct mem_cgroup
*memcg
;
992 struct lruvec
*lruvec
;
994 if (mem_cgroup_disabled()) {
995 lruvec
= &zone
->lruvec
;
999 memcg
= page
->mem_cgroup
;
1001 * Swapcache readahead pages are added to the LRU - and
1002 * possibly migrated - before they are charged.
1005 memcg
= root_mem_cgroup
;
1007 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1008 lruvec
= &mz
->lruvec
;
1011 * Since a node can be onlined after the mem_cgroup was created,
1012 * we have to be prepared to initialize lruvec->zone here;
1013 * and if offlined then reonlined, we need to reinitialize it.
1015 if (unlikely(lruvec
->zone
!= zone
))
1016 lruvec
->zone
= zone
;
1021 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1022 * @lruvec: mem_cgroup per zone lru vector
1023 * @lru: index of lru list the page is sitting on
1024 * @nr_pages: positive when adding or negative when removing
1026 * This function must be called under lru_lock, just before a page is added
1027 * to or just after a page is removed from an lru list (that ordering being
1028 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1030 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1033 struct mem_cgroup_per_zone
*mz
;
1034 unsigned long *lru_size
;
1038 __update_lru_size(lruvec
, lru
, nr_pages
);
1040 if (mem_cgroup_disabled())
1043 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1044 lru_size
= mz
->lru_size
+ lru
;
1045 empty
= list_empty(lruvec
->lists
+ lru
);
1048 *lru_size
+= nr_pages
;
1051 if (WARN_ONCE(size
< 0 || empty
!= !size
,
1052 "%s(%p, %d, %d): lru_size %ld but %sempty\n",
1053 __func__
, lruvec
, lru
, nr_pages
, size
, empty
? "" : "not ")) {
1059 *lru_size
+= nr_pages
;
1062 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1064 struct mem_cgroup
*task_memcg
;
1065 struct task_struct
*p
;
1068 p
= find_lock_task_mm(task
);
1070 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1074 * All threads may have already detached their mm's, but the oom
1075 * killer still needs to detect if they have already been oom
1076 * killed to prevent needlessly killing additional tasks.
1079 task_memcg
= mem_cgroup_from_task(task
);
1080 css_get(&task_memcg
->css
);
1083 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1084 css_put(&task_memcg
->css
);
1089 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1090 * @memcg: the memory cgroup
1092 * Returns the maximum amount of memory @mem can be charged with, in
1095 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1097 unsigned long margin
= 0;
1098 unsigned long count
;
1099 unsigned long limit
;
1101 count
= page_counter_read(&memcg
->memory
);
1102 limit
= READ_ONCE(memcg
->memory
.limit
);
1104 margin
= limit
- count
;
1106 if (do_memsw_account()) {
1107 count
= page_counter_read(&memcg
->memsw
);
1108 limit
= READ_ONCE(memcg
->memsw
.limit
);
1110 margin
= min(margin
, limit
- count
);
1119 * A routine for checking "mem" is under move_account() or not.
1121 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1122 * moving cgroups. This is for waiting at high-memory pressure
1125 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1127 struct mem_cgroup
*from
;
1128 struct mem_cgroup
*to
;
1131 * Unlike task_move routines, we access mc.to, mc.from not under
1132 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1134 spin_lock(&mc
.lock
);
1140 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1141 mem_cgroup_is_descendant(to
, memcg
);
1143 spin_unlock(&mc
.lock
);
1147 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1149 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1150 if (mem_cgroup_under_move(memcg
)) {
1152 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1153 /* moving charge context might have finished. */
1156 finish_wait(&mc
.waitq
, &wait
);
1163 #define K(x) ((x) << (PAGE_SHIFT-10))
1165 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1166 * @memcg: The memory cgroup that went over limit
1167 * @p: Task that is going to be killed
1169 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1172 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1174 struct mem_cgroup
*iter
;
1180 pr_info("Task in ");
1181 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1182 pr_cont(" killed as a result of limit of ");
1184 pr_info("Memory limit reached of cgroup ");
1187 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1192 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1193 K((u64
)page_counter_read(&memcg
->memory
)),
1194 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1195 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1196 K((u64
)page_counter_read(&memcg
->memsw
)),
1197 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1198 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1199 K((u64
)page_counter_read(&memcg
->kmem
)),
1200 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1202 for_each_mem_cgroup_tree(iter
, memcg
) {
1203 pr_info("Memory cgroup stats for ");
1204 pr_cont_cgroup_path(iter
->css
.cgroup
);
1207 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1208 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1210 pr_cont(" %s:%luKB", mem_cgroup_stat_names
[i
],
1211 K(mem_cgroup_read_stat(iter
, i
)));
1214 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1215 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1216 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1223 * This function returns the number of memcg under hierarchy tree. Returns
1224 * 1(self count) if no children.
1226 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1229 struct mem_cgroup
*iter
;
1231 for_each_mem_cgroup_tree(iter
, memcg
)
1237 * Return the memory (and swap, if configured) limit for a memcg.
1239 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1241 unsigned long limit
;
1243 limit
= memcg
->memory
.limit
;
1244 if (mem_cgroup_swappiness(memcg
)) {
1245 unsigned long memsw_limit
;
1246 unsigned long swap_limit
;
1248 memsw_limit
= memcg
->memsw
.limit
;
1249 swap_limit
= memcg
->swap
.limit
;
1250 swap_limit
= min(swap_limit
, (unsigned long)total_swap_pages
);
1251 limit
= min(limit
+ swap_limit
, memsw_limit
);
1256 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1259 struct oom_control oc
= {
1263 .gfp_mask
= gfp_mask
,
1266 struct mem_cgroup
*iter
;
1267 unsigned long chosen_points
= 0;
1268 unsigned long totalpages
;
1269 unsigned int points
= 0;
1270 struct task_struct
*chosen
= NULL
;
1272 mutex_lock(&oom_lock
);
1275 * If current has a pending SIGKILL or is exiting, then automatically
1276 * select it. The goal is to allow it to allocate so that it may
1277 * quickly exit and free its memory.
1279 if (task_will_free_mem(current
)) {
1280 mark_oom_victim(current
);
1281 wake_oom_reaper(current
);
1285 check_panic_on_oom(&oc
, CONSTRAINT_MEMCG
);
1286 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1287 for_each_mem_cgroup_tree(iter
, memcg
) {
1288 struct css_task_iter it
;
1289 struct task_struct
*task
;
1291 css_task_iter_start(&iter
->css
, &it
);
1292 while ((task
= css_task_iter_next(&it
))) {
1293 switch (oom_scan_process_thread(&oc
, task
)) {
1294 case OOM_SCAN_SELECT
:
1296 put_task_struct(chosen
);
1298 chosen_points
= ULONG_MAX
;
1299 get_task_struct(chosen
);
1301 case OOM_SCAN_CONTINUE
:
1303 case OOM_SCAN_ABORT
:
1304 css_task_iter_end(&it
);
1305 mem_cgroup_iter_break(memcg
, iter
);
1307 put_task_struct(chosen
);
1308 /* Set a dummy value to return "true". */
1309 chosen
= (void *) 1;
1314 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1315 if (!points
|| points
< chosen_points
)
1317 /* Prefer thread group leaders for display purposes */
1318 if (points
== chosen_points
&&
1319 thread_group_leader(chosen
))
1323 put_task_struct(chosen
);
1325 chosen_points
= points
;
1326 get_task_struct(chosen
);
1328 css_task_iter_end(&it
);
1332 points
= chosen_points
* 1000 / totalpages
;
1333 oom_kill_process(&oc
, chosen
, points
, totalpages
,
1334 "Memory cgroup out of memory");
1337 mutex_unlock(&oom_lock
);
1341 #if MAX_NUMNODES > 1
1344 * test_mem_cgroup_node_reclaimable
1345 * @memcg: the target memcg
1346 * @nid: the node ID to be checked.
1347 * @noswap : specify true here if the user wants flle only information.
1349 * This function returns whether the specified memcg contains any
1350 * reclaimable pages on a node. Returns true if there are any reclaimable
1351 * pages in the node.
1353 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1354 int nid
, bool noswap
)
1356 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1358 if (noswap
|| !total_swap_pages
)
1360 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1367 * Always updating the nodemask is not very good - even if we have an empty
1368 * list or the wrong list here, we can start from some node and traverse all
1369 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1372 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1376 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1377 * pagein/pageout changes since the last update.
1379 if (!atomic_read(&memcg
->numainfo_events
))
1381 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1384 /* make a nodemask where this memcg uses memory from */
1385 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1387 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1389 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1390 node_clear(nid
, memcg
->scan_nodes
);
1393 atomic_set(&memcg
->numainfo_events
, 0);
1394 atomic_set(&memcg
->numainfo_updating
, 0);
1398 * Selecting a node where we start reclaim from. Because what we need is just
1399 * reducing usage counter, start from anywhere is O,K. Considering
1400 * memory reclaim from current node, there are pros. and cons.
1402 * Freeing memory from current node means freeing memory from a node which
1403 * we'll use or we've used. So, it may make LRU bad. And if several threads
1404 * hit limits, it will see a contention on a node. But freeing from remote
1405 * node means more costs for memory reclaim because of memory latency.
1407 * Now, we use round-robin. Better algorithm is welcomed.
1409 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1413 mem_cgroup_may_update_nodemask(memcg
);
1414 node
= memcg
->last_scanned_node
;
1416 node
= next_node_in(node
, memcg
->scan_nodes
);
1418 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1419 * last time it really checked all the LRUs due to rate limiting.
1420 * Fallback to the current node in that case for simplicity.
1422 if (unlikely(node
== MAX_NUMNODES
))
1423 node
= numa_node_id();
1425 memcg
->last_scanned_node
= node
;
1429 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1435 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1438 unsigned long *total_scanned
)
1440 struct mem_cgroup
*victim
= NULL
;
1443 unsigned long excess
;
1444 unsigned long nr_scanned
;
1445 struct mem_cgroup_reclaim_cookie reclaim
= {
1450 excess
= soft_limit_excess(root_memcg
);
1453 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1458 * If we have not been able to reclaim
1459 * anything, it might because there are
1460 * no reclaimable pages under this hierarchy
1465 * We want to do more targeted reclaim.
1466 * excess >> 2 is not to excessive so as to
1467 * reclaim too much, nor too less that we keep
1468 * coming back to reclaim from this cgroup
1470 if (total
>= (excess
>> 2) ||
1471 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1476 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1478 *total_scanned
+= nr_scanned
;
1479 if (!soft_limit_excess(root_memcg
))
1482 mem_cgroup_iter_break(root_memcg
, victim
);
1486 #ifdef CONFIG_LOCKDEP
1487 static struct lockdep_map memcg_oom_lock_dep_map
= {
1488 .name
= "memcg_oom_lock",
1492 static DEFINE_SPINLOCK(memcg_oom_lock
);
1495 * Check OOM-Killer is already running under our hierarchy.
1496 * If someone is running, return false.
1498 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1500 struct mem_cgroup
*iter
, *failed
= NULL
;
1502 spin_lock(&memcg_oom_lock
);
1504 for_each_mem_cgroup_tree(iter
, memcg
) {
1505 if (iter
->oom_lock
) {
1507 * this subtree of our hierarchy is already locked
1508 * so we cannot give a lock.
1511 mem_cgroup_iter_break(memcg
, iter
);
1514 iter
->oom_lock
= true;
1519 * OK, we failed to lock the whole subtree so we have
1520 * to clean up what we set up to the failing subtree
1522 for_each_mem_cgroup_tree(iter
, memcg
) {
1523 if (iter
== failed
) {
1524 mem_cgroup_iter_break(memcg
, iter
);
1527 iter
->oom_lock
= false;
1530 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1532 spin_unlock(&memcg_oom_lock
);
1537 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1539 struct mem_cgroup
*iter
;
1541 spin_lock(&memcg_oom_lock
);
1542 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1543 for_each_mem_cgroup_tree(iter
, memcg
)
1544 iter
->oom_lock
= false;
1545 spin_unlock(&memcg_oom_lock
);
1548 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1550 struct mem_cgroup
*iter
;
1552 spin_lock(&memcg_oom_lock
);
1553 for_each_mem_cgroup_tree(iter
, memcg
)
1555 spin_unlock(&memcg_oom_lock
);
1558 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1560 struct mem_cgroup
*iter
;
1563 * When a new child is created while the hierarchy is under oom,
1564 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1566 spin_lock(&memcg_oom_lock
);
1567 for_each_mem_cgroup_tree(iter
, memcg
)
1568 if (iter
->under_oom
> 0)
1570 spin_unlock(&memcg_oom_lock
);
1573 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1575 struct oom_wait_info
{
1576 struct mem_cgroup
*memcg
;
1580 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1581 unsigned mode
, int sync
, void *arg
)
1583 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1584 struct mem_cgroup
*oom_wait_memcg
;
1585 struct oom_wait_info
*oom_wait_info
;
1587 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1588 oom_wait_memcg
= oom_wait_info
->memcg
;
1590 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1591 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1593 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1596 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1599 * For the following lockless ->under_oom test, the only required
1600 * guarantee is that it must see the state asserted by an OOM when
1601 * this function is called as a result of userland actions
1602 * triggered by the notification of the OOM. This is trivially
1603 * achieved by invoking mem_cgroup_mark_under_oom() before
1604 * triggering notification.
1606 if (memcg
&& memcg
->under_oom
)
1607 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1610 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1612 if (!current
->memcg_may_oom
)
1615 * We are in the middle of the charge context here, so we
1616 * don't want to block when potentially sitting on a callstack
1617 * that holds all kinds of filesystem and mm locks.
1619 * Also, the caller may handle a failed allocation gracefully
1620 * (like optional page cache readahead) and so an OOM killer
1621 * invocation might not even be necessary.
1623 * That's why we don't do anything here except remember the
1624 * OOM context and then deal with it at the end of the page
1625 * fault when the stack is unwound, the locks are released,
1626 * and when we know whether the fault was overall successful.
1628 css_get(&memcg
->css
);
1629 current
->memcg_in_oom
= memcg
;
1630 current
->memcg_oom_gfp_mask
= mask
;
1631 current
->memcg_oom_order
= order
;
1635 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1636 * @handle: actually kill/wait or just clean up the OOM state
1638 * This has to be called at the end of a page fault if the memcg OOM
1639 * handler was enabled.
1641 * Memcg supports userspace OOM handling where failed allocations must
1642 * sleep on a waitqueue until the userspace task resolves the
1643 * situation. Sleeping directly in the charge context with all kinds
1644 * of locks held is not a good idea, instead we remember an OOM state
1645 * in the task and mem_cgroup_oom_synchronize() has to be called at
1646 * the end of the page fault to complete the OOM handling.
1648 * Returns %true if an ongoing memcg OOM situation was detected and
1649 * completed, %false otherwise.
1651 bool mem_cgroup_oom_synchronize(bool handle
)
1653 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1654 struct oom_wait_info owait
;
1657 /* OOM is global, do not handle */
1661 if (!handle
|| oom_killer_disabled
)
1664 owait
.memcg
= memcg
;
1665 owait
.wait
.flags
= 0;
1666 owait
.wait
.func
= memcg_oom_wake_function
;
1667 owait
.wait
.private = current
;
1668 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1670 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1671 mem_cgroup_mark_under_oom(memcg
);
1673 locked
= mem_cgroup_oom_trylock(memcg
);
1676 mem_cgroup_oom_notify(memcg
);
1678 if (locked
&& !memcg
->oom_kill_disable
) {
1679 mem_cgroup_unmark_under_oom(memcg
);
1680 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1681 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1682 current
->memcg_oom_order
);
1685 mem_cgroup_unmark_under_oom(memcg
);
1686 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1690 mem_cgroup_oom_unlock(memcg
);
1692 * There is no guarantee that an OOM-lock contender
1693 * sees the wakeups triggered by the OOM kill
1694 * uncharges. Wake any sleepers explicitely.
1696 memcg_oom_recover(memcg
);
1699 current
->memcg_in_oom
= NULL
;
1700 css_put(&memcg
->css
);
1705 * lock_page_memcg - lock a page->mem_cgroup binding
1708 * This function protects unlocked LRU pages from being moved to
1709 * another cgroup and stabilizes their page->mem_cgroup binding.
1711 void lock_page_memcg(struct page
*page
)
1713 struct mem_cgroup
*memcg
;
1714 unsigned long flags
;
1717 * The RCU lock is held throughout the transaction. The fast
1718 * path can get away without acquiring the memcg->move_lock
1719 * because page moving starts with an RCU grace period.
1723 if (mem_cgroup_disabled())
1726 memcg
= page
->mem_cgroup
;
1727 if (unlikely(!memcg
))
1730 if (atomic_read(&memcg
->moving_account
) <= 0)
1733 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1734 if (memcg
!= page
->mem_cgroup
) {
1735 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1740 * When charge migration first begins, we can have locked and
1741 * unlocked page stat updates happening concurrently. Track
1742 * the task who has the lock for unlock_page_memcg().
1744 memcg
->move_lock_task
= current
;
1745 memcg
->move_lock_flags
= flags
;
1749 EXPORT_SYMBOL(lock_page_memcg
);
1752 * unlock_page_memcg - unlock a page->mem_cgroup binding
1755 void unlock_page_memcg(struct page
*page
)
1757 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
1759 if (memcg
&& memcg
->move_lock_task
== current
) {
1760 unsigned long flags
= memcg
->move_lock_flags
;
1762 memcg
->move_lock_task
= NULL
;
1763 memcg
->move_lock_flags
= 0;
1765 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1770 EXPORT_SYMBOL(unlock_page_memcg
);
1773 * size of first charge trial. "32" comes from vmscan.c's magic value.
1774 * TODO: maybe necessary to use big numbers in big irons.
1776 #define CHARGE_BATCH 32U
1777 struct memcg_stock_pcp
{
1778 struct mem_cgroup
*cached
; /* this never be root cgroup */
1779 unsigned int nr_pages
;
1780 struct work_struct work
;
1781 unsigned long flags
;
1782 #define FLUSHING_CACHED_CHARGE 0
1784 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1785 static DEFINE_MUTEX(percpu_charge_mutex
);
1788 * consume_stock: Try to consume stocked charge on this cpu.
1789 * @memcg: memcg to consume from.
1790 * @nr_pages: how many pages to charge.
1792 * The charges will only happen if @memcg matches the current cpu's memcg
1793 * stock, and at least @nr_pages are available in that stock. Failure to
1794 * service an allocation will refill the stock.
1796 * returns true if successful, false otherwise.
1798 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1800 struct memcg_stock_pcp
*stock
;
1803 if (nr_pages
> CHARGE_BATCH
)
1806 stock
= &get_cpu_var(memcg_stock
);
1807 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
1808 stock
->nr_pages
-= nr_pages
;
1811 put_cpu_var(memcg_stock
);
1816 * Returns stocks cached in percpu and reset cached information.
1818 static void drain_stock(struct memcg_stock_pcp
*stock
)
1820 struct mem_cgroup
*old
= stock
->cached
;
1822 if (stock
->nr_pages
) {
1823 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
1824 if (do_memsw_account())
1825 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
1826 css_put_many(&old
->css
, stock
->nr_pages
);
1827 stock
->nr_pages
= 0;
1829 stock
->cached
= NULL
;
1833 * This must be called under preempt disabled or must be called by
1834 * a thread which is pinned to local cpu.
1836 static void drain_local_stock(struct work_struct
*dummy
)
1838 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
1840 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
1844 * Cache charges(val) to local per_cpu area.
1845 * This will be consumed by consume_stock() function, later.
1847 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1849 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
1851 if (stock
->cached
!= memcg
) { /* reset if necessary */
1853 stock
->cached
= memcg
;
1855 stock
->nr_pages
+= nr_pages
;
1856 put_cpu_var(memcg_stock
);
1860 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1861 * of the hierarchy under it.
1863 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
1867 /* If someone's already draining, avoid adding running more workers. */
1868 if (!mutex_trylock(&percpu_charge_mutex
))
1870 /* Notify other cpus that system-wide "drain" is running */
1873 for_each_online_cpu(cpu
) {
1874 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1875 struct mem_cgroup
*memcg
;
1877 memcg
= stock
->cached
;
1878 if (!memcg
|| !stock
->nr_pages
)
1880 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
1882 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
1884 drain_local_stock(&stock
->work
);
1886 schedule_work_on(cpu
, &stock
->work
);
1891 mutex_unlock(&percpu_charge_mutex
);
1894 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
1895 unsigned long action
,
1898 int cpu
= (unsigned long)hcpu
;
1899 struct memcg_stock_pcp
*stock
;
1901 if (action
== CPU_ONLINE
)
1904 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
1907 stock
= &per_cpu(memcg_stock
, cpu
);
1912 static void reclaim_high(struct mem_cgroup
*memcg
,
1913 unsigned int nr_pages
,
1917 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
1919 mem_cgroup_events(memcg
, MEMCG_HIGH
, 1);
1920 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
1921 } while ((memcg
= parent_mem_cgroup(memcg
)));
1924 static void high_work_func(struct work_struct
*work
)
1926 struct mem_cgroup
*memcg
;
1928 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
1929 reclaim_high(memcg
, CHARGE_BATCH
, GFP_KERNEL
);
1933 * Scheduled by try_charge() to be executed from the userland return path
1934 * and reclaims memory over the high limit.
1936 void mem_cgroup_handle_over_high(void)
1938 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
1939 struct mem_cgroup
*memcg
;
1941 if (likely(!nr_pages
))
1944 memcg
= get_mem_cgroup_from_mm(current
->mm
);
1945 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
1946 css_put(&memcg
->css
);
1947 current
->memcg_nr_pages_over_high
= 0;
1950 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1951 unsigned int nr_pages
)
1953 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
1954 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1955 struct mem_cgroup
*mem_over_limit
;
1956 struct page_counter
*counter
;
1957 unsigned long nr_reclaimed
;
1958 bool may_swap
= true;
1959 bool drained
= false;
1961 if (mem_cgroup_is_root(memcg
))
1964 if (consume_stock(memcg
, nr_pages
))
1967 if (!do_memsw_account() ||
1968 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
1969 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
1971 if (do_memsw_account())
1972 page_counter_uncharge(&memcg
->memsw
, batch
);
1973 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
1975 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
1979 if (batch
> nr_pages
) {
1985 * Unlike in global OOM situations, memcg is not in a physical
1986 * memory shortage. Allow dying and OOM-killed tasks to
1987 * bypass the last charges so that they can exit quickly and
1988 * free their memory.
1990 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
1991 fatal_signal_pending(current
) ||
1992 current
->flags
& PF_EXITING
))
1995 if (unlikely(task_in_memcg_oom(current
)))
1998 if (!gfpflags_allow_blocking(gfp_mask
))
2001 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
2003 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2004 gfp_mask
, may_swap
);
2006 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2010 drain_all_stock(mem_over_limit
);
2015 if (gfp_mask
& __GFP_NORETRY
)
2018 * Even though the limit is exceeded at this point, reclaim
2019 * may have been able to free some pages. Retry the charge
2020 * before killing the task.
2022 * Only for regular pages, though: huge pages are rather
2023 * unlikely to succeed so close to the limit, and we fall back
2024 * to regular pages anyway in case of failure.
2026 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2029 * At task move, charge accounts can be doubly counted. So, it's
2030 * better to wait until the end of task_move if something is going on.
2032 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2038 if (gfp_mask
& __GFP_NOFAIL
)
2041 if (fatal_signal_pending(current
))
2044 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
2046 mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2047 get_order(nr_pages
* PAGE_SIZE
));
2049 if (!(gfp_mask
& __GFP_NOFAIL
))
2053 * The allocation either can't fail or will lead to more memory
2054 * being freed very soon. Allow memory usage go over the limit
2055 * temporarily by force charging it.
2057 page_counter_charge(&memcg
->memory
, nr_pages
);
2058 if (do_memsw_account())
2059 page_counter_charge(&memcg
->memsw
, nr_pages
);
2060 css_get_many(&memcg
->css
, nr_pages
);
2065 css_get_many(&memcg
->css
, batch
);
2066 if (batch
> nr_pages
)
2067 refill_stock(memcg
, batch
- nr_pages
);
2070 * If the hierarchy is above the normal consumption range, schedule
2071 * reclaim on returning to userland. We can perform reclaim here
2072 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2073 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2074 * not recorded as it most likely matches current's and won't
2075 * change in the meantime. As high limit is checked again before
2076 * reclaim, the cost of mismatch is negligible.
2079 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2080 /* Don't bother a random interrupted task */
2081 if (in_interrupt()) {
2082 schedule_work(&memcg
->high_work
);
2085 current
->memcg_nr_pages_over_high
+= batch
;
2086 set_notify_resume(current
);
2089 } while ((memcg
= parent_mem_cgroup(memcg
)));
2094 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2096 if (mem_cgroup_is_root(memcg
))
2099 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2100 if (do_memsw_account())
2101 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2103 css_put_many(&memcg
->css
, nr_pages
);
2106 static void lock_page_lru(struct page
*page
, int *isolated
)
2108 struct zone
*zone
= page_zone(page
);
2110 spin_lock_irq(&zone
->lru_lock
);
2111 if (PageLRU(page
)) {
2112 struct lruvec
*lruvec
;
2114 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2116 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2122 static void unlock_page_lru(struct page
*page
, int isolated
)
2124 struct zone
*zone
= page_zone(page
);
2127 struct lruvec
*lruvec
;
2129 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2130 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2132 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2134 spin_unlock_irq(&zone
->lru_lock
);
2137 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2142 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2145 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2146 * may already be on some other mem_cgroup's LRU. Take care of it.
2149 lock_page_lru(page
, &isolated
);
2152 * Nobody should be changing or seriously looking at
2153 * page->mem_cgroup at this point:
2155 * - the page is uncharged
2157 * - the page is off-LRU
2159 * - an anonymous fault has exclusive page access, except for
2160 * a locked page table
2162 * - a page cache insertion, a swapin fault, or a migration
2163 * have the page locked
2165 page
->mem_cgroup
= memcg
;
2168 unlock_page_lru(page
, isolated
);
2172 static int memcg_alloc_cache_id(void)
2177 id
= ida_simple_get(&memcg_cache_ida
,
2178 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2182 if (id
< memcg_nr_cache_ids
)
2186 * There's no space for the new id in memcg_caches arrays,
2187 * so we have to grow them.
2189 down_write(&memcg_cache_ids_sem
);
2191 size
= 2 * (id
+ 1);
2192 if (size
< MEMCG_CACHES_MIN_SIZE
)
2193 size
= MEMCG_CACHES_MIN_SIZE
;
2194 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2195 size
= MEMCG_CACHES_MAX_SIZE
;
2197 err
= memcg_update_all_caches(size
);
2199 err
= memcg_update_all_list_lrus(size
);
2201 memcg_nr_cache_ids
= size
;
2203 up_write(&memcg_cache_ids_sem
);
2206 ida_simple_remove(&memcg_cache_ida
, id
);
2212 static void memcg_free_cache_id(int id
)
2214 ida_simple_remove(&memcg_cache_ida
, id
);
2217 struct memcg_kmem_cache_create_work
{
2218 struct mem_cgroup
*memcg
;
2219 struct kmem_cache
*cachep
;
2220 struct work_struct work
;
2223 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2225 struct memcg_kmem_cache_create_work
*cw
=
2226 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2227 struct mem_cgroup
*memcg
= cw
->memcg
;
2228 struct kmem_cache
*cachep
= cw
->cachep
;
2230 memcg_create_kmem_cache(memcg
, cachep
);
2232 css_put(&memcg
->css
);
2237 * Enqueue the creation of a per-memcg kmem_cache.
2239 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2240 struct kmem_cache
*cachep
)
2242 struct memcg_kmem_cache_create_work
*cw
;
2244 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2248 css_get(&memcg
->css
);
2251 cw
->cachep
= cachep
;
2252 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2254 schedule_work(&cw
->work
);
2257 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2258 struct kmem_cache
*cachep
)
2261 * We need to stop accounting when we kmalloc, because if the
2262 * corresponding kmalloc cache is not yet created, the first allocation
2263 * in __memcg_schedule_kmem_cache_create will recurse.
2265 * However, it is better to enclose the whole function. Depending on
2266 * the debugging options enabled, INIT_WORK(), for instance, can
2267 * trigger an allocation. This too, will make us recurse. Because at
2268 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2269 * the safest choice is to do it like this, wrapping the whole function.
2271 current
->memcg_kmem_skip_account
= 1;
2272 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2273 current
->memcg_kmem_skip_account
= 0;
2276 static inline bool memcg_kmem_bypass(void)
2278 if (in_interrupt() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
2284 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2285 * @cachep: the original global kmem cache
2287 * Return the kmem_cache we're supposed to use for a slab allocation.
2288 * We try to use the current memcg's version of the cache.
2290 * If the cache does not exist yet, if we are the first user of it, we
2291 * create it asynchronously in a workqueue and let the current allocation
2292 * go through with the original cache.
2294 * This function takes a reference to the cache it returns to assure it
2295 * won't get destroyed while we are working with it. Once the caller is
2296 * done with it, memcg_kmem_put_cache() must be called to release the
2299 struct kmem_cache
*memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2301 struct mem_cgroup
*memcg
;
2302 struct kmem_cache
*memcg_cachep
;
2305 VM_BUG_ON(!is_root_cache(cachep
));
2307 if (memcg_kmem_bypass())
2310 if (current
->memcg_kmem_skip_account
)
2313 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2314 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2318 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2319 if (likely(memcg_cachep
))
2320 return memcg_cachep
;
2323 * If we are in a safe context (can wait, and not in interrupt
2324 * context), we could be be predictable and return right away.
2325 * This would guarantee that the allocation being performed
2326 * already belongs in the new cache.
2328 * However, there are some clashes that can arrive from locking.
2329 * For instance, because we acquire the slab_mutex while doing
2330 * memcg_create_kmem_cache, this means no further allocation
2331 * could happen with the slab_mutex held. So it's better to
2334 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2336 css_put(&memcg
->css
);
2341 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2342 * @cachep: the cache returned by memcg_kmem_get_cache
2344 void memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2346 if (!is_root_cache(cachep
))
2347 css_put(&cachep
->memcg_params
.memcg
->css
);
2351 * memcg_kmem_charge: charge a kmem page
2352 * @page: page to charge
2353 * @gfp: reclaim mode
2354 * @order: allocation order
2355 * @memcg: memory cgroup to charge
2357 * Returns 0 on success, an error code on failure.
2359 int memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2360 struct mem_cgroup
*memcg
)
2362 unsigned int nr_pages
= 1 << order
;
2363 struct page_counter
*counter
;
2366 ret
= try_charge(memcg
, gfp
, nr_pages
);
2370 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2371 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2372 cancel_charge(memcg
, nr_pages
);
2376 page
->mem_cgroup
= memcg
;
2382 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2383 * @page: page to charge
2384 * @gfp: reclaim mode
2385 * @order: allocation order
2387 * Returns 0 on success, an error code on failure.
2389 int memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2391 struct mem_cgroup
*memcg
;
2394 if (memcg_kmem_bypass())
2397 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2398 if (!mem_cgroup_is_root(memcg
))
2399 ret
= memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2400 css_put(&memcg
->css
);
2404 * memcg_kmem_uncharge: uncharge a kmem page
2405 * @page: page to uncharge
2406 * @order: allocation order
2408 void memcg_kmem_uncharge(struct page
*page
, int order
)
2410 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2411 unsigned int nr_pages
= 1 << order
;
2416 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2418 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2419 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2421 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2422 if (do_memsw_account())
2423 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2425 page
->mem_cgroup
= NULL
;
2426 css_put_many(&memcg
->css
, nr_pages
);
2428 #endif /* !CONFIG_SLOB */
2430 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2433 * Because tail pages are not marked as "used", set it. We're under
2434 * zone->lru_lock and migration entries setup in all page mappings.
2436 void mem_cgroup_split_huge_fixup(struct page
*head
)
2440 if (mem_cgroup_disabled())
2443 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2444 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2446 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2449 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2451 #ifdef CONFIG_MEMCG_SWAP
2452 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2455 int val
= (charge
) ? 1 : -1;
2456 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2460 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2461 * @entry: swap entry to be moved
2462 * @from: mem_cgroup which the entry is moved from
2463 * @to: mem_cgroup which the entry is moved to
2465 * It succeeds only when the swap_cgroup's record for this entry is the same
2466 * as the mem_cgroup's id of @from.
2468 * Returns 0 on success, -EINVAL on failure.
2470 * The caller must have charged to @to, IOW, called page_counter_charge() about
2471 * both res and memsw, and called css_get().
2473 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2474 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2476 unsigned short old_id
, new_id
;
2478 old_id
= mem_cgroup_id(from
);
2479 new_id
= mem_cgroup_id(to
);
2481 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2482 mem_cgroup_swap_statistics(from
, false);
2483 mem_cgroup_swap_statistics(to
, true);
2489 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2490 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2496 static DEFINE_MUTEX(memcg_limit_mutex
);
2498 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2499 unsigned long limit
)
2501 unsigned long curusage
;
2502 unsigned long oldusage
;
2503 bool enlarge
= false;
2508 * For keeping hierarchical_reclaim simple, how long we should retry
2509 * is depends on callers. We set our retry-count to be function
2510 * of # of children which we should visit in this loop.
2512 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2513 mem_cgroup_count_children(memcg
);
2515 oldusage
= page_counter_read(&memcg
->memory
);
2518 if (signal_pending(current
)) {
2523 mutex_lock(&memcg_limit_mutex
);
2524 if (limit
> memcg
->memsw
.limit
) {
2525 mutex_unlock(&memcg_limit_mutex
);
2529 if (limit
> memcg
->memory
.limit
)
2531 ret
= page_counter_limit(&memcg
->memory
, limit
);
2532 mutex_unlock(&memcg_limit_mutex
);
2537 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2539 curusage
= page_counter_read(&memcg
->memory
);
2540 /* Usage is reduced ? */
2541 if (curusage
>= oldusage
)
2544 oldusage
= curusage
;
2545 } while (retry_count
);
2547 if (!ret
&& enlarge
)
2548 memcg_oom_recover(memcg
);
2553 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2554 unsigned long limit
)
2556 unsigned long curusage
;
2557 unsigned long oldusage
;
2558 bool enlarge
= false;
2562 /* see mem_cgroup_resize_res_limit */
2563 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2564 mem_cgroup_count_children(memcg
);
2566 oldusage
= page_counter_read(&memcg
->memsw
);
2569 if (signal_pending(current
)) {
2574 mutex_lock(&memcg_limit_mutex
);
2575 if (limit
< memcg
->memory
.limit
) {
2576 mutex_unlock(&memcg_limit_mutex
);
2580 if (limit
> memcg
->memsw
.limit
)
2582 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2583 mutex_unlock(&memcg_limit_mutex
);
2588 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2590 curusage
= page_counter_read(&memcg
->memsw
);
2591 /* Usage is reduced ? */
2592 if (curusage
>= oldusage
)
2595 oldusage
= curusage
;
2596 } while (retry_count
);
2598 if (!ret
&& enlarge
)
2599 memcg_oom_recover(memcg
);
2604 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
2606 unsigned long *total_scanned
)
2608 unsigned long nr_reclaimed
= 0;
2609 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
2610 unsigned long reclaimed
;
2612 struct mem_cgroup_tree_per_zone
*mctz
;
2613 unsigned long excess
;
2614 unsigned long nr_scanned
;
2619 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
2621 * This loop can run a while, specially if mem_cgroup's continuously
2622 * keep exceeding their soft limit and putting the system under
2629 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2634 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
2635 gfp_mask
, &nr_scanned
);
2636 nr_reclaimed
+= reclaimed
;
2637 *total_scanned
+= nr_scanned
;
2638 spin_lock_irq(&mctz
->lock
);
2639 __mem_cgroup_remove_exceeded(mz
, mctz
);
2642 * If we failed to reclaim anything from this memory cgroup
2643 * it is time to move on to the next cgroup
2647 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2649 excess
= soft_limit_excess(mz
->memcg
);
2651 * One school of thought says that we should not add
2652 * back the node to the tree if reclaim returns 0.
2653 * But our reclaim could return 0, simply because due
2654 * to priority we are exposing a smaller subset of
2655 * memory to reclaim from. Consider this as a longer
2658 /* If excess == 0, no tree ops */
2659 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2660 spin_unlock_irq(&mctz
->lock
);
2661 css_put(&mz
->memcg
->css
);
2664 * Could not reclaim anything and there are no more
2665 * mem cgroups to try or we seem to be looping without
2666 * reclaiming anything.
2668 if (!nr_reclaimed
&&
2670 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2672 } while (!nr_reclaimed
);
2674 css_put(&next_mz
->memcg
->css
);
2675 return nr_reclaimed
;
2679 * Test whether @memcg has children, dead or alive. Note that this
2680 * function doesn't care whether @memcg has use_hierarchy enabled and
2681 * returns %true if there are child csses according to the cgroup
2682 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2684 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2689 ret
= css_next_child(NULL
, &memcg
->css
);
2695 * Reclaims as many pages from the given memcg as possible.
2697 * Caller is responsible for holding css reference for memcg.
2699 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2701 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2703 /* we call try-to-free pages for make this cgroup empty */
2704 lru_add_drain_all();
2705 /* try to free all pages in this cgroup */
2706 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2709 if (signal_pending(current
))
2712 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2716 /* maybe some writeback is necessary */
2717 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2725 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
2726 char *buf
, size_t nbytes
,
2729 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2731 if (mem_cgroup_is_root(memcg
))
2733 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
2736 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
2739 return mem_cgroup_from_css(css
)->use_hierarchy
;
2742 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
2743 struct cftype
*cft
, u64 val
)
2746 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2747 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
2749 if (memcg
->use_hierarchy
== val
)
2753 * If parent's use_hierarchy is set, we can't make any modifications
2754 * in the child subtrees. If it is unset, then the change can
2755 * occur, provided the current cgroup has no children.
2757 * For the root cgroup, parent_mem is NULL, we allow value to be
2758 * set if there are no children.
2760 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
2761 (val
== 1 || val
== 0)) {
2762 if (!memcg_has_children(memcg
))
2763 memcg
->use_hierarchy
= val
;
2772 static void tree_stat(struct mem_cgroup
*memcg
, unsigned long *stat
)
2774 struct mem_cgroup
*iter
;
2777 memset(stat
, 0, sizeof(*stat
) * MEMCG_NR_STAT
);
2779 for_each_mem_cgroup_tree(iter
, memcg
) {
2780 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
2781 stat
[i
] += mem_cgroup_read_stat(iter
, i
);
2785 static void tree_events(struct mem_cgroup
*memcg
, unsigned long *events
)
2787 struct mem_cgroup
*iter
;
2790 memset(events
, 0, sizeof(*events
) * MEMCG_NR_EVENTS
);
2792 for_each_mem_cgroup_tree(iter
, memcg
) {
2793 for (i
= 0; i
< MEMCG_NR_EVENTS
; i
++)
2794 events
[i
] += mem_cgroup_read_events(iter
, i
);
2798 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
2800 unsigned long val
= 0;
2802 if (mem_cgroup_is_root(memcg
)) {
2803 struct mem_cgroup
*iter
;
2805 for_each_mem_cgroup_tree(iter
, memcg
) {
2806 val
+= mem_cgroup_read_stat(iter
,
2807 MEM_CGROUP_STAT_CACHE
);
2808 val
+= mem_cgroup_read_stat(iter
,
2809 MEM_CGROUP_STAT_RSS
);
2811 val
+= mem_cgroup_read_stat(iter
,
2812 MEM_CGROUP_STAT_SWAP
);
2816 val
= page_counter_read(&memcg
->memory
);
2818 val
= page_counter_read(&memcg
->memsw
);
2831 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
2834 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2835 struct page_counter
*counter
;
2837 switch (MEMFILE_TYPE(cft
->private)) {
2839 counter
= &memcg
->memory
;
2842 counter
= &memcg
->memsw
;
2845 counter
= &memcg
->kmem
;
2848 counter
= &memcg
->tcpmem
;
2854 switch (MEMFILE_ATTR(cft
->private)) {
2856 if (counter
== &memcg
->memory
)
2857 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
2858 if (counter
== &memcg
->memsw
)
2859 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
2860 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
2862 return (u64
)counter
->limit
* PAGE_SIZE
;
2864 return (u64
)counter
->watermark
* PAGE_SIZE
;
2866 return counter
->failcnt
;
2867 case RES_SOFT_LIMIT
:
2868 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
2875 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2879 if (cgroup_memory_nokmem
)
2882 BUG_ON(memcg
->kmemcg_id
>= 0);
2883 BUG_ON(memcg
->kmem_state
);
2885 memcg_id
= memcg_alloc_cache_id();
2889 static_branch_inc(&memcg_kmem_enabled_key
);
2891 * A memory cgroup is considered kmem-online as soon as it gets
2892 * kmemcg_id. Setting the id after enabling static branching will
2893 * guarantee no one starts accounting before all call sites are
2896 memcg
->kmemcg_id
= memcg_id
;
2897 memcg
->kmem_state
= KMEM_ONLINE
;
2902 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2904 struct cgroup_subsys_state
*css
;
2905 struct mem_cgroup
*parent
, *child
;
2908 if (memcg
->kmem_state
!= KMEM_ONLINE
)
2911 * Clear the online state before clearing memcg_caches array
2912 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2913 * guarantees that no cache will be created for this cgroup
2914 * after we are done (see memcg_create_kmem_cache()).
2916 memcg
->kmem_state
= KMEM_ALLOCATED
;
2918 memcg_deactivate_kmem_caches(memcg
);
2920 kmemcg_id
= memcg
->kmemcg_id
;
2921 BUG_ON(kmemcg_id
< 0);
2923 parent
= parent_mem_cgroup(memcg
);
2925 parent
= root_mem_cgroup
;
2928 * Change kmemcg_id of this cgroup and all its descendants to the
2929 * parent's id, and then move all entries from this cgroup's list_lrus
2930 * to ones of the parent. After we have finished, all list_lrus
2931 * corresponding to this cgroup are guaranteed to remain empty. The
2932 * ordering is imposed by list_lru_node->lock taken by
2933 * memcg_drain_all_list_lrus().
2935 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2936 css_for_each_descendant_pre(css
, &memcg
->css
) {
2937 child
= mem_cgroup_from_css(css
);
2938 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
2939 child
->kmemcg_id
= parent
->kmemcg_id
;
2940 if (!memcg
->use_hierarchy
)
2945 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
2947 memcg_free_cache_id(kmemcg_id
);
2950 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2952 /* css_alloc() failed, offlining didn't happen */
2953 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
2954 memcg_offline_kmem(memcg
);
2956 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
2957 memcg_destroy_kmem_caches(memcg
);
2958 static_branch_dec(&memcg_kmem_enabled_key
);
2959 WARN_ON(page_counter_read(&memcg
->kmem
));
2963 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2967 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2970 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2973 #endif /* !CONFIG_SLOB */
2975 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2976 unsigned long limit
)
2980 mutex_lock(&memcg_limit_mutex
);
2981 ret
= page_counter_limit(&memcg
->kmem
, limit
);
2982 mutex_unlock(&memcg_limit_mutex
);
2986 static int memcg_update_tcp_limit(struct mem_cgroup
*memcg
, unsigned long limit
)
2990 mutex_lock(&memcg_limit_mutex
);
2992 ret
= page_counter_limit(&memcg
->tcpmem
, limit
);
2996 if (!memcg
->tcpmem_active
) {
2998 * The active flag needs to be written after the static_key
2999 * update. This is what guarantees that the socket activation
3000 * function is the last one to run. See sock_update_memcg() for
3001 * details, and note that we don't mark any socket as belonging
3002 * to this memcg until that flag is up.
3004 * We need to do this, because static_keys will span multiple
3005 * sites, but we can't control their order. If we mark a socket
3006 * as accounted, but the accounting functions are not patched in
3007 * yet, we'll lose accounting.
3009 * We never race with the readers in sock_update_memcg(),
3010 * because when this value change, the code to process it is not
3013 static_branch_inc(&memcg_sockets_enabled_key
);
3014 memcg
->tcpmem_active
= true;
3017 mutex_unlock(&memcg_limit_mutex
);
3022 * The user of this function is...
3025 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3026 char *buf
, size_t nbytes
, loff_t off
)
3028 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3029 unsigned long nr_pages
;
3032 buf
= strstrip(buf
);
3033 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3037 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3039 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3043 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3045 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
3048 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
3051 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
3054 ret
= memcg_update_tcp_limit(memcg
, nr_pages
);
3058 case RES_SOFT_LIMIT
:
3059 memcg
->soft_limit
= nr_pages
;
3063 return ret
?: nbytes
;
3066 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3067 size_t nbytes
, loff_t off
)
3069 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3070 struct page_counter
*counter
;
3072 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3074 counter
= &memcg
->memory
;
3077 counter
= &memcg
->memsw
;
3080 counter
= &memcg
->kmem
;
3083 counter
= &memcg
->tcpmem
;
3089 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3091 page_counter_reset_watermark(counter
);
3094 counter
->failcnt
= 0;
3103 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3106 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3110 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3111 struct cftype
*cft
, u64 val
)
3113 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3115 if (val
& ~MOVE_MASK
)
3119 * No kind of locking is needed in here, because ->can_attach() will
3120 * check this value once in the beginning of the process, and then carry
3121 * on with stale data. This means that changes to this value will only
3122 * affect task migrations starting after the change.
3124 memcg
->move_charge_at_immigrate
= val
;
3128 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3129 struct cftype
*cft
, u64 val
)
3136 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3140 unsigned int lru_mask
;
3143 static const struct numa_stat stats
[] = {
3144 { "total", LRU_ALL
},
3145 { "file", LRU_ALL_FILE
},
3146 { "anon", LRU_ALL_ANON
},
3147 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3149 const struct numa_stat
*stat
;
3152 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3154 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3155 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3156 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3157 for_each_node_state(nid
, N_MEMORY
) {
3158 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3160 seq_printf(m
, " N%d=%lu", nid
, nr
);
3165 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3166 struct mem_cgroup
*iter
;
3169 for_each_mem_cgroup_tree(iter
, memcg
)
3170 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3171 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3172 for_each_node_state(nid
, N_MEMORY
) {
3174 for_each_mem_cgroup_tree(iter
, memcg
)
3175 nr
+= mem_cgroup_node_nr_lru_pages(
3176 iter
, nid
, stat
->lru_mask
);
3177 seq_printf(m
, " N%d=%lu", nid
, nr
);
3184 #endif /* CONFIG_NUMA */
3186 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3188 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3189 unsigned long memory
, memsw
;
3190 struct mem_cgroup
*mi
;
3193 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3194 MEM_CGROUP_STAT_NSTATS
);
3195 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3196 MEM_CGROUP_EVENTS_NSTATS
);
3197 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3199 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3200 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3202 seq_printf(m
, "%s %lu\n", mem_cgroup_stat_names
[i
],
3203 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3206 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3207 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3208 mem_cgroup_read_events(memcg
, i
));
3210 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3211 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3212 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3214 /* Hierarchical information */
3215 memory
= memsw
= PAGE_COUNTER_MAX
;
3216 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3217 memory
= min(memory
, mi
->memory
.limit
);
3218 memsw
= min(memsw
, mi
->memsw
.limit
);
3220 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3221 (u64
)memory
* PAGE_SIZE
);
3222 if (do_memsw_account())
3223 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3224 (u64
)memsw
* PAGE_SIZE
);
3226 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3227 unsigned long long val
= 0;
3229 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3231 for_each_mem_cgroup_tree(mi
, memcg
)
3232 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3233 seq_printf(m
, "total_%s %llu\n", mem_cgroup_stat_names
[i
], val
);
3236 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3237 unsigned long long val
= 0;
3239 for_each_mem_cgroup_tree(mi
, memcg
)
3240 val
+= mem_cgroup_read_events(mi
, i
);
3241 seq_printf(m
, "total_%s %llu\n",
3242 mem_cgroup_events_names
[i
], val
);
3245 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3246 unsigned long long val
= 0;
3248 for_each_mem_cgroup_tree(mi
, memcg
)
3249 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3250 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3253 #ifdef CONFIG_DEBUG_VM
3256 struct mem_cgroup_per_zone
*mz
;
3257 struct zone_reclaim_stat
*rstat
;
3258 unsigned long recent_rotated
[2] = {0, 0};
3259 unsigned long recent_scanned
[2] = {0, 0};
3261 for_each_online_node(nid
)
3262 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3263 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
3264 rstat
= &mz
->lruvec
.reclaim_stat
;
3266 recent_rotated
[0] += rstat
->recent_rotated
[0];
3267 recent_rotated
[1] += rstat
->recent_rotated
[1];
3268 recent_scanned
[0] += rstat
->recent_scanned
[0];
3269 recent_scanned
[1] += rstat
->recent_scanned
[1];
3271 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3272 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3273 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3274 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3281 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3284 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3286 return mem_cgroup_swappiness(memcg
);
3289 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3290 struct cftype
*cft
, u64 val
)
3292 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3298 memcg
->swappiness
= val
;
3300 vm_swappiness
= val
;
3305 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3307 struct mem_cgroup_threshold_ary
*t
;
3308 unsigned long usage
;
3313 t
= rcu_dereference(memcg
->thresholds
.primary
);
3315 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3320 usage
= mem_cgroup_usage(memcg
, swap
);
3323 * current_threshold points to threshold just below or equal to usage.
3324 * If it's not true, a threshold was crossed after last
3325 * call of __mem_cgroup_threshold().
3327 i
= t
->current_threshold
;
3330 * Iterate backward over array of thresholds starting from
3331 * current_threshold and check if a threshold is crossed.
3332 * If none of thresholds below usage is crossed, we read
3333 * only one element of the array here.
3335 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3336 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3338 /* i = current_threshold + 1 */
3342 * Iterate forward over array of thresholds starting from
3343 * current_threshold+1 and check if a threshold is crossed.
3344 * If none of thresholds above usage is crossed, we read
3345 * only one element of the array here.
3347 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3348 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3350 /* Update current_threshold */
3351 t
->current_threshold
= i
- 1;
3356 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3359 __mem_cgroup_threshold(memcg
, false);
3360 if (do_memsw_account())
3361 __mem_cgroup_threshold(memcg
, true);
3363 memcg
= parent_mem_cgroup(memcg
);
3367 static int compare_thresholds(const void *a
, const void *b
)
3369 const struct mem_cgroup_threshold
*_a
= a
;
3370 const struct mem_cgroup_threshold
*_b
= b
;
3372 if (_a
->threshold
> _b
->threshold
)
3375 if (_a
->threshold
< _b
->threshold
)
3381 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3383 struct mem_cgroup_eventfd_list
*ev
;
3385 spin_lock(&memcg_oom_lock
);
3387 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3388 eventfd_signal(ev
->eventfd
, 1);
3390 spin_unlock(&memcg_oom_lock
);
3394 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3396 struct mem_cgroup
*iter
;
3398 for_each_mem_cgroup_tree(iter
, memcg
)
3399 mem_cgroup_oom_notify_cb(iter
);
3402 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3403 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3405 struct mem_cgroup_thresholds
*thresholds
;
3406 struct mem_cgroup_threshold_ary
*new;
3407 unsigned long threshold
;
3408 unsigned long usage
;
3411 ret
= page_counter_memparse(args
, "-1", &threshold
);
3415 mutex_lock(&memcg
->thresholds_lock
);
3418 thresholds
= &memcg
->thresholds
;
3419 usage
= mem_cgroup_usage(memcg
, false);
3420 } else if (type
== _MEMSWAP
) {
3421 thresholds
= &memcg
->memsw_thresholds
;
3422 usage
= mem_cgroup_usage(memcg
, true);
3426 /* Check if a threshold crossed before adding a new one */
3427 if (thresholds
->primary
)
3428 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3430 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3432 /* Allocate memory for new array of thresholds */
3433 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3441 /* Copy thresholds (if any) to new array */
3442 if (thresholds
->primary
) {
3443 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3444 sizeof(struct mem_cgroup_threshold
));
3447 /* Add new threshold */
3448 new->entries
[size
- 1].eventfd
= eventfd
;
3449 new->entries
[size
- 1].threshold
= threshold
;
3451 /* Sort thresholds. Registering of new threshold isn't time-critical */
3452 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3453 compare_thresholds
, NULL
);
3455 /* Find current threshold */
3456 new->current_threshold
= -1;
3457 for (i
= 0; i
< size
; i
++) {
3458 if (new->entries
[i
].threshold
<= usage
) {
3460 * new->current_threshold will not be used until
3461 * rcu_assign_pointer(), so it's safe to increment
3464 ++new->current_threshold
;
3469 /* Free old spare buffer and save old primary buffer as spare */
3470 kfree(thresholds
->spare
);
3471 thresholds
->spare
= thresholds
->primary
;
3473 rcu_assign_pointer(thresholds
->primary
, new);
3475 /* To be sure that nobody uses thresholds */
3479 mutex_unlock(&memcg
->thresholds_lock
);
3484 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3485 struct eventfd_ctx
*eventfd
, const char *args
)
3487 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3490 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3491 struct eventfd_ctx
*eventfd
, const char *args
)
3493 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3496 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3497 struct eventfd_ctx
*eventfd
, enum res_type type
)
3499 struct mem_cgroup_thresholds
*thresholds
;
3500 struct mem_cgroup_threshold_ary
*new;
3501 unsigned long usage
;
3504 mutex_lock(&memcg
->thresholds_lock
);
3507 thresholds
= &memcg
->thresholds
;
3508 usage
= mem_cgroup_usage(memcg
, false);
3509 } else if (type
== _MEMSWAP
) {
3510 thresholds
= &memcg
->memsw_thresholds
;
3511 usage
= mem_cgroup_usage(memcg
, true);
3515 if (!thresholds
->primary
)
3518 /* Check if a threshold crossed before removing */
3519 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3521 /* Calculate new number of threshold */
3523 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3524 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3528 new = thresholds
->spare
;
3530 /* Set thresholds array to NULL if we don't have thresholds */
3539 /* Copy thresholds and find current threshold */
3540 new->current_threshold
= -1;
3541 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3542 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3545 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3546 if (new->entries
[j
].threshold
<= usage
) {
3548 * new->current_threshold will not be used
3549 * until rcu_assign_pointer(), so it's safe to increment
3552 ++new->current_threshold
;
3558 /* Swap primary and spare array */
3559 thresholds
->spare
= thresholds
->primary
;
3561 rcu_assign_pointer(thresholds
->primary
, new);
3563 /* To be sure that nobody uses thresholds */
3566 /* If all events are unregistered, free the spare array */
3568 kfree(thresholds
->spare
);
3569 thresholds
->spare
= NULL
;
3572 mutex_unlock(&memcg
->thresholds_lock
);
3575 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3576 struct eventfd_ctx
*eventfd
)
3578 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3581 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3582 struct eventfd_ctx
*eventfd
)
3584 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3587 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3588 struct eventfd_ctx
*eventfd
, const char *args
)
3590 struct mem_cgroup_eventfd_list
*event
;
3592 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3596 spin_lock(&memcg_oom_lock
);
3598 event
->eventfd
= eventfd
;
3599 list_add(&event
->list
, &memcg
->oom_notify
);
3601 /* already in OOM ? */
3602 if (memcg
->under_oom
)
3603 eventfd_signal(eventfd
, 1);
3604 spin_unlock(&memcg_oom_lock
);
3609 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3610 struct eventfd_ctx
*eventfd
)
3612 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3614 spin_lock(&memcg_oom_lock
);
3616 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3617 if (ev
->eventfd
== eventfd
) {
3618 list_del(&ev
->list
);
3623 spin_unlock(&memcg_oom_lock
);
3626 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3628 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3630 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3631 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3635 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3636 struct cftype
*cft
, u64 val
)
3638 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3640 /* cannot set to root cgroup and only 0 and 1 are allowed */
3641 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3644 memcg
->oom_kill_disable
= val
;
3646 memcg_oom_recover(memcg
);
3651 #ifdef CONFIG_CGROUP_WRITEBACK
3653 struct list_head
*mem_cgroup_cgwb_list(struct mem_cgroup
*memcg
)
3655 return &memcg
->cgwb_list
;
3658 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3660 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3663 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3665 wb_domain_exit(&memcg
->cgwb_domain
);
3668 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3670 wb_domain_size_changed(&memcg
->cgwb_domain
);
3673 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3675 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3677 if (!memcg
->css
.parent
)
3680 return &memcg
->cgwb_domain
;
3684 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3685 * @wb: bdi_writeback in question
3686 * @pfilepages: out parameter for number of file pages
3687 * @pheadroom: out parameter for number of allocatable pages according to memcg
3688 * @pdirty: out parameter for number of dirty pages
3689 * @pwriteback: out parameter for number of pages under writeback
3691 * Determine the numbers of file, headroom, dirty, and writeback pages in
3692 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3693 * is a bit more involved.
3695 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3696 * headroom is calculated as the lowest headroom of itself and the
3697 * ancestors. Note that this doesn't consider the actual amount of
3698 * available memory in the system. The caller should further cap
3699 * *@pheadroom accordingly.
3701 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
3702 unsigned long *pheadroom
, unsigned long *pdirty
,
3703 unsigned long *pwriteback
)
3705 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3706 struct mem_cgroup
*parent
;
3708 *pdirty
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
3710 /* this should eventually include NR_UNSTABLE_NFS */
3711 *pwriteback
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
3712 *pfilepages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
3713 (1 << LRU_ACTIVE_FILE
));
3714 *pheadroom
= PAGE_COUNTER_MAX
;
3716 while ((parent
= parent_mem_cgroup(memcg
))) {
3717 unsigned long ceiling
= min(memcg
->memory
.limit
, memcg
->high
);
3718 unsigned long used
= page_counter_read(&memcg
->memory
);
3720 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
3725 #else /* CONFIG_CGROUP_WRITEBACK */
3727 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3732 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3736 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3740 #endif /* CONFIG_CGROUP_WRITEBACK */
3743 * DO NOT USE IN NEW FILES.
3745 * "cgroup.event_control" implementation.
3747 * This is way over-engineered. It tries to support fully configurable
3748 * events for each user. Such level of flexibility is completely
3749 * unnecessary especially in the light of the planned unified hierarchy.
3751 * Please deprecate this and replace with something simpler if at all
3756 * Unregister event and free resources.
3758 * Gets called from workqueue.
3760 static void memcg_event_remove(struct work_struct
*work
)
3762 struct mem_cgroup_event
*event
=
3763 container_of(work
, struct mem_cgroup_event
, remove
);
3764 struct mem_cgroup
*memcg
= event
->memcg
;
3766 remove_wait_queue(event
->wqh
, &event
->wait
);
3768 event
->unregister_event(memcg
, event
->eventfd
);
3770 /* Notify userspace the event is going away. */
3771 eventfd_signal(event
->eventfd
, 1);
3773 eventfd_ctx_put(event
->eventfd
);
3775 css_put(&memcg
->css
);
3779 * Gets called on POLLHUP on eventfd when user closes it.
3781 * Called with wqh->lock held and interrupts disabled.
3783 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
3784 int sync
, void *key
)
3786 struct mem_cgroup_event
*event
=
3787 container_of(wait
, struct mem_cgroup_event
, wait
);
3788 struct mem_cgroup
*memcg
= event
->memcg
;
3789 unsigned long flags
= (unsigned long)key
;
3791 if (flags
& POLLHUP
) {
3793 * If the event has been detached at cgroup removal, we
3794 * can simply return knowing the other side will cleanup
3797 * We can't race against event freeing since the other
3798 * side will require wqh->lock via remove_wait_queue(),
3801 spin_lock(&memcg
->event_list_lock
);
3802 if (!list_empty(&event
->list
)) {
3803 list_del_init(&event
->list
);
3805 * We are in atomic context, but cgroup_event_remove()
3806 * may sleep, so we have to call it in workqueue.
3808 schedule_work(&event
->remove
);
3810 spin_unlock(&memcg
->event_list_lock
);
3816 static void memcg_event_ptable_queue_proc(struct file
*file
,
3817 wait_queue_head_t
*wqh
, poll_table
*pt
)
3819 struct mem_cgroup_event
*event
=
3820 container_of(pt
, struct mem_cgroup_event
, pt
);
3823 add_wait_queue(wqh
, &event
->wait
);
3827 * DO NOT USE IN NEW FILES.
3829 * Parse input and register new cgroup event handler.
3831 * Input must be in format '<event_fd> <control_fd> <args>'.
3832 * Interpretation of args is defined by control file implementation.
3834 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
3835 char *buf
, size_t nbytes
, loff_t off
)
3837 struct cgroup_subsys_state
*css
= of_css(of
);
3838 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3839 struct mem_cgroup_event
*event
;
3840 struct cgroup_subsys_state
*cfile_css
;
3841 unsigned int efd
, cfd
;
3848 buf
= strstrip(buf
);
3850 efd
= simple_strtoul(buf
, &endp
, 10);
3855 cfd
= simple_strtoul(buf
, &endp
, 10);
3856 if ((*endp
!= ' ') && (*endp
!= '\0'))
3860 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
3864 event
->memcg
= memcg
;
3865 INIT_LIST_HEAD(&event
->list
);
3866 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
3867 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
3868 INIT_WORK(&event
->remove
, memcg_event_remove
);
3876 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
3877 if (IS_ERR(event
->eventfd
)) {
3878 ret
= PTR_ERR(event
->eventfd
);
3885 goto out_put_eventfd
;
3888 /* the process need read permission on control file */
3889 /* AV: shouldn't we check that it's been opened for read instead? */
3890 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
3895 * Determine the event callbacks and set them in @event. This used
3896 * to be done via struct cftype but cgroup core no longer knows
3897 * about these events. The following is crude but the whole thing
3898 * is for compatibility anyway.
3900 * DO NOT ADD NEW FILES.
3902 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
3904 if (!strcmp(name
, "memory.usage_in_bytes")) {
3905 event
->register_event
= mem_cgroup_usage_register_event
;
3906 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
3907 } else if (!strcmp(name
, "memory.oom_control")) {
3908 event
->register_event
= mem_cgroup_oom_register_event
;
3909 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
3910 } else if (!strcmp(name
, "memory.pressure_level")) {
3911 event
->register_event
= vmpressure_register_event
;
3912 event
->unregister_event
= vmpressure_unregister_event
;
3913 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
3914 event
->register_event
= memsw_cgroup_usage_register_event
;
3915 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
3922 * Verify @cfile should belong to @css. Also, remaining events are
3923 * automatically removed on cgroup destruction but the removal is
3924 * asynchronous, so take an extra ref on @css.
3926 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
3927 &memory_cgrp_subsys
);
3929 if (IS_ERR(cfile_css
))
3931 if (cfile_css
!= css
) {
3936 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
3940 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
3942 spin_lock(&memcg
->event_list_lock
);
3943 list_add(&event
->list
, &memcg
->event_list
);
3944 spin_unlock(&memcg
->event_list_lock
);
3956 eventfd_ctx_put(event
->eventfd
);
3965 static struct cftype mem_cgroup_legacy_files
[] = {
3967 .name
= "usage_in_bytes",
3968 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
3969 .read_u64
= mem_cgroup_read_u64
,
3972 .name
= "max_usage_in_bytes",
3973 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
3974 .write
= mem_cgroup_reset
,
3975 .read_u64
= mem_cgroup_read_u64
,
3978 .name
= "limit_in_bytes",
3979 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
3980 .write
= mem_cgroup_write
,
3981 .read_u64
= mem_cgroup_read_u64
,
3984 .name
= "soft_limit_in_bytes",
3985 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
3986 .write
= mem_cgroup_write
,
3987 .read_u64
= mem_cgroup_read_u64
,
3991 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
3992 .write
= mem_cgroup_reset
,
3993 .read_u64
= mem_cgroup_read_u64
,
3997 .seq_show
= memcg_stat_show
,
4000 .name
= "force_empty",
4001 .write
= mem_cgroup_force_empty_write
,
4004 .name
= "use_hierarchy",
4005 .write_u64
= mem_cgroup_hierarchy_write
,
4006 .read_u64
= mem_cgroup_hierarchy_read
,
4009 .name
= "cgroup.event_control", /* XXX: for compat */
4010 .write
= memcg_write_event_control
,
4011 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
4014 .name
= "swappiness",
4015 .read_u64
= mem_cgroup_swappiness_read
,
4016 .write_u64
= mem_cgroup_swappiness_write
,
4019 .name
= "move_charge_at_immigrate",
4020 .read_u64
= mem_cgroup_move_charge_read
,
4021 .write_u64
= mem_cgroup_move_charge_write
,
4024 .name
= "oom_control",
4025 .seq_show
= mem_cgroup_oom_control_read
,
4026 .write_u64
= mem_cgroup_oom_control_write
,
4027 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4030 .name
= "pressure_level",
4034 .name
= "numa_stat",
4035 .seq_show
= memcg_numa_stat_show
,
4039 .name
= "kmem.limit_in_bytes",
4040 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4041 .write
= mem_cgroup_write
,
4042 .read_u64
= mem_cgroup_read_u64
,
4045 .name
= "kmem.usage_in_bytes",
4046 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4047 .read_u64
= mem_cgroup_read_u64
,
4050 .name
= "kmem.failcnt",
4051 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4052 .write
= mem_cgroup_reset
,
4053 .read_u64
= mem_cgroup_read_u64
,
4056 .name
= "kmem.max_usage_in_bytes",
4057 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4058 .write
= mem_cgroup_reset
,
4059 .read_u64
= mem_cgroup_read_u64
,
4061 #ifdef CONFIG_SLABINFO
4063 .name
= "kmem.slabinfo",
4064 .seq_start
= slab_start
,
4065 .seq_next
= slab_next
,
4066 .seq_stop
= slab_stop
,
4067 .seq_show
= memcg_slab_show
,
4071 .name
= "kmem.tcp.limit_in_bytes",
4072 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4073 .write
= mem_cgroup_write
,
4074 .read_u64
= mem_cgroup_read_u64
,
4077 .name
= "kmem.tcp.usage_in_bytes",
4078 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4079 .read_u64
= mem_cgroup_read_u64
,
4082 .name
= "kmem.tcp.failcnt",
4083 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4084 .write
= mem_cgroup_reset
,
4085 .read_u64
= mem_cgroup_read_u64
,
4088 .name
= "kmem.tcp.max_usage_in_bytes",
4089 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4090 .write
= mem_cgroup_reset
,
4091 .read_u64
= mem_cgroup_read_u64
,
4093 { }, /* terminate */
4097 * Private memory cgroup IDR
4099 * Swap-out records and page cache shadow entries need to store memcg
4100 * references in constrained space, so we maintain an ID space that is
4101 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4102 * memory-controlled cgroups to 64k.
4104 * However, there usually are many references to the oflline CSS after
4105 * the cgroup has been destroyed, such as page cache or reclaimable
4106 * slab objects, that don't need to hang on to the ID. We want to keep
4107 * those dead CSS from occupying IDs, or we might quickly exhaust the
4108 * relatively small ID space and prevent the creation of new cgroups
4109 * even when there are much fewer than 64k cgroups - possibly none.
4111 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4112 * be freed and recycled when it's no longer needed, which is usually
4113 * when the CSS is offlined.
4115 * The only exception to that are records of swapped out tmpfs/shmem
4116 * pages that need to be attributed to live ancestors on swapin. But
4117 * those references are manageable from userspace.
4120 static DEFINE_IDR(mem_cgroup_idr
);
4122 static void mem_cgroup_id_get(struct mem_cgroup
*memcg
)
4124 atomic_inc(&memcg
->id
.ref
);
4127 static void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
4129 if (atomic_dec_and_test(&memcg
->id
.ref
)) {
4130 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4133 /* Memcg ID pins CSS */
4134 css_put(&memcg
->css
);
4139 * mem_cgroup_from_id - look up a memcg from a memcg id
4140 * @id: the memcg id to look up
4142 * Caller must hold rcu_read_lock().
4144 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
4146 WARN_ON_ONCE(!rcu_read_lock_held());
4147 return idr_find(&mem_cgroup_idr
, id
);
4150 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4152 struct mem_cgroup_per_node
*pn
;
4153 struct mem_cgroup_per_zone
*mz
;
4154 int zone
, tmp
= node
;
4156 * This routine is called against possible nodes.
4157 * But it's BUG to call kmalloc() against offline node.
4159 * TODO: this routine can waste much memory for nodes which will
4160 * never be onlined. It's better to use memory hotplug callback
4163 if (!node_state(node
, N_NORMAL_MEMORY
))
4165 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4169 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4170 mz
= &pn
->zoneinfo
[zone
];
4171 lruvec_init(&mz
->lruvec
);
4172 mz
->usage_in_excess
= 0;
4173 mz
->on_tree
= false;
4176 memcg
->nodeinfo
[node
] = pn
;
4180 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4182 kfree(memcg
->nodeinfo
[node
]);
4185 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4189 memcg_wb_domain_exit(memcg
);
4191 free_mem_cgroup_per_zone_info(memcg
, node
);
4192 free_percpu(memcg
->stat
);
4196 static struct mem_cgroup
*mem_cgroup_alloc(void)
4198 struct mem_cgroup
*memcg
;
4202 size
= sizeof(struct mem_cgroup
);
4203 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4205 memcg
= kzalloc(size
, GFP_KERNEL
);
4209 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
4210 1, MEM_CGROUP_ID_MAX
,
4212 if (memcg
->id
.id
< 0)
4215 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4220 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
4223 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4226 INIT_WORK(&memcg
->high_work
, high_work_func
);
4227 memcg
->last_scanned_node
= MAX_NUMNODES
;
4228 INIT_LIST_HEAD(&memcg
->oom_notify
);
4229 mutex_init(&memcg
->thresholds_lock
);
4230 spin_lock_init(&memcg
->move_lock
);
4231 vmpressure_init(&memcg
->vmpressure
);
4232 INIT_LIST_HEAD(&memcg
->event_list
);
4233 spin_lock_init(&memcg
->event_list_lock
);
4234 memcg
->socket_pressure
= jiffies
;
4236 memcg
->kmemcg_id
= -1;
4238 #ifdef CONFIG_CGROUP_WRITEBACK
4239 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4241 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
4244 if (memcg
->id
.id
> 0)
4245 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4246 mem_cgroup_free(memcg
);
4250 static struct cgroup_subsys_state
* __ref
4251 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4253 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
4254 struct mem_cgroup
*memcg
;
4255 long error
= -ENOMEM
;
4257 memcg
= mem_cgroup_alloc();
4259 return ERR_PTR(error
);
4261 memcg
->high
= PAGE_COUNTER_MAX
;
4262 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4264 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4265 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4267 if (parent
&& parent
->use_hierarchy
) {
4268 memcg
->use_hierarchy
= true;
4269 page_counter_init(&memcg
->memory
, &parent
->memory
);
4270 page_counter_init(&memcg
->swap
, &parent
->swap
);
4271 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4272 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4273 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
4275 page_counter_init(&memcg
->memory
, NULL
);
4276 page_counter_init(&memcg
->swap
, NULL
);
4277 page_counter_init(&memcg
->memsw
, NULL
);
4278 page_counter_init(&memcg
->kmem
, NULL
);
4279 page_counter_init(&memcg
->tcpmem
, NULL
);
4281 * Deeper hierachy with use_hierarchy == false doesn't make
4282 * much sense so let cgroup subsystem know about this
4283 * unfortunate state in our controller.
4285 if (parent
!= root_mem_cgroup
)
4286 memory_cgrp_subsys
.broken_hierarchy
= true;
4289 /* The following stuff does not apply to the root */
4291 root_mem_cgroup
= memcg
;
4295 error
= memcg_online_kmem(memcg
);
4299 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4300 static_branch_inc(&memcg_sockets_enabled_key
);
4304 mem_cgroup_free(memcg
);
4305 return ERR_PTR(-ENOMEM
);
4308 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4310 /* Online state pins memcg ID, memcg ID pins CSS */
4311 mem_cgroup_id_get(mem_cgroup_from_css(css
));
4316 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4318 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4319 struct mem_cgroup_event
*event
, *tmp
;
4322 * Unregister events and notify userspace.
4323 * Notify userspace about cgroup removing only after rmdir of cgroup
4324 * directory to avoid race between userspace and kernelspace.
4326 spin_lock(&memcg
->event_list_lock
);
4327 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4328 list_del_init(&event
->list
);
4329 schedule_work(&event
->remove
);
4331 spin_unlock(&memcg
->event_list_lock
);
4333 memcg_offline_kmem(memcg
);
4334 wb_memcg_offline(memcg
);
4336 mem_cgroup_id_put(memcg
);
4339 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4341 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4343 invalidate_reclaim_iterators(memcg
);
4346 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4348 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4350 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4351 static_branch_dec(&memcg_sockets_enabled_key
);
4353 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
4354 static_branch_dec(&memcg_sockets_enabled_key
);
4356 vmpressure_cleanup(&memcg
->vmpressure
);
4357 cancel_work_sync(&memcg
->high_work
);
4358 mem_cgroup_remove_from_trees(memcg
);
4359 memcg_free_kmem(memcg
);
4360 mem_cgroup_free(memcg
);
4364 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4365 * @css: the target css
4367 * Reset the states of the mem_cgroup associated with @css. This is
4368 * invoked when the userland requests disabling on the default hierarchy
4369 * but the memcg is pinned through dependency. The memcg should stop
4370 * applying policies and should revert to the vanilla state as it may be
4371 * made visible again.
4373 * The current implementation only resets the essential configurations.
4374 * This needs to be expanded to cover all the visible parts.
4376 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4378 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4380 page_counter_limit(&memcg
->memory
, PAGE_COUNTER_MAX
);
4381 page_counter_limit(&memcg
->swap
, PAGE_COUNTER_MAX
);
4382 page_counter_limit(&memcg
->memsw
, PAGE_COUNTER_MAX
);
4383 page_counter_limit(&memcg
->kmem
, PAGE_COUNTER_MAX
);
4384 page_counter_limit(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
4386 memcg
->high
= PAGE_COUNTER_MAX
;
4387 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4388 memcg_wb_domain_size_changed(memcg
);
4392 /* Handlers for move charge at task migration. */
4393 static int mem_cgroup_do_precharge(unsigned long count
)
4397 /* Try a single bulk charge without reclaim first, kswapd may wake */
4398 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4400 mc
.precharge
+= count
;
4404 /* Try charges one by one with reclaim */
4406 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4420 enum mc_target_type
{
4426 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4427 unsigned long addr
, pte_t ptent
)
4429 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4431 if (!page
|| !page_mapped(page
))
4433 if (PageAnon(page
)) {
4434 if (!(mc
.flags
& MOVE_ANON
))
4437 if (!(mc
.flags
& MOVE_FILE
))
4440 if (!get_page_unless_zero(page
))
4447 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4448 pte_t ptent
, swp_entry_t
*entry
)
4450 struct page
*page
= NULL
;
4451 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4453 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4456 * Because lookup_swap_cache() updates some statistics counter,
4457 * we call find_get_page() with swapper_space directly.
4459 page
= find_get_page(swap_address_space(ent
), ent
.val
);
4460 if (do_memsw_account())
4461 entry
->val
= ent
.val
;
4466 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4467 pte_t ptent
, swp_entry_t
*entry
)
4473 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4474 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4476 struct page
*page
= NULL
;
4477 struct address_space
*mapping
;
4480 if (!vma
->vm_file
) /* anonymous vma */
4482 if (!(mc
.flags
& MOVE_FILE
))
4485 mapping
= vma
->vm_file
->f_mapping
;
4486 pgoff
= linear_page_index(vma
, addr
);
4488 /* page is moved even if it's not RSS of this task(page-faulted). */
4490 /* shmem/tmpfs may report page out on swap: account for that too. */
4491 if (shmem_mapping(mapping
)) {
4492 page
= find_get_entry(mapping
, pgoff
);
4493 if (radix_tree_exceptional_entry(page
)) {
4494 swp_entry_t swp
= radix_to_swp_entry(page
);
4495 if (do_memsw_account())
4497 page
= find_get_page(swap_address_space(swp
), swp
.val
);
4500 page
= find_get_page(mapping
, pgoff
);
4502 page
= find_get_page(mapping
, pgoff
);
4508 * mem_cgroup_move_account - move account of the page
4510 * @compound: charge the page as compound or small page
4511 * @from: mem_cgroup which the page is moved from.
4512 * @to: mem_cgroup which the page is moved to. @from != @to.
4514 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4516 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4519 static int mem_cgroup_move_account(struct page
*page
,
4521 struct mem_cgroup
*from
,
4522 struct mem_cgroup
*to
)
4524 unsigned long flags
;
4525 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
4529 VM_BUG_ON(from
== to
);
4530 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4531 VM_BUG_ON(compound
&& !PageTransHuge(page
));
4534 * Prevent mem_cgroup_migrate() from looking at
4535 * page->mem_cgroup of its source page while we change it.
4538 if (!trylock_page(page
))
4542 if (page
->mem_cgroup
!= from
)
4545 anon
= PageAnon(page
);
4547 spin_lock_irqsave(&from
->move_lock
, flags
);
4549 if (!anon
&& page_mapped(page
)) {
4550 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4552 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4557 * move_lock grabbed above and caller set from->moving_account, so
4558 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4559 * So mapping should be stable for dirty pages.
4561 if (!anon
&& PageDirty(page
)) {
4562 struct address_space
*mapping
= page_mapping(page
);
4564 if (mapping_cap_account_dirty(mapping
)) {
4565 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4567 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4572 if (PageWriteback(page
)) {
4573 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4575 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4580 * It is safe to change page->mem_cgroup here because the page
4581 * is referenced, charged, and isolated - we can't race with
4582 * uncharging, charging, migration, or LRU putback.
4585 /* caller should have done css_get */
4586 page
->mem_cgroup
= to
;
4587 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4591 local_irq_disable();
4592 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
4593 memcg_check_events(to
, page
);
4594 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
4595 memcg_check_events(from
, page
);
4604 * get_mctgt_type - get target type of moving charge
4605 * @vma: the vma the pte to be checked belongs
4606 * @addr: the address corresponding to the pte to be checked
4607 * @ptent: the pte to be checked
4608 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4611 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4612 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4613 * move charge. if @target is not NULL, the page is stored in target->page
4614 * with extra refcnt got(Callers should handle it).
4615 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4616 * target for charge migration. if @target is not NULL, the entry is stored
4619 * Called with pte lock held.
4622 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4623 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4625 struct page
*page
= NULL
;
4626 enum mc_target_type ret
= MC_TARGET_NONE
;
4627 swp_entry_t ent
= { .val
= 0 };
4629 if (pte_present(ptent
))
4630 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4631 else if (is_swap_pte(ptent
))
4632 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
4633 else if (pte_none(ptent
))
4634 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4636 if (!page
&& !ent
.val
)
4640 * Do only loose check w/o serialization.
4641 * mem_cgroup_move_account() checks the page is valid or
4642 * not under LRU exclusion.
4644 if (page
->mem_cgroup
== mc
.from
) {
4645 ret
= MC_TARGET_PAGE
;
4647 target
->page
= page
;
4649 if (!ret
|| !target
)
4652 /* There is a swap entry and a page doesn't exist or isn't charged */
4653 if (ent
.val
&& !ret
&&
4654 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4655 ret
= MC_TARGET_SWAP
;
4662 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4664 * We don't consider swapping or file mapped pages because THP does not
4665 * support them for now.
4666 * Caller should make sure that pmd_trans_huge(pmd) is true.
4668 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4669 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4671 struct page
*page
= NULL
;
4672 enum mc_target_type ret
= MC_TARGET_NONE
;
4674 page
= pmd_page(pmd
);
4675 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4676 if (!(mc
.flags
& MOVE_ANON
))
4678 if (page
->mem_cgroup
== mc
.from
) {
4679 ret
= MC_TARGET_PAGE
;
4682 target
->page
= page
;
4688 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4689 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4691 return MC_TARGET_NONE
;
4695 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4696 unsigned long addr
, unsigned long end
,
4697 struct mm_walk
*walk
)
4699 struct vm_area_struct
*vma
= walk
->vma
;
4703 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4705 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4706 mc
.precharge
+= HPAGE_PMD_NR
;
4711 if (pmd_trans_unstable(pmd
))
4713 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4714 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4715 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4716 mc
.precharge
++; /* increment precharge temporarily */
4717 pte_unmap_unlock(pte
- 1, ptl
);
4723 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4725 unsigned long precharge
;
4727 struct mm_walk mem_cgroup_count_precharge_walk
= {
4728 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4731 down_read(&mm
->mmap_sem
);
4732 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk
);
4733 up_read(&mm
->mmap_sem
);
4735 precharge
= mc
.precharge
;
4741 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4743 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4745 VM_BUG_ON(mc
.moving_task
);
4746 mc
.moving_task
= current
;
4747 return mem_cgroup_do_precharge(precharge
);
4750 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4751 static void __mem_cgroup_clear_mc(void)
4753 struct mem_cgroup
*from
= mc
.from
;
4754 struct mem_cgroup
*to
= mc
.to
;
4756 /* we must uncharge all the leftover precharges from mc.to */
4758 cancel_charge(mc
.to
, mc
.precharge
);
4762 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4763 * we must uncharge here.
4765 if (mc
.moved_charge
) {
4766 cancel_charge(mc
.from
, mc
.moved_charge
);
4767 mc
.moved_charge
= 0;
4769 /* we must fixup refcnts and charges */
4770 if (mc
.moved_swap
) {
4771 /* uncharge swap account from the old cgroup */
4772 if (!mem_cgroup_is_root(mc
.from
))
4773 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4776 * we charged both to->memory and to->memsw, so we
4777 * should uncharge to->memory.
4779 if (!mem_cgroup_is_root(mc
.to
))
4780 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4782 css_put_many(&mc
.from
->css
, mc
.moved_swap
);
4784 /* we've already done css_get(mc.to) */
4787 memcg_oom_recover(from
);
4788 memcg_oom_recover(to
);
4789 wake_up_all(&mc
.waitq
);
4792 static void mem_cgroup_clear_mc(void)
4794 struct mm_struct
*mm
= mc
.mm
;
4797 * we must clear moving_task before waking up waiters at the end of
4800 mc
.moving_task
= NULL
;
4801 __mem_cgroup_clear_mc();
4802 spin_lock(&mc
.lock
);
4806 spin_unlock(&mc
.lock
);
4811 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4813 struct cgroup_subsys_state
*css
;
4814 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
4815 struct mem_cgroup
*from
;
4816 struct task_struct
*leader
, *p
;
4817 struct mm_struct
*mm
;
4818 unsigned long move_flags
;
4821 /* charge immigration isn't supported on the default hierarchy */
4822 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4826 * Multi-process migrations only happen on the default hierarchy
4827 * where charge immigration is not used. Perform charge
4828 * immigration if @tset contains a leader and whine if there are
4832 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
4835 memcg
= mem_cgroup_from_css(css
);
4841 * We are now commited to this value whatever it is. Changes in this
4842 * tunable will only affect upcoming migrations, not the current one.
4843 * So we need to save it, and keep it going.
4845 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
4849 from
= mem_cgroup_from_task(p
);
4851 VM_BUG_ON(from
== memcg
);
4853 mm
= get_task_mm(p
);
4856 /* We move charges only when we move a owner of the mm */
4857 if (mm
->owner
== p
) {
4860 VM_BUG_ON(mc
.precharge
);
4861 VM_BUG_ON(mc
.moved_charge
);
4862 VM_BUG_ON(mc
.moved_swap
);
4864 spin_lock(&mc
.lock
);
4868 mc
.flags
= move_flags
;
4869 spin_unlock(&mc
.lock
);
4870 /* We set mc.moving_task later */
4872 ret
= mem_cgroup_precharge_mc(mm
);
4874 mem_cgroup_clear_mc();
4881 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4884 mem_cgroup_clear_mc();
4887 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4888 unsigned long addr
, unsigned long end
,
4889 struct mm_walk
*walk
)
4892 struct vm_area_struct
*vma
= walk
->vma
;
4895 enum mc_target_type target_type
;
4896 union mc_target target
;
4899 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4901 if (mc
.precharge
< HPAGE_PMD_NR
) {
4905 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
4906 if (target_type
== MC_TARGET_PAGE
) {
4908 if (!isolate_lru_page(page
)) {
4909 if (!mem_cgroup_move_account(page
, true,
4911 mc
.precharge
-= HPAGE_PMD_NR
;
4912 mc
.moved_charge
+= HPAGE_PMD_NR
;
4914 putback_lru_page(page
);
4922 if (pmd_trans_unstable(pmd
))
4925 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4926 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4927 pte_t ptent
= *(pte
++);
4933 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
4934 case MC_TARGET_PAGE
:
4937 * We can have a part of the split pmd here. Moving it
4938 * can be done but it would be too convoluted so simply
4939 * ignore such a partial THP and keep it in original
4940 * memcg. There should be somebody mapping the head.
4942 if (PageTransCompound(page
))
4944 if (isolate_lru_page(page
))
4946 if (!mem_cgroup_move_account(page
, false,
4949 /* we uncharge from mc.from later. */
4952 putback_lru_page(page
);
4953 put
: /* get_mctgt_type() gets the page */
4956 case MC_TARGET_SWAP
:
4958 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
4960 /* we fixup refcnts and charges later. */
4968 pte_unmap_unlock(pte
- 1, ptl
);
4973 * We have consumed all precharges we got in can_attach().
4974 * We try charge one by one, but don't do any additional
4975 * charges to mc.to if we have failed in charge once in attach()
4978 ret
= mem_cgroup_do_precharge(1);
4986 static void mem_cgroup_move_charge(void)
4988 struct mm_walk mem_cgroup_move_charge_walk
= {
4989 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4993 lru_add_drain_all();
4995 * Signal lock_page_memcg() to take the memcg's move_lock
4996 * while we're moving its pages to another memcg. Then wait
4997 * for already started RCU-only updates to finish.
4999 atomic_inc(&mc
.from
->moving_account
);
5002 if (unlikely(!down_read_trylock(&mc
.mm
->mmap_sem
))) {
5004 * Someone who are holding the mmap_sem might be waiting in
5005 * waitq. So we cancel all extra charges, wake up all waiters,
5006 * and retry. Because we cancel precharges, we might not be able
5007 * to move enough charges, but moving charge is a best-effort
5008 * feature anyway, so it wouldn't be a big problem.
5010 __mem_cgroup_clear_mc();
5015 * When we have consumed all precharges and failed in doing
5016 * additional charge, the page walk just aborts.
5018 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk
);
5019 up_read(&mc
.mm
->mmap_sem
);
5020 atomic_dec(&mc
.from
->moving_account
);
5023 static void mem_cgroup_move_task(void)
5026 mem_cgroup_move_charge();
5027 mem_cgroup_clear_mc();
5030 #else /* !CONFIG_MMU */
5031 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5035 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5038 static void mem_cgroup_move_task(void)
5044 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5045 * to verify whether we're attached to the default hierarchy on each mount
5048 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5051 * use_hierarchy is forced on the default hierarchy. cgroup core
5052 * guarantees that @root doesn't have any children, so turning it
5053 * on for the root memcg is enough.
5055 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5056 root_mem_cgroup
->use_hierarchy
= true;
5058 root_mem_cgroup
->use_hierarchy
= false;
5061 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5064 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5066 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
5069 static int memory_low_show(struct seq_file
*m
, void *v
)
5071 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5072 unsigned long low
= READ_ONCE(memcg
->low
);
5074 if (low
== PAGE_COUNTER_MAX
)
5075 seq_puts(m
, "max\n");
5077 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
5082 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5083 char *buf
, size_t nbytes
, loff_t off
)
5085 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5089 buf
= strstrip(buf
);
5090 err
= page_counter_memparse(buf
, "max", &low
);
5099 static int memory_high_show(struct seq_file
*m
, void *v
)
5101 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5102 unsigned long high
= READ_ONCE(memcg
->high
);
5104 if (high
== PAGE_COUNTER_MAX
)
5105 seq_puts(m
, "max\n");
5107 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
5112 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5113 char *buf
, size_t nbytes
, loff_t off
)
5115 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5116 unsigned long nr_pages
;
5120 buf
= strstrip(buf
);
5121 err
= page_counter_memparse(buf
, "max", &high
);
5127 nr_pages
= page_counter_read(&memcg
->memory
);
5128 if (nr_pages
> high
)
5129 try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
5132 memcg_wb_domain_size_changed(memcg
);
5136 static int memory_max_show(struct seq_file
*m
, void *v
)
5138 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5139 unsigned long max
= READ_ONCE(memcg
->memory
.limit
);
5141 if (max
== PAGE_COUNTER_MAX
)
5142 seq_puts(m
, "max\n");
5144 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5149 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5150 char *buf
, size_t nbytes
, loff_t off
)
5152 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5153 unsigned int nr_reclaims
= MEM_CGROUP_RECLAIM_RETRIES
;
5154 bool drained
= false;
5158 buf
= strstrip(buf
);
5159 err
= page_counter_memparse(buf
, "max", &max
);
5163 xchg(&memcg
->memory
.limit
, max
);
5166 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
5168 if (nr_pages
<= max
)
5171 if (signal_pending(current
)) {
5177 drain_all_stock(memcg
);
5183 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
5189 mem_cgroup_events(memcg
, MEMCG_OOM
, 1);
5190 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
5194 memcg_wb_domain_size_changed(memcg
);
5198 static int memory_events_show(struct seq_file
*m
, void *v
)
5200 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5202 seq_printf(m
, "low %lu\n", mem_cgroup_read_events(memcg
, MEMCG_LOW
));
5203 seq_printf(m
, "high %lu\n", mem_cgroup_read_events(memcg
, MEMCG_HIGH
));
5204 seq_printf(m
, "max %lu\n", mem_cgroup_read_events(memcg
, MEMCG_MAX
));
5205 seq_printf(m
, "oom %lu\n", mem_cgroup_read_events(memcg
, MEMCG_OOM
));
5210 static int memory_stat_show(struct seq_file
*m
, void *v
)
5212 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5213 unsigned long stat
[MEMCG_NR_STAT
];
5214 unsigned long events
[MEMCG_NR_EVENTS
];
5218 * Provide statistics on the state of the memory subsystem as
5219 * well as cumulative event counters that show past behavior.
5221 * This list is ordered following a combination of these gradients:
5222 * 1) generic big picture -> specifics and details
5223 * 2) reflecting userspace activity -> reflecting kernel heuristics
5225 * Current memory state:
5228 tree_stat(memcg
, stat
);
5229 tree_events(memcg
, events
);
5231 seq_printf(m
, "anon %llu\n",
5232 (u64
)stat
[MEM_CGROUP_STAT_RSS
] * PAGE_SIZE
);
5233 seq_printf(m
, "file %llu\n",
5234 (u64
)stat
[MEM_CGROUP_STAT_CACHE
] * PAGE_SIZE
);
5235 seq_printf(m
, "kernel_stack %llu\n",
5236 (u64
)stat
[MEMCG_KERNEL_STACK
] * PAGE_SIZE
);
5237 seq_printf(m
, "slab %llu\n",
5238 (u64
)(stat
[MEMCG_SLAB_RECLAIMABLE
] +
5239 stat
[MEMCG_SLAB_UNRECLAIMABLE
]) * PAGE_SIZE
);
5240 seq_printf(m
, "sock %llu\n",
5241 (u64
)stat
[MEMCG_SOCK
] * PAGE_SIZE
);
5243 seq_printf(m
, "file_mapped %llu\n",
5244 (u64
)stat
[MEM_CGROUP_STAT_FILE_MAPPED
] * PAGE_SIZE
);
5245 seq_printf(m
, "file_dirty %llu\n",
5246 (u64
)stat
[MEM_CGROUP_STAT_DIRTY
] * PAGE_SIZE
);
5247 seq_printf(m
, "file_writeback %llu\n",
5248 (u64
)stat
[MEM_CGROUP_STAT_WRITEBACK
] * PAGE_SIZE
);
5250 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5251 struct mem_cgroup
*mi
;
5252 unsigned long val
= 0;
5254 for_each_mem_cgroup_tree(mi
, memcg
)
5255 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
));
5256 seq_printf(m
, "%s %llu\n",
5257 mem_cgroup_lru_names
[i
], (u64
)val
* PAGE_SIZE
);
5260 seq_printf(m
, "slab_reclaimable %llu\n",
5261 (u64
)stat
[MEMCG_SLAB_RECLAIMABLE
] * PAGE_SIZE
);
5262 seq_printf(m
, "slab_unreclaimable %llu\n",
5263 (u64
)stat
[MEMCG_SLAB_UNRECLAIMABLE
] * PAGE_SIZE
);
5265 /* Accumulated memory events */
5267 seq_printf(m
, "pgfault %lu\n",
5268 events
[MEM_CGROUP_EVENTS_PGFAULT
]);
5269 seq_printf(m
, "pgmajfault %lu\n",
5270 events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
5275 static struct cftype memory_files
[] = {
5278 .flags
= CFTYPE_NOT_ON_ROOT
,
5279 .read_u64
= memory_current_read
,
5283 .flags
= CFTYPE_NOT_ON_ROOT
,
5284 .seq_show
= memory_low_show
,
5285 .write
= memory_low_write
,
5289 .flags
= CFTYPE_NOT_ON_ROOT
,
5290 .seq_show
= memory_high_show
,
5291 .write
= memory_high_write
,
5295 .flags
= CFTYPE_NOT_ON_ROOT
,
5296 .seq_show
= memory_max_show
,
5297 .write
= memory_max_write
,
5301 .flags
= CFTYPE_NOT_ON_ROOT
,
5302 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
5303 .seq_show
= memory_events_show
,
5307 .flags
= CFTYPE_NOT_ON_ROOT
,
5308 .seq_show
= memory_stat_show
,
5313 struct cgroup_subsys memory_cgrp_subsys
= {
5314 .css_alloc
= mem_cgroup_css_alloc
,
5315 .css_online
= mem_cgroup_css_online
,
5316 .css_offline
= mem_cgroup_css_offline
,
5317 .css_released
= mem_cgroup_css_released
,
5318 .css_free
= mem_cgroup_css_free
,
5319 .css_reset
= mem_cgroup_css_reset
,
5320 .can_attach
= mem_cgroup_can_attach
,
5321 .cancel_attach
= mem_cgroup_cancel_attach
,
5322 .post_attach
= mem_cgroup_move_task
,
5323 .bind
= mem_cgroup_bind
,
5324 .dfl_cftypes
= memory_files
,
5325 .legacy_cftypes
= mem_cgroup_legacy_files
,
5330 * mem_cgroup_low - check if memory consumption is below the normal range
5331 * @root: the highest ancestor to consider
5332 * @memcg: the memory cgroup to check
5334 * Returns %true if memory consumption of @memcg, and that of all
5335 * configurable ancestors up to @root, is below the normal range.
5337 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5339 if (mem_cgroup_disabled())
5343 * The toplevel group doesn't have a configurable range, so
5344 * it's never low when looked at directly, and it is not
5345 * considered an ancestor when assessing the hierarchy.
5348 if (memcg
== root_mem_cgroup
)
5351 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5354 while (memcg
!= root
) {
5355 memcg
= parent_mem_cgroup(memcg
);
5357 if (memcg
== root_mem_cgroup
)
5360 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5367 * mem_cgroup_try_charge - try charging a page
5368 * @page: page to charge
5369 * @mm: mm context of the victim
5370 * @gfp_mask: reclaim mode
5371 * @memcgp: charged memcg return
5372 * @compound: charge the page as compound or small page
5374 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5375 * pages according to @gfp_mask if necessary.
5377 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5378 * Otherwise, an error code is returned.
5380 * After page->mapping has been set up, the caller must finalize the
5381 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5382 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5384 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5385 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
5388 struct mem_cgroup
*memcg
= NULL
;
5389 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5392 if (mem_cgroup_disabled())
5395 if (PageSwapCache(page
)) {
5397 * Every swap fault against a single page tries to charge the
5398 * page, bail as early as possible. shmem_unuse() encounters
5399 * already charged pages, too. The USED bit is protected by
5400 * the page lock, which serializes swap cache removal, which
5401 * in turn serializes uncharging.
5403 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5404 if (page
->mem_cgroup
)
5407 if (do_swap_account
) {
5408 swp_entry_t ent
= { .val
= page_private(page
), };
5409 unsigned short id
= lookup_swap_cgroup_id(ent
);
5412 memcg
= mem_cgroup_from_id(id
);
5413 if (memcg
&& !css_tryget_online(&memcg
->css
))
5420 memcg
= get_mem_cgroup_from_mm(mm
);
5422 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5424 css_put(&memcg
->css
);
5431 * mem_cgroup_commit_charge - commit a page charge
5432 * @page: page to charge
5433 * @memcg: memcg to charge the page to
5434 * @lrucare: page might be on LRU already
5435 * @compound: charge the page as compound or small page
5437 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5438 * after page->mapping has been set up. This must happen atomically
5439 * as part of the page instantiation, i.e. under the page table lock
5440 * for anonymous pages, under the page lock for page and swap cache.
5442 * In addition, the page must not be on the LRU during the commit, to
5443 * prevent racing with task migration. If it might be, use @lrucare.
5445 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5447 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5448 bool lrucare
, bool compound
)
5450 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5452 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5453 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5455 if (mem_cgroup_disabled())
5458 * Swap faults will attempt to charge the same page multiple
5459 * times. But reuse_swap_page() might have removed the page
5460 * from swapcache already, so we can't check PageSwapCache().
5465 commit_charge(page
, memcg
, lrucare
);
5467 local_irq_disable();
5468 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
5469 memcg_check_events(memcg
, page
);
5472 if (do_memsw_account() && PageSwapCache(page
)) {
5473 swp_entry_t entry
= { .val
= page_private(page
) };
5475 * The swap entry might not get freed for a long time,
5476 * let's not wait for it. The page already received a
5477 * memory+swap charge, drop the swap entry duplicate.
5479 mem_cgroup_uncharge_swap(entry
);
5484 * mem_cgroup_cancel_charge - cancel a page charge
5485 * @page: page to charge
5486 * @memcg: memcg to charge the page to
5487 * @compound: charge the page as compound or small page
5489 * Cancel a charge transaction started by mem_cgroup_try_charge().
5491 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5494 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5496 if (mem_cgroup_disabled())
5499 * Swap faults will attempt to charge the same page multiple
5500 * times. But reuse_swap_page() might have removed the page
5501 * from swapcache already, so we can't check PageSwapCache().
5506 cancel_charge(memcg
, nr_pages
);
5509 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5510 unsigned long nr_anon
, unsigned long nr_file
,
5511 unsigned long nr_huge
, unsigned long nr_kmem
,
5512 struct page
*dummy_page
)
5514 unsigned long nr_pages
= nr_anon
+ nr_file
+ nr_kmem
;
5515 unsigned long flags
;
5517 if (!mem_cgroup_is_root(memcg
)) {
5518 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5519 if (do_memsw_account())
5520 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5521 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && nr_kmem
)
5522 page_counter_uncharge(&memcg
->kmem
, nr_kmem
);
5523 memcg_oom_recover(memcg
);
5526 local_irq_save(flags
);
5527 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5528 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5529 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5530 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5531 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5532 memcg_check_events(memcg
, dummy_page
);
5533 local_irq_restore(flags
);
5535 if (!mem_cgroup_is_root(memcg
))
5536 css_put_many(&memcg
->css
, nr_pages
);
5539 static void uncharge_list(struct list_head
*page_list
)
5541 struct mem_cgroup
*memcg
= NULL
;
5542 unsigned long nr_anon
= 0;
5543 unsigned long nr_file
= 0;
5544 unsigned long nr_huge
= 0;
5545 unsigned long nr_kmem
= 0;
5546 unsigned long pgpgout
= 0;
5547 struct list_head
*next
;
5551 * Note that the list can be a single page->lru; hence the
5552 * do-while loop instead of a simple list_for_each_entry().
5554 next
= page_list
->next
;
5556 page
= list_entry(next
, struct page
, lru
);
5557 next
= page
->lru
.next
;
5559 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5560 VM_BUG_ON_PAGE(page_count(page
), page
);
5562 if (!page
->mem_cgroup
)
5566 * Nobody should be changing or seriously looking at
5567 * page->mem_cgroup at this point, we have fully
5568 * exclusive access to the page.
5571 if (memcg
!= page
->mem_cgroup
) {
5573 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5574 nr_huge
, nr_kmem
, page
);
5575 pgpgout
= nr_anon
= nr_file
=
5576 nr_huge
= nr_kmem
= 0;
5578 memcg
= page
->mem_cgroup
;
5581 if (!PageKmemcg(page
)) {
5582 unsigned int nr_pages
= 1;
5584 if (PageTransHuge(page
)) {
5585 nr_pages
<<= compound_order(page
);
5586 nr_huge
+= nr_pages
;
5589 nr_anon
+= nr_pages
;
5591 nr_file
+= nr_pages
;
5594 nr_kmem
+= 1 << compound_order(page
);
5596 page
->mem_cgroup
= NULL
;
5597 } while (next
!= page_list
);
5600 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5601 nr_huge
, nr_kmem
, page
);
5605 * mem_cgroup_uncharge - uncharge a page
5606 * @page: page to uncharge
5608 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5609 * mem_cgroup_commit_charge().
5611 void mem_cgroup_uncharge(struct page
*page
)
5613 if (mem_cgroup_disabled())
5616 /* Don't touch page->lru of any random page, pre-check: */
5617 if (!page
->mem_cgroup
)
5620 INIT_LIST_HEAD(&page
->lru
);
5621 uncharge_list(&page
->lru
);
5625 * mem_cgroup_uncharge_list - uncharge a list of page
5626 * @page_list: list of pages to uncharge
5628 * Uncharge a list of pages previously charged with
5629 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5631 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5633 if (mem_cgroup_disabled())
5636 if (!list_empty(page_list
))
5637 uncharge_list(page_list
);
5641 * mem_cgroup_migrate - charge a page's replacement
5642 * @oldpage: currently circulating page
5643 * @newpage: replacement page
5645 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5646 * be uncharged upon free.
5648 * Both pages must be locked, @newpage->mapping must be set up.
5650 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
5652 struct mem_cgroup
*memcg
;
5653 unsigned int nr_pages
;
5655 unsigned long flags
;
5657 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5658 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5659 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5660 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5663 if (mem_cgroup_disabled())
5666 /* Page cache replacement: new page already charged? */
5667 if (newpage
->mem_cgroup
)
5670 /* Swapcache readahead pages can get replaced before being charged */
5671 memcg
= oldpage
->mem_cgroup
;
5675 /* Force-charge the new page. The old one will be freed soon */
5676 compound
= PageTransHuge(newpage
);
5677 nr_pages
= compound
? hpage_nr_pages(newpage
) : 1;
5679 page_counter_charge(&memcg
->memory
, nr_pages
);
5680 if (do_memsw_account())
5681 page_counter_charge(&memcg
->memsw
, nr_pages
);
5682 css_get_many(&memcg
->css
, nr_pages
);
5684 commit_charge(newpage
, memcg
, false);
5686 local_irq_save(flags
);
5687 mem_cgroup_charge_statistics(memcg
, newpage
, compound
, nr_pages
);
5688 memcg_check_events(memcg
, newpage
);
5689 local_irq_restore(flags
);
5692 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
5693 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
5695 void sock_update_memcg(struct sock
*sk
)
5697 struct mem_cgroup
*memcg
;
5699 /* Socket cloning can throw us here with sk_cgrp already
5700 * filled. It won't however, necessarily happen from
5701 * process context. So the test for root memcg given
5702 * the current task's memcg won't help us in this case.
5704 * Respecting the original socket's memcg is a better
5705 * decision in this case.
5708 BUG_ON(mem_cgroup_is_root(sk
->sk_memcg
));
5709 css_get(&sk
->sk_memcg
->css
);
5714 memcg
= mem_cgroup_from_task(current
);
5715 if (memcg
== root_mem_cgroup
)
5717 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
5719 if (css_tryget_online(&memcg
->css
))
5720 sk
->sk_memcg
= memcg
;
5724 EXPORT_SYMBOL(sock_update_memcg
);
5726 void sock_release_memcg(struct sock
*sk
)
5728 WARN_ON(!sk
->sk_memcg
);
5729 css_put(&sk
->sk_memcg
->css
);
5733 * mem_cgroup_charge_skmem - charge socket memory
5734 * @memcg: memcg to charge
5735 * @nr_pages: number of pages to charge
5737 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5738 * @memcg's configured limit, %false if the charge had to be forced.
5740 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5742 gfp_t gfp_mask
= GFP_KERNEL
;
5744 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5745 struct page_counter
*fail
;
5747 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
5748 memcg
->tcpmem_pressure
= 0;
5751 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
5752 memcg
->tcpmem_pressure
= 1;
5756 /* Don't block in the packet receive path */
5758 gfp_mask
= GFP_NOWAIT
;
5760 this_cpu_add(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5762 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
5765 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
5770 * mem_cgroup_uncharge_skmem - uncharge socket memory
5771 * @memcg - memcg to uncharge
5772 * @nr_pages - number of pages to uncharge
5774 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5776 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5777 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
5781 this_cpu_sub(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5783 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5784 css_put_many(&memcg
->css
, nr_pages
);
5787 static int __init
cgroup_memory(char *s
)
5791 while ((token
= strsep(&s
, ",")) != NULL
) {
5794 if (!strcmp(token
, "nosocket"))
5795 cgroup_memory_nosocket
= true;
5796 if (!strcmp(token
, "nokmem"))
5797 cgroup_memory_nokmem
= true;
5801 __setup("cgroup.memory=", cgroup_memory
);
5804 * subsys_initcall() for memory controller.
5806 * Some parts like hotcpu_notifier() have to be initialized from this context
5807 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5808 * everything that doesn't depend on a specific mem_cgroup structure should
5809 * be initialized from here.
5811 static int __init
mem_cgroup_init(void)
5815 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5817 for_each_possible_cpu(cpu
)
5818 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5821 for_each_node(node
) {
5822 struct mem_cgroup_tree_per_node
*rtpn
;
5825 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5826 node_online(node
) ? node
: NUMA_NO_NODE
);
5828 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5829 struct mem_cgroup_tree_per_zone
*rtpz
;
5831 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
5832 rtpz
->rb_root
= RB_ROOT
;
5833 spin_lock_init(&rtpz
->lock
);
5835 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5840 subsys_initcall(mem_cgroup_init
);
5842 #ifdef CONFIG_MEMCG_SWAP
5844 * mem_cgroup_swapout - transfer a memsw charge to swap
5845 * @page: page whose memsw charge to transfer
5846 * @entry: swap entry to move the charge to
5848 * Transfer the memsw charge of @page to @entry.
5850 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5852 struct mem_cgroup
*memcg
;
5853 unsigned short oldid
;
5855 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5856 VM_BUG_ON_PAGE(page_count(page
), page
);
5858 if (!do_memsw_account())
5861 memcg
= page
->mem_cgroup
;
5863 /* Readahead page, never charged */
5867 mem_cgroup_id_get(memcg
);
5868 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5869 VM_BUG_ON_PAGE(oldid
, page
);
5870 mem_cgroup_swap_statistics(memcg
, true);
5872 page
->mem_cgroup
= NULL
;
5874 if (!mem_cgroup_is_root(memcg
))
5875 page_counter_uncharge(&memcg
->memory
, 1);
5878 * Interrupts should be disabled here because the caller holds the
5879 * mapping->tree_lock lock which is taken with interrupts-off. It is
5880 * important here to have the interrupts disabled because it is the
5881 * only synchronisation we have for udpating the per-CPU variables.
5883 VM_BUG_ON(!irqs_disabled());
5884 mem_cgroup_charge_statistics(memcg
, page
, false, -1);
5885 memcg_check_events(memcg
, page
);
5887 if (!mem_cgroup_is_root(memcg
))
5888 css_put(&memcg
->css
);
5892 * mem_cgroup_try_charge_swap - try charging a swap entry
5893 * @page: page being added to swap
5894 * @entry: swap entry to charge
5896 * Try to charge @entry to the memcg that @page belongs to.
5898 * Returns 0 on success, -ENOMEM on failure.
5900 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
5902 struct mem_cgroup
*memcg
;
5903 struct page_counter
*counter
;
5904 unsigned short oldid
;
5906 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
5909 memcg
= page
->mem_cgroup
;
5911 /* Readahead page, never charged */
5915 if (!mem_cgroup_is_root(memcg
) &&
5916 !page_counter_try_charge(&memcg
->swap
, 1, &counter
))
5919 mem_cgroup_id_get(memcg
);
5920 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5921 VM_BUG_ON_PAGE(oldid
, page
);
5922 mem_cgroup_swap_statistics(memcg
, true);
5928 * mem_cgroup_uncharge_swap - uncharge a swap entry
5929 * @entry: swap entry to uncharge
5931 * Drop the swap charge associated with @entry.
5933 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5935 struct mem_cgroup
*memcg
;
5938 if (!do_swap_account
)
5941 id
= swap_cgroup_record(entry
, 0);
5943 memcg
= mem_cgroup_from_id(id
);
5945 if (!mem_cgroup_is_root(memcg
)) {
5946 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5947 page_counter_uncharge(&memcg
->swap
, 1);
5949 page_counter_uncharge(&memcg
->memsw
, 1);
5951 mem_cgroup_swap_statistics(memcg
, false);
5952 mem_cgroup_id_put(memcg
);
5957 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
5959 long nr_swap_pages
= get_nr_swap_pages();
5961 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5962 return nr_swap_pages
;
5963 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5964 nr_swap_pages
= min_t(long, nr_swap_pages
,
5965 READ_ONCE(memcg
->swap
.limit
) -
5966 page_counter_read(&memcg
->swap
));
5967 return nr_swap_pages
;
5970 bool mem_cgroup_swap_full(struct page
*page
)
5972 struct mem_cgroup
*memcg
;
5974 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5978 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5981 memcg
= page
->mem_cgroup
;
5985 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5986 if (page_counter_read(&memcg
->swap
) * 2 >= memcg
->swap
.limit
)
5992 /* for remember boot option*/
5993 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5994 static int really_do_swap_account __initdata
= 1;
5996 static int really_do_swap_account __initdata
;
5999 static int __init
enable_swap_account(char *s
)
6001 if (!strcmp(s
, "1"))
6002 really_do_swap_account
= 1;
6003 else if (!strcmp(s
, "0"))
6004 really_do_swap_account
= 0;
6007 __setup("swapaccount=", enable_swap_account
);
6009 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
6012 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6014 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
6017 static int swap_max_show(struct seq_file
*m
, void *v
)
6019 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
6020 unsigned long max
= READ_ONCE(memcg
->swap
.limit
);
6022 if (max
== PAGE_COUNTER_MAX
)
6023 seq_puts(m
, "max\n");
6025 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
6030 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
6031 char *buf
, size_t nbytes
, loff_t off
)
6033 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6037 buf
= strstrip(buf
);
6038 err
= page_counter_memparse(buf
, "max", &max
);
6042 mutex_lock(&memcg_limit_mutex
);
6043 err
= page_counter_limit(&memcg
->swap
, max
);
6044 mutex_unlock(&memcg_limit_mutex
);
6051 static struct cftype swap_files
[] = {
6053 .name
= "swap.current",
6054 .flags
= CFTYPE_NOT_ON_ROOT
,
6055 .read_u64
= swap_current_read
,
6059 .flags
= CFTYPE_NOT_ON_ROOT
,
6060 .seq_show
= swap_max_show
,
6061 .write
= swap_max_write
,
6066 static struct cftype memsw_cgroup_files
[] = {
6068 .name
= "memsw.usage_in_bytes",
6069 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6070 .read_u64
= mem_cgroup_read_u64
,
6073 .name
= "memsw.max_usage_in_bytes",
6074 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6075 .write
= mem_cgroup_reset
,
6076 .read_u64
= mem_cgroup_read_u64
,
6079 .name
= "memsw.limit_in_bytes",
6080 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6081 .write
= mem_cgroup_write
,
6082 .read_u64
= mem_cgroup_read_u64
,
6085 .name
= "memsw.failcnt",
6086 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6087 .write
= mem_cgroup_reset
,
6088 .read_u64
= mem_cgroup_read_u64
,
6090 { }, /* terminate */
6093 static int __init
mem_cgroup_swap_init(void)
6095 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6096 do_swap_account
= 1;
6097 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
6099 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
6100 memsw_cgroup_files
));
6104 subsys_initcall(mem_cgroup_swap_init
);
6106 #endif /* CONFIG_MEMCG_SWAP */