1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
65 #include <linux/tracehook.h>
71 #include <asm/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
76 EXPORT_SYMBOL(memory_cgrp_subsys
);
78 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
80 #define MEM_CGROUP_RECLAIM_RETRIES 5
82 /* Socket memory accounting disabled? */
83 static bool cgroup_memory_nosocket
;
85 /* Kernel memory accounting disabled? */
86 static bool cgroup_memory_nokmem
;
88 /* Whether the swap controller is active */
89 #ifdef CONFIG_MEMCG_SWAP
90 int do_swap_account __read_mostly
;
92 #define do_swap_account 0
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys
) && do_swap_account
;
101 static const char * const mem_cgroup_stat_names
[] = {
111 static const char * const mem_cgroup_events_names
[] = {
118 static const char * const mem_cgroup_lru_names
[] = {
126 #define THRESHOLDS_EVENTS_TARGET 128
127 #define SOFTLIMIT_EVENTS_TARGET 1024
128 #define NUMAINFO_EVENTS_TARGET 1024
131 * Cgroups above their limits are maintained in a RB-Tree, independent of
132 * their hierarchy representation
135 struct mem_cgroup_tree_per_zone
{
136 struct rb_root rb_root
;
140 struct mem_cgroup_tree_per_node
{
141 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
144 struct mem_cgroup_tree
{
145 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
148 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
151 struct mem_cgroup_eventfd_list
{
152 struct list_head list
;
153 struct eventfd_ctx
*eventfd
;
157 * cgroup_event represents events which userspace want to receive.
159 struct mem_cgroup_event
{
161 * memcg which the event belongs to.
163 struct mem_cgroup
*memcg
;
165 * eventfd to signal userspace about the event.
167 struct eventfd_ctx
*eventfd
;
169 * Each of these stored in a list by the cgroup.
171 struct list_head list
;
173 * register_event() callback will be used to add new userspace
174 * waiter for changes related to this event. Use eventfd_signal()
175 * on eventfd to send notification to userspace.
177 int (*register_event
)(struct mem_cgroup
*memcg
,
178 struct eventfd_ctx
*eventfd
, const char *args
);
180 * unregister_event() callback will be called when userspace closes
181 * the eventfd or on cgroup removing. This callback must be set,
182 * if you want provide notification functionality.
184 void (*unregister_event
)(struct mem_cgroup
*memcg
,
185 struct eventfd_ctx
*eventfd
);
187 * All fields below needed to unregister event when
188 * userspace closes eventfd.
191 wait_queue_head_t
*wqh
;
193 struct work_struct remove
;
196 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
197 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
199 /* Stuffs for move charges at task migration. */
201 * Types of charges to be moved.
203 #define MOVE_ANON 0x1U
204 #define MOVE_FILE 0x2U
205 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
207 /* "mc" and its members are protected by cgroup_mutex */
208 static struct move_charge_struct
{
209 spinlock_t lock
; /* for from, to */
210 struct mm_struct
*mm
;
211 struct mem_cgroup
*from
;
212 struct mem_cgroup
*to
;
214 unsigned long precharge
;
215 unsigned long moved_charge
;
216 unsigned long moved_swap
;
217 struct task_struct
*moving_task
; /* a task moving charges */
218 wait_queue_head_t waitq
; /* a waitq for other context */
220 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
221 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
225 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
226 * limit reclaim to prevent infinite loops, if they ever occur.
228 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
229 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
232 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
233 MEM_CGROUP_CHARGE_TYPE_ANON
,
234 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
235 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
239 /* for encoding cft->private value on file */
248 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
249 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
250 #define MEMFILE_ATTR(val) ((val) & 0xffff)
251 /* Used for OOM nofiier */
252 #define OOM_CONTROL (0)
254 /* Some nice accessors for the vmpressure. */
255 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
258 memcg
= root_mem_cgroup
;
259 return &memcg
->vmpressure
;
262 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
264 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
267 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
269 return (memcg
== root_mem_cgroup
);
274 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
275 * The main reason for not using cgroup id for this:
276 * this works better in sparse environments, where we have a lot of memcgs,
277 * but only a few kmem-limited. Or also, if we have, for instance, 200
278 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
279 * 200 entry array for that.
281 * The current size of the caches array is stored in memcg_nr_cache_ids. It
282 * will double each time we have to increase it.
284 static DEFINE_IDA(memcg_cache_ida
);
285 int memcg_nr_cache_ids
;
287 /* Protects memcg_nr_cache_ids */
288 static DECLARE_RWSEM(memcg_cache_ids_sem
);
290 void memcg_get_cache_ids(void)
292 down_read(&memcg_cache_ids_sem
);
295 void memcg_put_cache_ids(void)
297 up_read(&memcg_cache_ids_sem
);
301 * MIN_SIZE is different than 1, because we would like to avoid going through
302 * the alloc/free process all the time. In a small machine, 4 kmem-limited
303 * cgroups is a reasonable guess. In the future, it could be a parameter or
304 * tunable, but that is strictly not necessary.
306 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
307 * this constant directly from cgroup, but it is understandable that this is
308 * better kept as an internal representation in cgroup.c. In any case, the
309 * cgrp_id space is not getting any smaller, and we don't have to necessarily
310 * increase ours as well if it increases.
312 #define MEMCG_CACHES_MIN_SIZE 4
313 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
316 * A lot of the calls to the cache allocation functions are expected to be
317 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
318 * conditional to this static branch, we'll have to allow modules that does
319 * kmem_cache_alloc and the such to see this symbol as well
321 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
322 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
324 #endif /* !CONFIG_SLOB */
326 static struct mem_cgroup_per_zone
*
327 mem_cgroup_zone_zoneinfo(struct mem_cgroup
*memcg
, struct zone
*zone
)
329 int nid
= zone_to_nid(zone
);
330 int zid
= zone_idx(zone
);
332 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
336 * mem_cgroup_css_from_page - css of the memcg associated with a page
337 * @page: page of interest
339 * If memcg is bound to the default hierarchy, css of the memcg associated
340 * with @page is returned. The returned css remains associated with @page
341 * until it is released.
343 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
346 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
348 struct mem_cgroup
*memcg
;
350 memcg
= page
->mem_cgroup
;
352 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
353 memcg
= root_mem_cgroup
;
359 * page_cgroup_ino - return inode number of the memcg a page is charged to
362 * Look up the closest online ancestor of the memory cgroup @page is charged to
363 * and return its inode number or 0 if @page is not charged to any cgroup. It
364 * is safe to call this function without holding a reference to @page.
366 * Note, this function is inherently racy, because there is nothing to prevent
367 * the cgroup inode from getting torn down and potentially reallocated a moment
368 * after page_cgroup_ino() returns, so it only should be used by callers that
369 * do not care (such as procfs interfaces).
371 ino_t
page_cgroup_ino(struct page
*page
)
373 struct mem_cgroup
*memcg
;
374 unsigned long ino
= 0;
377 memcg
= READ_ONCE(page
->mem_cgroup
);
378 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
379 memcg
= parent_mem_cgroup(memcg
);
381 ino
= cgroup_ino(memcg
->css
.cgroup
);
386 static struct mem_cgroup_per_zone
*
387 mem_cgroup_page_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
389 int nid
= page_to_nid(page
);
390 int zid
= page_zonenum(page
);
392 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
395 static struct mem_cgroup_tree_per_zone
*
396 soft_limit_tree_node_zone(int nid
, int zid
)
398 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
401 static struct mem_cgroup_tree_per_zone
*
402 soft_limit_tree_from_page(struct page
*page
)
404 int nid
= page_to_nid(page
);
405 int zid
= page_zonenum(page
);
407 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
410 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone
*mz
,
411 struct mem_cgroup_tree_per_zone
*mctz
,
412 unsigned long new_usage_in_excess
)
414 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
415 struct rb_node
*parent
= NULL
;
416 struct mem_cgroup_per_zone
*mz_node
;
421 mz
->usage_in_excess
= new_usage_in_excess
;
422 if (!mz
->usage_in_excess
)
426 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
428 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
431 * We can't avoid mem cgroups that are over their soft
432 * limit by the same amount
434 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
437 rb_link_node(&mz
->tree_node
, parent
, p
);
438 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
442 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
443 struct mem_cgroup_tree_per_zone
*mctz
)
447 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
451 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
452 struct mem_cgroup_tree_per_zone
*mctz
)
456 spin_lock_irqsave(&mctz
->lock
, flags
);
457 __mem_cgroup_remove_exceeded(mz
, mctz
);
458 spin_unlock_irqrestore(&mctz
->lock
, flags
);
461 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
463 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
464 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
465 unsigned long excess
= 0;
467 if (nr_pages
> soft_limit
)
468 excess
= nr_pages
- soft_limit
;
473 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
475 unsigned long excess
;
476 struct mem_cgroup_per_zone
*mz
;
477 struct mem_cgroup_tree_per_zone
*mctz
;
479 mctz
= soft_limit_tree_from_page(page
);
481 * Necessary to update all ancestors when hierarchy is used.
482 * because their event counter is not touched.
484 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
485 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
486 excess
= soft_limit_excess(memcg
);
488 * We have to update the tree if mz is on RB-tree or
489 * mem is over its softlimit.
491 if (excess
|| mz
->on_tree
) {
494 spin_lock_irqsave(&mctz
->lock
, flags
);
495 /* if on-tree, remove it */
497 __mem_cgroup_remove_exceeded(mz
, mctz
);
499 * Insert again. mz->usage_in_excess will be updated.
500 * If excess is 0, no tree ops.
502 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
503 spin_unlock_irqrestore(&mctz
->lock
, flags
);
508 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
510 struct mem_cgroup_tree_per_zone
*mctz
;
511 struct mem_cgroup_per_zone
*mz
;
515 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
516 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
517 mctz
= soft_limit_tree_node_zone(nid
, zid
);
518 mem_cgroup_remove_exceeded(mz
, mctz
);
523 static struct mem_cgroup_per_zone
*
524 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
526 struct rb_node
*rightmost
= NULL
;
527 struct mem_cgroup_per_zone
*mz
;
531 rightmost
= rb_last(&mctz
->rb_root
);
533 goto done
; /* Nothing to reclaim from */
535 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
537 * Remove the node now but someone else can add it back,
538 * we will to add it back at the end of reclaim to its correct
539 * position in the tree.
541 __mem_cgroup_remove_exceeded(mz
, mctz
);
542 if (!soft_limit_excess(mz
->memcg
) ||
543 !css_tryget_online(&mz
->memcg
->css
))
549 static struct mem_cgroup_per_zone
*
550 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
552 struct mem_cgroup_per_zone
*mz
;
554 spin_lock_irq(&mctz
->lock
);
555 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
556 spin_unlock_irq(&mctz
->lock
);
561 * Return page count for single (non recursive) @memcg.
563 * Implementation Note: reading percpu statistics for memcg.
565 * Both of vmstat[] and percpu_counter has threshold and do periodic
566 * synchronization to implement "quick" read. There are trade-off between
567 * reading cost and precision of value. Then, we may have a chance to implement
568 * a periodic synchronization of counter in memcg's counter.
570 * But this _read() function is used for user interface now. The user accounts
571 * memory usage by memory cgroup and he _always_ requires exact value because
572 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
573 * have to visit all online cpus and make sum. So, for now, unnecessary
574 * synchronization is not implemented. (just implemented for cpu hotplug)
576 * If there are kernel internal actions which can make use of some not-exact
577 * value, and reading all cpu value can be performance bottleneck in some
578 * common workload, threshold and synchronization as vmstat[] should be
582 mem_cgroup_read_stat(struct mem_cgroup
*memcg
, enum mem_cgroup_stat_index idx
)
587 /* Per-cpu values can be negative, use a signed accumulator */
588 for_each_possible_cpu(cpu
)
589 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
591 * Summing races with updates, so val may be negative. Avoid exposing
592 * transient negative values.
599 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
600 enum mem_cgroup_events_index idx
)
602 unsigned long val
= 0;
605 for_each_possible_cpu(cpu
)
606 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
610 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
612 bool compound
, int nr_pages
)
615 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
616 * counted as CACHE even if it's on ANON LRU.
619 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
622 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
626 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
627 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
631 /* pagein of a big page is an event. So, ignore page size */
633 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
635 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
636 nr_pages
= -nr_pages
; /* for event */
639 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
642 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
643 int nid
, unsigned int lru_mask
)
645 unsigned long nr
= 0;
648 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
650 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
651 struct mem_cgroup_per_zone
*mz
;
655 if (!(BIT(lru
) & lru_mask
))
657 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
658 nr
+= mz
->lru_size
[lru
];
664 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
665 unsigned int lru_mask
)
667 unsigned long nr
= 0;
670 for_each_node_state(nid
, N_MEMORY
)
671 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
675 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
676 enum mem_cgroup_events_target target
)
678 unsigned long val
, next
;
680 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
681 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
682 /* from time_after() in jiffies.h */
683 if ((long)next
- (long)val
< 0) {
685 case MEM_CGROUP_TARGET_THRESH
:
686 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
688 case MEM_CGROUP_TARGET_SOFTLIMIT
:
689 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
691 case MEM_CGROUP_TARGET_NUMAINFO
:
692 next
= val
+ NUMAINFO_EVENTS_TARGET
;
697 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
704 * Check events in order.
707 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
709 /* threshold event is triggered in finer grain than soft limit */
710 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
711 MEM_CGROUP_TARGET_THRESH
))) {
713 bool do_numainfo __maybe_unused
;
715 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
716 MEM_CGROUP_TARGET_SOFTLIMIT
);
718 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
719 MEM_CGROUP_TARGET_NUMAINFO
);
721 mem_cgroup_threshold(memcg
);
722 if (unlikely(do_softlimit
))
723 mem_cgroup_update_tree(memcg
, page
);
725 if (unlikely(do_numainfo
))
726 atomic_inc(&memcg
->numainfo_events
);
731 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
734 * mm_update_next_owner() may clear mm->owner to NULL
735 * if it races with swapoff, page migration, etc.
736 * So this can be called with p == NULL.
741 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
743 EXPORT_SYMBOL(mem_cgroup_from_task
);
745 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
747 struct mem_cgroup
*memcg
= NULL
;
752 * Page cache insertions can happen withou an
753 * actual mm context, e.g. during disk probing
754 * on boot, loopback IO, acct() writes etc.
757 memcg
= root_mem_cgroup
;
759 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
760 if (unlikely(!memcg
))
761 memcg
= root_mem_cgroup
;
763 } while (!css_tryget_online(&memcg
->css
));
769 * mem_cgroup_iter - iterate over memory cgroup hierarchy
770 * @root: hierarchy root
771 * @prev: previously returned memcg, NULL on first invocation
772 * @reclaim: cookie for shared reclaim walks, NULL for full walks
774 * Returns references to children of the hierarchy below @root, or
775 * @root itself, or %NULL after a full round-trip.
777 * Caller must pass the return value in @prev on subsequent
778 * invocations for reference counting, or use mem_cgroup_iter_break()
779 * to cancel a hierarchy walk before the round-trip is complete.
781 * Reclaimers can specify a zone and a priority level in @reclaim to
782 * divide up the memcgs in the hierarchy among all concurrent
783 * reclaimers operating on the same zone and priority.
785 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
786 struct mem_cgroup
*prev
,
787 struct mem_cgroup_reclaim_cookie
*reclaim
)
789 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
790 struct cgroup_subsys_state
*css
= NULL
;
791 struct mem_cgroup
*memcg
= NULL
;
792 struct mem_cgroup
*pos
= NULL
;
794 if (mem_cgroup_disabled())
798 root
= root_mem_cgroup
;
800 if (prev
&& !reclaim
)
803 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
812 struct mem_cgroup_per_zone
*mz
;
814 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
815 iter
= &mz
->iter
[reclaim
->priority
];
817 if (prev
&& reclaim
->generation
!= iter
->generation
)
821 pos
= READ_ONCE(iter
->position
);
822 if (!pos
|| css_tryget(&pos
->css
))
825 * css reference reached zero, so iter->position will
826 * be cleared by ->css_released. However, we should not
827 * rely on this happening soon, because ->css_released
828 * is called from a work queue, and by busy-waiting we
829 * might block it. So we clear iter->position right
832 (void)cmpxchg(&iter
->position
, pos
, NULL
);
840 css
= css_next_descendant_pre(css
, &root
->css
);
843 * Reclaimers share the hierarchy walk, and a
844 * new one might jump in right at the end of
845 * the hierarchy - make sure they see at least
846 * one group and restart from the beginning.
854 * Verify the css and acquire a reference. The root
855 * is provided by the caller, so we know it's alive
856 * and kicking, and don't take an extra reference.
858 memcg
= mem_cgroup_from_css(css
);
860 if (css
== &root
->css
)
871 * The position could have already been updated by a competing
872 * thread, so check that the value hasn't changed since we read
873 * it to avoid reclaiming from the same cgroup twice.
875 (void)cmpxchg(&iter
->position
, pos
, memcg
);
883 reclaim
->generation
= iter
->generation
;
889 if (prev
&& prev
!= root
)
896 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
897 * @root: hierarchy root
898 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
900 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
901 struct mem_cgroup
*prev
)
904 root
= root_mem_cgroup
;
905 if (prev
&& prev
!= root
)
909 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
911 struct mem_cgroup
*memcg
= dead_memcg
;
912 struct mem_cgroup_reclaim_iter
*iter
;
913 struct mem_cgroup_per_zone
*mz
;
917 while ((memcg
= parent_mem_cgroup(memcg
))) {
919 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
920 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
921 for (i
= 0; i
<= DEF_PRIORITY
; i
++) {
923 cmpxchg(&iter
->position
,
932 * Iteration constructs for visiting all cgroups (under a tree). If
933 * loops are exited prematurely (break), mem_cgroup_iter_break() must
934 * be used for reference counting.
936 #define for_each_mem_cgroup_tree(iter, root) \
937 for (iter = mem_cgroup_iter(root, NULL, NULL); \
939 iter = mem_cgroup_iter(root, iter, NULL))
941 #define for_each_mem_cgroup(iter) \
942 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
944 iter = mem_cgroup_iter(NULL, iter, NULL))
947 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
948 * @zone: zone of the wanted lruvec
949 * @memcg: memcg of the wanted lruvec
951 * Returns the lru list vector holding pages for the given @zone and
952 * @mem. This can be the global zone lruvec, if the memory controller
955 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
956 struct mem_cgroup
*memcg
)
958 struct mem_cgroup_per_zone
*mz
;
959 struct lruvec
*lruvec
;
961 if (mem_cgroup_disabled()) {
962 lruvec
= &zone
->lruvec
;
966 mz
= mem_cgroup_zone_zoneinfo(memcg
, zone
);
967 lruvec
= &mz
->lruvec
;
970 * Since a node can be onlined after the mem_cgroup was created,
971 * we have to be prepared to initialize lruvec->zone here;
972 * and if offlined then reonlined, we need to reinitialize it.
974 if (unlikely(lruvec
->zone
!= zone
))
980 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
982 * @zone: zone of the page
984 * This function is only safe when following the LRU page isolation
985 * and putback protocol: the LRU lock must be held, and the page must
986 * either be PageLRU() or the caller must have isolated/allocated it.
988 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
990 struct mem_cgroup_per_zone
*mz
;
991 struct mem_cgroup
*memcg
;
992 struct lruvec
*lruvec
;
994 if (mem_cgroup_disabled()) {
995 lruvec
= &zone
->lruvec
;
999 memcg
= page
->mem_cgroup
;
1001 * Swapcache readahead pages are added to the LRU - and
1002 * possibly migrated - before they are charged.
1005 memcg
= root_mem_cgroup
;
1007 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1008 lruvec
= &mz
->lruvec
;
1011 * Since a node can be onlined after the mem_cgroup was created,
1012 * we have to be prepared to initialize lruvec->zone here;
1013 * and if offlined then reonlined, we need to reinitialize it.
1015 if (unlikely(lruvec
->zone
!= zone
))
1016 lruvec
->zone
= zone
;
1021 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1022 * @lruvec: mem_cgroup per zone lru vector
1023 * @lru: index of lru list the page is sitting on
1024 * @nr_pages: positive when adding or negative when removing
1026 * This function must be called under lru_lock, just before a page is added
1027 * to or just after a page is removed from an lru list (that ordering being
1028 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1030 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1033 struct mem_cgroup_per_zone
*mz
;
1034 unsigned long *lru_size
;
1038 __update_lru_size(lruvec
, lru
, nr_pages
);
1040 if (mem_cgroup_disabled())
1043 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1044 lru_size
= mz
->lru_size
+ lru
;
1045 empty
= list_empty(lruvec
->lists
+ lru
);
1048 *lru_size
+= nr_pages
;
1051 if (WARN_ONCE(size
< 0 || empty
!= !size
,
1052 "%s(%p, %d, %d): lru_size %ld but %sempty\n",
1053 __func__
, lruvec
, lru
, nr_pages
, size
, empty
? "" : "not ")) {
1059 *lru_size
+= nr_pages
;
1062 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1064 struct mem_cgroup
*task_memcg
;
1065 struct task_struct
*p
;
1068 p
= find_lock_task_mm(task
);
1070 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1074 * All threads may have already detached their mm's, but the oom
1075 * killer still needs to detect if they have already been oom
1076 * killed to prevent needlessly killing additional tasks.
1079 task_memcg
= mem_cgroup_from_task(task
);
1080 css_get(&task_memcg
->css
);
1083 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1084 css_put(&task_memcg
->css
);
1089 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1090 * @memcg: the memory cgroup
1092 * Returns the maximum amount of memory @mem can be charged with, in
1095 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1097 unsigned long margin
= 0;
1098 unsigned long count
;
1099 unsigned long limit
;
1101 count
= page_counter_read(&memcg
->memory
);
1102 limit
= READ_ONCE(memcg
->memory
.limit
);
1104 margin
= limit
- count
;
1106 if (do_memsw_account()) {
1107 count
= page_counter_read(&memcg
->memsw
);
1108 limit
= READ_ONCE(memcg
->memsw
.limit
);
1110 margin
= min(margin
, limit
- count
);
1117 * A routine for checking "mem" is under move_account() or not.
1119 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1120 * moving cgroups. This is for waiting at high-memory pressure
1123 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1125 struct mem_cgroup
*from
;
1126 struct mem_cgroup
*to
;
1129 * Unlike task_move routines, we access mc.to, mc.from not under
1130 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1132 spin_lock(&mc
.lock
);
1138 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1139 mem_cgroup_is_descendant(to
, memcg
);
1141 spin_unlock(&mc
.lock
);
1145 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1147 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1148 if (mem_cgroup_under_move(memcg
)) {
1150 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1151 /* moving charge context might have finished. */
1154 finish_wait(&mc
.waitq
, &wait
);
1161 #define K(x) ((x) << (PAGE_SHIFT-10))
1163 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1164 * @memcg: The memory cgroup that went over limit
1165 * @p: Task that is going to be killed
1167 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1170 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1172 struct mem_cgroup
*iter
;
1178 pr_info("Task in ");
1179 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1180 pr_cont(" killed as a result of limit of ");
1182 pr_info("Memory limit reached of cgroup ");
1185 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1190 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1191 K((u64
)page_counter_read(&memcg
->memory
)),
1192 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1193 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1194 K((u64
)page_counter_read(&memcg
->memsw
)),
1195 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1196 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1197 K((u64
)page_counter_read(&memcg
->kmem
)),
1198 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1200 for_each_mem_cgroup_tree(iter
, memcg
) {
1201 pr_info("Memory cgroup stats for ");
1202 pr_cont_cgroup_path(iter
->css
.cgroup
);
1205 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1206 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1208 pr_cont(" %s:%luKB", mem_cgroup_stat_names
[i
],
1209 K(mem_cgroup_read_stat(iter
, i
)));
1212 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1213 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1214 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1221 * This function returns the number of memcg under hierarchy tree. Returns
1222 * 1(self count) if no children.
1224 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1227 struct mem_cgroup
*iter
;
1229 for_each_mem_cgroup_tree(iter
, memcg
)
1235 * Return the memory (and swap, if configured) limit for a memcg.
1237 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1239 unsigned long limit
;
1241 limit
= memcg
->memory
.limit
;
1242 if (mem_cgroup_swappiness(memcg
)) {
1243 unsigned long memsw_limit
;
1244 unsigned long swap_limit
;
1246 memsw_limit
= memcg
->memsw
.limit
;
1247 swap_limit
= memcg
->swap
.limit
;
1248 swap_limit
= min(swap_limit
, (unsigned long)total_swap_pages
);
1249 limit
= min(limit
+ swap_limit
, memsw_limit
);
1254 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1257 struct oom_control oc
= {
1260 .gfp_mask
= gfp_mask
,
1263 struct mem_cgroup
*iter
;
1264 unsigned long chosen_points
= 0;
1265 unsigned long totalpages
;
1266 unsigned int points
= 0;
1267 struct task_struct
*chosen
= NULL
;
1269 mutex_lock(&oom_lock
);
1272 * If current has a pending SIGKILL or is exiting, then automatically
1273 * select it. The goal is to allow it to allocate so that it may
1274 * quickly exit and free its memory.
1276 if (fatal_signal_pending(current
) || task_will_free_mem(current
)) {
1277 mark_oom_victim(current
);
1278 try_oom_reaper(current
);
1282 check_panic_on_oom(&oc
, CONSTRAINT_MEMCG
, memcg
);
1283 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1284 for_each_mem_cgroup_tree(iter
, memcg
) {
1285 struct css_task_iter it
;
1286 struct task_struct
*task
;
1288 css_task_iter_start(&iter
->css
, &it
);
1289 while ((task
= css_task_iter_next(&it
))) {
1290 switch (oom_scan_process_thread(&oc
, task
, totalpages
)) {
1291 case OOM_SCAN_SELECT
:
1293 put_task_struct(chosen
);
1295 chosen_points
= ULONG_MAX
;
1296 get_task_struct(chosen
);
1298 case OOM_SCAN_CONTINUE
:
1300 case OOM_SCAN_ABORT
:
1301 css_task_iter_end(&it
);
1302 mem_cgroup_iter_break(memcg
, iter
);
1304 put_task_struct(chosen
);
1309 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1310 if (!points
|| points
< chosen_points
)
1312 /* Prefer thread group leaders for display purposes */
1313 if (points
== chosen_points
&&
1314 thread_group_leader(chosen
))
1318 put_task_struct(chosen
);
1320 chosen_points
= points
;
1321 get_task_struct(chosen
);
1323 css_task_iter_end(&it
);
1327 points
= chosen_points
* 1000 / totalpages
;
1328 oom_kill_process(&oc
, chosen
, points
, totalpages
, memcg
,
1329 "Memory cgroup out of memory");
1332 mutex_unlock(&oom_lock
);
1336 #if MAX_NUMNODES > 1
1339 * test_mem_cgroup_node_reclaimable
1340 * @memcg: the target memcg
1341 * @nid: the node ID to be checked.
1342 * @noswap : specify true here if the user wants flle only information.
1344 * This function returns whether the specified memcg contains any
1345 * reclaimable pages on a node. Returns true if there are any reclaimable
1346 * pages in the node.
1348 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1349 int nid
, bool noswap
)
1351 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1353 if (noswap
|| !total_swap_pages
)
1355 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1362 * Always updating the nodemask is not very good - even if we have an empty
1363 * list or the wrong list here, we can start from some node and traverse all
1364 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1367 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1371 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1372 * pagein/pageout changes since the last update.
1374 if (!atomic_read(&memcg
->numainfo_events
))
1376 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1379 /* make a nodemask where this memcg uses memory from */
1380 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1382 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1384 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1385 node_clear(nid
, memcg
->scan_nodes
);
1388 atomic_set(&memcg
->numainfo_events
, 0);
1389 atomic_set(&memcg
->numainfo_updating
, 0);
1393 * Selecting a node where we start reclaim from. Because what we need is just
1394 * reducing usage counter, start from anywhere is O,K. Considering
1395 * memory reclaim from current node, there are pros. and cons.
1397 * Freeing memory from current node means freeing memory from a node which
1398 * we'll use or we've used. So, it may make LRU bad. And if several threads
1399 * hit limits, it will see a contention on a node. But freeing from remote
1400 * node means more costs for memory reclaim because of memory latency.
1402 * Now, we use round-robin. Better algorithm is welcomed.
1404 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1408 mem_cgroup_may_update_nodemask(memcg
);
1409 node
= memcg
->last_scanned_node
;
1411 node
= next_node_in(node
, memcg
->scan_nodes
);
1413 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1414 * last time it really checked all the LRUs due to rate limiting.
1415 * Fallback to the current node in that case for simplicity.
1417 if (unlikely(node
== MAX_NUMNODES
))
1418 node
= numa_node_id();
1420 memcg
->last_scanned_node
= node
;
1424 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1430 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1433 unsigned long *total_scanned
)
1435 struct mem_cgroup
*victim
= NULL
;
1438 unsigned long excess
;
1439 unsigned long nr_scanned
;
1440 struct mem_cgroup_reclaim_cookie reclaim
= {
1445 excess
= soft_limit_excess(root_memcg
);
1448 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1453 * If we have not been able to reclaim
1454 * anything, it might because there are
1455 * no reclaimable pages under this hierarchy
1460 * We want to do more targeted reclaim.
1461 * excess >> 2 is not to excessive so as to
1462 * reclaim too much, nor too less that we keep
1463 * coming back to reclaim from this cgroup
1465 if (total
>= (excess
>> 2) ||
1466 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1471 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1473 *total_scanned
+= nr_scanned
;
1474 if (!soft_limit_excess(root_memcg
))
1477 mem_cgroup_iter_break(root_memcg
, victim
);
1481 #ifdef CONFIG_LOCKDEP
1482 static struct lockdep_map memcg_oom_lock_dep_map
= {
1483 .name
= "memcg_oom_lock",
1487 static DEFINE_SPINLOCK(memcg_oom_lock
);
1490 * Check OOM-Killer is already running under our hierarchy.
1491 * If someone is running, return false.
1493 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1495 struct mem_cgroup
*iter
, *failed
= NULL
;
1497 spin_lock(&memcg_oom_lock
);
1499 for_each_mem_cgroup_tree(iter
, memcg
) {
1500 if (iter
->oom_lock
) {
1502 * this subtree of our hierarchy is already locked
1503 * so we cannot give a lock.
1506 mem_cgroup_iter_break(memcg
, iter
);
1509 iter
->oom_lock
= true;
1514 * OK, we failed to lock the whole subtree so we have
1515 * to clean up what we set up to the failing subtree
1517 for_each_mem_cgroup_tree(iter
, memcg
) {
1518 if (iter
== failed
) {
1519 mem_cgroup_iter_break(memcg
, iter
);
1522 iter
->oom_lock
= false;
1525 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1527 spin_unlock(&memcg_oom_lock
);
1532 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1534 struct mem_cgroup
*iter
;
1536 spin_lock(&memcg_oom_lock
);
1537 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1538 for_each_mem_cgroup_tree(iter
, memcg
)
1539 iter
->oom_lock
= false;
1540 spin_unlock(&memcg_oom_lock
);
1543 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1545 struct mem_cgroup
*iter
;
1547 spin_lock(&memcg_oom_lock
);
1548 for_each_mem_cgroup_tree(iter
, memcg
)
1550 spin_unlock(&memcg_oom_lock
);
1553 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1555 struct mem_cgroup
*iter
;
1558 * When a new child is created while the hierarchy is under oom,
1559 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1561 spin_lock(&memcg_oom_lock
);
1562 for_each_mem_cgroup_tree(iter
, memcg
)
1563 if (iter
->under_oom
> 0)
1565 spin_unlock(&memcg_oom_lock
);
1568 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1570 struct oom_wait_info
{
1571 struct mem_cgroup
*memcg
;
1575 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1576 unsigned mode
, int sync
, void *arg
)
1578 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1579 struct mem_cgroup
*oom_wait_memcg
;
1580 struct oom_wait_info
*oom_wait_info
;
1582 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1583 oom_wait_memcg
= oom_wait_info
->memcg
;
1585 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1586 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1588 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1591 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1594 * For the following lockless ->under_oom test, the only required
1595 * guarantee is that it must see the state asserted by an OOM when
1596 * this function is called as a result of userland actions
1597 * triggered by the notification of the OOM. This is trivially
1598 * achieved by invoking mem_cgroup_mark_under_oom() before
1599 * triggering notification.
1601 if (memcg
&& memcg
->under_oom
)
1602 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1605 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1607 if (!current
->memcg_may_oom
)
1610 * We are in the middle of the charge context here, so we
1611 * don't want to block when potentially sitting on a callstack
1612 * that holds all kinds of filesystem and mm locks.
1614 * Also, the caller may handle a failed allocation gracefully
1615 * (like optional page cache readahead) and so an OOM killer
1616 * invocation might not even be necessary.
1618 * That's why we don't do anything here except remember the
1619 * OOM context and then deal with it at the end of the page
1620 * fault when the stack is unwound, the locks are released,
1621 * and when we know whether the fault was overall successful.
1623 css_get(&memcg
->css
);
1624 current
->memcg_in_oom
= memcg
;
1625 current
->memcg_oom_gfp_mask
= mask
;
1626 current
->memcg_oom_order
= order
;
1630 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1631 * @handle: actually kill/wait or just clean up the OOM state
1633 * This has to be called at the end of a page fault if the memcg OOM
1634 * handler was enabled.
1636 * Memcg supports userspace OOM handling where failed allocations must
1637 * sleep on a waitqueue until the userspace task resolves the
1638 * situation. Sleeping directly in the charge context with all kinds
1639 * of locks held is not a good idea, instead we remember an OOM state
1640 * in the task and mem_cgroup_oom_synchronize() has to be called at
1641 * the end of the page fault to complete the OOM handling.
1643 * Returns %true if an ongoing memcg OOM situation was detected and
1644 * completed, %false otherwise.
1646 bool mem_cgroup_oom_synchronize(bool handle
)
1648 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1649 struct oom_wait_info owait
;
1652 /* OOM is global, do not handle */
1656 if (!handle
|| oom_killer_disabled
)
1659 owait
.memcg
= memcg
;
1660 owait
.wait
.flags
= 0;
1661 owait
.wait
.func
= memcg_oom_wake_function
;
1662 owait
.wait
.private = current
;
1663 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1665 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1666 mem_cgroup_mark_under_oom(memcg
);
1668 locked
= mem_cgroup_oom_trylock(memcg
);
1671 mem_cgroup_oom_notify(memcg
);
1673 if (locked
&& !memcg
->oom_kill_disable
) {
1674 mem_cgroup_unmark_under_oom(memcg
);
1675 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1676 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1677 current
->memcg_oom_order
);
1680 mem_cgroup_unmark_under_oom(memcg
);
1681 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1685 mem_cgroup_oom_unlock(memcg
);
1687 * There is no guarantee that an OOM-lock contender
1688 * sees the wakeups triggered by the OOM kill
1689 * uncharges. Wake any sleepers explicitely.
1691 memcg_oom_recover(memcg
);
1694 current
->memcg_in_oom
= NULL
;
1695 css_put(&memcg
->css
);
1700 * lock_page_memcg - lock a page->mem_cgroup binding
1703 * This function protects unlocked LRU pages from being moved to
1704 * another cgroup and stabilizes their page->mem_cgroup binding.
1706 void lock_page_memcg(struct page
*page
)
1708 struct mem_cgroup
*memcg
;
1709 unsigned long flags
;
1712 * The RCU lock is held throughout the transaction. The fast
1713 * path can get away without acquiring the memcg->move_lock
1714 * because page moving starts with an RCU grace period.
1718 if (mem_cgroup_disabled())
1721 memcg
= page
->mem_cgroup
;
1722 if (unlikely(!memcg
))
1725 if (atomic_read(&memcg
->moving_account
) <= 0)
1728 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1729 if (memcg
!= page
->mem_cgroup
) {
1730 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1735 * When charge migration first begins, we can have locked and
1736 * unlocked page stat updates happening concurrently. Track
1737 * the task who has the lock for unlock_page_memcg().
1739 memcg
->move_lock_task
= current
;
1740 memcg
->move_lock_flags
= flags
;
1744 EXPORT_SYMBOL(lock_page_memcg
);
1747 * unlock_page_memcg - unlock a page->mem_cgroup binding
1750 void unlock_page_memcg(struct page
*page
)
1752 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
1754 if (memcg
&& memcg
->move_lock_task
== current
) {
1755 unsigned long flags
= memcg
->move_lock_flags
;
1757 memcg
->move_lock_task
= NULL
;
1758 memcg
->move_lock_flags
= 0;
1760 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1765 EXPORT_SYMBOL(unlock_page_memcg
);
1768 * size of first charge trial. "32" comes from vmscan.c's magic value.
1769 * TODO: maybe necessary to use big numbers in big irons.
1771 #define CHARGE_BATCH 32U
1772 struct memcg_stock_pcp
{
1773 struct mem_cgroup
*cached
; /* this never be root cgroup */
1774 unsigned int nr_pages
;
1775 struct work_struct work
;
1776 unsigned long flags
;
1777 #define FLUSHING_CACHED_CHARGE 0
1779 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1780 static DEFINE_MUTEX(percpu_charge_mutex
);
1783 * consume_stock: Try to consume stocked charge on this cpu.
1784 * @memcg: memcg to consume from.
1785 * @nr_pages: how many pages to charge.
1787 * The charges will only happen if @memcg matches the current cpu's memcg
1788 * stock, and at least @nr_pages are available in that stock. Failure to
1789 * service an allocation will refill the stock.
1791 * returns true if successful, false otherwise.
1793 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1795 struct memcg_stock_pcp
*stock
;
1798 if (nr_pages
> CHARGE_BATCH
)
1801 stock
= &get_cpu_var(memcg_stock
);
1802 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
1803 stock
->nr_pages
-= nr_pages
;
1806 put_cpu_var(memcg_stock
);
1811 * Returns stocks cached in percpu and reset cached information.
1813 static void drain_stock(struct memcg_stock_pcp
*stock
)
1815 struct mem_cgroup
*old
= stock
->cached
;
1817 if (stock
->nr_pages
) {
1818 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
1819 if (do_memsw_account())
1820 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
1821 css_put_many(&old
->css
, stock
->nr_pages
);
1822 stock
->nr_pages
= 0;
1824 stock
->cached
= NULL
;
1828 * This must be called under preempt disabled or must be called by
1829 * a thread which is pinned to local cpu.
1831 static void drain_local_stock(struct work_struct
*dummy
)
1833 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
1835 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
1839 * Cache charges(val) to local per_cpu area.
1840 * This will be consumed by consume_stock() function, later.
1842 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1844 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
1846 if (stock
->cached
!= memcg
) { /* reset if necessary */
1848 stock
->cached
= memcg
;
1850 stock
->nr_pages
+= nr_pages
;
1851 put_cpu_var(memcg_stock
);
1855 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1856 * of the hierarchy under it.
1858 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
1862 /* If someone's already draining, avoid adding running more workers. */
1863 if (!mutex_trylock(&percpu_charge_mutex
))
1865 /* Notify other cpus that system-wide "drain" is running */
1868 for_each_online_cpu(cpu
) {
1869 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1870 struct mem_cgroup
*memcg
;
1872 memcg
= stock
->cached
;
1873 if (!memcg
|| !stock
->nr_pages
)
1875 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
1877 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
1879 drain_local_stock(&stock
->work
);
1881 schedule_work_on(cpu
, &stock
->work
);
1886 mutex_unlock(&percpu_charge_mutex
);
1889 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
1890 unsigned long action
,
1893 int cpu
= (unsigned long)hcpu
;
1894 struct memcg_stock_pcp
*stock
;
1896 if (action
== CPU_ONLINE
)
1899 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
1902 stock
= &per_cpu(memcg_stock
, cpu
);
1907 static void reclaim_high(struct mem_cgroup
*memcg
,
1908 unsigned int nr_pages
,
1912 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
1914 mem_cgroup_events(memcg
, MEMCG_HIGH
, 1);
1915 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
1916 } while ((memcg
= parent_mem_cgroup(memcg
)));
1919 static void high_work_func(struct work_struct
*work
)
1921 struct mem_cgroup
*memcg
;
1923 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
1924 reclaim_high(memcg
, CHARGE_BATCH
, GFP_KERNEL
);
1928 * Scheduled by try_charge() to be executed from the userland return path
1929 * and reclaims memory over the high limit.
1931 void mem_cgroup_handle_over_high(void)
1933 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
1934 struct mem_cgroup
*memcg
;
1936 if (likely(!nr_pages
))
1939 memcg
= get_mem_cgroup_from_mm(current
->mm
);
1940 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
1941 css_put(&memcg
->css
);
1942 current
->memcg_nr_pages_over_high
= 0;
1945 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1946 unsigned int nr_pages
)
1948 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
1949 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1950 struct mem_cgroup
*mem_over_limit
;
1951 struct page_counter
*counter
;
1952 unsigned long nr_reclaimed
;
1953 bool may_swap
= true;
1954 bool drained
= false;
1956 if (mem_cgroup_is_root(memcg
))
1959 if (consume_stock(memcg
, nr_pages
))
1962 if (!do_memsw_account() ||
1963 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
1964 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
1966 if (do_memsw_account())
1967 page_counter_uncharge(&memcg
->memsw
, batch
);
1968 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
1970 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
1974 if (batch
> nr_pages
) {
1980 * Unlike in global OOM situations, memcg is not in a physical
1981 * memory shortage. Allow dying and OOM-killed tasks to
1982 * bypass the last charges so that they can exit quickly and
1983 * free their memory.
1985 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
1986 fatal_signal_pending(current
) ||
1987 current
->flags
& PF_EXITING
))
1990 if (unlikely(task_in_memcg_oom(current
)))
1993 if (!gfpflags_allow_blocking(gfp_mask
))
1996 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
1998 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
1999 gfp_mask
, may_swap
);
2001 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2005 drain_all_stock(mem_over_limit
);
2010 if (gfp_mask
& __GFP_NORETRY
)
2013 * Even though the limit is exceeded at this point, reclaim
2014 * may have been able to free some pages. Retry the charge
2015 * before killing the task.
2017 * Only for regular pages, though: huge pages are rather
2018 * unlikely to succeed so close to the limit, and we fall back
2019 * to regular pages anyway in case of failure.
2021 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2024 * At task move, charge accounts can be doubly counted. So, it's
2025 * better to wait until the end of task_move if something is going on.
2027 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2033 if (gfp_mask
& __GFP_NOFAIL
)
2036 if (fatal_signal_pending(current
))
2039 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
2041 mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2042 get_order(nr_pages
* PAGE_SIZE
));
2044 if (!(gfp_mask
& __GFP_NOFAIL
))
2048 * The allocation either can't fail or will lead to more memory
2049 * being freed very soon. Allow memory usage go over the limit
2050 * temporarily by force charging it.
2052 page_counter_charge(&memcg
->memory
, nr_pages
);
2053 if (do_memsw_account())
2054 page_counter_charge(&memcg
->memsw
, nr_pages
);
2055 css_get_many(&memcg
->css
, nr_pages
);
2060 css_get_many(&memcg
->css
, batch
);
2061 if (batch
> nr_pages
)
2062 refill_stock(memcg
, batch
- nr_pages
);
2065 * If the hierarchy is above the normal consumption range, schedule
2066 * reclaim on returning to userland. We can perform reclaim here
2067 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2068 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2069 * not recorded as it most likely matches current's and won't
2070 * change in the meantime. As high limit is checked again before
2071 * reclaim, the cost of mismatch is negligible.
2074 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2075 /* Don't bother a random interrupted task */
2076 if (in_interrupt()) {
2077 schedule_work(&memcg
->high_work
);
2080 current
->memcg_nr_pages_over_high
+= batch
;
2081 set_notify_resume(current
);
2084 } while ((memcg
= parent_mem_cgroup(memcg
)));
2089 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2091 if (mem_cgroup_is_root(memcg
))
2094 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2095 if (do_memsw_account())
2096 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2098 css_put_many(&memcg
->css
, nr_pages
);
2101 static void lock_page_lru(struct page
*page
, int *isolated
)
2103 struct zone
*zone
= page_zone(page
);
2105 spin_lock_irq(&zone
->lru_lock
);
2106 if (PageLRU(page
)) {
2107 struct lruvec
*lruvec
;
2109 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2111 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2117 static void unlock_page_lru(struct page
*page
, int isolated
)
2119 struct zone
*zone
= page_zone(page
);
2122 struct lruvec
*lruvec
;
2124 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2125 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2127 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2129 spin_unlock_irq(&zone
->lru_lock
);
2132 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2137 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2140 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2141 * may already be on some other mem_cgroup's LRU. Take care of it.
2144 lock_page_lru(page
, &isolated
);
2147 * Nobody should be changing or seriously looking at
2148 * page->mem_cgroup at this point:
2150 * - the page is uncharged
2152 * - the page is off-LRU
2154 * - an anonymous fault has exclusive page access, except for
2155 * a locked page table
2157 * - a page cache insertion, a swapin fault, or a migration
2158 * have the page locked
2160 page
->mem_cgroup
= memcg
;
2163 unlock_page_lru(page
, isolated
);
2167 static int memcg_alloc_cache_id(void)
2172 id
= ida_simple_get(&memcg_cache_ida
,
2173 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2177 if (id
< memcg_nr_cache_ids
)
2181 * There's no space for the new id in memcg_caches arrays,
2182 * so we have to grow them.
2184 down_write(&memcg_cache_ids_sem
);
2186 size
= 2 * (id
+ 1);
2187 if (size
< MEMCG_CACHES_MIN_SIZE
)
2188 size
= MEMCG_CACHES_MIN_SIZE
;
2189 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2190 size
= MEMCG_CACHES_MAX_SIZE
;
2192 err
= memcg_update_all_caches(size
);
2194 err
= memcg_update_all_list_lrus(size
);
2196 memcg_nr_cache_ids
= size
;
2198 up_write(&memcg_cache_ids_sem
);
2201 ida_simple_remove(&memcg_cache_ida
, id
);
2207 static void memcg_free_cache_id(int id
)
2209 ida_simple_remove(&memcg_cache_ida
, id
);
2212 struct memcg_kmem_cache_create_work
{
2213 struct mem_cgroup
*memcg
;
2214 struct kmem_cache
*cachep
;
2215 struct work_struct work
;
2218 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2220 struct memcg_kmem_cache_create_work
*cw
=
2221 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2222 struct mem_cgroup
*memcg
= cw
->memcg
;
2223 struct kmem_cache
*cachep
= cw
->cachep
;
2225 memcg_create_kmem_cache(memcg
, cachep
);
2227 css_put(&memcg
->css
);
2232 * Enqueue the creation of a per-memcg kmem_cache.
2234 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2235 struct kmem_cache
*cachep
)
2237 struct memcg_kmem_cache_create_work
*cw
;
2239 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2243 css_get(&memcg
->css
);
2246 cw
->cachep
= cachep
;
2247 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2249 schedule_work(&cw
->work
);
2252 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2253 struct kmem_cache
*cachep
)
2256 * We need to stop accounting when we kmalloc, because if the
2257 * corresponding kmalloc cache is not yet created, the first allocation
2258 * in __memcg_schedule_kmem_cache_create will recurse.
2260 * However, it is better to enclose the whole function. Depending on
2261 * the debugging options enabled, INIT_WORK(), for instance, can
2262 * trigger an allocation. This too, will make us recurse. Because at
2263 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2264 * the safest choice is to do it like this, wrapping the whole function.
2266 current
->memcg_kmem_skip_account
= 1;
2267 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2268 current
->memcg_kmem_skip_account
= 0;
2272 * Return the kmem_cache we're supposed to use for a slab allocation.
2273 * We try to use the current memcg's version of the cache.
2275 * If the cache does not exist yet, if we are the first user of it,
2276 * we either create it immediately, if possible, or create it asynchronously
2278 * In the latter case, we will let the current allocation go through with
2279 * the original cache.
2281 * Can't be called in interrupt context or from kernel threads.
2282 * This function needs to be called with rcu_read_lock() held.
2284 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2286 struct mem_cgroup
*memcg
;
2287 struct kmem_cache
*memcg_cachep
;
2290 VM_BUG_ON(!is_root_cache(cachep
));
2292 if (cachep
->flags
& SLAB_ACCOUNT
)
2293 gfp
|= __GFP_ACCOUNT
;
2295 if (!(gfp
& __GFP_ACCOUNT
))
2298 if (current
->memcg_kmem_skip_account
)
2301 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2302 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2306 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2307 if (likely(memcg_cachep
))
2308 return memcg_cachep
;
2311 * If we are in a safe context (can wait, and not in interrupt
2312 * context), we could be be predictable and return right away.
2313 * This would guarantee that the allocation being performed
2314 * already belongs in the new cache.
2316 * However, there are some clashes that can arrive from locking.
2317 * For instance, because we acquire the slab_mutex while doing
2318 * memcg_create_kmem_cache, this means no further allocation
2319 * could happen with the slab_mutex held. So it's better to
2322 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2324 css_put(&memcg
->css
);
2328 void __memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2330 if (!is_root_cache(cachep
))
2331 css_put(&cachep
->memcg_params
.memcg
->css
);
2334 int __memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2335 struct mem_cgroup
*memcg
)
2337 unsigned int nr_pages
= 1 << order
;
2338 struct page_counter
*counter
;
2341 ret
= try_charge(memcg
, gfp
, nr_pages
);
2345 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2346 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2347 cancel_charge(memcg
, nr_pages
);
2351 page
->mem_cgroup
= memcg
;
2356 int __memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2358 struct mem_cgroup
*memcg
;
2361 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2362 if (!mem_cgroup_is_root(memcg
))
2363 ret
= __memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2364 css_put(&memcg
->css
);
2368 void __memcg_kmem_uncharge(struct page
*page
, int order
)
2370 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2371 unsigned int nr_pages
= 1 << order
;
2376 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2378 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2379 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2381 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2382 if (do_memsw_account())
2383 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2385 page
->mem_cgroup
= NULL
;
2386 css_put_many(&memcg
->css
, nr_pages
);
2388 #endif /* !CONFIG_SLOB */
2390 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2393 * Because tail pages are not marked as "used", set it. We're under
2394 * zone->lru_lock and migration entries setup in all page mappings.
2396 void mem_cgroup_split_huge_fixup(struct page
*head
)
2400 if (mem_cgroup_disabled())
2403 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2404 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2406 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2409 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2411 #ifdef CONFIG_MEMCG_SWAP
2412 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2415 int val
= (charge
) ? 1 : -1;
2416 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2420 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2421 * @entry: swap entry to be moved
2422 * @from: mem_cgroup which the entry is moved from
2423 * @to: mem_cgroup which the entry is moved to
2425 * It succeeds only when the swap_cgroup's record for this entry is the same
2426 * as the mem_cgroup's id of @from.
2428 * Returns 0 on success, -EINVAL on failure.
2430 * The caller must have charged to @to, IOW, called page_counter_charge() about
2431 * both res and memsw, and called css_get().
2433 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2434 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2436 unsigned short old_id
, new_id
;
2438 old_id
= mem_cgroup_id(from
);
2439 new_id
= mem_cgroup_id(to
);
2441 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2442 mem_cgroup_swap_statistics(from
, false);
2443 mem_cgroup_swap_statistics(to
, true);
2449 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2450 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2456 static DEFINE_MUTEX(memcg_limit_mutex
);
2458 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2459 unsigned long limit
)
2461 unsigned long curusage
;
2462 unsigned long oldusage
;
2463 bool enlarge
= false;
2468 * For keeping hierarchical_reclaim simple, how long we should retry
2469 * is depends on callers. We set our retry-count to be function
2470 * of # of children which we should visit in this loop.
2472 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2473 mem_cgroup_count_children(memcg
);
2475 oldusage
= page_counter_read(&memcg
->memory
);
2478 if (signal_pending(current
)) {
2483 mutex_lock(&memcg_limit_mutex
);
2484 if (limit
> memcg
->memsw
.limit
) {
2485 mutex_unlock(&memcg_limit_mutex
);
2489 if (limit
> memcg
->memory
.limit
)
2491 ret
= page_counter_limit(&memcg
->memory
, limit
);
2492 mutex_unlock(&memcg_limit_mutex
);
2497 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2499 curusage
= page_counter_read(&memcg
->memory
);
2500 /* Usage is reduced ? */
2501 if (curusage
>= oldusage
)
2504 oldusage
= curusage
;
2505 } while (retry_count
);
2507 if (!ret
&& enlarge
)
2508 memcg_oom_recover(memcg
);
2513 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2514 unsigned long limit
)
2516 unsigned long curusage
;
2517 unsigned long oldusage
;
2518 bool enlarge
= false;
2522 /* see mem_cgroup_resize_res_limit */
2523 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2524 mem_cgroup_count_children(memcg
);
2526 oldusage
= page_counter_read(&memcg
->memsw
);
2529 if (signal_pending(current
)) {
2534 mutex_lock(&memcg_limit_mutex
);
2535 if (limit
< memcg
->memory
.limit
) {
2536 mutex_unlock(&memcg_limit_mutex
);
2540 if (limit
> memcg
->memsw
.limit
)
2542 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2543 mutex_unlock(&memcg_limit_mutex
);
2548 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2550 curusage
= page_counter_read(&memcg
->memsw
);
2551 /* Usage is reduced ? */
2552 if (curusage
>= oldusage
)
2555 oldusage
= curusage
;
2556 } while (retry_count
);
2558 if (!ret
&& enlarge
)
2559 memcg_oom_recover(memcg
);
2564 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
2566 unsigned long *total_scanned
)
2568 unsigned long nr_reclaimed
= 0;
2569 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
2570 unsigned long reclaimed
;
2572 struct mem_cgroup_tree_per_zone
*mctz
;
2573 unsigned long excess
;
2574 unsigned long nr_scanned
;
2579 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
2581 * This loop can run a while, specially if mem_cgroup's continuously
2582 * keep exceeding their soft limit and putting the system under
2589 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2594 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
2595 gfp_mask
, &nr_scanned
);
2596 nr_reclaimed
+= reclaimed
;
2597 *total_scanned
+= nr_scanned
;
2598 spin_lock_irq(&mctz
->lock
);
2599 __mem_cgroup_remove_exceeded(mz
, mctz
);
2602 * If we failed to reclaim anything from this memory cgroup
2603 * it is time to move on to the next cgroup
2607 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2609 excess
= soft_limit_excess(mz
->memcg
);
2611 * One school of thought says that we should not add
2612 * back the node to the tree if reclaim returns 0.
2613 * But our reclaim could return 0, simply because due
2614 * to priority we are exposing a smaller subset of
2615 * memory to reclaim from. Consider this as a longer
2618 /* If excess == 0, no tree ops */
2619 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2620 spin_unlock_irq(&mctz
->lock
);
2621 css_put(&mz
->memcg
->css
);
2624 * Could not reclaim anything and there are no more
2625 * mem cgroups to try or we seem to be looping without
2626 * reclaiming anything.
2628 if (!nr_reclaimed
&&
2630 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2632 } while (!nr_reclaimed
);
2634 css_put(&next_mz
->memcg
->css
);
2635 return nr_reclaimed
;
2639 * Test whether @memcg has children, dead or alive. Note that this
2640 * function doesn't care whether @memcg has use_hierarchy enabled and
2641 * returns %true if there are child csses according to the cgroup
2642 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2644 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2649 ret
= css_next_child(NULL
, &memcg
->css
);
2655 * Reclaims as many pages from the given memcg as possible.
2657 * Caller is responsible for holding css reference for memcg.
2659 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2661 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2663 /* we call try-to-free pages for make this cgroup empty */
2664 lru_add_drain_all();
2665 /* try to free all pages in this cgroup */
2666 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2669 if (signal_pending(current
))
2672 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2676 /* maybe some writeback is necessary */
2677 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2685 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
2686 char *buf
, size_t nbytes
,
2689 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2691 if (mem_cgroup_is_root(memcg
))
2693 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
2696 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
2699 return mem_cgroup_from_css(css
)->use_hierarchy
;
2702 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
2703 struct cftype
*cft
, u64 val
)
2706 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2707 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
2709 if (memcg
->use_hierarchy
== val
)
2713 * If parent's use_hierarchy is set, we can't make any modifications
2714 * in the child subtrees. If it is unset, then the change can
2715 * occur, provided the current cgroup has no children.
2717 * For the root cgroup, parent_mem is NULL, we allow value to be
2718 * set if there are no children.
2720 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
2721 (val
== 1 || val
== 0)) {
2722 if (!memcg_has_children(memcg
))
2723 memcg
->use_hierarchy
= val
;
2732 static void tree_stat(struct mem_cgroup
*memcg
, unsigned long *stat
)
2734 struct mem_cgroup
*iter
;
2737 memset(stat
, 0, sizeof(*stat
) * MEMCG_NR_STAT
);
2739 for_each_mem_cgroup_tree(iter
, memcg
) {
2740 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
2741 stat
[i
] += mem_cgroup_read_stat(iter
, i
);
2745 static void tree_events(struct mem_cgroup
*memcg
, unsigned long *events
)
2747 struct mem_cgroup
*iter
;
2750 memset(events
, 0, sizeof(*events
) * MEMCG_NR_EVENTS
);
2752 for_each_mem_cgroup_tree(iter
, memcg
) {
2753 for (i
= 0; i
< MEMCG_NR_EVENTS
; i
++)
2754 events
[i
] += mem_cgroup_read_events(iter
, i
);
2758 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
2760 unsigned long val
= 0;
2762 if (mem_cgroup_is_root(memcg
)) {
2763 struct mem_cgroup
*iter
;
2765 for_each_mem_cgroup_tree(iter
, memcg
) {
2766 val
+= mem_cgroup_read_stat(iter
,
2767 MEM_CGROUP_STAT_CACHE
);
2768 val
+= mem_cgroup_read_stat(iter
,
2769 MEM_CGROUP_STAT_RSS
);
2771 val
+= mem_cgroup_read_stat(iter
,
2772 MEM_CGROUP_STAT_SWAP
);
2776 val
= page_counter_read(&memcg
->memory
);
2778 val
= page_counter_read(&memcg
->memsw
);
2791 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
2794 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2795 struct page_counter
*counter
;
2797 switch (MEMFILE_TYPE(cft
->private)) {
2799 counter
= &memcg
->memory
;
2802 counter
= &memcg
->memsw
;
2805 counter
= &memcg
->kmem
;
2808 counter
= &memcg
->tcpmem
;
2814 switch (MEMFILE_ATTR(cft
->private)) {
2816 if (counter
== &memcg
->memory
)
2817 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
2818 if (counter
== &memcg
->memsw
)
2819 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
2820 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
2822 return (u64
)counter
->limit
* PAGE_SIZE
;
2824 return (u64
)counter
->watermark
* PAGE_SIZE
;
2826 return counter
->failcnt
;
2827 case RES_SOFT_LIMIT
:
2828 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
2835 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2839 if (cgroup_memory_nokmem
)
2842 BUG_ON(memcg
->kmemcg_id
>= 0);
2843 BUG_ON(memcg
->kmem_state
);
2845 memcg_id
= memcg_alloc_cache_id();
2849 static_branch_inc(&memcg_kmem_enabled_key
);
2851 * A memory cgroup is considered kmem-online as soon as it gets
2852 * kmemcg_id. Setting the id after enabling static branching will
2853 * guarantee no one starts accounting before all call sites are
2856 memcg
->kmemcg_id
= memcg_id
;
2857 memcg
->kmem_state
= KMEM_ONLINE
;
2862 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2864 struct cgroup_subsys_state
*css
;
2865 struct mem_cgroup
*parent
, *child
;
2868 if (memcg
->kmem_state
!= KMEM_ONLINE
)
2871 * Clear the online state before clearing memcg_caches array
2872 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2873 * guarantees that no cache will be created for this cgroup
2874 * after we are done (see memcg_create_kmem_cache()).
2876 memcg
->kmem_state
= KMEM_ALLOCATED
;
2878 memcg_deactivate_kmem_caches(memcg
);
2880 kmemcg_id
= memcg
->kmemcg_id
;
2881 BUG_ON(kmemcg_id
< 0);
2883 parent
= parent_mem_cgroup(memcg
);
2885 parent
= root_mem_cgroup
;
2888 * Change kmemcg_id of this cgroup and all its descendants to the
2889 * parent's id, and then move all entries from this cgroup's list_lrus
2890 * to ones of the parent. After we have finished, all list_lrus
2891 * corresponding to this cgroup are guaranteed to remain empty. The
2892 * ordering is imposed by list_lru_node->lock taken by
2893 * memcg_drain_all_list_lrus().
2895 css_for_each_descendant_pre(css
, &memcg
->css
) {
2896 child
= mem_cgroup_from_css(css
);
2897 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
2898 child
->kmemcg_id
= parent
->kmemcg_id
;
2899 if (!memcg
->use_hierarchy
)
2902 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
2904 memcg_free_cache_id(kmemcg_id
);
2907 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2909 /* css_alloc() failed, offlining didn't happen */
2910 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
2911 memcg_offline_kmem(memcg
);
2913 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
2914 memcg_destroy_kmem_caches(memcg
);
2915 static_branch_dec(&memcg_kmem_enabled_key
);
2916 WARN_ON(page_counter_read(&memcg
->kmem
));
2920 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2924 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2927 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2930 #endif /* !CONFIG_SLOB */
2932 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2933 unsigned long limit
)
2937 mutex_lock(&memcg_limit_mutex
);
2938 ret
= page_counter_limit(&memcg
->kmem
, limit
);
2939 mutex_unlock(&memcg_limit_mutex
);
2943 static int memcg_update_tcp_limit(struct mem_cgroup
*memcg
, unsigned long limit
)
2947 mutex_lock(&memcg_limit_mutex
);
2949 ret
= page_counter_limit(&memcg
->tcpmem
, limit
);
2953 if (!memcg
->tcpmem_active
) {
2955 * The active flag needs to be written after the static_key
2956 * update. This is what guarantees that the socket activation
2957 * function is the last one to run. See sock_update_memcg() for
2958 * details, and note that we don't mark any socket as belonging
2959 * to this memcg until that flag is up.
2961 * We need to do this, because static_keys will span multiple
2962 * sites, but we can't control their order. If we mark a socket
2963 * as accounted, but the accounting functions are not patched in
2964 * yet, we'll lose accounting.
2966 * We never race with the readers in sock_update_memcg(),
2967 * because when this value change, the code to process it is not
2970 static_branch_inc(&memcg_sockets_enabled_key
);
2971 memcg
->tcpmem_active
= true;
2974 mutex_unlock(&memcg_limit_mutex
);
2979 * The user of this function is...
2982 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
2983 char *buf
, size_t nbytes
, loff_t off
)
2985 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2986 unsigned long nr_pages
;
2989 buf
= strstrip(buf
);
2990 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
2994 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
2996 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3000 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3002 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
3005 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
3008 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
3011 ret
= memcg_update_tcp_limit(memcg
, nr_pages
);
3015 case RES_SOFT_LIMIT
:
3016 memcg
->soft_limit
= nr_pages
;
3020 return ret
?: nbytes
;
3023 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3024 size_t nbytes
, loff_t off
)
3026 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3027 struct page_counter
*counter
;
3029 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3031 counter
= &memcg
->memory
;
3034 counter
= &memcg
->memsw
;
3037 counter
= &memcg
->kmem
;
3040 counter
= &memcg
->tcpmem
;
3046 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3048 page_counter_reset_watermark(counter
);
3051 counter
->failcnt
= 0;
3060 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3063 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3067 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3068 struct cftype
*cft
, u64 val
)
3070 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3072 if (val
& ~MOVE_MASK
)
3076 * No kind of locking is needed in here, because ->can_attach() will
3077 * check this value once in the beginning of the process, and then carry
3078 * on with stale data. This means that changes to this value will only
3079 * affect task migrations starting after the change.
3081 memcg
->move_charge_at_immigrate
= val
;
3085 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3086 struct cftype
*cft
, u64 val
)
3093 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3097 unsigned int lru_mask
;
3100 static const struct numa_stat stats
[] = {
3101 { "total", LRU_ALL
},
3102 { "file", LRU_ALL_FILE
},
3103 { "anon", LRU_ALL_ANON
},
3104 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3106 const struct numa_stat
*stat
;
3109 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3111 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3112 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3113 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3114 for_each_node_state(nid
, N_MEMORY
) {
3115 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3117 seq_printf(m
, " N%d=%lu", nid
, nr
);
3122 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3123 struct mem_cgroup
*iter
;
3126 for_each_mem_cgroup_tree(iter
, memcg
)
3127 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3128 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3129 for_each_node_state(nid
, N_MEMORY
) {
3131 for_each_mem_cgroup_tree(iter
, memcg
)
3132 nr
+= mem_cgroup_node_nr_lru_pages(
3133 iter
, nid
, stat
->lru_mask
);
3134 seq_printf(m
, " N%d=%lu", nid
, nr
);
3141 #endif /* CONFIG_NUMA */
3143 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3145 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3146 unsigned long memory
, memsw
;
3147 struct mem_cgroup
*mi
;
3150 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3151 MEM_CGROUP_STAT_NSTATS
);
3152 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3153 MEM_CGROUP_EVENTS_NSTATS
);
3154 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3156 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3157 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3159 seq_printf(m
, "%s %lu\n", mem_cgroup_stat_names
[i
],
3160 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3163 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3164 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3165 mem_cgroup_read_events(memcg
, i
));
3167 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3168 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3169 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3171 /* Hierarchical information */
3172 memory
= memsw
= PAGE_COUNTER_MAX
;
3173 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3174 memory
= min(memory
, mi
->memory
.limit
);
3175 memsw
= min(memsw
, mi
->memsw
.limit
);
3177 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3178 (u64
)memory
* PAGE_SIZE
);
3179 if (do_memsw_account())
3180 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3181 (u64
)memsw
* PAGE_SIZE
);
3183 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3184 unsigned long long val
= 0;
3186 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3188 for_each_mem_cgroup_tree(mi
, memcg
)
3189 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3190 seq_printf(m
, "total_%s %llu\n", mem_cgroup_stat_names
[i
], val
);
3193 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3194 unsigned long long val
= 0;
3196 for_each_mem_cgroup_tree(mi
, memcg
)
3197 val
+= mem_cgroup_read_events(mi
, i
);
3198 seq_printf(m
, "total_%s %llu\n",
3199 mem_cgroup_events_names
[i
], val
);
3202 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3203 unsigned long long val
= 0;
3205 for_each_mem_cgroup_tree(mi
, memcg
)
3206 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3207 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3210 #ifdef CONFIG_DEBUG_VM
3213 struct mem_cgroup_per_zone
*mz
;
3214 struct zone_reclaim_stat
*rstat
;
3215 unsigned long recent_rotated
[2] = {0, 0};
3216 unsigned long recent_scanned
[2] = {0, 0};
3218 for_each_online_node(nid
)
3219 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3220 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
3221 rstat
= &mz
->lruvec
.reclaim_stat
;
3223 recent_rotated
[0] += rstat
->recent_rotated
[0];
3224 recent_rotated
[1] += rstat
->recent_rotated
[1];
3225 recent_scanned
[0] += rstat
->recent_scanned
[0];
3226 recent_scanned
[1] += rstat
->recent_scanned
[1];
3228 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3229 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3230 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3231 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3238 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3241 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3243 return mem_cgroup_swappiness(memcg
);
3246 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3247 struct cftype
*cft
, u64 val
)
3249 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3255 memcg
->swappiness
= val
;
3257 vm_swappiness
= val
;
3262 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3264 struct mem_cgroup_threshold_ary
*t
;
3265 unsigned long usage
;
3270 t
= rcu_dereference(memcg
->thresholds
.primary
);
3272 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3277 usage
= mem_cgroup_usage(memcg
, swap
);
3280 * current_threshold points to threshold just below or equal to usage.
3281 * If it's not true, a threshold was crossed after last
3282 * call of __mem_cgroup_threshold().
3284 i
= t
->current_threshold
;
3287 * Iterate backward over array of thresholds starting from
3288 * current_threshold and check if a threshold is crossed.
3289 * If none of thresholds below usage is crossed, we read
3290 * only one element of the array here.
3292 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3293 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3295 /* i = current_threshold + 1 */
3299 * Iterate forward over array of thresholds starting from
3300 * current_threshold+1 and check if a threshold is crossed.
3301 * If none of thresholds above usage is crossed, we read
3302 * only one element of the array here.
3304 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3305 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3307 /* Update current_threshold */
3308 t
->current_threshold
= i
- 1;
3313 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3316 __mem_cgroup_threshold(memcg
, false);
3317 if (do_memsw_account())
3318 __mem_cgroup_threshold(memcg
, true);
3320 memcg
= parent_mem_cgroup(memcg
);
3324 static int compare_thresholds(const void *a
, const void *b
)
3326 const struct mem_cgroup_threshold
*_a
= a
;
3327 const struct mem_cgroup_threshold
*_b
= b
;
3329 if (_a
->threshold
> _b
->threshold
)
3332 if (_a
->threshold
< _b
->threshold
)
3338 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3340 struct mem_cgroup_eventfd_list
*ev
;
3342 spin_lock(&memcg_oom_lock
);
3344 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3345 eventfd_signal(ev
->eventfd
, 1);
3347 spin_unlock(&memcg_oom_lock
);
3351 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3353 struct mem_cgroup
*iter
;
3355 for_each_mem_cgroup_tree(iter
, memcg
)
3356 mem_cgroup_oom_notify_cb(iter
);
3359 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3360 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3362 struct mem_cgroup_thresholds
*thresholds
;
3363 struct mem_cgroup_threshold_ary
*new;
3364 unsigned long threshold
;
3365 unsigned long usage
;
3368 ret
= page_counter_memparse(args
, "-1", &threshold
);
3372 mutex_lock(&memcg
->thresholds_lock
);
3375 thresholds
= &memcg
->thresholds
;
3376 usage
= mem_cgroup_usage(memcg
, false);
3377 } else if (type
== _MEMSWAP
) {
3378 thresholds
= &memcg
->memsw_thresholds
;
3379 usage
= mem_cgroup_usage(memcg
, true);
3383 /* Check if a threshold crossed before adding a new one */
3384 if (thresholds
->primary
)
3385 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3387 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3389 /* Allocate memory for new array of thresholds */
3390 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3398 /* Copy thresholds (if any) to new array */
3399 if (thresholds
->primary
) {
3400 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3401 sizeof(struct mem_cgroup_threshold
));
3404 /* Add new threshold */
3405 new->entries
[size
- 1].eventfd
= eventfd
;
3406 new->entries
[size
- 1].threshold
= threshold
;
3408 /* Sort thresholds. Registering of new threshold isn't time-critical */
3409 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3410 compare_thresholds
, NULL
);
3412 /* Find current threshold */
3413 new->current_threshold
= -1;
3414 for (i
= 0; i
< size
; i
++) {
3415 if (new->entries
[i
].threshold
<= usage
) {
3417 * new->current_threshold will not be used until
3418 * rcu_assign_pointer(), so it's safe to increment
3421 ++new->current_threshold
;
3426 /* Free old spare buffer and save old primary buffer as spare */
3427 kfree(thresholds
->spare
);
3428 thresholds
->spare
= thresholds
->primary
;
3430 rcu_assign_pointer(thresholds
->primary
, new);
3432 /* To be sure that nobody uses thresholds */
3436 mutex_unlock(&memcg
->thresholds_lock
);
3441 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3442 struct eventfd_ctx
*eventfd
, const char *args
)
3444 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3447 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3448 struct eventfd_ctx
*eventfd
, const char *args
)
3450 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3453 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3454 struct eventfd_ctx
*eventfd
, enum res_type type
)
3456 struct mem_cgroup_thresholds
*thresholds
;
3457 struct mem_cgroup_threshold_ary
*new;
3458 unsigned long usage
;
3461 mutex_lock(&memcg
->thresholds_lock
);
3464 thresholds
= &memcg
->thresholds
;
3465 usage
= mem_cgroup_usage(memcg
, false);
3466 } else if (type
== _MEMSWAP
) {
3467 thresholds
= &memcg
->memsw_thresholds
;
3468 usage
= mem_cgroup_usage(memcg
, true);
3472 if (!thresholds
->primary
)
3475 /* Check if a threshold crossed before removing */
3476 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3478 /* Calculate new number of threshold */
3480 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3481 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3485 new = thresholds
->spare
;
3487 /* Set thresholds array to NULL if we don't have thresholds */
3496 /* Copy thresholds and find current threshold */
3497 new->current_threshold
= -1;
3498 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3499 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3502 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3503 if (new->entries
[j
].threshold
<= usage
) {
3505 * new->current_threshold will not be used
3506 * until rcu_assign_pointer(), so it's safe to increment
3509 ++new->current_threshold
;
3515 /* Swap primary and spare array */
3516 thresholds
->spare
= thresholds
->primary
;
3518 rcu_assign_pointer(thresholds
->primary
, new);
3520 /* To be sure that nobody uses thresholds */
3523 /* If all events are unregistered, free the spare array */
3525 kfree(thresholds
->spare
);
3526 thresholds
->spare
= NULL
;
3529 mutex_unlock(&memcg
->thresholds_lock
);
3532 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3533 struct eventfd_ctx
*eventfd
)
3535 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3538 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3539 struct eventfd_ctx
*eventfd
)
3541 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3544 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3545 struct eventfd_ctx
*eventfd
, const char *args
)
3547 struct mem_cgroup_eventfd_list
*event
;
3549 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3553 spin_lock(&memcg_oom_lock
);
3555 event
->eventfd
= eventfd
;
3556 list_add(&event
->list
, &memcg
->oom_notify
);
3558 /* already in OOM ? */
3559 if (memcg
->under_oom
)
3560 eventfd_signal(eventfd
, 1);
3561 spin_unlock(&memcg_oom_lock
);
3566 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3567 struct eventfd_ctx
*eventfd
)
3569 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3571 spin_lock(&memcg_oom_lock
);
3573 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3574 if (ev
->eventfd
== eventfd
) {
3575 list_del(&ev
->list
);
3580 spin_unlock(&memcg_oom_lock
);
3583 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3585 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3587 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3588 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3592 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3593 struct cftype
*cft
, u64 val
)
3595 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3597 /* cannot set to root cgroup and only 0 and 1 are allowed */
3598 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3601 memcg
->oom_kill_disable
= val
;
3603 memcg_oom_recover(memcg
);
3608 #ifdef CONFIG_CGROUP_WRITEBACK
3610 struct list_head
*mem_cgroup_cgwb_list(struct mem_cgroup
*memcg
)
3612 return &memcg
->cgwb_list
;
3615 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3617 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3620 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3622 wb_domain_exit(&memcg
->cgwb_domain
);
3625 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3627 wb_domain_size_changed(&memcg
->cgwb_domain
);
3630 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3632 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3634 if (!memcg
->css
.parent
)
3637 return &memcg
->cgwb_domain
;
3641 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3642 * @wb: bdi_writeback in question
3643 * @pfilepages: out parameter for number of file pages
3644 * @pheadroom: out parameter for number of allocatable pages according to memcg
3645 * @pdirty: out parameter for number of dirty pages
3646 * @pwriteback: out parameter for number of pages under writeback
3648 * Determine the numbers of file, headroom, dirty, and writeback pages in
3649 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3650 * is a bit more involved.
3652 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3653 * headroom is calculated as the lowest headroom of itself and the
3654 * ancestors. Note that this doesn't consider the actual amount of
3655 * available memory in the system. The caller should further cap
3656 * *@pheadroom accordingly.
3658 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
3659 unsigned long *pheadroom
, unsigned long *pdirty
,
3660 unsigned long *pwriteback
)
3662 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3663 struct mem_cgroup
*parent
;
3665 *pdirty
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
3667 /* this should eventually include NR_UNSTABLE_NFS */
3668 *pwriteback
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
3669 *pfilepages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
3670 (1 << LRU_ACTIVE_FILE
));
3671 *pheadroom
= PAGE_COUNTER_MAX
;
3673 while ((parent
= parent_mem_cgroup(memcg
))) {
3674 unsigned long ceiling
= min(memcg
->memory
.limit
, memcg
->high
);
3675 unsigned long used
= page_counter_read(&memcg
->memory
);
3677 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
3682 #else /* CONFIG_CGROUP_WRITEBACK */
3684 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3689 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3693 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3697 #endif /* CONFIG_CGROUP_WRITEBACK */
3700 * DO NOT USE IN NEW FILES.
3702 * "cgroup.event_control" implementation.
3704 * This is way over-engineered. It tries to support fully configurable
3705 * events for each user. Such level of flexibility is completely
3706 * unnecessary especially in the light of the planned unified hierarchy.
3708 * Please deprecate this and replace with something simpler if at all
3713 * Unregister event and free resources.
3715 * Gets called from workqueue.
3717 static void memcg_event_remove(struct work_struct
*work
)
3719 struct mem_cgroup_event
*event
=
3720 container_of(work
, struct mem_cgroup_event
, remove
);
3721 struct mem_cgroup
*memcg
= event
->memcg
;
3723 remove_wait_queue(event
->wqh
, &event
->wait
);
3725 event
->unregister_event(memcg
, event
->eventfd
);
3727 /* Notify userspace the event is going away. */
3728 eventfd_signal(event
->eventfd
, 1);
3730 eventfd_ctx_put(event
->eventfd
);
3732 css_put(&memcg
->css
);
3736 * Gets called on POLLHUP on eventfd when user closes it.
3738 * Called with wqh->lock held and interrupts disabled.
3740 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
3741 int sync
, void *key
)
3743 struct mem_cgroup_event
*event
=
3744 container_of(wait
, struct mem_cgroup_event
, wait
);
3745 struct mem_cgroup
*memcg
= event
->memcg
;
3746 unsigned long flags
= (unsigned long)key
;
3748 if (flags
& POLLHUP
) {
3750 * If the event has been detached at cgroup removal, we
3751 * can simply return knowing the other side will cleanup
3754 * We can't race against event freeing since the other
3755 * side will require wqh->lock via remove_wait_queue(),
3758 spin_lock(&memcg
->event_list_lock
);
3759 if (!list_empty(&event
->list
)) {
3760 list_del_init(&event
->list
);
3762 * We are in atomic context, but cgroup_event_remove()
3763 * may sleep, so we have to call it in workqueue.
3765 schedule_work(&event
->remove
);
3767 spin_unlock(&memcg
->event_list_lock
);
3773 static void memcg_event_ptable_queue_proc(struct file
*file
,
3774 wait_queue_head_t
*wqh
, poll_table
*pt
)
3776 struct mem_cgroup_event
*event
=
3777 container_of(pt
, struct mem_cgroup_event
, pt
);
3780 add_wait_queue(wqh
, &event
->wait
);
3784 * DO NOT USE IN NEW FILES.
3786 * Parse input and register new cgroup event handler.
3788 * Input must be in format '<event_fd> <control_fd> <args>'.
3789 * Interpretation of args is defined by control file implementation.
3791 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
3792 char *buf
, size_t nbytes
, loff_t off
)
3794 struct cgroup_subsys_state
*css
= of_css(of
);
3795 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3796 struct mem_cgroup_event
*event
;
3797 struct cgroup_subsys_state
*cfile_css
;
3798 unsigned int efd
, cfd
;
3805 buf
= strstrip(buf
);
3807 efd
= simple_strtoul(buf
, &endp
, 10);
3812 cfd
= simple_strtoul(buf
, &endp
, 10);
3813 if ((*endp
!= ' ') && (*endp
!= '\0'))
3817 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
3821 event
->memcg
= memcg
;
3822 INIT_LIST_HEAD(&event
->list
);
3823 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
3824 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
3825 INIT_WORK(&event
->remove
, memcg_event_remove
);
3833 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
3834 if (IS_ERR(event
->eventfd
)) {
3835 ret
= PTR_ERR(event
->eventfd
);
3842 goto out_put_eventfd
;
3845 /* the process need read permission on control file */
3846 /* AV: shouldn't we check that it's been opened for read instead? */
3847 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
3852 * Determine the event callbacks and set them in @event. This used
3853 * to be done via struct cftype but cgroup core no longer knows
3854 * about these events. The following is crude but the whole thing
3855 * is for compatibility anyway.
3857 * DO NOT ADD NEW FILES.
3859 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
3861 if (!strcmp(name
, "memory.usage_in_bytes")) {
3862 event
->register_event
= mem_cgroup_usage_register_event
;
3863 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
3864 } else if (!strcmp(name
, "memory.oom_control")) {
3865 event
->register_event
= mem_cgroup_oom_register_event
;
3866 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
3867 } else if (!strcmp(name
, "memory.pressure_level")) {
3868 event
->register_event
= vmpressure_register_event
;
3869 event
->unregister_event
= vmpressure_unregister_event
;
3870 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
3871 event
->register_event
= memsw_cgroup_usage_register_event
;
3872 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
3879 * Verify @cfile should belong to @css. Also, remaining events are
3880 * automatically removed on cgroup destruction but the removal is
3881 * asynchronous, so take an extra ref on @css.
3883 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
3884 &memory_cgrp_subsys
);
3886 if (IS_ERR(cfile_css
))
3888 if (cfile_css
!= css
) {
3893 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
3897 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
3899 spin_lock(&memcg
->event_list_lock
);
3900 list_add(&event
->list
, &memcg
->event_list
);
3901 spin_unlock(&memcg
->event_list_lock
);
3913 eventfd_ctx_put(event
->eventfd
);
3922 static struct cftype mem_cgroup_legacy_files
[] = {
3924 .name
= "usage_in_bytes",
3925 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
3926 .read_u64
= mem_cgroup_read_u64
,
3929 .name
= "max_usage_in_bytes",
3930 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
3931 .write
= mem_cgroup_reset
,
3932 .read_u64
= mem_cgroup_read_u64
,
3935 .name
= "limit_in_bytes",
3936 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
3937 .write
= mem_cgroup_write
,
3938 .read_u64
= mem_cgroup_read_u64
,
3941 .name
= "soft_limit_in_bytes",
3942 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
3943 .write
= mem_cgroup_write
,
3944 .read_u64
= mem_cgroup_read_u64
,
3948 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
3949 .write
= mem_cgroup_reset
,
3950 .read_u64
= mem_cgroup_read_u64
,
3954 .seq_show
= memcg_stat_show
,
3957 .name
= "force_empty",
3958 .write
= mem_cgroup_force_empty_write
,
3961 .name
= "use_hierarchy",
3962 .write_u64
= mem_cgroup_hierarchy_write
,
3963 .read_u64
= mem_cgroup_hierarchy_read
,
3966 .name
= "cgroup.event_control", /* XXX: for compat */
3967 .write
= memcg_write_event_control
,
3968 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
3971 .name
= "swappiness",
3972 .read_u64
= mem_cgroup_swappiness_read
,
3973 .write_u64
= mem_cgroup_swappiness_write
,
3976 .name
= "move_charge_at_immigrate",
3977 .read_u64
= mem_cgroup_move_charge_read
,
3978 .write_u64
= mem_cgroup_move_charge_write
,
3981 .name
= "oom_control",
3982 .seq_show
= mem_cgroup_oom_control_read
,
3983 .write_u64
= mem_cgroup_oom_control_write
,
3984 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
3987 .name
= "pressure_level",
3991 .name
= "numa_stat",
3992 .seq_show
= memcg_numa_stat_show
,
3996 .name
= "kmem.limit_in_bytes",
3997 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
3998 .write
= mem_cgroup_write
,
3999 .read_u64
= mem_cgroup_read_u64
,
4002 .name
= "kmem.usage_in_bytes",
4003 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4004 .read_u64
= mem_cgroup_read_u64
,
4007 .name
= "kmem.failcnt",
4008 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4009 .write
= mem_cgroup_reset
,
4010 .read_u64
= mem_cgroup_read_u64
,
4013 .name
= "kmem.max_usage_in_bytes",
4014 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4015 .write
= mem_cgroup_reset
,
4016 .read_u64
= mem_cgroup_read_u64
,
4018 #ifdef CONFIG_SLABINFO
4020 .name
= "kmem.slabinfo",
4021 .seq_start
= slab_start
,
4022 .seq_next
= slab_next
,
4023 .seq_stop
= slab_stop
,
4024 .seq_show
= memcg_slab_show
,
4028 .name
= "kmem.tcp.limit_in_bytes",
4029 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4030 .write
= mem_cgroup_write
,
4031 .read_u64
= mem_cgroup_read_u64
,
4034 .name
= "kmem.tcp.usage_in_bytes",
4035 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4036 .read_u64
= mem_cgroup_read_u64
,
4039 .name
= "kmem.tcp.failcnt",
4040 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4041 .write
= mem_cgroup_reset
,
4042 .read_u64
= mem_cgroup_read_u64
,
4045 .name
= "kmem.tcp.max_usage_in_bytes",
4046 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4047 .write
= mem_cgroup_reset
,
4048 .read_u64
= mem_cgroup_read_u64
,
4050 { }, /* terminate */
4053 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4055 struct mem_cgroup_per_node
*pn
;
4056 struct mem_cgroup_per_zone
*mz
;
4057 int zone
, tmp
= node
;
4059 * This routine is called against possible nodes.
4060 * But it's BUG to call kmalloc() against offline node.
4062 * TODO: this routine can waste much memory for nodes which will
4063 * never be onlined. It's better to use memory hotplug callback
4066 if (!node_state(node
, N_NORMAL_MEMORY
))
4068 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4072 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4073 mz
= &pn
->zoneinfo
[zone
];
4074 lruvec_init(&mz
->lruvec
);
4075 mz
->usage_in_excess
= 0;
4076 mz
->on_tree
= false;
4079 memcg
->nodeinfo
[node
] = pn
;
4083 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4085 kfree(memcg
->nodeinfo
[node
]);
4088 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4092 memcg_wb_domain_exit(memcg
);
4094 free_mem_cgroup_per_zone_info(memcg
, node
);
4095 free_percpu(memcg
->stat
);
4099 static struct mem_cgroup
*mem_cgroup_alloc(void)
4101 struct mem_cgroup
*memcg
;
4105 size
= sizeof(struct mem_cgroup
);
4106 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4108 memcg
= kzalloc(size
, GFP_KERNEL
);
4112 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4117 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
4120 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4123 INIT_WORK(&memcg
->high_work
, high_work_func
);
4124 memcg
->last_scanned_node
= MAX_NUMNODES
;
4125 INIT_LIST_HEAD(&memcg
->oom_notify
);
4126 mutex_init(&memcg
->thresholds_lock
);
4127 spin_lock_init(&memcg
->move_lock
);
4128 vmpressure_init(&memcg
->vmpressure
);
4129 INIT_LIST_HEAD(&memcg
->event_list
);
4130 spin_lock_init(&memcg
->event_list_lock
);
4131 memcg
->socket_pressure
= jiffies
;
4133 memcg
->kmemcg_id
= -1;
4135 #ifdef CONFIG_CGROUP_WRITEBACK
4136 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4140 mem_cgroup_free(memcg
);
4144 static struct cgroup_subsys_state
* __ref
4145 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4147 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
4148 struct mem_cgroup
*memcg
;
4149 long error
= -ENOMEM
;
4151 memcg
= mem_cgroup_alloc();
4153 return ERR_PTR(error
);
4155 memcg
->high
= PAGE_COUNTER_MAX
;
4156 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4158 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4159 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4161 if (parent
&& parent
->use_hierarchy
) {
4162 memcg
->use_hierarchy
= true;
4163 page_counter_init(&memcg
->memory
, &parent
->memory
);
4164 page_counter_init(&memcg
->swap
, &parent
->swap
);
4165 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4166 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4167 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
4169 page_counter_init(&memcg
->memory
, NULL
);
4170 page_counter_init(&memcg
->swap
, NULL
);
4171 page_counter_init(&memcg
->memsw
, NULL
);
4172 page_counter_init(&memcg
->kmem
, NULL
);
4173 page_counter_init(&memcg
->tcpmem
, NULL
);
4175 * Deeper hierachy with use_hierarchy == false doesn't make
4176 * much sense so let cgroup subsystem know about this
4177 * unfortunate state in our controller.
4179 if (parent
!= root_mem_cgroup
)
4180 memory_cgrp_subsys
.broken_hierarchy
= true;
4183 /* The following stuff does not apply to the root */
4185 root_mem_cgroup
= memcg
;
4189 error
= memcg_online_kmem(memcg
);
4193 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4194 static_branch_inc(&memcg_sockets_enabled_key
);
4198 mem_cgroup_free(memcg
);
4203 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4205 if (css
->id
> MEM_CGROUP_ID_MAX
)
4211 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4213 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4214 struct mem_cgroup_event
*event
, *tmp
;
4217 * Unregister events and notify userspace.
4218 * Notify userspace about cgroup removing only after rmdir of cgroup
4219 * directory to avoid race between userspace and kernelspace.
4221 spin_lock(&memcg
->event_list_lock
);
4222 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4223 list_del_init(&event
->list
);
4224 schedule_work(&event
->remove
);
4226 spin_unlock(&memcg
->event_list_lock
);
4228 memcg_offline_kmem(memcg
);
4229 wb_memcg_offline(memcg
);
4232 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4234 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4236 invalidate_reclaim_iterators(memcg
);
4239 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4241 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4243 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4244 static_branch_dec(&memcg_sockets_enabled_key
);
4246 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
4247 static_branch_dec(&memcg_sockets_enabled_key
);
4249 vmpressure_cleanup(&memcg
->vmpressure
);
4250 cancel_work_sync(&memcg
->high_work
);
4251 mem_cgroup_remove_from_trees(memcg
);
4252 memcg_free_kmem(memcg
);
4253 mem_cgroup_free(memcg
);
4257 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4258 * @css: the target css
4260 * Reset the states of the mem_cgroup associated with @css. This is
4261 * invoked when the userland requests disabling on the default hierarchy
4262 * but the memcg is pinned through dependency. The memcg should stop
4263 * applying policies and should revert to the vanilla state as it may be
4264 * made visible again.
4266 * The current implementation only resets the essential configurations.
4267 * This needs to be expanded to cover all the visible parts.
4269 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4271 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4273 page_counter_limit(&memcg
->memory
, PAGE_COUNTER_MAX
);
4274 page_counter_limit(&memcg
->swap
, PAGE_COUNTER_MAX
);
4275 page_counter_limit(&memcg
->memsw
, PAGE_COUNTER_MAX
);
4276 page_counter_limit(&memcg
->kmem
, PAGE_COUNTER_MAX
);
4277 page_counter_limit(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
4279 memcg
->high
= PAGE_COUNTER_MAX
;
4280 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4281 memcg_wb_domain_size_changed(memcg
);
4285 /* Handlers for move charge at task migration. */
4286 static int mem_cgroup_do_precharge(unsigned long count
)
4290 /* Try a single bulk charge without reclaim first, kswapd may wake */
4291 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4293 mc
.precharge
+= count
;
4297 /* Try charges one by one with reclaim */
4299 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4309 * get_mctgt_type - get target type of moving charge
4310 * @vma: the vma the pte to be checked belongs
4311 * @addr: the address corresponding to the pte to be checked
4312 * @ptent: the pte to be checked
4313 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4316 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4317 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4318 * move charge. if @target is not NULL, the page is stored in target->page
4319 * with extra refcnt got(Callers should handle it).
4320 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4321 * target for charge migration. if @target is not NULL, the entry is stored
4324 * Called with pte lock held.
4331 enum mc_target_type
{
4337 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4338 unsigned long addr
, pte_t ptent
)
4340 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4342 if (!page
|| !page_mapped(page
))
4344 if (PageAnon(page
)) {
4345 if (!(mc
.flags
& MOVE_ANON
))
4348 if (!(mc
.flags
& MOVE_FILE
))
4351 if (!get_page_unless_zero(page
))
4358 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4359 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4361 struct page
*page
= NULL
;
4362 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4364 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4367 * Because lookup_swap_cache() updates some statistics counter,
4368 * we call find_get_page() with swapper_space directly.
4370 page
= find_get_page(swap_address_space(ent
), ent
.val
);
4371 if (do_memsw_account())
4372 entry
->val
= ent
.val
;
4377 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4378 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4384 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4385 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4387 struct page
*page
= NULL
;
4388 struct address_space
*mapping
;
4391 if (!vma
->vm_file
) /* anonymous vma */
4393 if (!(mc
.flags
& MOVE_FILE
))
4396 mapping
= vma
->vm_file
->f_mapping
;
4397 pgoff
= linear_page_index(vma
, addr
);
4399 /* page is moved even if it's not RSS of this task(page-faulted). */
4401 /* shmem/tmpfs may report page out on swap: account for that too. */
4402 if (shmem_mapping(mapping
)) {
4403 page
= find_get_entry(mapping
, pgoff
);
4404 if (radix_tree_exceptional_entry(page
)) {
4405 swp_entry_t swp
= radix_to_swp_entry(page
);
4406 if (do_memsw_account())
4408 page
= find_get_page(swap_address_space(swp
), swp
.val
);
4411 page
= find_get_page(mapping
, pgoff
);
4413 page
= find_get_page(mapping
, pgoff
);
4419 * mem_cgroup_move_account - move account of the page
4421 * @nr_pages: number of regular pages (>1 for huge pages)
4422 * @from: mem_cgroup which the page is moved from.
4423 * @to: mem_cgroup which the page is moved to. @from != @to.
4425 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4427 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4430 static int mem_cgroup_move_account(struct page
*page
,
4432 struct mem_cgroup
*from
,
4433 struct mem_cgroup
*to
)
4435 unsigned long flags
;
4436 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
4440 VM_BUG_ON(from
== to
);
4441 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4442 VM_BUG_ON(compound
&& !PageTransHuge(page
));
4445 * Prevent mem_cgroup_migrate() from looking at
4446 * page->mem_cgroup of its source page while we change it.
4449 if (!trylock_page(page
))
4453 if (page
->mem_cgroup
!= from
)
4456 anon
= PageAnon(page
);
4458 spin_lock_irqsave(&from
->move_lock
, flags
);
4460 if (!anon
&& page_mapped(page
)) {
4461 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4463 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4468 * move_lock grabbed above and caller set from->moving_account, so
4469 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4470 * So mapping should be stable for dirty pages.
4472 if (!anon
&& PageDirty(page
)) {
4473 struct address_space
*mapping
= page_mapping(page
);
4475 if (mapping_cap_account_dirty(mapping
)) {
4476 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4478 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4483 if (PageWriteback(page
)) {
4484 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4486 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4491 * It is safe to change page->mem_cgroup here because the page
4492 * is referenced, charged, and isolated - we can't race with
4493 * uncharging, charging, migration, or LRU putback.
4496 /* caller should have done css_get */
4497 page
->mem_cgroup
= to
;
4498 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4502 local_irq_disable();
4503 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
4504 memcg_check_events(to
, page
);
4505 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
4506 memcg_check_events(from
, page
);
4514 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4515 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4517 struct page
*page
= NULL
;
4518 enum mc_target_type ret
= MC_TARGET_NONE
;
4519 swp_entry_t ent
= { .val
= 0 };
4521 if (pte_present(ptent
))
4522 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4523 else if (is_swap_pte(ptent
))
4524 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
4525 else if (pte_none(ptent
))
4526 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4528 if (!page
&& !ent
.val
)
4532 * Do only loose check w/o serialization.
4533 * mem_cgroup_move_account() checks the page is valid or
4534 * not under LRU exclusion.
4536 if (page
->mem_cgroup
== mc
.from
) {
4537 ret
= MC_TARGET_PAGE
;
4539 target
->page
= page
;
4541 if (!ret
|| !target
)
4544 /* There is a swap entry and a page doesn't exist or isn't charged */
4545 if (ent
.val
&& !ret
&&
4546 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4547 ret
= MC_TARGET_SWAP
;
4554 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4556 * We don't consider swapping or file mapped pages because THP does not
4557 * support them for now.
4558 * Caller should make sure that pmd_trans_huge(pmd) is true.
4560 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4561 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4563 struct page
*page
= NULL
;
4564 enum mc_target_type ret
= MC_TARGET_NONE
;
4566 page
= pmd_page(pmd
);
4567 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4568 if (!(mc
.flags
& MOVE_ANON
))
4570 if (page
->mem_cgroup
== mc
.from
) {
4571 ret
= MC_TARGET_PAGE
;
4574 target
->page
= page
;
4580 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4581 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4583 return MC_TARGET_NONE
;
4587 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4588 unsigned long addr
, unsigned long end
,
4589 struct mm_walk
*walk
)
4591 struct vm_area_struct
*vma
= walk
->vma
;
4595 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4597 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4598 mc
.precharge
+= HPAGE_PMD_NR
;
4603 if (pmd_trans_unstable(pmd
))
4605 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4606 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4607 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4608 mc
.precharge
++; /* increment precharge temporarily */
4609 pte_unmap_unlock(pte
- 1, ptl
);
4615 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4617 unsigned long precharge
;
4619 struct mm_walk mem_cgroup_count_precharge_walk
= {
4620 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4623 down_read(&mm
->mmap_sem
);
4624 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk
);
4625 up_read(&mm
->mmap_sem
);
4627 precharge
= mc
.precharge
;
4633 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4635 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4637 VM_BUG_ON(mc
.moving_task
);
4638 mc
.moving_task
= current
;
4639 return mem_cgroup_do_precharge(precharge
);
4642 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4643 static void __mem_cgroup_clear_mc(void)
4645 struct mem_cgroup
*from
= mc
.from
;
4646 struct mem_cgroup
*to
= mc
.to
;
4648 /* we must uncharge all the leftover precharges from mc.to */
4650 cancel_charge(mc
.to
, mc
.precharge
);
4654 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4655 * we must uncharge here.
4657 if (mc
.moved_charge
) {
4658 cancel_charge(mc
.from
, mc
.moved_charge
);
4659 mc
.moved_charge
= 0;
4661 /* we must fixup refcnts and charges */
4662 if (mc
.moved_swap
) {
4663 /* uncharge swap account from the old cgroup */
4664 if (!mem_cgroup_is_root(mc
.from
))
4665 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4668 * we charged both to->memory and to->memsw, so we
4669 * should uncharge to->memory.
4671 if (!mem_cgroup_is_root(mc
.to
))
4672 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4674 css_put_many(&mc
.from
->css
, mc
.moved_swap
);
4676 /* we've already done css_get(mc.to) */
4679 memcg_oom_recover(from
);
4680 memcg_oom_recover(to
);
4681 wake_up_all(&mc
.waitq
);
4684 static void mem_cgroup_clear_mc(void)
4686 struct mm_struct
*mm
= mc
.mm
;
4689 * we must clear moving_task before waking up waiters at the end of
4692 mc
.moving_task
= NULL
;
4693 __mem_cgroup_clear_mc();
4694 spin_lock(&mc
.lock
);
4698 spin_unlock(&mc
.lock
);
4703 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4705 struct cgroup_subsys_state
*css
;
4706 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
4707 struct mem_cgroup
*from
;
4708 struct task_struct
*leader
, *p
;
4709 struct mm_struct
*mm
;
4710 unsigned long move_flags
;
4713 /* charge immigration isn't supported on the default hierarchy */
4714 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4718 * Multi-process migrations only happen on the default hierarchy
4719 * where charge immigration is not used. Perform charge
4720 * immigration if @tset contains a leader and whine if there are
4724 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
4727 memcg
= mem_cgroup_from_css(css
);
4733 * We are now commited to this value whatever it is. Changes in this
4734 * tunable will only affect upcoming migrations, not the current one.
4735 * So we need to save it, and keep it going.
4737 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
4741 from
= mem_cgroup_from_task(p
);
4743 VM_BUG_ON(from
== memcg
);
4745 mm
= get_task_mm(p
);
4748 /* We move charges only when we move a owner of the mm */
4749 if (mm
->owner
== p
) {
4752 VM_BUG_ON(mc
.precharge
);
4753 VM_BUG_ON(mc
.moved_charge
);
4754 VM_BUG_ON(mc
.moved_swap
);
4756 spin_lock(&mc
.lock
);
4760 mc
.flags
= move_flags
;
4761 spin_unlock(&mc
.lock
);
4762 /* We set mc.moving_task later */
4764 ret
= mem_cgroup_precharge_mc(mm
);
4766 mem_cgroup_clear_mc();
4773 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4776 mem_cgroup_clear_mc();
4779 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4780 unsigned long addr
, unsigned long end
,
4781 struct mm_walk
*walk
)
4784 struct vm_area_struct
*vma
= walk
->vma
;
4787 enum mc_target_type target_type
;
4788 union mc_target target
;
4791 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4793 if (mc
.precharge
< HPAGE_PMD_NR
) {
4797 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
4798 if (target_type
== MC_TARGET_PAGE
) {
4800 if (!isolate_lru_page(page
)) {
4801 if (!mem_cgroup_move_account(page
, true,
4803 mc
.precharge
-= HPAGE_PMD_NR
;
4804 mc
.moved_charge
+= HPAGE_PMD_NR
;
4806 putback_lru_page(page
);
4814 if (pmd_trans_unstable(pmd
))
4817 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4818 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4819 pte_t ptent
= *(pte
++);
4825 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
4826 case MC_TARGET_PAGE
:
4829 * We can have a part of the split pmd here. Moving it
4830 * can be done but it would be too convoluted so simply
4831 * ignore such a partial THP and keep it in original
4832 * memcg. There should be somebody mapping the head.
4834 if (PageTransCompound(page
))
4836 if (isolate_lru_page(page
))
4838 if (!mem_cgroup_move_account(page
, false,
4841 /* we uncharge from mc.from later. */
4844 putback_lru_page(page
);
4845 put
: /* get_mctgt_type() gets the page */
4848 case MC_TARGET_SWAP
:
4850 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
4852 /* we fixup refcnts and charges later. */
4860 pte_unmap_unlock(pte
- 1, ptl
);
4865 * We have consumed all precharges we got in can_attach().
4866 * We try charge one by one, but don't do any additional
4867 * charges to mc.to if we have failed in charge once in attach()
4870 ret
= mem_cgroup_do_precharge(1);
4878 static void mem_cgroup_move_charge(void)
4880 struct mm_walk mem_cgroup_move_charge_walk
= {
4881 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4885 lru_add_drain_all();
4887 * Signal lock_page_memcg() to take the memcg's move_lock
4888 * while we're moving its pages to another memcg. Then wait
4889 * for already started RCU-only updates to finish.
4891 atomic_inc(&mc
.from
->moving_account
);
4894 if (unlikely(!down_read_trylock(&mc
.mm
->mmap_sem
))) {
4896 * Someone who are holding the mmap_sem might be waiting in
4897 * waitq. So we cancel all extra charges, wake up all waiters,
4898 * and retry. Because we cancel precharges, we might not be able
4899 * to move enough charges, but moving charge is a best-effort
4900 * feature anyway, so it wouldn't be a big problem.
4902 __mem_cgroup_clear_mc();
4907 * When we have consumed all precharges and failed in doing
4908 * additional charge, the page walk just aborts.
4910 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk
);
4911 up_read(&mc
.mm
->mmap_sem
);
4912 atomic_dec(&mc
.from
->moving_account
);
4915 static void mem_cgroup_move_task(void)
4918 mem_cgroup_move_charge();
4919 mem_cgroup_clear_mc();
4922 #else /* !CONFIG_MMU */
4923 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4927 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4930 static void mem_cgroup_move_task(void)
4936 * Cgroup retains root cgroups across [un]mount cycles making it necessary
4937 * to verify whether we're attached to the default hierarchy on each mount
4940 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
4943 * use_hierarchy is forced on the default hierarchy. cgroup core
4944 * guarantees that @root doesn't have any children, so turning it
4945 * on for the root memcg is enough.
4947 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4948 root_mem_cgroup
->use_hierarchy
= true;
4950 root_mem_cgroup
->use_hierarchy
= false;
4953 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
4956 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4958 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
4961 static int memory_low_show(struct seq_file
*m
, void *v
)
4963 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
4964 unsigned long low
= READ_ONCE(memcg
->low
);
4966 if (low
== PAGE_COUNTER_MAX
)
4967 seq_puts(m
, "max\n");
4969 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
4974 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
4975 char *buf
, size_t nbytes
, loff_t off
)
4977 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
4981 buf
= strstrip(buf
);
4982 err
= page_counter_memparse(buf
, "max", &low
);
4991 static int memory_high_show(struct seq_file
*m
, void *v
)
4993 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
4994 unsigned long high
= READ_ONCE(memcg
->high
);
4996 if (high
== PAGE_COUNTER_MAX
)
4997 seq_puts(m
, "max\n");
4999 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
5004 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5005 char *buf
, size_t nbytes
, loff_t off
)
5007 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5008 unsigned long nr_pages
;
5012 buf
= strstrip(buf
);
5013 err
= page_counter_memparse(buf
, "max", &high
);
5019 nr_pages
= page_counter_read(&memcg
->memory
);
5020 if (nr_pages
> high
)
5021 try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
5024 memcg_wb_domain_size_changed(memcg
);
5028 static int memory_max_show(struct seq_file
*m
, void *v
)
5030 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5031 unsigned long max
= READ_ONCE(memcg
->memory
.limit
);
5033 if (max
== PAGE_COUNTER_MAX
)
5034 seq_puts(m
, "max\n");
5036 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5041 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5042 char *buf
, size_t nbytes
, loff_t off
)
5044 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5045 unsigned int nr_reclaims
= MEM_CGROUP_RECLAIM_RETRIES
;
5046 bool drained
= false;
5050 buf
= strstrip(buf
);
5051 err
= page_counter_memparse(buf
, "max", &max
);
5055 xchg(&memcg
->memory
.limit
, max
);
5058 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
5060 if (nr_pages
<= max
)
5063 if (signal_pending(current
)) {
5069 drain_all_stock(memcg
);
5075 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
5081 mem_cgroup_events(memcg
, MEMCG_OOM
, 1);
5082 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
5086 memcg_wb_domain_size_changed(memcg
);
5090 static int memory_events_show(struct seq_file
*m
, void *v
)
5092 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5094 seq_printf(m
, "low %lu\n", mem_cgroup_read_events(memcg
, MEMCG_LOW
));
5095 seq_printf(m
, "high %lu\n", mem_cgroup_read_events(memcg
, MEMCG_HIGH
));
5096 seq_printf(m
, "max %lu\n", mem_cgroup_read_events(memcg
, MEMCG_MAX
));
5097 seq_printf(m
, "oom %lu\n", mem_cgroup_read_events(memcg
, MEMCG_OOM
));
5102 static int memory_stat_show(struct seq_file
*m
, void *v
)
5104 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5105 unsigned long stat
[MEMCG_NR_STAT
];
5106 unsigned long events
[MEMCG_NR_EVENTS
];
5110 * Provide statistics on the state of the memory subsystem as
5111 * well as cumulative event counters that show past behavior.
5113 * This list is ordered following a combination of these gradients:
5114 * 1) generic big picture -> specifics and details
5115 * 2) reflecting userspace activity -> reflecting kernel heuristics
5117 * Current memory state:
5120 tree_stat(memcg
, stat
);
5121 tree_events(memcg
, events
);
5123 seq_printf(m
, "anon %llu\n",
5124 (u64
)stat
[MEM_CGROUP_STAT_RSS
] * PAGE_SIZE
);
5125 seq_printf(m
, "file %llu\n",
5126 (u64
)stat
[MEM_CGROUP_STAT_CACHE
] * PAGE_SIZE
);
5127 seq_printf(m
, "kernel_stack %llu\n",
5128 (u64
)stat
[MEMCG_KERNEL_STACK
] * PAGE_SIZE
);
5129 seq_printf(m
, "slab %llu\n",
5130 (u64
)(stat
[MEMCG_SLAB_RECLAIMABLE
] +
5131 stat
[MEMCG_SLAB_UNRECLAIMABLE
]) * PAGE_SIZE
);
5132 seq_printf(m
, "sock %llu\n",
5133 (u64
)stat
[MEMCG_SOCK
] * PAGE_SIZE
);
5135 seq_printf(m
, "file_mapped %llu\n",
5136 (u64
)stat
[MEM_CGROUP_STAT_FILE_MAPPED
] * PAGE_SIZE
);
5137 seq_printf(m
, "file_dirty %llu\n",
5138 (u64
)stat
[MEM_CGROUP_STAT_DIRTY
] * PAGE_SIZE
);
5139 seq_printf(m
, "file_writeback %llu\n",
5140 (u64
)stat
[MEM_CGROUP_STAT_WRITEBACK
] * PAGE_SIZE
);
5142 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5143 struct mem_cgroup
*mi
;
5144 unsigned long val
= 0;
5146 for_each_mem_cgroup_tree(mi
, memcg
)
5147 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
));
5148 seq_printf(m
, "%s %llu\n",
5149 mem_cgroup_lru_names
[i
], (u64
)val
* PAGE_SIZE
);
5152 seq_printf(m
, "slab_reclaimable %llu\n",
5153 (u64
)stat
[MEMCG_SLAB_RECLAIMABLE
] * PAGE_SIZE
);
5154 seq_printf(m
, "slab_unreclaimable %llu\n",
5155 (u64
)stat
[MEMCG_SLAB_UNRECLAIMABLE
] * PAGE_SIZE
);
5157 /* Accumulated memory events */
5159 seq_printf(m
, "pgfault %lu\n",
5160 events
[MEM_CGROUP_EVENTS_PGFAULT
]);
5161 seq_printf(m
, "pgmajfault %lu\n",
5162 events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
5167 static struct cftype memory_files
[] = {
5170 .flags
= CFTYPE_NOT_ON_ROOT
,
5171 .read_u64
= memory_current_read
,
5175 .flags
= CFTYPE_NOT_ON_ROOT
,
5176 .seq_show
= memory_low_show
,
5177 .write
= memory_low_write
,
5181 .flags
= CFTYPE_NOT_ON_ROOT
,
5182 .seq_show
= memory_high_show
,
5183 .write
= memory_high_write
,
5187 .flags
= CFTYPE_NOT_ON_ROOT
,
5188 .seq_show
= memory_max_show
,
5189 .write
= memory_max_write
,
5193 .flags
= CFTYPE_NOT_ON_ROOT
,
5194 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
5195 .seq_show
= memory_events_show
,
5199 .flags
= CFTYPE_NOT_ON_ROOT
,
5200 .seq_show
= memory_stat_show
,
5205 struct cgroup_subsys memory_cgrp_subsys
= {
5206 .css_alloc
= mem_cgroup_css_alloc
,
5207 .css_online
= mem_cgroup_css_online
,
5208 .css_offline
= mem_cgroup_css_offline
,
5209 .css_released
= mem_cgroup_css_released
,
5210 .css_free
= mem_cgroup_css_free
,
5211 .css_reset
= mem_cgroup_css_reset
,
5212 .can_attach
= mem_cgroup_can_attach
,
5213 .cancel_attach
= mem_cgroup_cancel_attach
,
5214 .post_attach
= mem_cgroup_move_task
,
5215 .bind
= mem_cgroup_bind
,
5216 .dfl_cftypes
= memory_files
,
5217 .legacy_cftypes
= mem_cgroup_legacy_files
,
5222 * mem_cgroup_low - check if memory consumption is below the normal range
5223 * @root: the highest ancestor to consider
5224 * @memcg: the memory cgroup to check
5226 * Returns %true if memory consumption of @memcg, and that of all
5227 * configurable ancestors up to @root, is below the normal range.
5229 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5231 if (mem_cgroup_disabled())
5235 * The toplevel group doesn't have a configurable range, so
5236 * it's never low when looked at directly, and it is not
5237 * considered an ancestor when assessing the hierarchy.
5240 if (memcg
== root_mem_cgroup
)
5243 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5246 while (memcg
!= root
) {
5247 memcg
= parent_mem_cgroup(memcg
);
5249 if (memcg
== root_mem_cgroup
)
5252 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5259 * mem_cgroup_try_charge - try charging a page
5260 * @page: page to charge
5261 * @mm: mm context of the victim
5262 * @gfp_mask: reclaim mode
5263 * @memcgp: charged memcg return
5265 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5266 * pages according to @gfp_mask if necessary.
5268 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5269 * Otherwise, an error code is returned.
5271 * After page->mapping has been set up, the caller must finalize the
5272 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5273 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5275 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5276 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
5279 struct mem_cgroup
*memcg
= NULL
;
5280 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5283 if (mem_cgroup_disabled())
5286 if (PageSwapCache(page
)) {
5288 * Every swap fault against a single page tries to charge the
5289 * page, bail as early as possible. shmem_unuse() encounters
5290 * already charged pages, too. The USED bit is protected by
5291 * the page lock, which serializes swap cache removal, which
5292 * in turn serializes uncharging.
5294 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5295 if (page
->mem_cgroup
)
5298 if (do_swap_account
) {
5299 swp_entry_t ent
= { .val
= page_private(page
), };
5300 unsigned short id
= lookup_swap_cgroup_id(ent
);
5303 memcg
= mem_cgroup_from_id(id
);
5304 if (memcg
&& !css_tryget_online(&memcg
->css
))
5311 memcg
= get_mem_cgroup_from_mm(mm
);
5313 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5315 css_put(&memcg
->css
);
5322 * mem_cgroup_commit_charge - commit a page charge
5323 * @page: page to charge
5324 * @memcg: memcg to charge the page to
5325 * @lrucare: page might be on LRU already
5327 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5328 * after page->mapping has been set up. This must happen atomically
5329 * as part of the page instantiation, i.e. under the page table lock
5330 * for anonymous pages, under the page lock for page and swap cache.
5332 * In addition, the page must not be on the LRU during the commit, to
5333 * prevent racing with task migration. If it might be, use @lrucare.
5335 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5337 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5338 bool lrucare
, bool compound
)
5340 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5342 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5343 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5345 if (mem_cgroup_disabled())
5348 * Swap faults will attempt to charge the same page multiple
5349 * times. But reuse_swap_page() might have removed the page
5350 * from swapcache already, so we can't check PageSwapCache().
5355 commit_charge(page
, memcg
, lrucare
);
5357 local_irq_disable();
5358 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
5359 memcg_check_events(memcg
, page
);
5362 if (do_memsw_account() && PageSwapCache(page
)) {
5363 swp_entry_t entry
= { .val
= page_private(page
) };
5365 * The swap entry might not get freed for a long time,
5366 * let's not wait for it. The page already received a
5367 * memory+swap charge, drop the swap entry duplicate.
5369 mem_cgroup_uncharge_swap(entry
);
5374 * mem_cgroup_cancel_charge - cancel a page charge
5375 * @page: page to charge
5376 * @memcg: memcg to charge the page to
5378 * Cancel a charge transaction started by mem_cgroup_try_charge().
5380 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5383 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5385 if (mem_cgroup_disabled())
5388 * Swap faults will attempt to charge the same page multiple
5389 * times. But reuse_swap_page() might have removed the page
5390 * from swapcache already, so we can't check PageSwapCache().
5395 cancel_charge(memcg
, nr_pages
);
5398 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5399 unsigned long nr_anon
, unsigned long nr_file
,
5400 unsigned long nr_huge
, struct page
*dummy_page
)
5402 unsigned long nr_pages
= nr_anon
+ nr_file
;
5403 unsigned long flags
;
5405 if (!mem_cgroup_is_root(memcg
)) {
5406 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5407 if (do_memsw_account())
5408 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5409 memcg_oom_recover(memcg
);
5412 local_irq_save(flags
);
5413 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5414 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5415 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5416 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5417 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5418 memcg_check_events(memcg
, dummy_page
);
5419 local_irq_restore(flags
);
5421 if (!mem_cgroup_is_root(memcg
))
5422 css_put_many(&memcg
->css
, nr_pages
);
5425 static void uncharge_list(struct list_head
*page_list
)
5427 struct mem_cgroup
*memcg
= NULL
;
5428 unsigned long nr_anon
= 0;
5429 unsigned long nr_file
= 0;
5430 unsigned long nr_huge
= 0;
5431 unsigned long pgpgout
= 0;
5432 struct list_head
*next
;
5436 * Note that the list can be a single page->lru; hence the
5437 * do-while loop instead of a simple list_for_each_entry().
5439 next
= page_list
->next
;
5441 unsigned int nr_pages
= 1;
5443 page
= list_entry(next
, struct page
, lru
);
5444 next
= page
->lru
.next
;
5446 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5447 VM_BUG_ON_PAGE(page_count(page
), page
);
5449 if (!page
->mem_cgroup
)
5453 * Nobody should be changing or seriously looking at
5454 * page->mem_cgroup at this point, we have fully
5455 * exclusive access to the page.
5458 if (memcg
!= page
->mem_cgroup
) {
5460 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5462 pgpgout
= nr_anon
= nr_file
= nr_huge
= 0;
5464 memcg
= page
->mem_cgroup
;
5467 if (PageTransHuge(page
)) {
5468 nr_pages
<<= compound_order(page
);
5469 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5470 nr_huge
+= nr_pages
;
5474 nr_anon
+= nr_pages
;
5476 nr_file
+= nr_pages
;
5478 page
->mem_cgroup
= NULL
;
5481 } while (next
!= page_list
);
5484 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5489 * mem_cgroup_uncharge - uncharge a page
5490 * @page: page to uncharge
5492 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5493 * mem_cgroup_commit_charge().
5495 void mem_cgroup_uncharge(struct page
*page
)
5497 if (mem_cgroup_disabled())
5500 /* Don't touch page->lru of any random page, pre-check: */
5501 if (!page
->mem_cgroup
)
5504 INIT_LIST_HEAD(&page
->lru
);
5505 uncharge_list(&page
->lru
);
5509 * mem_cgroup_uncharge_list - uncharge a list of page
5510 * @page_list: list of pages to uncharge
5512 * Uncharge a list of pages previously charged with
5513 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5515 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5517 if (mem_cgroup_disabled())
5520 if (!list_empty(page_list
))
5521 uncharge_list(page_list
);
5525 * mem_cgroup_migrate - charge a page's replacement
5526 * @oldpage: currently circulating page
5527 * @newpage: replacement page
5529 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5530 * be uncharged upon free.
5532 * Both pages must be locked, @newpage->mapping must be set up.
5534 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
5536 struct mem_cgroup
*memcg
;
5537 unsigned int nr_pages
;
5540 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5541 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5542 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5543 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5546 if (mem_cgroup_disabled())
5549 /* Page cache replacement: new page already charged? */
5550 if (newpage
->mem_cgroup
)
5553 /* Swapcache readahead pages can get replaced before being charged */
5554 memcg
= oldpage
->mem_cgroup
;
5558 /* Force-charge the new page. The old one will be freed soon */
5559 compound
= PageTransHuge(newpage
);
5560 nr_pages
= compound
? hpage_nr_pages(newpage
) : 1;
5562 page_counter_charge(&memcg
->memory
, nr_pages
);
5563 if (do_memsw_account())
5564 page_counter_charge(&memcg
->memsw
, nr_pages
);
5565 css_get_many(&memcg
->css
, nr_pages
);
5567 commit_charge(newpage
, memcg
, false);
5569 local_irq_disable();
5570 mem_cgroup_charge_statistics(memcg
, newpage
, compound
, nr_pages
);
5571 memcg_check_events(memcg
, newpage
);
5575 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
5576 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
5578 void sock_update_memcg(struct sock
*sk
)
5580 struct mem_cgroup
*memcg
;
5582 /* Socket cloning can throw us here with sk_cgrp already
5583 * filled. It won't however, necessarily happen from
5584 * process context. So the test for root memcg given
5585 * the current task's memcg won't help us in this case.
5587 * Respecting the original socket's memcg is a better
5588 * decision in this case.
5591 BUG_ON(mem_cgroup_is_root(sk
->sk_memcg
));
5592 css_get(&sk
->sk_memcg
->css
);
5597 memcg
= mem_cgroup_from_task(current
);
5598 if (memcg
== root_mem_cgroup
)
5600 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
5602 if (css_tryget_online(&memcg
->css
))
5603 sk
->sk_memcg
= memcg
;
5607 EXPORT_SYMBOL(sock_update_memcg
);
5609 void sock_release_memcg(struct sock
*sk
)
5611 WARN_ON(!sk
->sk_memcg
);
5612 css_put(&sk
->sk_memcg
->css
);
5616 * mem_cgroup_charge_skmem - charge socket memory
5617 * @memcg: memcg to charge
5618 * @nr_pages: number of pages to charge
5620 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5621 * @memcg's configured limit, %false if the charge had to be forced.
5623 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5625 gfp_t gfp_mask
= GFP_KERNEL
;
5627 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5628 struct page_counter
*fail
;
5630 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
5631 memcg
->tcpmem_pressure
= 0;
5634 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
5635 memcg
->tcpmem_pressure
= 1;
5639 /* Don't block in the packet receive path */
5641 gfp_mask
= GFP_NOWAIT
;
5643 this_cpu_add(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5645 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
5648 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
5653 * mem_cgroup_uncharge_skmem - uncharge socket memory
5654 * @memcg - memcg to uncharge
5655 * @nr_pages - number of pages to uncharge
5657 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5659 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5660 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
5664 this_cpu_sub(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5666 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5667 css_put_many(&memcg
->css
, nr_pages
);
5670 static int __init
cgroup_memory(char *s
)
5674 while ((token
= strsep(&s
, ",")) != NULL
) {
5677 if (!strcmp(token
, "nosocket"))
5678 cgroup_memory_nosocket
= true;
5679 if (!strcmp(token
, "nokmem"))
5680 cgroup_memory_nokmem
= true;
5684 __setup("cgroup.memory=", cgroup_memory
);
5687 * subsys_initcall() for memory controller.
5689 * Some parts like hotcpu_notifier() have to be initialized from this context
5690 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5691 * everything that doesn't depend on a specific mem_cgroup structure should
5692 * be initialized from here.
5694 static int __init
mem_cgroup_init(void)
5698 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5700 for_each_possible_cpu(cpu
)
5701 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5704 for_each_node(node
) {
5705 struct mem_cgroup_tree_per_node
*rtpn
;
5708 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5709 node_online(node
) ? node
: NUMA_NO_NODE
);
5711 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5712 struct mem_cgroup_tree_per_zone
*rtpz
;
5714 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
5715 rtpz
->rb_root
= RB_ROOT
;
5716 spin_lock_init(&rtpz
->lock
);
5718 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5723 subsys_initcall(mem_cgroup_init
);
5725 #ifdef CONFIG_MEMCG_SWAP
5727 * mem_cgroup_swapout - transfer a memsw charge to swap
5728 * @page: page whose memsw charge to transfer
5729 * @entry: swap entry to move the charge to
5731 * Transfer the memsw charge of @page to @entry.
5733 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5735 struct mem_cgroup
*memcg
;
5736 unsigned short oldid
;
5738 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5739 VM_BUG_ON_PAGE(page_count(page
), page
);
5741 if (!do_memsw_account())
5744 memcg
= page
->mem_cgroup
;
5746 /* Readahead page, never charged */
5750 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5751 VM_BUG_ON_PAGE(oldid
, page
);
5752 mem_cgroup_swap_statistics(memcg
, true);
5754 page
->mem_cgroup
= NULL
;
5756 if (!mem_cgroup_is_root(memcg
))
5757 page_counter_uncharge(&memcg
->memory
, 1);
5760 * Interrupts should be disabled here because the caller holds the
5761 * mapping->tree_lock lock which is taken with interrupts-off. It is
5762 * important here to have the interrupts disabled because it is the
5763 * only synchronisation we have for udpating the per-CPU variables.
5765 VM_BUG_ON(!irqs_disabled());
5766 mem_cgroup_charge_statistics(memcg
, page
, false, -1);
5767 memcg_check_events(memcg
, page
);
5771 * mem_cgroup_try_charge_swap - try charging a swap entry
5772 * @page: page being added to swap
5773 * @entry: swap entry to charge
5775 * Try to charge @entry to the memcg that @page belongs to.
5777 * Returns 0 on success, -ENOMEM on failure.
5779 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
5781 struct mem_cgroup
*memcg
;
5782 struct page_counter
*counter
;
5783 unsigned short oldid
;
5785 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
5788 memcg
= page
->mem_cgroup
;
5790 /* Readahead page, never charged */
5794 if (!mem_cgroup_is_root(memcg
) &&
5795 !page_counter_try_charge(&memcg
->swap
, 1, &counter
))
5798 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5799 VM_BUG_ON_PAGE(oldid
, page
);
5800 mem_cgroup_swap_statistics(memcg
, true);
5802 css_get(&memcg
->css
);
5807 * mem_cgroup_uncharge_swap - uncharge a swap entry
5808 * @entry: swap entry to uncharge
5810 * Drop the swap charge associated with @entry.
5812 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5814 struct mem_cgroup
*memcg
;
5817 if (!do_swap_account
)
5820 id
= swap_cgroup_record(entry
, 0);
5822 memcg
= mem_cgroup_from_id(id
);
5824 if (!mem_cgroup_is_root(memcg
)) {
5825 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5826 page_counter_uncharge(&memcg
->swap
, 1);
5828 page_counter_uncharge(&memcg
->memsw
, 1);
5830 mem_cgroup_swap_statistics(memcg
, false);
5831 css_put(&memcg
->css
);
5836 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
5838 long nr_swap_pages
= get_nr_swap_pages();
5840 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5841 return nr_swap_pages
;
5842 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5843 nr_swap_pages
= min_t(long, nr_swap_pages
,
5844 READ_ONCE(memcg
->swap
.limit
) -
5845 page_counter_read(&memcg
->swap
));
5846 return nr_swap_pages
;
5849 bool mem_cgroup_swap_full(struct page
*page
)
5851 struct mem_cgroup
*memcg
;
5853 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5857 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5860 memcg
= page
->mem_cgroup
;
5864 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5865 if (page_counter_read(&memcg
->swap
) * 2 >= memcg
->swap
.limit
)
5871 /* for remember boot option*/
5872 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5873 static int really_do_swap_account __initdata
= 1;
5875 static int really_do_swap_account __initdata
;
5878 static int __init
enable_swap_account(char *s
)
5880 if (!strcmp(s
, "1"))
5881 really_do_swap_account
= 1;
5882 else if (!strcmp(s
, "0"))
5883 really_do_swap_account
= 0;
5886 __setup("swapaccount=", enable_swap_account
);
5888 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
5891 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5893 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
5896 static int swap_max_show(struct seq_file
*m
, void *v
)
5898 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5899 unsigned long max
= READ_ONCE(memcg
->swap
.limit
);
5901 if (max
== PAGE_COUNTER_MAX
)
5902 seq_puts(m
, "max\n");
5904 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5909 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
5910 char *buf
, size_t nbytes
, loff_t off
)
5912 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5916 buf
= strstrip(buf
);
5917 err
= page_counter_memparse(buf
, "max", &max
);
5921 mutex_lock(&memcg_limit_mutex
);
5922 err
= page_counter_limit(&memcg
->swap
, max
);
5923 mutex_unlock(&memcg_limit_mutex
);
5930 static struct cftype swap_files
[] = {
5932 .name
= "swap.current",
5933 .flags
= CFTYPE_NOT_ON_ROOT
,
5934 .read_u64
= swap_current_read
,
5938 .flags
= CFTYPE_NOT_ON_ROOT
,
5939 .seq_show
= swap_max_show
,
5940 .write
= swap_max_write
,
5945 static struct cftype memsw_cgroup_files
[] = {
5947 .name
= "memsw.usage_in_bytes",
5948 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5949 .read_u64
= mem_cgroup_read_u64
,
5952 .name
= "memsw.max_usage_in_bytes",
5953 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5954 .write
= mem_cgroup_reset
,
5955 .read_u64
= mem_cgroup_read_u64
,
5958 .name
= "memsw.limit_in_bytes",
5959 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5960 .write
= mem_cgroup_write
,
5961 .read_u64
= mem_cgroup_read_u64
,
5964 .name
= "memsw.failcnt",
5965 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5966 .write
= mem_cgroup_reset
,
5967 .read_u64
= mem_cgroup_read_u64
,
5969 { }, /* terminate */
5972 static int __init
mem_cgroup_swap_init(void)
5974 if (!mem_cgroup_disabled() && really_do_swap_account
) {
5975 do_swap_account
= 1;
5976 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
5978 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
5979 memsw_cgroup_files
));
5983 subsys_initcall(mem_cgroup_swap_init
);
5985 #endif /* CONFIG_MEMCG_SWAP */