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 when a page is added to or removed from an
1029 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1032 struct mem_cgroup_per_zone
*mz
;
1033 unsigned long *lru_size
;
1035 if (mem_cgroup_disabled())
1038 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1039 lru_size
= mz
->lru_size
+ lru
;
1040 *lru_size
+= nr_pages
;
1041 VM_BUG_ON((long)(*lru_size
) < 0);
1044 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1046 struct mem_cgroup
*task_memcg
;
1047 struct task_struct
*p
;
1050 p
= find_lock_task_mm(task
);
1052 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1056 * All threads may have already detached their mm's, but the oom
1057 * killer still needs to detect if they have already been oom
1058 * killed to prevent needlessly killing additional tasks.
1061 task_memcg
= mem_cgroup_from_task(task
);
1062 css_get(&task_memcg
->css
);
1065 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1066 css_put(&task_memcg
->css
);
1071 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1072 * @memcg: the memory cgroup
1074 * Returns the maximum amount of memory @mem can be charged with, in
1077 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1079 unsigned long margin
= 0;
1080 unsigned long count
;
1081 unsigned long limit
;
1083 count
= page_counter_read(&memcg
->memory
);
1084 limit
= READ_ONCE(memcg
->memory
.limit
);
1086 margin
= limit
- count
;
1088 if (do_memsw_account()) {
1089 count
= page_counter_read(&memcg
->memsw
);
1090 limit
= READ_ONCE(memcg
->memsw
.limit
);
1092 margin
= min(margin
, limit
- count
);
1099 * A routine for checking "mem" is under move_account() or not.
1101 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1102 * moving cgroups. This is for waiting at high-memory pressure
1105 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1107 struct mem_cgroup
*from
;
1108 struct mem_cgroup
*to
;
1111 * Unlike task_move routines, we access mc.to, mc.from not under
1112 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1114 spin_lock(&mc
.lock
);
1120 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1121 mem_cgroup_is_descendant(to
, memcg
);
1123 spin_unlock(&mc
.lock
);
1127 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1129 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1130 if (mem_cgroup_under_move(memcg
)) {
1132 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1133 /* moving charge context might have finished. */
1136 finish_wait(&mc
.waitq
, &wait
);
1143 #define K(x) ((x) << (PAGE_SHIFT-10))
1145 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1146 * @memcg: The memory cgroup that went over limit
1147 * @p: Task that is going to be killed
1149 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1152 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1154 struct mem_cgroup
*iter
;
1160 pr_info("Task in ");
1161 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1162 pr_cont(" killed as a result of limit of ");
1164 pr_info("Memory limit reached of cgroup ");
1167 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1172 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1173 K((u64
)page_counter_read(&memcg
->memory
)),
1174 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1175 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1176 K((u64
)page_counter_read(&memcg
->memsw
)),
1177 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1178 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1179 K((u64
)page_counter_read(&memcg
->kmem
)),
1180 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1182 for_each_mem_cgroup_tree(iter
, memcg
) {
1183 pr_info("Memory cgroup stats for ");
1184 pr_cont_cgroup_path(iter
->css
.cgroup
);
1187 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1188 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1190 pr_cont(" %s:%luKB", mem_cgroup_stat_names
[i
],
1191 K(mem_cgroup_read_stat(iter
, i
)));
1194 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1195 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1196 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1203 * This function returns the number of memcg under hierarchy tree. Returns
1204 * 1(self count) if no children.
1206 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1209 struct mem_cgroup
*iter
;
1211 for_each_mem_cgroup_tree(iter
, memcg
)
1217 * Return the memory (and swap, if configured) limit for a memcg.
1219 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1221 unsigned long limit
;
1223 limit
= memcg
->memory
.limit
;
1224 if (mem_cgroup_swappiness(memcg
)) {
1225 unsigned long memsw_limit
;
1226 unsigned long swap_limit
;
1228 memsw_limit
= memcg
->memsw
.limit
;
1229 swap_limit
= memcg
->swap
.limit
;
1230 swap_limit
= min(swap_limit
, (unsigned long)total_swap_pages
);
1231 limit
= min(limit
+ swap_limit
, memsw_limit
);
1236 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1239 struct oom_control oc
= {
1242 .gfp_mask
= gfp_mask
,
1245 struct mem_cgroup
*iter
;
1246 unsigned long chosen_points
= 0;
1247 unsigned long totalpages
;
1248 unsigned int points
= 0;
1249 struct task_struct
*chosen
= NULL
;
1251 mutex_lock(&oom_lock
);
1254 * If current has a pending SIGKILL or is exiting, then automatically
1255 * select it. The goal is to allow it to allocate so that it may
1256 * quickly exit and free its memory.
1258 if (fatal_signal_pending(current
) || task_will_free_mem(current
)) {
1259 mark_oom_victim(current
);
1263 check_panic_on_oom(&oc
, CONSTRAINT_MEMCG
, memcg
);
1264 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1265 for_each_mem_cgroup_tree(iter
, memcg
) {
1266 struct css_task_iter it
;
1267 struct task_struct
*task
;
1269 css_task_iter_start(&iter
->css
, &it
);
1270 while ((task
= css_task_iter_next(&it
))) {
1271 switch (oom_scan_process_thread(&oc
, task
, totalpages
)) {
1272 case OOM_SCAN_SELECT
:
1274 put_task_struct(chosen
);
1276 chosen_points
= ULONG_MAX
;
1277 get_task_struct(chosen
);
1279 case OOM_SCAN_CONTINUE
:
1281 case OOM_SCAN_ABORT
:
1282 css_task_iter_end(&it
);
1283 mem_cgroup_iter_break(memcg
, iter
);
1285 put_task_struct(chosen
);
1290 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1291 if (!points
|| points
< chosen_points
)
1293 /* Prefer thread group leaders for display purposes */
1294 if (points
== chosen_points
&&
1295 thread_group_leader(chosen
))
1299 put_task_struct(chosen
);
1301 chosen_points
= points
;
1302 get_task_struct(chosen
);
1304 css_task_iter_end(&it
);
1308 points
= chosen_points
* 1000 / totalpages
;
1309 oom_kill_process(&oc
, chosen
, points
, totalpages
, memcg
,
1310 "Memory cgroup out of memory");
1313 mutex_unlock(&oom_lock
);
1317 #if MAX_NUMNODES > 1
1320 * test_mem_cgroup_node_reclaimable
1321 * @memcg: the target memcg
1322 * @nid: the node ID to be checked.
1323 * @noswap : specify true here if the user wants flle only information.
1325 * This function returns whether the specified memcg contains any
1326 * reclaimable pages on a node. Returns true if there are any reclaimable
1327 * pages in the node.
1329 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1330 int nid
, bool noswap
)
1332 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1334 if (noswap
|| !total_swap_pages
)
1336 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1343 * Always updating the nodemask is not very good - even if we have an empty
1344 * list or the wrong list here, we can start from some node and traverse all
1345 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1348 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1352 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1353 * pagein/pageout changes since the last update.
1355 if (!atomic_read(&memcg
->numainfo_events
))
1357 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1360 /* make a nodemask where this memcg uses memory from */
1361 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1363 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1365 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1366 node_clear(nid
, memcg
->scan_nodes
);
1369 atomic_set(&memcg
->numainfo_events
, 0);
1370 atomic_set(&memcg
->numainfo_updating
, 0);
1374 * Selecting a node where we start reclaim from. Because what we need is just
1375 * reducing usage counter, start from anywhere is O,K. Considering
1376 * memory reclaim from current node, there are pros. and cons.
1378 * Freeing memory from current node means freeing memory from a node which
1379 * we'll use or we've used. So, it may make LRU bad. And if several threads
1380 * hit limits, it will see a contention on a node. But freeing from remote
1381 * node means more costs for memory reclaim because of memory latency.
1383 * Now, we use round-robin. Better algorithm is welcomed.
1385 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1389 mem_cgroup_may_update_nodemask(memcg
);
1390 node
= memcg
->last_scanned_node
;
1392 node
= next_node_in(node
, memcg
->scan_nodes
);
1394 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1395 * last time it really checked all the LRUs due to rate limiting.
1396 * Fallback to the current node in that case for simplicity.
1398 if (unlikely(node
== MAX_NUMNODES
))
1399 node
= numa_node_id();
1401 memcg
->last_scanned_node
= node
;
1405 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1411 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1414 unsigned long *total_scanned
)
1416 struct mem_cgroup
*victim
= NULL
;
1419 unsigned long excess
;
1420 unsigned long nr_scanned
;
1421 struct mem_cgroup_reclaim_cookie reclaim
= {
1426 excess
= soft_limit_excess(root_memcg
);
1429 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1434 * If we have not been able to reclaim
1435 * anything, it might because there are
1436 * no reclaimable pages under this hierarchy
1441 * We want to do more targeted reclaim.
1442 * excess >> 2 is not to excessive so as to
1443 * reclaim too much, nor too less that we keep
1444 * coming back to reclaim from this cgroup
1446 if (total
>= (excess
>> 2) ||
1447 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1452 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1454 *total_scanned
+= nr_scanned
;
1455 if (!soft_limit_excess(root_memcg
))
1458 mem_cgroup_iter_break(root_memcg
, victim
);
1462 #ifdef CONFIG_LOCKDEP
1463 static struct lockdep_map memcg_oom_lock_dep_map
= {
1464 .name
= "memcg_oom_lock",
1468 static DEFINE_SPINLOCK(memcg_oom_lock
);
1471 * Check OOM-Killer is already running under our hierarchy.
1472 * If someone is running, return false.
1474 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1476 struct mem_cgroup
*iter
, *failed
= NULL
;
1478 spin_lock(&memcg_oom_lock
);
1480 for_each_mem_cgroup_tree(iter
, memcg
) {
1481 if (iter
->oom_lock
) {
1483 * this subtree of our hierarchy is already locked
1484 * so we cannot give a lock.
1487 mem_cgroup_iter_break(memcg
, iter
);
1490 iter
->oom_lock
= true;
1495 * OK, we failed to lock the whole subtree so we have
1496 * to clean up what we set up to the failing subtree
1498 for_each_mem_cgroup_tree(iter
, memcg
) {
1499 if (iter
== failed
) {
1500 mem_cgroup_iter_break(memcg
, iter
);
1503 iter
->oom_lock
= false;
1506 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1508 spin_unlock(&memcg_oom_lock
);
1513 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1515 struct mem_cgroup
*iter
;
1517 spin_lock(&memcg_oom_lock
);
1518 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1519 for_each_mem_cgroup_tree(iter
, memcg
)
1520 iter
->oom_lock
= false;
1521 spin_unlock(&memcg_oom_lock
);
1524 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1526 struct mem_cgroup
*iter
;
1528 spin_lock(&memcg_oom_lock
);
1529 for_each_mem_cgroup_tree(iter
, memcg
)
1531 spin_unlock(&memcg_oom_lock
);
1534 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1536 struct mem_cgroup
*iter
;
1539 * When a new child is created while the hierarchy is under oom,
1540 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1542 spin_lock(&memcg_oom_lock
);
1543 for_each_mem_cgroup_tree(iter
, memcg
)
1544 if (iter
->under_oom
> 0)
1546 spin_unlock(&memcg_oom_lock
);
1549 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1551 struct oom_wait_info
{
1552 struct mem_cgroup
*memcg
;
1556 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1557 unsigned mode
, int sync
, void *arg
)
1559 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1560 struct mem_cgroup
*oom_wait_memcg
;
1561 struct oom_wait_info
*oom_wait_info
;
1563 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1564 oom_wait_memcg
= oom_wait_info
->memcg
;
1566 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1567 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1569 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1572 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1575 * For the following lockless ->under_oom test, the only required
1576 * guarantee is that it must see the state asserted by an OOM when
1577 * this function is called as a result of userland actions
1578 * triggered by the notification of the OOM. This is trivially
1579 * achieved by invoking mem_cgroup_mark_under_oom() before
1580 * triggering notification.
1582 if (memcg
&& memcg
->under_oom
)
1583 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1586 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1588 if (!current
->memcg_may_oom
)
1591 * We are in the middle of the charge context here, so we
1592 * don't want to block when potentially sitting on a callstack
1593 * that holds all kinds of filesystem and mm locks.
1595 * Also, the caller may handle a failed allocation gracefully
1596 * (like optional page cache readahead) and so an OOM killer
1597 * invocation might not even be necessary.
1599 * That's why we don't do anything here except remember the
1600 * OOM context and then deal with it at the end of the page
1601 * fault when the stack is unwound, the locks are released,
1602 * and when we know whether the fault was overall successful.
1604 css_get(&memcg
->css
);
1605 current
->memcg_in_oom
= memcg
;
1606 current
->memcg_oom_gfp_mask
= mask
;
1607 current
->memcg_oom_order
= order
;
1611 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1612 * @handle: actually kill/wait or just clean up the OOM state
1614 * This has to be called at the end of a page fault if the memcg OOM
1615 * handler was enabled.
1617 * Memcg supports userspace OOM handling where failed allocations must
1618 * sleep on a waitqueue until the userspace task resolves the
1619 * situation. Sleeping directly in the charge context with all kinds
1620 * of locks held is not a good idea, instead we remember an OOM state
1621 * in the task and mem_cgroup_oom_synchronize() has to be called at
1622 * the end of the page fault to complete the OOM handling.
1624 * Returns %true if an ongoing memcg OOM situation was detected and
1625 * completed, %false otherwise.
1627 bool mem_cgroup_oom_synchronize(bool handle
)
1629 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1630 struct oom_wait_info owait
;
1633 /* OOM is global, do not handle */
1637 if (!handle
|| oom_killer_disabled
)
1640 owait
.memcg
= memcg
;
1641 owait
.wait
.flags
= 0;
1642 owait
.wait
.func
= memcg_oom_wake_function
;
1643 owait
.wait
.private = current
;
1644 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1646 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1647 mem_cgroup_mark_under_oom(memcg
);
1649 locked
= mem_cgroup_oom_trylock(memcg
);
1652 mem_cgroup_oom_notify(memcg
);
1654 if (locked
&& !memcg
->oom_kill_disable
) {
1655 mem_cgroup_unmark_under_oom(memcg
);
1656 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1657 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1658 current
->memcg_oom_order
);
1661 mem_cgroup_unmark_under_oom(memcg
);
1662 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1666 mem_cgroup_oom_unlock(memcg
);
1668 * There is no guarantee that an OOM-lock contender
1669 * sees the wakeups triggered by the OOM kill
1670 * uncharges. Wake any sleepers explicitely.
1672 memcg_oom_recover(memcg
);
1675 current
->memcg_in_oom
= NULL
;
1676 css_put(&memcg
->css
);
1681 * lock_page_memcg - lock a page->mem_cgroup binding
1684 * This function protects unlocked LRU pages from being moved to
1685 * another cgroup and stabilizes their page->mem_cgroup binding.
1687 void lock_page_memcg(struct page
*page
)
1689 struct mem_cgroup
*memcg
;
1690 unsigned long flags
;
1693 * The RCU lock is held throughout the transaction. The fast
1694 * path can get away without acquiring the memcg->move_lock
1695 * because page moving starts with an RCU grace period.
1699 if (mem_cgroup_disabled())
1702 memcg
= page
->mem_cgroup
;
1703 if (unlikely(!memcg
))
1706 if (atomic_read(&memcg
->moving_account
) <= 0)
1709 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1710 if (memcg
!= page
->mem_cgroup
) {
1711 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1716 * When charge migration first begins, we can have locked and
1717 * unlocked page stat updates happening concurrently. Track
1718 * the task who has the lock for unlock_page_memcg().
1720 memcg
->move_lock_task
= current
;
1721 memcg
->move_lock_flags
= flags
;
1725 EXPORT_SYMBOL(lock_page_memcg
);
1728 * unlock_page_memcg - unlock a page->mem_cgroup binding
1731 void unlock_page_memcg(struct page
*page
)
1733 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
1735 if (memcg
&& memcg
->move_lock_task
== current
) {
1736 unsigned long flags
= memcg
->move_lock_flags
;
1738 memcg
->move_lock_task
= NULL
;
1739 memcg
->move_lock_flags
= 0;
1741 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1746 EXPORT_SYMBOL(unlock_page_memcg
);
1749 * size of first charge trial. "32" comes from vmscan.c's magic value.
1750 * TODO: maybe necessary to use big numbers in big irons.
1752 #define CHARGE_BATCH 32U
1753 struct memcg_stock_pcp
{
1754 struct mem_cgroup
*cached
; /* this never be root cgroup */
1755 unsigned int nr_pages
;
1756 struct work_struct work
;
1757 unsigned long flags
;
1758 #define FLUSHING_CACHED_CHARGE 0
1760 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1761 static DEFINE_MUTEX(percpu_charge_mutex
);
1764 * consume_stock: Try to consume stocked charge on this cpu.
1765 * @memcg: memcg to consume from.
1766 * @nr_pages: how many pages to charge.
1768 * The charges will only happen if @memcg matches the current cpu's memcg
1769 * stock, and at least @nr_pages are available in that stock. Failure to
1770 * service an allocation will refill the stock.
1772 * returns true if successful, false otherwise.
1774 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1776 struct memcg_stock_pcp
*stock
;
1779 if (nr_pages
> CHARGE_BATCH
)
1782 stock
= &get_cpu_var(memcg_stock
);
1783 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
1784 stock
->nr_pages
-= nr_pages
;
1787 put_cpu_var(memcg_stock
);
1792 * Returns stocks cached in percpu and reset cached information.
1794 static void drain_stock(struct memcg_stock_pcp
*stock
)
1796 struct mem_cgroup
*old
= stock
->cached
;
1798 if (stock
->nr_pages
) {
1799 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
1800 if (do_memsw_account())
1801 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
1802 css_put_many(&old
->css
, stock
->nr_pages
);
1803 stock
->nr_pages
= 0;
1805 stock
->cached
= NULL
;
1809 * This must be called under preempt disabled or must be called by
1810 * a thread which is pinned to local cpu.
1812 static void drain_local_stock(struct work_struct
*dummy
)
1814 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
1816 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
1820 * Cache charges(val) to local per_cpu area.
1821 * This will be consumed by consume_stock() function, later.
1823 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1825 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
1827 if (stock
->cached
!= memcg
) { /* reset if necessary */
1829 stock
->cached
= memcg
;
1831 stock
->nr_pages
+= nr_pages
;
1832 put_cpu_var(memcg_stock
);
1836 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1837 * of the hierarchy under it.
1839 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
1843 /* If someone's already draining, avoid adding running more workers. */
1844 if (!mutex_trylock(&percpu_charge_mutex
))
1846 /* Notify other cpus that system-wide "drain" is running */
1849 for_each_online_cpu(cpu
) {
1850 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1851 struct mem_cgroup
*memcg
;
1853 memcg
= stock
->cached
;
1854 if (!memcg
|| !stock
->nr_pages
)
1856 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
1858 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
1860 drain_local_stock(&stock
->work
);
1862 schedule_work_on(cpu
, &stock
->work
);
1867 mutex_unlock(&percpu_charge_mutex
);
1870 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
1871 unsigned long action
,
1874 int cpu
= (unsigned long)hcpu
;
1875 struct memcg_stock_pcp
*stock
;
1877 if (action
== CPU_ONLINE
)
1880 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
1883 stock
= &per_cpu(memcg_stock
, cpu
);
1888 static void reclaim_high(struct mem_cgroup
*memcg
,
1889 unsigned int nr_pages
,
1893 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
1895 mem_cgroup_events(memcg
, MEMCG_HIGH
, 1);
1896 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
1897 } while ((memcg
= parent_mem_cgroup(memcg
)));
1900 static void high_work_func(struct work_struct
*work
)
1902 struct mem_cgroup
*memcg
;
1904 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
1905 reclaim_high(memcg
, CHARGE_BATCH
, GFP_KERNEL
);
1909 * Scheduled by try_charge() to be executed from the userland return path
1910 * and reclaims memory over the high limit.
1912 void mem_cgroup_handle_over_high(void)
1914 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
1915 struct mem_cgroup
*memcg
;
1917 if (likely(!nr_pages
))
1920 memcg
= get_mem_cgroup_from_mm(current
->mm
);
1921 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
1922 css_put(&memcg
->css
);
1923 current
->memcg_nr_pages_over_high
= 0;
1926 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1927 unsigned int nr_pages
)
1929 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
1930 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1931 struct mem_cgroup
*mem_over_limit
;
1932 struct page_counter
*counter
;
1933 unsigned long nr_reclaimed
;
1934 bool may_swap
= true;
1935 bool drained
= false;
1937 if (mem_cgroup_is_root(memcg
))
1940 if (consume_stock(memcg
, nr_pages
))
1943 if (!do_memsw_account() ||
1944 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
1945 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
1947 if (do_memsw_account())
1948 page_counter_uncharge(&memcg
->memsw
, batch
);
1949 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
1951 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
1955 if (batch
> nr_pages
) {
1961 * Unlike in global OOM situations, memcg is not in a physical
1962 * memory shortage. Allow dying and OOM-killed tasks to
1963 * bypass the last charges so that they can exit quickly and
1964 * free their memory.
1966 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
1967 fatal_signal_pending(current
) ||
1968 current
->flags
& PF_EXITING
))
1971 if (unlikely(task_in_memcg_oom(current
)))
1974 if (!gfpflags_allow_blocking(gfp_mask
))
1977 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
1979 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
1980 gfp_mask
, may_swap
);
1982 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
1986 drain_all_stock(mem_over_limit
);
1991 if (gfp_mask
& __GFP_NORETRY
)
1994 * Even though the limit is exceeded at this point, reclaim
1995 * may have been able to free some pages. Retry the charge
1996 * before killing the task.
1998 * Only for regular pages, though: huge pages are rather
1999 * unlikely to succeed so close to the limit, and we fall back
2000 * to regular pages anyway in case of failure.
2002 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2005 * At task move, charge accounts can be doubly counted. So, it's
2006 * better to wait until the end of task_move if something is going on.
2008 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2014 if (gfp_mask
& __GFP_NOFAIL
)
2017 if (fatal_signal_pending(current
))
2020 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
2022 mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2023 get_order(nr_pages
* PAGE_SIZE
));
2025 if (!(gfp_mask
& __GFP_NOFAIL
))
2029 * The allocation either can't fail or will lead to more memory
2030 * being freed very soon. Allow memory usage go over the limit
2031 * temporarily by force charging it.
2033 page_counter_charge(&memcg
->memory
, nr_pages
);
2034 if (do_memsw_account())
2035 page_counter_charge(&memcg
->memsw
, nr_pages
);
2036 css_get_many(&memcg
->css
, nr_pages
);
2041 css_get_many(&memcg
->css
, batch
);
2042 if (batch
> nr_pages
)
2043 refill_stock(memcg
, batch
- nr_pages
);
2046 * If the hierarchy is above the normal consumption range, schedule
2047 * reclaim on returning to userland. We can perform reclaim here
2048 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2049 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2050 * not recorded as it most likely matches current's and won't
2051 * change in the meantime. As high limit is checked again before
2052 * reclaim, the cost of mismatch is negligible.
2055 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2056 /* Don't bother a random interrupted task */
2057 if (in_interrupt()) {
2058 schedule_work(&memcg
->high_work
);
2061 current
->memcg_nr_pages_over_high
+= batch
;
2062 set_notify_resume(current
);
2065 } while ((memcg
= parent_mem_cgroup(memcg
)));
2070 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2072 if (mem_cgroup_is_root(memcg
))
2075 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2076 if (do_memsw_account())
2077 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2079 css_put_many(&memcg
->css
, nr_pages
);
2082 static void lock_page_lru(struct page
*page
, int *isolated
)
2084 struct zone
*zone
= page_zone(page
);
2086 spin_lock_irq(&zone
->lru_lock
);
2087 if (PageLRU(page
)) {
2088 struct lruvec
*lruvec
;
2090 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2092 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2098 static void unlock_page_lru(struct page
*page
, int isolated
)
2100 struct zone
*zone
= page_zone(page
);
2103 struct lruvec
*lruvec
;
2105 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2106 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2108 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2110 spin_unlock_irq(&zone
->lru_lock
);
2113 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2118 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2121 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2122 * may already be on some other mem_cgroup's LRU. Take care of it.
2125 lock_page_lru(page
, &isolated
);
2128 * Nobody should be changing or seriously looking at
2129 * page->mem_cgroup at this point:
2131 * - the page is uncharged
2133 * - the page is off-LRU
2135 * - an anonymous fault has exclusive page access, except for
2136 * a locked page table
2138 * - a page cache insertion, a swapin fault, or a migration
2139 * have the page locked
2141 page
->mem_cgroup
= memcg
;
2144 unlock_page_lru(page
, isolated
);
2148 static int memcg_alloc_cache_id(void)
2153 id
= ida_simple_get(&memcg_cache_ida
,
2154 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2158 if (id
< memcg_nr_cache_ids
)
2162 * There's no space for the new id in memcg_caches arrays,
2163 * so we have to grow them.
2165 down_write(&memcg_cache_ids_sem
);
2167 size
= 2 * (id
+ 1);
2168 if (size
< MEMCG_CACHES_MIN_SIZE
)
2169 size
= MEMCG_CACHES_MIN_SIZE
;
2170 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2171 size
= MEMCG_CACHES_MAX_SIZE
;
2173 err
= memcg_update_all_caches(size
);
2175 err
= memcg_update_all_list_lrus(size
);
2177 memcg_nr_cache_ids
= size
;
2179 up_write(&memcg_cache_ids_sem
);
2182 ida_simple_remove(&memcg_cache_ida
, id
);
2188 static void memcg_free_cache_id(int id
)
2190 ida_simple_remove(&memcg_cache_ida
, id
);
2193 struct memcg_kmem_cache_create_work
{
2194 struct mem_cgroup
*memcg
;
2195 struct kmem_cache
*cachep
;
2196 struct work_struct work
;
2199 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2201 struct memcg_kmem_cache_create_work
*cw
=
2202 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2203 struct mem_cgroup
*memcg
= cw
->memcg
;
2204 struct kmem_cache
*cachep
= cw
->cachep
;
2206 memcg_create_kmem_cache(memcg
, cachep
);
2208 css_put(&memcg
->css
);
2213 * Enqueue the creation of a per-memcg kmem_cache.
2215 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2216 struct kmem_cache
*cachep
)
2218 struct memcg_kmem_cache_create_work
*cw
;
2220 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2224 css_get(&memcg
->css
);
2227 cw
->cachep
= cachep
;
2228 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2230 schedule_work(&cw
->work
);
2233 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2234 struct kmem_cache
*cachep
)
2237 * We need to stop accounting when we kmalloc, because if the
2238 * corresponding kmalloc cache is not yet created, the first allocation
2239 * in __memcg_schedule_kmem_cache_create will recurse.
2241 * However, it is better to enclose the whole function. Depending on
2242 * the debugging options enabled, INIT_WORK(), for instance, can
2243 * trigger an allocation. This too, will make us recurse. Because at
2244 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2245 * the safest choice is to do it like this, wrapping the whole function.
2247 current
->memcg_kmem_skip_account
= 1;
2248 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2249 current
->memcg_kmem_skip_account
= 0;
2253 * Return the kmem_cache we're supposed to use for a slab allocation.
2254 * We try to use the current memcg's version of the cache.
2256 * If the cache does not exist yet, if we are the first user of it,
2257 * we either create it immediately, if possible, or create it asynchronously
2259 * In the latter case, we will let the current allocation go through with
2260 * the original cache.
2262 * Can't be called in interrupt context or from kernel threads.
2263 * This function needs to be called with rcu_read_lock() held.
2265 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2267 struct mem_cgroup
*memcg
;
2268 struct kmem_cache
*memcg_cachep
;
2271 VM_BUG_ON(!is_root_cache(cachep
));
2273 if (cachep
->flags
& SLAB_ACCOUNT
)
2274 gfp
|= __GFP_ACCOUNT
;
2276 if (!(gfp
& __GFP_ACCOUNT
))
2279 if (current
->memcg_kmem_skip_account
)
2282 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2283 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2287 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2288 if (likely(memcg_cachep
))
2289 return memcg_cachep
;
2292 * If we are in a safe context (can wait, and not in interrupt
2293 * context), we could be be predictable and return right away.
2294 * This would guarantee that the allocation being performed
2295 * already belongs in the new cache.
2297 * However, there are some clashes that can arrive from locking.
2298 * For instance, because we acquire the slab_mutex while doing
2299 * memcg_create_kmem_cache, this means no further allocation
2300 * could happen with the slab_mutex held. So it's better to
2303 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2305 css_put(&memcg
->css
);
2309 void __memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2311 if (!is_root_cache(cachep
))
2312 css_put(&cachep
->memcg_params
.memcg
->css
);
2315 int __memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2316 struct mem_cgroup
*memcg
)
2318 unsigned int nr_pages
= 1 << order
;
2319 struct page_counter
*counter
;
2322 ret
= try_charge(memcg
, gfp
, nr_pages
);
2326 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2327 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2328 cancel_charge(memcg
, nr_pages
);
2332 page
->mem_cgroup
= memcg
;
2337 int __memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2339 struct mem_cgroup
*memcg
;
2342 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2343 if (!mem_cgroup_is_root(memcg
))
2344 ret
= __memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2345 css_put(&memcg
->css
);
2349 void __memcg_kmem_uncharge(struct page
*page
, int order
)
2351 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2352 unsigned int nr_pages
= 1 << order
;
2357 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2359 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2360 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2362 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2363 if (do_memsw_account())
2364 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2366 page
->mem_cgroup
= NULL
;
2367 css_put_many(&memcg
->css
, nr_pages
);
2369 #endif /* !CONFIG_SLOB */
2371 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2374 * Because tail pages are not marked as "used", set it. We're under
2375 * zone->lru_lock and migration entries setup in all page mappings.
2377 void mem_cgroup_split_huge_fixup(struct page
*head
)
2381 if (mem_cgroup_disabled())
2384 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2385 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2387 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2390 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2392 #ifdef CONFIG_MEMCG_SWAP
2393 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2396 int val
= (charge
) ? 1 : -1;
2397 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2401 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2402 * @entry: swap entry to be moved
2403 * @from: mem_cgroup which the entry is moved from
2404 * @to: mem_cgroup which the entry is moved to
2406 * It succeeds only when the swap_cgroup's record for this entry is the same
2407 * as the mem_cgroup's id of @from.
2409 * Returns 0 on success, -EINVAL on failure.
2411 * The caller must have charged to @to, IOW, called page_counter_charge() about
2412 * both res and memsw, and called css_get().
2414 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2415 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2417 unsigned short old_id
, new_id
;
2419 old_id
= mem_cgroup_id(from
);
2420 new_id
= mem_cgroup_id(to
);
2422 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2423 mem_cgroup_swap_statistics(from
, false);
2424 mem_cgroup_swap_statistics(to
, true);
2430 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2431 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2437 static DEFINE_MUTEX(memcg_limit_mutex
);
2439 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2440 unsigned long limit
)
2442 unsigned long curusage
;
2443 unsigned long oldusage
;
2444 bool enlarge
= false;
2449 * For keeping hierarchical_reclaim simple, how long we should retry
2450 * is depends on callers. We set our retry-count to be function
2451 * of # of children which we should visit in this loop.
2453 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2454 mem_cgroup_count_children(memcg
);
2456 oldusage
= page_counter_read(&memcg
->memory
);
2459 if (signal_pending(current
)) {
2464 mutex_lock(&memcg_limit_mutex
);
2465 if (limit
> memcg
->memsw
.limit
) {
2466 mutex_unlock(&memcg_limit_mutex
);
2470 if (limit
> memcg
->memory
.limit
)
2472 ret
= page_counter_limit(&memcg
->memory
, limit
);
2473 mutex_unlock(&memcg_limit_mutex
);
2478 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2480 curusage
= page_counter_read(&memcg
->memory
);
2481 /* Usage is reduced ? */
2482 if (curusage
>= oldusage
)
2485 oldusage
= curusage
;
2486 } while (retry_count
);
2488 if (!ret
&& enlarge
)
2489 memcg_oom_recover(memcg
);
2494 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2495 unsigned long limit
)
2497 unsigned long curusage
;
2498 unsigned long oldusage
;
2499 bool enlarge
= false;
2503 /* see mem_cgroup_resize_res_limit */
2504 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2505 mem_cgroup_count_children(memcg
);
2507 oldusage
= page_counter_read(&memcg
->memsw
);
2510 if (signal_pending(current
)) {
2515 mutex_lock(&memcg_limit_mutex
);
2516 if (limit
< memcg
->memory
.limit
) {
2517 mutex_unlock(&memcg_limit_mutex
);
2521 if (limit
> memcg
->memsw
.limit
)
2523 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2524 mutex_unlock(&memcg_limit_mutex
);
2529 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2531 curusage
= page_counter_read(&memcg
->memsw
);
2532 /* Usage is reduced ? */
2533 if (curusage
>= oldusage
)
2536 oldusage
= curusage
;
2537 } while (retry_count
);
2539 if (!ret
&& enlarge
)
2540 memcg_oom_recover(memcg
);
2545 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
2547 unsigned long *total_scanned
)
2549 unsigned long nr_reclaimed
= 0;
2550 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
2551 unsigned long reclaimed
;
2553 struct mem_cgroup_tree_per_zone
*mctz
;
2554 unsigned long excess
;
2555 unsigned long nr_scanned
;
2560 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
2562 * This loop can run a while, specially if mem_cgroup's continuously
2563 * keep exceeding their soft limit and putting the system under
2570 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2575 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
2576 gfp_mask
, &nr_scanned
);
2577 nr_reclaimed
+= reclaimed
;
2578 *total_scanned
+= nr_scanned
;
2579 spin_lock_irq(&mctz
->lock
);
2580 __mem_cgroup_remove_exceeded(mz
, mctz
);
2583 * If we failed to reclaim anything from this memory cgroup
2584 * it is time to move on to the next cgroup
2588 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2590 excess
= soft_limit_excess(mz
->memcg
);
2592 * One school of thought says that we should not add
2593 * back the node to the tree if reclaim returns 0.
2594 * But our reclaim could return 0, simply because due
2595 * to priority we are exposing a smaller subset of
2596 * memory to reclaim from. Consider this as a longer
2599 /* If excess == 0, no tree ops */
2600 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2601 spin_unlock_irq(&mctz
->lock
);
2602 css_put(&mz
->memcg
->css
);
2605 * Could not reclaim anything and there are no more
2606 * mem cgroups to try or we seem to be looping without
2607 * reclaiming anything.
2609 if (!nr_reclaimed
&&
2611 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2613 } while (!nr_reclaimed
);
2615 css_put(&next_mz
->memcg
->css
);
2616 return nr_reclaimed
;
2620 * Test whether @memcg has children, dead or alive. Note that this
2621 * function doesn't care whether @memcg has use_hierarchy enabled and
2622 * returns %true if there are child csses according to the cgroup
2623 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2625 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2630 ret
= css_next_child(NULL
, &memcg
->css
);
2636 * Reclaims as many pages from the given memcg as possible and moves
2637 * the rest to the parent.
2639 * Caller is responsible for holding css reference for memcg.
2641 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2643 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2645 /* we call try-to-free pages for make this cgroup empty */
2646 lru_add_drain_all();
2647 /* try to free all pages in this cgroup */
2648 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2651 if (signal_pending(current
))
2654 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2658 /* maybe some writeback is necessary */
2659 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2667 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
2668 char *buf
, size_t nbytes
,
2671 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2673 if (mem_cgroup_is_root(memcg
))
2675 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
2678 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
2681 return mem_cgroup_from_css(css
)->use_hierarchy
;
2684 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
2685 struct cftype
*cft
, u64 val
)
2688 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2689 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
2691 if (memcg
->use_hierarchy
== val
)
2695 * If parent's use_hierarchy is set, we can't make any modifications
2696 * in the child subtrees. If it is unset, then the change can
2697 * occur, provided the current cgroup has no children.
2699 * For the root cgroup, parent_mem is NULL, we allow value to be
2700 * set if there are no children.
2702 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
2703 (val
== 1 || val
== 0)) {
2704 if (!memcg_has_children(memcg
))
2705 memcg
->use_hierarchy
= val
;
2714 static void tree_stat(struct mem_cgroup
*memcg
, unsigned long *stat
)
2716 struct mem_cgroup
*iter
;
2719 memset(stat
, 0, sizeof(*stat
) * MEMCG_NR_STAT
);
2721 for_each_mem_cgroup_tree(iter
, memcg
) {
2722 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
2723 stat
[i
] += mem_cgroup_read_stat(iter
, i
);
2727 static void tree_events(struct mem_cgroup
*memcg
, unsigned long *events
)
2729 struct mem_cgroup
*iter
;
2732 memset(events
, 0, sizeof(*events
) * MEMCG_NR_EVENTS
);
2734 for_each_mem_cgroup_tree(iter
, memcg
) {
2735 for (i
= 0; i
< MEMCG_NR_EVENTS
; i
++)
2736 events
[i
] += mem_cgroup_read_events(iter
, i
);
2740 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
2742 unsigned long val
= 0;
2744 if (mem_cgroup_is_root(memcg
)) {
2745 struct mem_cgroup
*iter
;
2747 for_each_mem_cgroup_tree(iter
, memcg
) {
2748 val
+= mem_cgroup_read_stat(iter
,
2749 MEM_CGROUP_STAT_CACHE
);
2750 val
+= mem_cgroup_read_stat(iter
,
2751 MEM_CGROUP_STAT_RSS
);
2753 val
+= mem_cgroup_read_stat(iter
,
2754 MEM_CGROUP_STAT_SWAP
);
2758 val
= page_counter_read(&memcg
->memory
);
2760 val
= page_counter_read(&memcg
->memsw
);
2773 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
2776 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2777 struct page_counter
*counter
;
2779 switch (MEMFILE_TYPE(cft
->private)) {
2781 counter
= &memcg
->memory
;
2784 counter
= &memcg
->memsw
;
2787 counter
= &memcg
->kmem
;
2790 counter
= &memcg
->tcpmem
;
2796 switch (MEMFILE_ATTR(cft
->private)) {
2798 if (counter
== &memcg
->memory
)
2799 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
2800 if (counter
== &memcg
->memsw
)
2801 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
2802 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
2804 return (u64
)counter
->limit
* PAGE_SIZE
;
2806 return (u64
)counter
->watermark
* PAGE_SIZE
;
2808 return counter
->failcnt
;
2809 case RES_SOFT_LIMIT
:
2810 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
2817 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2821 if (cgroup_memory_nokmem
)
2824 BUG_ON(memcg
->kmemcg_id
>= 0);
2825 BUG_ON(memcg
->kmem_state
);
2827 memcg_id
= memcg_alloc_cache_id();
2831 static_branch_inc(&memcg_kmem_enabled_key
);
2833 * A memory cgroup is considered kmem-online as soon as it gets
2834 * kmemcg_id. Setting the id after enabling static branching will
2835 * guarantee no one starts accounting before all call sites are
2838 memcg
->kmemcg_id
= memcg_id
;
2839 memcg
->kmem_state
= KMEM_ONLINE
;
2844 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2846 struct cgroup_subsys_state
*css
;
2847 struct mem_cgroup
*parent
, *child
;
2850 if (memcg
->kmem_state
!= KMEM_ONLINE
)
2853 * Clear the online state before clearing memcg_caches array
2854 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2855 * guarantees that no cache will be created for this cgroup
2856 * after we are done (see memcg_create_kmem_cache()).
2858 memcg
->kmem_state
= KMEM_ALLOCATED
;
2860 memcg_deactivate_kmem_caches(memcg
);
2862 kmemcg_id
= memcg
->kmemcg_id
;
2863 BUG_ON(kmemcg_id
< 0);
2865 parent
= parent_mem_cgroup(memcg
);
2867 parent
= root_mem_cgroup
;
2870 * Change kmemcg_id of this cgroup and all its descendants to the
2871 * parent's id, and then move all entries from this cgroup's list_lrus
2872 * to ones of the parent. After we have finished, all list_lrus
2873 * corresponding to this cgroup are guaranteed to remain empty. The
2874 * ordering is imposed by list_lru_node->lock taken by
2875 * memcg_drain_all_list_lrus().
2877 css_for_each_descendant_pre(css
, &memcg
->css
) {
2878 child
= mem_cgroup_from_css(css
);
2879 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
2880 child
->kmemcg_id
= parent
->kmemcg_id
;
2881 if (!memcg
->use_hierarchy
)
2884 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
2886 memcg_free_cache_id(kmemcg_id
);
2889 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2891 /* css_alloc() failed, offlining didn't happen */
2892 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
2893 memcg_offline_kmem(memcg
);
2895 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
2896 memcg_destroy_kmem_caches(memcg
);
2897 static_branch_dec(&memcg_kmem_enabled_key
);
2898 WARN_ON(page_counter_read(&memcg
->kmem
));
2902 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2906 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2909 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2912 #endif /* !CONFIG_SLOB */
2914 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2915 unsigned long limit
)
2919 mutex_lock(&memcg_limit_mutex
);
2920 ret
= page_counter_limit(&memcg
->kmem
, limit
);
2921 mutex_unlock(&memcg_limit_mutex
);
2925 static int memcg_update_tcp_limit(struct mem_cgroup
*memcg
, unsigned long limit
)
2929 mutex_lock(&memcg_limit_mutex
);
2931 ret
= page_counter_limit(&memcg
->tcpmem
, limit
);
2935 if (!memcg
->tcpmem_active
) {
2937 * The active flag needs to be written after the static_key
2938 * update. This is what guarantees that the socket activation
2939 * function is the last one to run. See sock_update_memcg() for
2940 * details, and note that we don't mark any socket as belonging
2941 * to this memcg until that flag is up.
2943 * We need to do this, because static_keys will span multiple
2944 * sites, but we can't control their order. If we mark a socket
2945 * as accounted, but the accounting functions are not patched in
2946 * yet, we'll lose accounting.
2948 * We never race with the readers in sock_update_memcg(),
2949 * because when this value change, the code to process it is not
2952 static_branch_inc(&memcg_sockets_enabled_key
);
2953 memcg
->tcpmem_active
= true;
2956 mutex_unlock(&memcg_limit_mutex
);
2961 * The user of this function is...
2964 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
2965 char *buf
, size_t nbytes
, loff_t off
)
2967 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2968 unsigned long nr_pages
;
2971 buf
= strstrip(buf
);
2972 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
2976 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
2978 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
2982 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
2984 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
2987 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
2990 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
2993 ret
= memcg_update_tcp_limit(memcg
, nr_pages
);
2997 case RES_SOFT_LIMIT
:
2998 memcg
->soft_limit
= nr_pages
;
3002 return ret
?: nbytes
;
3005 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3006 size_t nbytes
, loff_t off
)
3008 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3009 struct page_counter
*counter
;
3011 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3013 counter
= &memcg
->memory
;
3016 counter
= &memcg
->memsw
;
3019 counter
= &memcg
->kmem
;
3022 counter
= &memcg
->tcpmem
;
3028 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3030 page_counter_reset_watermark(counter
);
3033 counter
->failcnt
= 0;
3042 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3045 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3049 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3050 struct cftype
*cft
, u64 val
)
3052 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3054 if (val
& ~MOVE_MASK
)
3058 * No kind of locking is needed in here, because ->can_attach() will
3059 * check this value once in the beginning of the process, and then carry
3060 * on with stale data. This means that changes to this value will only
3061 * affect task migrations starting after the change.
3063 memcg
->move_charge_at_immigrate
= val
;
3067 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3068 struct cftype
*cft
, u64 val
)
3075 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3079 unsigned int lru_mask
;
3082 static const struct numa_stat stats
[] = {
3083 { "total", LRU_ALL
},
3084 { "file", LRU_ALL_FILE
},
3085 { "anon", LRU_ALL_ANON
},
3086 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3088 const struct numa_stat
*stat
;
3091 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3093 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3094 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3095 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3096 for_each_node_state(nid
, N_MEMORY
) {
3097 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3099 seq_printf(m
, " N%d=%lu", nid
, nr
);
3104 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3105 struct mem_cgroup
*iter
;
3108 for_each_mem_cgroup_tree(iter
, memcg
)
3109 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3110 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3111 for_each_node_state(nid
, N_MEMORY
) {
3113 for_each_mem_cgroup_tree(iter
, memcg
)
3114 nr
+= mem_cgroup_node_nr_lru_pages(
3115 iter
, nid
, stat
->lru_mask
);
3116 seq_printf(m
, " N%d=%lu", nid
, nr
);
3123 #endif /* CONFIG_NUMA */
3125 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3127 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3128 unsigned long memory
, memsw
;
3129 struct mem_cgroup
*mi
;
3132 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3133 MEM_CGROUP_STAT_NSTATS
);
3134 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3135 MEM_CGROUP_EVENTS_NSTATS
);
3136 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3138 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3139 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3141 seq_printf(m
, "%s %lu\n", mem_cgroup_stat_names
[i
],
3142 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3145 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3146 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3147 mem_cgroup_read_events(memcg
, i
));
3149 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3150 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3151 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3153 /* Hierarchical information */
3154 memory
= memsw
= PAGE_COUNTER_MAX
;
3155 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3156 memory
= min(memory
, mi
->memory
.limit
);
3157 memsw
= min(memsw
, mi
->memsw
.limit
);
3159 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3160 (u64
)memory
* PAGE_SIZE
);
3161 if (do_memsw_account())
3162 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3163 (u64
)memsw
* PAGE_SIZE
);
3165 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3166 unsigned long long val
= 0;
3168 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3170 for_each_mem_cgroup_tree(mi
, memcg
)
3171 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3172 seq_printf(m
, "total_%s %llu\n", mem_cgroup_stat_names
[i
], val
);
3175 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3176 unsigned long long val
= 0;
3178 for_each_mem_cgroup_tree(mi
, memcg
)
3179 val
+= mem_cgroup_read_events(mi
, i
);
3180 seq_printf(m
, "total_%s %llu\n",
3181 mem_cgroup_events_names
[i
], val
);
3184 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3185 unsigned long long val
= 0;
3187 for_each_mem_cgroup_tree(mi
, memcg
)
3188 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3189 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3192 #ifdef CONFIG_DEBUG_VM
3195 struct mem_cgroup_per_zone
*mz
;
3196 struct zone_reclaim_stat
*rstat
;
3197 unsigned long recent_rotated
[2] = {0, 0};
3198 unsigned long recent_scanned
[2] = {0, 0};
3200 for_each_online_node(nid
)
3201 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3202 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
3203 rstat
= &mz
->lruvec
.reclaim_stat
;
3205 recent_rotated
[0] += rstat
->recent_rotated
[0];
3206 recent_rotated
[1] += rstat
->recent_rotated
[1];
3207 recent_scanned
[0] += rstat
->recent_scanned
[0];
3208 recent_scanned
[1] += rstat
->recent_scanned
[1];
3210 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3211 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3212 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3213 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3220 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3223 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3225 return mem_cgroup_swappiness(memcg
);
3228 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3229 struct cftype
*cft
, u64 val
)
3231 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3237 memcg
->swappiness
= val
;
3239 vm_swappiness
= val
;
3244 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3246 struct mem_cgroup_threshold_ary
*t
;
3247 unsigned long usage
;
3252 t
= rcu_dereference(memcg
->thresholds
.primary
);
3254 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3259 usage
= mem_cgroup_usage(memcg
, swap
);
3262 * current_threshold points to threshold just below or equal to usage.
3263 * If it's not true, a threshold was crossed after last
3264 * call of __mem_cgroup_threshold().
3266 i
= t
->current_threshold
;
3269 * Iterate backward over array of thresholds starting from
3270 * current_threshold and check if a threshold is crossed.
3271 * If none of thresholds below usage is crossed, we read
3272 * only one element of the array here.
3274 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3275 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3277 /* i = current_threshold + 1 */
3281 * Iterate forward over array of thresholds starting from
3282 * current_threshold+1 and check if a threshold is crossed.
3283 * If none of thresholds above usage is crossed, we read
3284 * only one element of the array here.
3286 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3287 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3289 /* Update current_threshold */
3290 t
->current_threshold
= i
- 1;
3295 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3298 __mem_cgroup_threshold(memcg
, false);
3299 if (do_memsw_account())
3300 __mem_cgroup_threshold(memcg
, true);
3302 memcg
= parent_mem_cgroup(memcg
);
3306 static int compare_thresholds(const void *a
, const void *b
)
3308 const struct mem_cgroup_threshold
*_a
= a
;
3309 const struct mem_cgroup_threshold
*_b
= b
;
3311 if (_a
->threshold
> _b
->threshold
)
3314 if (_a
->threshold
< _b
->threshold
)
3320 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3322 struct mem_cgroup_eventfd_list
*ev
;
3324 spin_lock(&memcg_oom_lock
);
3326 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3327 eventfd_signal(ev
->eventfd
, 1);
3329 spin_unlock(&memcg_oom_lock
);
3333 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3335 struct mem_cgroup
*iter
;
3337 for_each_mem_cgroup_tree(iter
, memcg
)
3338 mem_cgroup_oom_notify_cb(iter
);
3341 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3342 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3344 struct mem_cgroup_thresholds
*thresholds
;
3345 struct mem_cgroup_threshold_ary
*new;
3346 unsigned long threshold
;
3347 unsigned long usage
;
3350 ret
= page_counter_memparse(args
, "-1", &threshold
);
3354 mutex_lock(&memcg
->thresholds_lock
);
3357 thresholds
= &memcg
->thresholds
;
3358 usage
= mem_cgroup_usage(memcg
, false);
3359 } else if (type
== _MEMSWAP
) {
3360 thresholds
= &memcg
->memsw_thresholds
;
3361 usage
= mem_cgroup_usage(memcg
, true);
3365 /* Check if a threshold crossed before adding a new one */
3366 if (thresholds
->primary
)
3367 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3369 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3371 /* Allocate memory for new array of thresholds */
3372 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3380 /* Copy thresholds (if any) to new array */
3381 if (thresholds
->primary
) {
3382 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3383 sizeof(struct mem_cgroup_threshold
));
3386 /* Add new threshold */
3387 new->entries
[size
- 1].eventfd
= eventfd
;
3388 new->entries
[size
- 1].threshold
= threshold
;
3390 /* Sort thresholds. Registering of new threshold isn't time-critical */
3391 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3392 compare_thresholds
, NULL
);
3394 /* Find current threshold */
3395 new->current_threshold
= -1;
3396 for (i
= 0; i
< size
; i
++) {
3397 if (new->entries
[i
].threshold
<= usage
) {
3399 * new->current_threshold will not be used until
3400 * rcu_assign_pointer(), so it's safe to increment
3403 ++new->current_threshold
;
3408 /* Free old spare buffer and save old primary buffer as spare */
3409 kfree(thresholds
->spare
);
3410 thresholds
->spare
= thresholds
->primary
;
3412 rcu_assign_pointer(thresholds
->primary
, new);
3414 /* To be sure that nobody uses thresholds */
3418 mutex_unlock(&memcg
->thresholds_lock
);
3423 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3424 struct eventfd_ctx
*eventfd
, const char *args
)
3426 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3429 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3430 struct eventfd_ctx
*eventfd
, const char *args
)
3432 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3435 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3436 struct eventfd_ctx
*eventfd
, enum res_type type
)
3438 struct mem_cgroup_thresholds
*thresholds
;
3439 struct mem_cgroup_threshold_ary
*new;
3440 unsigned long usage
;
3443 mutex_lock(&memcg
->thresholds_lock
);
3446 thresholds
= &memcg
->thresholds
;
3447 usage
= mem_cgroup_usage(memcg
, false);
3448 } else if (type
== _MEMSWAP
) {
3449 thresholds
= &memcg
->memsw_thresholds
;
3450 usage
= mem_cgroup_usage(memcg
, true);
3454 if (!thresholds
->primary
)
3457 /* Check if a threshold crossed before removing */
3458 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3460 /* Calculate new number of threshold */
3462 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3463 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3467 new = thresholds
->spare
;
3469 /* Set thresholds array to NULL if we don't have thresholds */
3478 /* Copy thresholds and find current threshold */
3479 new->current_threshold
= -1;
3480 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3481 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3484 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3485 if (new->entries
[j
].threshold
<= usage
) {
3487 * new->current_threshold will not be used
3488 * until rcu_assign_pointer(), so it's safe to increment
3491 ++new->current_threshold
;
3497 /* Swap primary and spare array */
3498 thresholds
->spare
= thresholds
->primary
;
3500 rcu_assign_pointer(thresholds
->primary
, new);
3502 /* To be sure that nobody uses thresholds */
3505 /* If all events are unregistered, free the spare array */
3507 kfree(thresholds
->spare
);
3508 thresholds
->spare
= NULL
;
3511 mutex_unlock(&memcg
->thresholds_lock
);
3514 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3515 struct eventfd_ctx
*eventfd
)
3517 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3520 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3521 struct eventfd_ctx
*eventfd
)
3523 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3526 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3527 struct eventfd_ctx
*eventfd
, const char *args
)
3529 struct mem_cgroup_eventfd_list
*event
;
3531 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3535 spin_lock(&memcg_oom_lock
);
3537 event
->eventfd
= eventfd
;
3538 list_add(&event
->list
, &memcg
->oom_notify
);
3540 /* already in OOM ? */
3541 if (memcg
->under_oom
)
3542 eventfd_signal(eventfd
, 1);
3543 spin_unlock(&memcg_oom_lock
);
3548 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3549 struct eventfd_ctx
*eventfd
)
3551 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3553 spin_lock(&memcg_oom_lock
);
3555 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3556 if (ev
->eventfd
== eventfd
) {
3557 list_del(&ev
->list
);
3562 spin_unlock(&memcg_oom_lock
);
3565 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3567 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3569 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3570 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3574 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3575 struct cftype
*cft
, u64 val
)
3577 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3579 /* cannot set to root cgroup and only 0 and 1 are allowed */
3580 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3583 memcg
->oom_kill_disable
= val
;
3585 memcg_oom_recover(memcg
);
3590 #ifdef CONFIG_CGROUP_WRITEBACK
3592 struct list_head
*mem_cgroup_cgwb_list(struct mem_cgroup
*memcg
)
3594 return &memcg
->cgwb_list
;
3597 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3599 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3602 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3604 wb_domain_exit(&memcg
->cgwb_domain
);
3607 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3609 wb_domain_size_changed(&memcg
->cgwb_domain
);
3612 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3614 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3616 if (!memcg
->css
.parent
)
3619 return &memcg
->cgwb_domain
;
3623 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3624 * @wb: bdi_writeback in question
3625 * @pfilepages: out parameter for number of file pages
3626 * @pheadroom: out parameter for number of allocatable pages according to memcg
3627 * @pdirty: out parameter for number of dirty pages
3628 * @pwriteback: out parameter for number of pages under writeback
3630 * Determine the numbers of file, headroom, dirty, and writeback pages in
3631 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3632 * is a bit more involved.
3634 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3635 * headroom is calculated as the lowest headroom of itself and the
3636 * ancestors. Note that this doesn't consider the actual amount of
3637 * available memory in the system. The caller should further cap
3638 * *@pheadroom accordingly.
3640 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
3641 unsigned long *pheadroom
, unsigned long *pdirty
,
3642 unsigned long *pwriteback
)
3644 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3645 struct mem_cgroup
*parent
;
3647 *pdirty
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
3649 /* this should eventually include NR_UNSTABLE_NFS */
3650 *pwriteback
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
3651 *pfilepages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
3652 (1 << LRU_ACTIVE_FILE
));
3653 *pheadroom
= PAGE_COUNTER_MAX
;
3655 while ((parent
= parent_mem_cgroup(memcg
))) {
3656 unsigned long ceiling
= min(memcg
->memory
.limit
, memcg
->high
);
3657 unsigned long used
= page_counter_read(&memcg
->memory
);
3659 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
3664 #else /* CONFIG_CGROUP_WRITEBACK */
3666 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3671 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3675 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3679 #endif /* CONFIG_CGROUP_WRITEBACK */
3682 * DO NOT USE IN NEW FILES.
3684 * "cgroup.event_control" implementation.
3686 * This is way over-engineered. It tries to support fully configurable
3687 * events for each user. Such level of flexibility is completely
3688 * unnecessary especially in the light of the planned unified hierarchy.
3690 * Please deprecate this and replace with something simpler if at all
3695 * Unregister event and free resources.
3697 * Gets called from workqueue.
3699 static void memcg_event_remove(struct work_struct
*work
)
3701 struct mem_cgroup_event
*event
=
3702 container_of(work
, struct mem_cgroup_event
, remove
);
3703 struct mem_cgroup
*memcg
= event
->memcg
;
3705 remove_wait_queue(event
->wqh
, &event
->wait
);
3707 event
->unregister_event(memcg
, event
->eventfd
);
3709 /* Notify userspace the event is going away. */
3710 eventfd_signal(event
->eventfd
, 1);
3712 eventfd_ctx_put(event
->eventfd
);
3714 css_put(&memcg
->css
);
3718 * Gets called on POLLHUP on eventfd when user closes it.
3720 * Called with wqh->lock held and interrupts disabled.
3722 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
3723 int sync
, void *key
)
3725 struct mem_cgroup_event
*event
=
3726 container_of(wait
, struct mem_cgroup_event
, wait
);
3727 struct mem_cgroup
*memcg
= event
->memcg
;
3728 unsigned long flags
= (unsigned long)key
;
3730 if (flags
& POLLHUP
) {
3732 * If the event has been detached at cgroup removal, we
3733 * can simply return knowing the other side will cleanup
3736 * We can't race against event freeing since the other
3737 * side will require wqh->lock via remove_wait_queue(),
3740 spin_lock(&memcg
->event_list_lock
);
3741 if (!list_empty(&event
->list
)) {
3742 list_del_init(&event
->list
);
3744 * We are in atomic context, but cgroup_event_remove()
3745 * may sleep, so we have to call it in workqueue.
3747 schedule_work(&event
->remove
);
3749 spin_unlock(&memcg
->event_list_lock
);
3755 static void memcg_event_ptable_queue_proc(struct file
*file
,
3756 wait_queue_head_t
*wqh
, poll_table
*pt
)
3758 struct mem_cgroup_event
*event
=
3759 container_of(pt
, struct mem_cgroup_event
, pt
);
3762 add_wait_queue(wqh
, &event
->wait
);
3766 * DO NOT USE IN NEW FILES.
3768 * Parse input and register new cgroup event handler.
3770 * Input must be in format '<event_fd> <control_fd> <args>'.
3771 * Interpretation of args is defined by control file implementation.
3773 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
3774 char *buf
, size_t nbytes
, loff_t off
)
3776 struct cgroup_subsys_state
*css
= of_css(of
);
3777 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3778 struct mem_cgroup_event
*event
;
3779 struct cgroup_subsys_state
*cfile_css
;
3780 unsigned int efd
, cfd
;
3787 buf
= strstrip(buf
);
3789 efd
= simple_strtoul(buf
, &endp
, 10);
3794 cfd
= simple_strtoul(buf
, &endp
, 10);
3795 if ((*endp
!= ' ') && (*endp
!= '\0'))
3799 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
3803 event
->memcg
= memcg
;
3804 INIT_LIST_HEAD(&event
->list
);
3805 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
3806 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
3807 INIT_WORK(&event
->remove
, memcg_event_remove
);
3815 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
3816 if (IS_ERR(event
->eventfd
)) {
3817 ret
= PTR_ERR(event
->eventfd
);
3824 goto out_put_eventfd
;
3827 /* the process need read permission on control file */
3828 /* AV: shouldn't we check that it's been opened for read instead? */
3829 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
3834 * Determine the event callbacks and set them in @event. This used
3835 * to be done via struct cftype but cgroup core no longer knows
3836 * about these events. The following is crude but the whole thing
3837 * is for compatibility anyway.
3839 * DO NOT ADD NEW FILES.
3841 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
3843 if (!strcmp(name
, "memory.usage_in_bytes")) {
3844 event
->register_event
= mem_cgroup_usage_register_event
;
3845 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
3846 } else if (!strcmp(name
, "memory.oom_control")) {
3847 event
->register_event
= mem_cgroup_oom_register_event
;
3848 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
3849 } else if (!strcmp(name
, "memory.pressure_level")) {
3850 event
->register_event
= vmpressure_register_event
;
3851 event
->unregister_event
= vmpressure_unregister_event
;
3852 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
3853 event
->register_event
= memsw_cgroup_usage_register_event
;
3854 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
3861 * Verify @cfile should belong to @css. Also, remaining events are
3862 * automatically removed on cgroup destruction but the removal is
3863 * asynchronous, so take an extra ref on @css.
3865 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
3866 &memory_cgrp_subsys
);
3868 if (IS_ERR(cfile_css
))
3870 if (cfile_css
!= css
) {
3875 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
3879 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
3881 spin_lock(&memcg
->event_list_lock
);
3882 list_add(&event
->list
, &memcg
->event_list
);
3883 spin_unlock(&memcg
->event_list_lock
);
3895 eventfd_ctx_put(event
->eventfd
);
3904 static struct cftype mem_cgroup_legacy_files
[] = {
3906 .name
= "usage_in_bytes",
3907 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
3908 .read_u64
= mem_cgroup_read_u64
,
3911 .name
= "max_usage_in_bytes",
3912 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
3913 .write
= mem_cgroup_reset
,
3914 .read_u64
= mem_cgroup_read_u64
,
3917 .name
= "limit_in_bytes",
3918 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
3919 .write
= mem_cgroup_write
,
3920 .read_u64
= mem_cgroup_read_u64
,
3923 .name
= "soft_limit_in_bytes",
3924 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
3925 .write
= mem_cgroup_write
,
3926 .read_u64
= mem_cgroup_read_u64
,
3930 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
3931 .write
= mem_cgroup_reset
,
3932 .read_u64
= mem_cgroup_read_u64
,
3936 .seq_show
= memcg_stat_show
,
3939 .name
= "force_empty",
3940 .write
= mem_cgroup_force_empty_write
,
3943 .name
= "use_hierarchy",
3944 .write_u64
= mem_cgroup_hierarchy_write
,
3945 .read_u64
= mem_cgroup_hierarchy_read
,
3948 .name
= "cgroup.event_control", /* XXX: for compat */
3949 .write
= memcg_write_event_control
,
3950 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
3953 .name
= "swappiness",
3954 .read_u64
= mem_cgroup_swappiness_read
,
3955 .write_u64
= mem_cgroup_swappiness_write
,
3958 .name
= "move_charge_at_immigrate",
3959 .read_u64
= mem_cgroup_move_charge_read
,
3960 .write_u64
= mem_cgroup_move_charge_write
,
3963 .name
= "oom_control",
3964 .seq_show
= mem_cgroup_oom_control_read
,
3965 .write_u64
= mem_cgroup_oom_control_write
,
3966 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
3969 .name
= "pressure_level",
3973 .name
= "numa_stat",
3974 .seq_show
= memcg_numa_stat_show
,
3978 .name
= "kmem.limit_in_bytes",
3979 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
3980 .write
= mem_cgroup_write
,
3981 .read_u64
= mem_cgroup_read_u64
,
3984 .name
= "kmem.usage_in_bytes",
3985 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
3986 .read_u64
= mem_cgroup_read_u64
,
3989 .name
= "kmem.failcnt",
3990 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
3991 .write
= mem_cgroup_reset
,
3992 .read_u64
= mem_cgroup_read_u64
,
3995 .name
= "kmem.max_usage_in_bytes",
3996 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
3997 .write
= mem_cgroup_reset
,
3998 .read_u64
= mem_cgroup_read_u64
,
4000 #ifdef CONFIG_SLABINFO
4002 .name
= "kmem.slabinfo",
4003 .seq_start
= slab_start
,
4004 .seq_next
= slab_next
,
4005 .seq_stop
= slab_stop
,
4006 .seq_show
= memcg_slab_show
,
4010 .name
= "kmem.tcp.limit_in_bytes",
4011 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4012 .write
= mem_cgroup_write
,
4013 .read_u64
= mem_cgroup_read_u64
,
4016 .name
= "kmem.tcp.usage_in_bytes",
4017 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4018 .read_u64
= mem_cgroup_read_u64
,
4021 .name
= "kmem.tcp.failcnt",
4022 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4023 .write
= mem_cgroup_reset
,
4024 .read_u64
= mem_cgroup_read_u64
,
4027 .name
= "kmem.tcp.max_usage_in_bytes",
4028 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4029 .write
= mem_cgroup_reset
,
4030 .read_u64
= mem_cgroup_read_u64
,
4032 { }, /* terminate */
4035 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4037 struct mem_cgroup_per_node
*pn
;
4038 struct mem_cgroup_per_zone
*mz
;
4039 int zone
, tmp
= node
;
4041 * This routine is called against possible nodes.
4042 * But it's BUG to call kmalloc() against offline node.
4044 * TODO: this routine can waste much memory for nodes which will
4045 * never be onlined. It's better to use memory hotplug callback
4048 if (!node_state(node
, N_NORMAL_MEMORY
))
4050 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4054 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4055 mz
= &pn
->zoneinfo
[zone
];
4056 lruvec_init(&mz
->lruvec
);
4057 mz
->usage_in_excess
= 0;
4058 mz
->on_tree
= false;
4061 memcg
->nodeinfo
[node
] = pn
;
4065 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4067 kfree(memcg
->nodeinfo
[node
]);
4070 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4074 memcg_wb_domain_exit(memcg
);
4076 free_mem_cgroup_per_zone_info(memcg
, node
);
4077 free_percpu(memcg
->stat
);
4081 static struct mem_cgroup
*mem_cgroup_alloc(void)
4083 struct mem_cgroup
*memcg
;
4087 size
= sizeof(struct mem_cgroup
);
4088 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4090 memcg
= kzalloc(size
, GFP_KERNEL
);
4094 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4099 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
4102 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4105 INIT_WORK(&memcg
->high_work
, high_work_func
);
4106 memcg
->last_scanned_node
= MAX_NUMNODES
;
4107 INIT_LIST_HEAD(&memcg
->oom_notify
);
4108 mutex_init(&memcg
->thresholds_lock
);
4109 spin_lock_init(&memcg
->move_lock
);
4110 vmpressure_init(&memcg
->vmpressure
);
4111 INIT_LIST_HEAD(&memcg
->event_list
);
4112 spin_lock_init(&memcg
->event_list_lock
);
4113 memcg
->socket_pressure
= jiffies
;
4115 memcg
->kmemcg_id
= -1;
4117 #ifdef CONFIG_CGROUP_WRITEBACK
4118 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4122 mem_cgroup_free(memcg
);
4126 static struct cgroup_subsys_state
* __ref
4127 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4129 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
4130 struct mem_cgroup
*memcg
;
4131 long error
= -ENOMEM
;
4133 memcg
= mem_cgroup_alloc();
4135 return ERR_PTR(error
);
4137 memcg
->high
= PAGE_COUNTER_MAX
;
4138 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4140 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4141 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4143 if (parent
&& parent
->use_hierarchy
) {
4144 memcg
->use_hierarchy
= true;
4145 page_counter_init(&memcg
->memory
, &parent
->memory
);
4146 page_counter_init(&memcg
->swap
, &parent
->swap
);
4147 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4148 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4149 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
4151 page_counter_init(&memcg
->memory
, NULL
);
4152 page_counter_init(&memcg
->swap
, NULL
);
4153 page_counter_init(&memcg
->memsw
, NULL
);
4154 page_counter_init(&memcg
->kmem
, NULL
);
4155 page_counter_init(&memcg
->tcpmem
, NULL
);
4157 * Deeper hierachy with use_hierarchy == false doesn't make
4158 * much sense so let cgroup subsystem know about this
4159 * unfortunate state in our controller.
4161 if (parent
!= root_mem_cgroup
)
4162 memory_cgrp_subsys
.broken_hierarchy
= true;
4165 /* The following stuff does not apply to the root */
4167 root_mem_cgroup
= memcg
;
4171 error
= memcg_online_kmem(memcg
);
4175 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4176 static_branch_inc(&memcg_sockets_enabled_key
);
4180 mem_cgroup_free(memcg
);
4185 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4187 if (css
->id
> MEM_CGROUP_ID_MAX
)
4193 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4195 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4196 struct mem_cgroup_event
*event
, *tmp
;
4199 * Unregister events and notify userspace.
4200 * Notify userspace about cgroup removing only after rmdir of cgroup
4201 * directory to avoid race between userspace and kernelspace.
4203 spin_lock(&memcg
->event_list_lock
);
4204 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4205 list_del_init(&event
->list
);
4206 schedule_work(&event
->remove
);
4208 spin_unlock(&memcg
->event_list_lock
);
4210 memcg_offline_kmem(memcg
);
4211 wb_memcg_offline(memcg
);
4214 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4216 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4218 invalidate_reclaim_iterators(memcg
);
4221 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4223 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4225 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4226 static_branch_dec(&memcg_sockets_enabled_key
);
4228 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
4229 static_branch_dec(&memcg_sockets_enabled_key
);
4231 vmpressure_cleanup(&memcg
->vmpressure
);
4232 cancel_work_sync(&memcg
->high_work
);
4233 mem_cgroup_remove_from_trees(memcg
);
4234 memcg_free_kmem(memcg
);
4235 mem_cgroup_free(memcg
);
4239 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4240 * @css: the target css
4242 * Reset the states of the mem_cgroup associated with @css. This is
4243 * invoked when the userland requests disabling on the default hierarchy
4244 * but the memcg is pinned through dependency. The memcg should stop
4245 * applying policies and should revert to the vanilla state as it may be
4246 * made visible again.
4248 * The current implementation only resets the essential configurations.
4249 * This needs to be expanded to cover all the visible parts.
4251 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4253 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4255 page_counter_limit(&memcg
->memory
, PAGE_COUNTER_MAX
);
4256 page_counter_limit(&memcg
->swap
, PAGE_COUNTER_MAX
);
4257 page_counter_limit(&memcg
->memsw
, PAGE_COUNTER_MAX
);
4258 page_counter_limit(&memcg
->kmem
, PAGE_COUNTER_MAX
);
4259 page_counter_limit(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
4261 memcg
->high
= PAGE_COUNTER_MAX
;
4262 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4263 memcg_wb_domain_size_changed(memcg
);
4267 /* Handlers for move charge at task migration. */
4268 static int mem_cgroup_do_precharge(unsigned long count
)
4272 /* Try a single bulk charge without reclaim first, kswapd may wake */
4273 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4275 mc
.precharge
+= count
;
4279 /* Try charges one by one with reclaim */
4281 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4291 * get_mctgt_type - get target type of moving charge
4292 * @vma: the vma the pte to be checked belongs
4293 * @addr: the address corresponding to the pte to be checked
4294 * @ptent: the pte to be checked
4295 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4298 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4299 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4300 * move charge. if @target is not NULL, the page is stored in target->page
4301 * with extra refcnt got(Callers should handle it).
4302 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4303 * target for charge migration. if @target is not NULL, the entry is stored
4306 * Called with pte lock held.
4313 enum mc_target_type
{
4319 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4320 unsigned long addr
, pte_t ptent
)
4322 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4324 if (!page
|| !page_mapped(page
))
4326 if (PageAnon(page
)) {
4327 if (!(mc
.flags
& MOVE_ANON
))
4330 if (!(mc
.flags
& MOVE_FILE
))
4333 if (!get_page_unless_zero(page
))
4340 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4341 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4343 struct page
*page
= NULL
;
4344 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4346 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4349 * Because lookup_swap_cache() updates some statistics counter,
4350 * we call find_get_page() with swapper_space directly.
4352 page
= find_get_page(swap_address_space(ent
), ent
.val
);
4353 if (do_memsw_account())
4354 entry
->val
= ent
.val
;
4359 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4360 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4366 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4367 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4369 struct page
*page
= NULL
;
4370 struct address_space
*mapping
;
4373 if (!vma
->vm_file
) /* anonymous vma */
4375 if (!(mc
.flags
& MOVE_FILE
))
4378 mapping
= vma
->vm_file
->f_mapping
;
4379 pgoff
= linear_page_index(vma
, addr
);
4381 /* page is moved even if it's not RSS of this task(page-faulted). */
4383 /* shmem/tmpfs may report page out on swap: account for that too. */
4384 if (shmem_mapping(mapping
)) {
4385 page
= find_get_entry(mapping
, pgoff
);
4386 if (radix_tree_exceptional_entry(page
)) {
4387 swp_entry_t swp
= radix_to_swp_entry(page
);
4388 if (do_memsw_account())
4390 page
= find_get_page(swap_address_space(swp
), swp
.val
);
4393 page
= find_get_page(mapping
, pgoff
);
4395 page
= find_get_page(mapping
, pgoff
);
4401 * mem_cgroup_move_account - move account of the page
4403 * @nr_pages: number of regular pages (>1 for huge pages)
4404 * @from: mem_cgroup which the page is moved from.
4405 * @to: mem_cgroup which the page is moved to. @from != @to.
4407 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4409 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4412 static int mem_cgroup_move_account(struct page
*page
,
4414 struct mem_cgroup
*from
,
4415 struct mem_cgroup
*to
)
4417 unsigned long flags
;
4418 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
4422 VM_BUG_ON(from
== to
);
4423 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4424 VM_BUG_ON(compound
&& !PageTransHuge(page
));
4427 * Prevent mem_cgroup_migrate() from looking at
4428 * page->mem_cgroup of its source page while we change it.
4431 if (!trylock_page(page
))
4435 if (page
->mem_cgroup
!= from
)
4438 anon
= PageAnon(page
);
4440 spin_lock_irqsave(&from
->move_lock
, flags
);
4442 if (!anon
&& page_mapped(page
)) {
4443 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4445 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4450 * move_lock grabbed above and caller set from->moving_account, so
4451 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4452 * So mapping should be stable for dirty pages.
4454 if (!anon
&& PageDirty(page
)) {
4455 struct address_space
*mapping
= page_mapping(page
);
4457 if (mapping_cap_account_dirty(mapping
)) {
4458 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4460 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4465 if (PageWriteback(page
)) {
4466 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4468 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4473 * It is safe to change page->mem_cgroup here because the page
4474 * is referenced, charged, and isolated - we can't race with
4475 * uncharging, charging, migration, or LRU putback.
4478 /* caller should have done css_get */
4479 page
->mem_cgroup
= to
;
4480 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4484 local_irq_disable();
4485 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
4486 memcg_check_events(to
, page
);
4487 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
4488 memcg_check_events(from
, page
);
4496 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4497 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4499 struct page
*page
= NULL
;
4500 enum mc_target_type ret
= MC_TARGET_NONE
;
4501 swp_entry_t ent
= { .val
= 0 };
4503 if (pte_present(ptent
))
4504 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4505 else if (is_swap_pte(ptent
))
4506 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
4507 else if (pte_none(ptent
))
4508 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4510 if (!page
&& !ent
.val
)
4514 * Do only loose check w/o serialization.
4515 * mem_cgroup_move_account() checks the page is valid or
4516 * not under LRU exclusion.
4518 if (page
->mem_cgroup
== mc
.from
) {
4519 ret
= MC_TARGET_PAGE
;
4521 target
->page
= page
;
4523 if (!ret
|| !target
)
4526 /* There is a swap entry and a page doesn't exist or isn't charged */
4527 if (ent
.val
&& !ret
&&
4528 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4529 ret
= MC_TARGET_SWAP
;
4536 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4538 * We don't consider swapping or file mapped pages because THP does not
4539 * support them for now.
4540 * Caller should make sure that pmd_trans_huge(pmd) is true.
4542 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4543 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4545 struct page
*page
= NULL
;
4546 enum mc_target_type ret
= MC_TARGET_NONE
;
4548 page
= pmd_page(pmd
);
4549 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4550 if (!(mc
.flags
& MOVE_ANON
))
4552 if (page
->mem_cgroup
== mc
.from
) {
4553 ret
= MC_TARGET_PAGE
;
4556 target
->page
= page
;
4562 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4563 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4565 return MC_TARGET_NONE
;
4569 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4570 unsigned long addr
, unsigned long end
,
4571 struct mm_walk
*walk
)
4573 struct vm_area_struct
*vma
= walk
->vma
;
4577 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4579 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4580 mc
.precharge
+= HPAGE_PMD_NR
;
4585 if (pmd_trans_unstable(pmd
))
4587 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4588 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4589 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4590 mc
.precharge
++; /* increment precharge temporarily */
4591 pte_unmap_unlock(pte
- 1, ptl
);
4597 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4599 unsigned long precharge
;
4601 struct mm_walk mem_cgroup_count_precharge_walk
= {
4602 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4605 down_read(&mm
->mmap_sem
);
4606 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk
);
4607 up_read(&mm
->mmap_sem
);
4609 precharge
= mc
.precharge
;
4615 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4617 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4619 VM_BUG_ON(mc
.moving_task
);
4620 mc
.moving_task
= current
;
4621 return mem_cgroup_do_precharge(precharge
);
4624 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4625 static void __mem_cgroup_clear_mc(void)
4627 struct mem_cgroup
*from
= mc
.from
;
4628 struct mem_cgroup
*to
= mc
.to
;
4630 /* we must uncharge all the leftover precharges from mc.to */
4632 cancel_charge(mc
.to
, mc
.precharge
);
4636 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4637 * we must uncharge here.
4639 if (mc
.moved_charge
) {
4640 cancel_charge(mc
.from
, mc
.moved_charge
);
4641 mc
.moved_charge
= 0;
4643 /* we must fixup refcnts and charges */
4644 if (mc
.moved_swap
) {
4645 /* uncharge swap account from the old cgroup */
4646 if (!mem_cgroup_is_root(mc
.from
))
4647 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4650 * we charged both to->memory and to->memsw, so we
4651 * should uncharge to->memory.
4653 if (!mem_cgroup_is_root(mc
.to
))
4654 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4656 css_put_many(&mc
.from
->css
, mc
.moved_swap
);
4658 /* we've already done css_get(mc.to) */
4661 memcg_oom_recover(from
);
4662 memcg_oom_recover(to
);
4663 wake_up_all(&mc
.waitq
);
4666 static void mem_cgroup_clear_mc(void)
4668 struct mm_struct
*mm
= mc
.mm
;
4671 * we must clear moving_task before waking up waiters at the end of
4674 mc
.moving_task
= NULL
;
4675 __mem_cgroup_clear_mc();
4676 spin_lock(&mc
.lock
);
4680 spin_unlock(&mc
.lock
);
4685 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4687 struct cgroup_subsys_state
*css
;
4688 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
4689 struct mem_cgroup
*from
;
4690 struct task_struct
*leader
, *p
;
4691 struct mm_struct
*mm
;
4692 unsigned long move_flags
;
4695 /* charge immigration isn't supported on the default hierarchy */
4696 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4700 * Multi-process migrations only happen on the default hierarchy
4701 * where charge immigration is not used. Perform charge
4702 * immigration if @tset contains a leader and whine if there are
4706 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
4709 memcg
= mem_cgroup_from_css(css
);
4715 * We are now commited to this value whatever it is. Changes in this
4716 * tunable will only affect upcoming migrations, not the current one.
4717 * So we need to save it, and keep it going.
4719 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
4723 from
= mem_cgroup_from_task(p
);
4725 VM_BUG_ON(from
== memcg
);
4727 mm
= get_task_mm(p
);
4730 /* We move charges only when we move a owner of the mm */
4731 if (mm
->owner
== p
) {
4734 VM_BUG_ON(mc
.precharge
);
4735 VM_BUG_ON(mc
.moved_charge
);
4736 VM_BUG_ON(mc
.moved_swap
);
4738 spin_lock(&mc
.lock
);
4742 mc
.flags
= move_flags
;
4743 spin_unlock(&mc
.lock
);
4744 /* We set mc.moving_task later */
4746 ret
= mem_cgroup_precharge_mc(mm
);
4748 mem_cgroup_clear_mc();
4755 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4758 mem_cgroup_clear_mc();
4761 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4762 unsigned long addr
, unsigned long end
,
4763 struct mm_walk
*walk
)
4766 struct vm_area_struct
*vma
= walk
->vma
;
4769 enum mc_target_type target_type
;
4770 union mc_target target
;
4773 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4775 if (mc
.precharge
< HPAGE_PMD_NR
) {
4779 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
4780 if (target_type
== MC_TARGET_PAGE
) {
4782 if (!isolate_lru_page(page
)) {
4783 if (!mem_cgroup_move_account(page
, true,
4785 mc
.precharge
-= HPAGE_PMD_NR
;
4786 mc
.moved_charge
+= HPAGE_PMD_NR
;
4788 putback_lru_page(page
);
4796 if (pmd_trans_unstable(pmd
))
4799 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4800 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4801 pte_t ptent
= *(pte
++);
4807 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
4808 case MC_TARGET_PAGE
:
4811 * We can have a part of the split pmd here. Moving it
4812 * can be done but it would be too convoluted so simply
4813 * ignore such a partial THP and keep it in original
4814 * memcg. There should be somebody mapping the head.
4816 if (PageTransCompound(page
))
4818 if (isolate_lru_page(page
))
4820 if (!mem_cgroup_move_account(page
, false,
4823 /* we uncharge from mc.from later. */
4826 putback_lru_page(page
);
4827 put
: /* get_mctgt_type() gets the page */
4830 case MC_TARGET_SWAP
:
4832 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
4834 /* we fixup refcnts and charges later. */
4842 pte_unmap_unlock(pte
- 1, ptl
);
4847 * We have consumed all precharges we got in can_attach().
4848 * We try charge one by one, but don't do any additional
4849 * charges to mc.to if we have failed in charge once in attach()
4852 ret
= mem_cgroup_do_precharge(1);
4860 static void mem_cgroup_move_charge(void)
4862 struct mm_walk mem_cgroup_move_charge_walk
= {
4863 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4867 lru_add_drain_all();
4869 * Signal lock_page_memcg() to take the memcg's move_lock
4870 * while we're moving its pages to another memcg. Then wait
4871 * for already started RCU-only updates to finish.
4873 atomic_inc(&mc
.from
->moving_account
);
4876 if (unlikely(!down_read_trylock(&mc
.mm
->mmap_sem
))) {
4878 * Someone who are holding the mmap_sem might be waiting in
4879 * waitq. So we cancel all extra charges, wake up all waiters,
4880 * and retry. Because we cancel precharges, we might not be able
4881 * to move enough charges, but moving charge is a best-effort
4882 * feature anyway, so it wouldn't be a big problem.
4884 __mem_cgroup_clear_mc();
4889 * When we have consumed all precharges and failed in doing
4890 * additional charge, the page walk just aborts.
4892 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk
);
4893 up_read(&mc
.mm
->mmap_sem
);
4894 atomic_dec(&mc
.from
->moving_account
);
4897 static void mem_cgroup_move_task(void)
4900 mem_cgroup_move_charge();
4901 mem_cgroup_clear_mc();
4904 #else /* !CONFIG_MMU */
4905 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4909 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4912 static void mem_cgroup_move_task(void)
4918 * Cgroup retains root cgroups across [un]mount cycles making it necessary
4919 * to verify whether we're attached to the default hierarchy on each mount
4922 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
4925 * use_hierarchy is forced on the default hierarchy. cgroup core
4926 * guarantees that @root doesn't have any children, so turning it
4927 * on for the root memcg is enough.
4929 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4930 root_mem_cgroup
->use_hierarchy
= true;
4932 root_mem_cgroup
->use_hierarchy
= false;
4935 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
4938 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4940 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
4943 static int memory_low_show(struct seq_file
*m
, void *v
)
4945 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
4946 unsigned long low
= READ_ONCE(memcg
->low
);
4948 if (low
== PAGE_COUNTER_MAX
)
4949 seq_puts(m
, "max\n");
4951 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
4956 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
4957 char *buf
, size_t nbytes
, loff_t off
)
4959 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
4963 buf
= strstrip(buf
);
4964 err
= page_counter_memparse(buf
, "max", &low
);
4973 static int memory_high_show(struct seq_file
*m
, void *v
)
4975 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
4976 unsigned long high
= READ_ONCE(memcg
->high
);
4978 if (high
== PAGE_COUNTER_MAX
)
4979 seq_puts(m
, "max\n");
4981 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
4986 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
4987 char *buf
, size_t nbytes
, loff_t off
)
4989 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
4990 unsigned long nr_pages
;
4994 buf
= strstrip(buf
);
4995 err
= page_counter_memparse(buf
, "max", &high
);
5001 nr_pages
= page_counter_read(&memcg
->memory
);
5002 if (nr_pages
> high
)
5003 try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
5006 memcg_wb_domain_size_changed(memcg
);
5010 static int memory_max_show(struct seq_file
*m
, void *v
)
5012 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5013 unsigned long max
= READ_ONCE(memcg
->memory
.limit
);
5015 if (max
== PAGE_COUNTER_MAX
)
5016 seq_puts(m
, "max\n");
5018 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5023 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5024 char *buf
, size_t nbytes
, loff_t off
)
5026 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5027 unsigned int nr_reclaims
= MEM_CGROUP_RECLAIM_RETRIES
;
5028 bool drained
= false;
5032 buf
= strstrip(buf
);
5033 err
= page_counter_memparse(buf
, "max", &max
);
5037 xchg(&memcg
->memory
.limit
, max
);
5040 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
5042 if (nr_pages
<= max
)
5045 if (signal_pending(current
)) {
5051 drain_all_stock(memcg
);
5057 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
5063 mem_cgroup_events(memcg
, MEMCG_OOM
, 1);
5064 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
5068 memcg_wb_domain_size_changed(memcg
);
5072 static int memory_events_show(struct seq_file
*m
, void *v
)
5074 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5076 seq_printf(m
, "low %lu\n", mem_cgroup_read_events(memcg
, MEMCG_LOW
));
5077 seq_printf(m
, "high %lu\n", mem_cgroup_read_events(memcg
, MEMCG_HIGH
));
5078 seq_printf(m
, "max %lu\n", mem_cgroup_read_events(memcg
, MEMCG_MAX
));
5079 seq_printf(m
, "oom %lu\n", mem_cgroup_read_events(memcg
, MEMCG_OOM
));
5084 static int memory_stat_show(struct seq_file
*m
, void *v
)
5086 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5087 unsigned long stat
[MEMCG_NR_STAT
];
5088 unsigned long events
[MEMCG_NR_EVENTS
];
5092 * Provide statistics on the state of the memory subsystem as
5093 * well as cumulative event counters that show past behavior.
5095 * This list is ordered following a combination of these gradients:
5096 * 1) generic big picture -> specifics and details
5097 * 2) reflecting userspace activity -> reflecting kernel heuristics
5099 * Current memory state:
5102 tree_stat(memcg
, stat
);
5103 tree_events(memcg
, events
);
5105 seq_printf(m
, "anon %llu\n",
5106 (u64
)stat
[MEM_CGROUP_STAT_RSS
] * PAGE_SIZE
);
5107 seq_printf(m
, "file %llu\n",
5108 (u64
)stat
[MEM_CGROUP_STAT_CACHE
] * PAGE_SIZE
);
5109 seq_printf(m
, "kernel_stack %llu\n",
5110 (u64
)stat
[MEMCG_KERNEL_STACK
] * PAGE_SIZE
);
5111 seq_printf(m
, "slab %llu\n",
5112 (u64
)(stat
[MEMCG_SLAB_RECLAIMABLE
] +
5113 stat
[MEMCG_SLAB_UNRECLAIMABLE
]) * PAGE_SIZE
);
5114 seq_printf(m
, "sock %llu\n",
5115 (u64
)stat
[MEMCG_SOCK
] * PAGE_SIZE
);
5117 seq_printf(m
, "file_mapped %llu\n",
5118 (u64
)stat
[MEM_CGROUP_STAT_FILE_MAPPED
] * PAGE_SIZE
);
5119 seq_printf(m
, "file_dirty %llu\n",
5120 (u64
)stat
[MEM_CGROUP_STAT_DIRTY
] * PAGE_SIZE
);
5121 seq_printf(m
, "file_writeback %llu\n",
5122 (u64
)stat
[MEM_CGROUP_STAT_WRITEBACK
] * PAGE_SIZE
);
5124 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5125 struct mem_cgroup
*mi
;
5126 unsigned long val
= 0;
5128 for_each_mem_cgroup_tree(mi
, memcg
)
5129 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
));
5130 seq_printf(m
, "%s %llu\n",
5131 mem_cgroup_lru_names
[i
], (u64
)val
* PAGE_SIZE
);
5134 seq_printf(m
, "slab_reclaimable %llu\n",
5135 (u64
)stat
[MEMCG_SLAB_RECLAIMABLE
] * PAGE_SIZE
);
5136 seq_printf(m
, "slab_unreclaimable %llu\n",
5137 (u64
)stat
[MEMCG_SLAB_UNRECLAIMABLE
] * PAGE_SIZE
);
5139 /* Accumulated memory events */
5141 seq_printf(m
, "pgfault %lu\n",
5142 events
[MEM_CGROUP_EVENTS_PGFAULT
]);
5143 seq_printf(m
, "pgmajfault %lu\n",
5144 events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
5149 static struct cftype memory_files
[] = {
5152 .flags
= CFTYPE_NOT_ON_ROOT
,
5153 .read_u64
= memory_current_read
,
5157 .flags
= CFTYPE_NOT_ON_ROOT
,
5158 .seq_show
= memory_low_show
,
5159 .write
= memory_low_write
,
5163 .flags
= CFTYPE_NOT_ON_ROOT
,
5164 .seq_show
= memory_high_show
,
5165 .write
= memory_high_write
,
5169 .flags
= CFTYPE_NOT_ON_ROOT
,
5170 .seq_show
= memory_max_show
,
5171 .write
= memory_max_write
,
5175 .flags
= CFTYPE_NOT_ON_ROOT
,
5176 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
5177 .seq_show
= memory_events_show
,
5181 .flags
= CFTYPE_NOT_ON_ROOT
,
5182 .seq_show
= memory_stat_show
,
5187 struct cgroup_subsys memory_cgrp_subsys
= {
5188 .css_alloc
= mem_cgroup_css_alloc
,
5189 .css_online
= mem_cgroup_css_online
,
5190 .css_offline
= mem_cgroup_css_offline
,
5191 .css_released
= mem_cgroup_css_released
,
5192 .css_free
= mem_cgroup_css_free
,
5193 .css_reset
= mem_cgroup_css_reset
,
5194 .can_attach
= mem_cgroup_can_attach
,
5195 .cancel_attach
= mem_cgroup_cancel_attach
,
5196 .post_attach
= mem_cgroup_move_task
,
5197 .bind
= mem_cgroup_bind
,
5198 .dfl_cftypes
= memory_files
,
5199 .legacy_cftypes
= mem_cgroup_legacy_files
,
5204 * mem_cgroup_low - check if memory consumption is below the normal range
5205 * @root: the highest ancestor to consider
5206 * @memcg: the memory cgroup to check
5208 * Returns %true if memory consumption of @memcg, and that of all
5209 * configurable ancestors up to @root, is below the normal range.
5211 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5213 if (mem_cgroup_disabled())
5217 * The toplevel group doesn't have a configurable range, so
5218 * it's never low when looked at directly, and it is not
5219 * considered an ancestor when assessing the hierarchy.
5222 if (memcg
== root_mem_cgroup
)
5225 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5228 while (memcg
!= root
) {
5229 memcg
= parent_mem_cgroup(memcg
);
5231 if (memcg
== root_mem_cgroup
)
5234 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5241 * mem_cgroup_try_charge - try charging a page
5242 * @page: page to charge
5243 * @mm: mm context of the victim
5244 * @gfp_mask: reclaim mode
5245 * @memcgp: charged memcg return
5247 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5248 * pages according to @gfp_mask if necessary.
5250 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5251 * Otherwise, an error code is returned.
5253 * After page->mapping has been set up, the caller must finalize the
5254 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5255 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5257 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5258 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
5261 struct mem_cgroup
*memcg
= NULL
;
5262 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5265 if (mem_cgroup_disabled())
5268 if (PageSwapCache(page
)) {
5270 * Every swap fault against a single page tries to charge the
5271 * page, bail as early as possible. shmem_unuse() encounters
5272 * already charged pages, too. The USED bit is protected by
5273 * the page lock, which serializes swap cache removal, which
5274 * in turn serializes uncharging.
5276 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5277 if (page
->mem_cgroup
)
5280 if (do_swap_account
) {
5281 swp_entry_t ent
= { .val
= page_private(page
), };
5282 unsigned short id
= lookup_swap_cgroup_id(ent
);
5285 memcg
= mem_cgroup_from_id(id
);
5286 if (memcg
&& !css_tryget_online(&memcg
->css
))
5293 memcg
= get_mem_cgroup_from_mm(mm
);
5295 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5297 css_put(&memcg
->css
);
5304 * mem_cgroup_commit_charge - commit a page charge
5305 * @page: page to charge
5306 * @memcg: memcg to charge the page to
5307 * @lrucare: page might be on LRU already
5309 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5310 * after page->mapping has been set up. This must happen atomically
5311 * as part of the page instantiation, i.e. under the page table lock
5312 * for anonymous pages, under the page lock for page and swap cache.
5314 * In addition, the page must not be on the LRU during the commit, to
5315 * prevent racing with task migration. If it might be, use @lrucare.
5317 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5319 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5320 bool lrucare
, bool compound
)
5322 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5324 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5325 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5327 if (mem_cgroup_disabled())
5330 * Swap faults will attempt to charge the same page multiple
5331 * times. But reuse_swap_page() might have removed the page
5332 * from swapcache already, so we can't check PageSwapCache().
5337 commit_charge(page
, memcg
, lrucare
);
5339 local_irq_disable();
5340 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
5341 memcg_check_events(memcg
, page
);
5344 if (do_memsw_account() && PageSwapCache(page
)) {
5345 swp_entry_t entry
= { .val
= page_private(page
) };
5347 * The swap entry might not get freed for a long time,
5348 * let's not wait for it. The page already received a
5349 * memory+swap charge, drop the swap entry duplicate.
5351 mem_cgroup_uncharge_swap(entry
);
5356 * mem_cgroup_cancel_charge - cancel a page charge
5357 * @page: page to charge
5358 * @memcg: memcg to charge the page to
5360 * Cancel a charge transaction started by mem_cgroup_try_charge().
5362 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5365 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5367 if (mem_cgroup_disabled())
5370 * Swap faults will attempt to charge the same page multiple
5371 * times. But reuse_swap_page() might have removed the page
5372 * from swapcache already, so we can't check PageSwapCache().
5377 cancel_charge(memcg
, nr_pages
);
5380 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5381 unsigned long nr_anon
, unsigned long nr_file
,
5382 unsigned long nr_huge
, struct page
*dummy_page
)
5384 unsigned long nr_pages
= nr_anon
+ nr_file
;
5385 unsigned long flags
;
5387 if (!mem_cgroup_is_root(memcg
)) {
5388 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5389 if (do_memsw_account())
5390 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5391 memcg_oom_recover(memcg
);
5394 local_irq_save(flags
);
5395 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5396 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5397 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5398 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5399 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5400 memcg_check_events(memcg
, dummy_page
);
5401 local_irq_restore(flags
);
5403 if (!mem_cgroup_is_root(memcg
))
5404 css_put_many(&memcg
->css
, nr_pages
);
5407 static void uncharge_list(struct list_head
*page_list
)
5409 struct mem_cgroup
*memcg
= NULL
;
5410 unsigned long nr_anon
= 0;
5411 unsigned long nr_file
= 0;
5412 unsigned long nr_huge
= 0;
5413 unsigned long pgpgout
= 0;
5414 struct list_head
*next
;
5418 * Note that the list can be a single page->lru; hence the
5419 * do-while loop instead of a simple list_for_each_entry().
5421 next
= page_list
->next
;
5423 unsigned int nr_pages
= 1;
5425 page
= list_entry(next
, struct page
, lru
);
5426 next
= page
->lru
.next
;
5428 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5429 VM_BUG_ON_PAGE(page_count(page
), page
);
5431 if (!page
->mem_cgroup
)
5435 * Nobody should be changing or seriously looking at
5436 * page->mem_cgroup at this point, we have fully
5437 * exclusive access to the page.
5440 if (memcg
!= page
->mem_cgroup
) {
5442 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5444 pgpgout
= nr_anon
= nr_file
= nr_huge
= 0;
5446 memcg
= page
->mem_cgroup
;
5449 if (PageTransHuge(page
)) {
5450 nr_pages
<<= compound_order(page
);
5451 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5452 nr_huge
+= nr_pages
;
5456 nr_anon
+= nr_pages
;
5458 nr_file
+= nr_pages
;
5460 page
->mem_cgroup
= NULL
;
5463 } while (next
!= page_list
);
5466 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5471 * mem_cgroup_uncharge - uncharge a page
5472 * @page: page to uncharge
5474 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5475 * mem_cgroup_commit_charge().
5477 void mem_cgroup_uncharge(struct page
*page
)
5479 if (mem_cgroup_disabled())
5482 /* Don't touch page->lru of any random page, pre-check: */
5483 if (!page
->mem_cgroup
)
5486 INIT_LIST_HEAD(&page
->lru
);
5487 uncharge_list(&page
->lru
);
5491 * mem_cgroup_uncharge_list - uncharge a list of page
5492 * @page_list: list of pages to uncharge
5494 * Uncharge a list of pages previously charged with
5495 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5497 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5499 if (mem_cgroup_disabled())
5502 if (!list_empty(page_list
))
5503 uncharge_list(page_list
);
5507 * mem_cgroup_migrate - charge a page's replacement
5508 * @oldpage: currently circulating page
5509 * @newpage: replacement page
5511 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5512 * be uncharged upon free.
5514 * Both pages must be locked, @newpage->mapping must be set up.
5516 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
5518 struct mem_cgroup
*memcg
;
5519 unsigned int nr_pages
;
5522 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5523 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5524 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5525 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5528 if (mem_cgroup_disabled())
5531 /* Page cache replacement: new page already charged? */
5532 if (newpage
->mem_cgroup
)
5535 /* Swapcache readahead pages can get replaced before being charged */
5536 memcg
= oldpage
->mem_cgroup
;
5540 /* Force-charge the new page. The old one will be freed soon */
5541 compound
= PageTransHuge(newpage
);
5542 nr_pages
= compound
? hpage_nr_pages(newpage
) : 1;
5544 page_counter_charge(&memcg
->memory
, nr_pages
);
5545 if (do_memsw_account())
5546 page_counter_charge(&memcg
->memsw
, nr_pages
);
5547 css_get_many(&memcg
->css
, nr_pages
);
5549 commit_charge(newpage
, memcg
, false);
5551 local_irq_disable();
5552 mem_cgroup_charge_statistics(memcg
, newpage
, compound
, nr_pages
);
5553 memcg_check_events(memcg
, newpage
);
5557 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
5558 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
5560 void sock_update_memcg(struct sock
*sk
)
5562 struct mem_cgroup
*memcg
;
5564 /* Socket cloning can throw us here with sk_cgrp already
5565 * filled. It won't however, necessarily happen from
5566 * process context. So the test for root memcg given
5567 * the current task's memcg won't help us in this case.
5569 * Respecting the original socket's memcg is a better
5570 * decision in this case.
5573 BUG_ON(mem_cgroup_is_root(sk
->sk_memcg
));
5574 css_get(&sk
->sk_memcg
->css
);
5579 memcg
= mem_cgroup_from_task(current
);
5580 if (memcg
== root_mem_cgroup
)
5582 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
5584 if (css_tryget_online(&memcg
->css
))
5585 sk
->sk_memcg
= memcg
;
5589 EXPORT_SYMBOL(sock_update_memcg
);
5591 void sock_release_memcg(struct sock
*sk
)
5593 WARN_ON(!sk
->sk_memcg
);
5594 css_put(&sk
->sk_memcg
->css
);
5598 * mem_cgroup_charge_skmem - charge socket memory
5599 * @memcg: memcg to charge
5600 * @nr_pages: number of pages to charge
5602 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5603 * @memcg's configured limit, %false if the charge had to be forced.
5605 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5607 gfp_t gfp_mask
= GFP_KERNEL
;
5609 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5610 struct page_counter
*fail
;
5612 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
5613 memcg
->tcpmem_pressure
= 0;
5616 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
5617 memcg
->tcpmem_pressure
= 1;
5621 /* Don't block in the packet receive path */
5623 gfp_mask
= GFP_NOWAIT
;
5625 this_cpu_add(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5627 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
5630 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
5635 * mem_cgroup_uncharge_skmem - uncharge socket memory
5636 * @memcg - memcg to uncharge
5637 * @nr_pages - number of pages to uncharge
5639 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5641 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5642 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
5646 this_cpu_sub(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5648 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5649 css_put_many(&memcg
->css
, nr_pages
);
5652 static int __init
cgroup_memory(char *s
)
5656 while ((token
= strsep(&s
, ",")) != NULL
) {
5659 if (!strcmp(token
, "nosocket"))
5660 cgroup_memory_nosocket
= true;
5661 if (!strcmp(token
, "nokmem"))
5662 cgroup_memory_nokmem
= true;
5666 __setup("cgroup.memory=", cgroup_memory
);
5669 * subsys_initcall() for memory controller.
5671 * Some parts like hotcpu_notifier() have to be initialized from this context
5672 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5673 * everything that doesn't depend on a specific mem_cgroup structure should
5674 * be initialized from here.
5676 static int __init
mem_cgroup_init(void)
5680 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5682 for_each_possible_cpu(cpu
)
5683 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5686 for_each_node(node
) {
5687 struct mem_cgroup_tree_per_node
*rtpn
;
5690 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5691 node_online(node
) ? node
: NUMA_NO_NODE
);
5693 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5694 struct mem_cgroup_tree_per_zone
*rtpz
;
5696 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
5697 rtpz
->rb_root
= RB_ROOT
;
5698 spin_lock_init(&rtpz
->lock
);
5700 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5705 subsys_initcall(mem_cgroup_init
);
5707 #ifdef CONFIG_MEMCG_SWAP
5709 * mem_cgroup_swapout - transfer a memsw charge to swap
5710 * @page: page whose memsw charge to transfer
5711 * @entry: swap entry to move the charge to
5713 * Transfer the memsw charge of @page to @entry.
5715 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5717 struct mem_cgroup
*memcg
;
5718 unsigned short oldid
;
5720 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5721 VM_BUG_ON_PAGE(page_count(page
), page
);
5723 if (!do_memsw_account())
5726 memcg
= page
->mem_cgroup
;
5728 /* Readahead page, never charged */
5732 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5733 VM_BUG_ON_PAGE(oldid
, page
);
5734 mem_cgroup_swap_statistics(memcg
, true);
5736 page
->mem_cgroup
= NULL
;
5738 if (!mem_cgroup_is_root(memcg
))
5739 page_counter_uncharge(&memcg
->memory
, 1);
5742 * Interrupts should be disabled here because the caller holds the
5743 * mapping->tree_lock lock which is taken with interrupts-off. It is
5744 * important here to have the interrupts disabled because it is the
5745 * only synchronisation we have for udpating the per-CPU variables.
5747 VM_BUG_ON(!irqs_disabled());
5748 mem_cgroup_charge_statistics(memcg
, page
, false, -1);
5749 memcg_check_events(memcg
, page
);
5753 * mem_cgroup_try_charge_swap - try charging a swap entry
5754 * @page: page being added to swap
5755 * @entry: swap entry to charge
5757 * Try to charge @entry to the memcg that @page belongs to.
5759 * Returns 0 on success, -ENOMEM on failure.
5761 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
5763 struct mem_cgroup
*memcg
;
5764 struct page_counter
*counter
;
5765 unsigned short oldid
;
5767 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
5770 memcg
= page
->mem_cgroup
;
5772 /* Readahead page, never charged */
5776 if (!mem_cgroup_is_root(memcg
) &&
5777 !page_counter_try_charge(&memcg
->swap
, 1, &counter
))
5780 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5781 VM_BUG_ON_PAGE(oldid
, page
);
5782 mem_cgroup_swap_statistics(memcg
, true);
5784 css_get(&memcg
->css
);
5789 * mem_cgroup_uncharge_swap - uncharge a swap entry
5790 * @entry: swap entry to uncharge
5792 * Drop the swap charge associated with @entry.
5794 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5796 struct mem_cgroup
*memcg
;
5799 if (!do_swap_account
)
5802 id
= swap_cgroup_record(entry
, 0);
5804 memcg
= mem_cgroup_from_id(id
);
5806 if (!mem_cgroup_is_root(memcg
)) {
5807 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5808 page_counter_uncharge(&memcg
->swap
, 1);
5810 page_counter_uncharge(&memcg
->memsw
, 1);
5812 mem_cgroup_swap_statistics(memcg
, false);
5813 css_put(&memcg
->css
);
5818 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
5820 long nr_swap_pages
= get_nr_swap_pages();
5822 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5823 return nr_swap_pages
;
5824 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5825 nr_swap_pages
= min_t(long, nr_swap_pages
,
5826 READ_ONCE(memcg
->swap
.limit
) -
5827 page_counter_read(&memcg
->swap
));
5828 return nr_swap_pages
;
5831 bool mem_cgroup_swap_full(struct page
*page
)
5833 struct mem_cgroup
*memcg
;
5835 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5839 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5842 memcg
= page
->mem_cgroup
;
5846 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5847 if (page_counter_read(&memcg
->swap
) * 2 >= memcg
->swap
.limit
)
5853 /* for remember boot option*/
5854 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5855 static int really_do_swap_account __initdata
= 1;
5857 static int really_do_swap_account __initdata
;
5860 static int __init
enable_swap_account(char *s
)
5862 if (!strcmp(s
, "1"))
5863 really_do_swap_account
= 1;
5864 else if (!strcmp(s
, "0"))
5865 really_do_swap_account
= 0;
5868 __setup("swapaccount=", enable_swap_account
);
5870 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
5873 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5875 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
5878 static int swap_max_show(struct seq_file
*m
, void *v
)
5880 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5881 unsigned long max
= READ_ONCE(memcg
->swap
.limit
);
5883 if (max
== PAGE_COUNTER_MAX
)
5884 seq_puts(m
, "max\n");
5886 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5891 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
5892 char *buf
, size_t nbytes
, loff_t off
)
5894 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5898 buf
= strstrip(buf
);
5899 err
= page_counter_memparse(buf
, "max", &max
);
5903 mutex_lock(&memcg_limit_mutex
);
5904 err
= page_counter_limit(&memcg
->swap
, max
);
5905 mutex_unlock(&memcg_limit_mutex
);
5912 static struct cftype swap_files
[] = {
5914 .name
= "swap.current",
5915 .flags
= CFTYPE_NOT_ON_ROOT
,
5916 .read_u64
= swap_current_read
,
5920 .flags
= CFTYPE_NOT_ON_ROOT
,
5921 .seq_show
= swap_max_show
,
5922 .write
= swap_max_write
,
5927 static struct cftype memsw_cgroup_files
[] = {
5929 .name
= "memsw.usage_in_bytes",
5930 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5931 .read_u64
= mem_cgroup_read_u64
,
5934 .name
= "memsw.max_usage_in_bytes",
5935 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5936 .write
= mem_cgroup_reset
,
5937 .read_u64
= mem_cgroup_read_u64
,
5940 .name
= "memsw.limit_in_bytes",
5941 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5942 .write
= mem_cgroup_write
,
5943 .read_u64
= mem_cgroup_read_u64
,
5946 .name
= "memsw.failcnt",
5947 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5948 .write
= mem_cgroup_reset
,
5949 .read_u64
= mem_cgroup_read_u64
,
5951 { }, /* terminate */
5954 static int __init
mem_cgroup_swap_init(void)
5956 if (!mem_cgroup_disabled() && really_do_swap_account
) {
5957 do_swap_account
= 1;
5958 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
5960 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
5961 memsw_cgroup_files
));
5965 subsys_initcall(mem_cgroup_swap_init
);
5967 #endif /* CONFIG_MEMCG_SWAP */