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 mem_cgroup
*from
;
211 struct mem_cgroup
*to
;
213 unsigned long precharge
;
214 unsigned long moved_charge
;
215 unsigned long moved_swap
;
216 struct task_struct
*moving_task
; /* a task moving charges */
217 wait_queue_head_t waitq
; /* a waitq for other context */
219 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
220 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
224 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
225 * limit reclaim to prevent infinite loops, if they ever occur.
227 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
228 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
231 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
232 MEM_CGROUP_CHARGE_TYPE_ANON
,
233 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
234 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
238 /* for encoding cft->private value on file */
247 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
248 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
249 #define MEMFILE_ATTR(val) ((val) & 0xffff)
250 /* Used for OOM nofiier */
251 #define OOM_CONTROL (0)
253 /* Some nice accessors for the vmpressure. */
254 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
257 memcg
= root_mem_cgroup
;
258 return &memcg
->vmpressure
;
261 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
263 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
266 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
268 return (memcg
== root_mem_cgroup
);
273 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
274 * The main reason for not using cgroup id for this:
275 * this works better in sparse environments, where we have a lot of memcgs,
276 * but only a few kmem-limited. Or also, if we have, for instance, 200
277 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
278 * 200 entry array for that.
280 * The current size of the caches array is stored in memcg_nr_cache_ids. It
281 * will double each time we have to increase it.
283 static DEFINE_IDA(memcg_cache_ida
);
284 int memcg_nr_cache_ids
;
286 /* Protects memcg_nr_cache_ids */
287 static DECLARE_RWSEM(memcg_cache_ids_sem
);
289 void memcg_get_cache_ids(void)
291 down_read(&memcg_cache_ids_sem
);
294 void memcg_put_cache_ids(void)
296 up_read(&memcg_cache_ids_sem
);
300 * MIN_SIZE is different than 1, because we would like to avoid going through
301 * the alloc/free process all the time. In a small machine, 4 kmem-limited
302 * cgroups is a reasonable guess. In the future, it could be a parameter or
303 * tunable, but that is strictly not necessary.
305 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
306 * this constant directly from cgroup, but it is understandable that this is
307 * better kept as an internal representation in cgroup.c. In any case, the
308 * cgrp_id space is not getting any smaller, and we don't have to necessarily
309 * increase ours as well if it increases.
311 #define MEMCG_CACHES_MIN_SIZE 4
312 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
315 * A lot of the calls to the cache allocation functions are expected to be
316 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
317 * conditional to this static branch, we'll have to allow modules that does
318 * kmem_cache_alloc and the such to see this symbol as well
320 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
321 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
323 #endif /* !CONFIG_SLOB */
325 static struct mem_cgroup_per_zone
*
326 mem_cgroup_zone_zoneinfo(struct mem_cgroup
*memcg
, struct zone
*zone
)
328 int nid
= zone_to_nid(zone
);
329 int zid
= zone_idx(zone
);
331 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
335 * mem_cgroup_css_from_page - css of the memcg associated with a page
336 * @page: page of interest
338 * If memcg is bound to the default hierarchy, css of the memcg associated
339 * with @page is returned. The returned css remains associated with @page
340 * until it is released.
342 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
345 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
347 struct mem_cgroup
*memcg
;
349 memcg
= page
->mem_cgroup
;
351 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
352 memcg
= root_mem_cgroup
;
358 * page_cgroup_ino - return inode number of the memcg a page is charged to
361 * Look up the closest online ancestor of the memory cgroup @page is charged to
362 * and return its inode number or 0 if @page is not charged to any cgroup. It
363 * is safe to call this function without holding a reference to @page.
365 * Note, this function is inherently racy, because there is nothing to prevent
366 * the cgroup inode from getting torn down and potentially reallocated a moment
367 * after page_cgroup_ino() returns, so it only should be used by callers that
368 * do not care (such as procfs interfaces).
370 ino_t
page_cgroup_ino(struct page
*page
)
372 struct mem_cgroup
*memcg
;
373 unsigned long ino
= 0;
376 memcg
= READ_ONCE(page
->mem_cgroup
);
377 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
378 memcg
= parent_mem_cgroup(memcg
);
380 ino
= cgroup_ino(memcg
->css
.cgroup
);
385 static struct mem_cgroup_per_zone
*
386 mem_cgroup_page_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
388 int nid
= page_to_nid(page
);
389 int zid
= page_zonenum(page
);
391 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
394 static struct mem_cgroup_tree_per_zone
*
395 soft_limit_tree_node_zone(int nid
, int zid
)
397 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
400 static struct mem_cgroup_tree_per_zone
*
401 soft_limit_tree_from_page(struct page
*page
)
403 int nid
= page_to_nid(page
);
404 int zid
= page_zonenum(page
);
406 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
409 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone
*mz
,
410 struct mem_cgroup_tree_per_zone
*mctz
,
411 unsigned long new_usage_in_excess
)
413 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
414 struct rb_node
*parent
= NULL
;
415 struct mem_cgroup_per_zone
*mz_node
;
420 mz
->usage_in_excess
= new_usage_in_excess
;
421 if (!mz
->usage_in_excess
)
425 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
427 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
430 * We can't avoid mem cgroups that are over their soft
431 * limit by the same amount
433 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
436 rb_link_node(&mz
->tree_node
, parent
, p
);
437 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
441 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
442 struct mem_cgroup_tree_per_zone
*mctz
)
446 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
450 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
451 struct mem_cgroup_tree_per_zone
*mctz
)
455 spin_lock_irqsave(&mctz
->lock
, flags
);
456 __mem_cgroup_remove_exceeded(mz
, mctz
);
457 spin_unlock_irqrestore(&mctz
->lock
, flags
);
460 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
462 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
463 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
464 unsigned long excess
= 0;
466 if (nr_pages
> soft_limit
)
467 excess
= nr_pages
- soft_limit
;
472 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
474 unsigned long excess
;
475 struct mem_cgroup_per_zone
*mz
;
476 struct mem_cgroup_tree_per_zone
*mctz
;
478 mctz
= soft_limit_tree_from_page(page
);
480 * Necessary to update all ancestors when hierarchy is used.
481 * because their event counter is not touched.
483 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
484 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
485 excess
= soft_limit_excess(memcg
);
487 * We have to update the tree if mz is on RB-tree or
488 * mem is over its softlimit.
490 if (excess
|| mz
->on_tree
) {
493 spin_lock_irqsave(&mctz
->lock
, flags
);
494 /* if on-tree, remove it */
496 __mem_cgroup_remove_exceeded(mz
, mctz
);
498 * Insert again. mz->usage_in_excess will be updated.
499 * If excess is 0, no tree ops.
501 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
502 spin_unlock_irqrestore(&mctz
->lock
, flags
);
507 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
509 struct mem_cgroup_tree_per_zone
*mctz
;
510 struct mem_cgroup_per_zone
*mz
;
514 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
515 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
516 mctz
= soft_limit_tree_node_zone(nid
, zid
);
517 mem_cgroup_remove_exceeded(mz
, mctz
);
522 static struct mem_cgroup_per_zone
*
523 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
525 struct rb_node
*rightmost
= NULL
;
526 struct mem_cgroup_per_zone
*mz
;
530 rightmost
= rb_last(&mctz
->rb_root
);
532 goto done
; /* Nothing to reclaim from */
534 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
536 * Remove the node now but someone else can add it back,
537 * we will to add it back at the end of reclaim to its correct
538 * position in the tree.
540 __mem_cgroup_remove_exceeded(mz
, mctz
);
541 if (!soft_limit_excess(mz
->memcg
) ||
542 !css_tryget_online(&mz
->memcg
->css
))
548 static struct mem_cgroup_per_zone
*
549 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
551 struct mem_cgroup_per_zone
*mz
;
553 spin_lock_irq(&mctz
->lock
);
554 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
555 spin_unlock_irq(&mctz
->lock
);
560 * Return page count for single (non recursive) @memcg.
562 * Implementation Note: reading percpu statistics for memcg.
564 * Both of vmstat[] and percpu_counter has threshold and do periodic
565 * synchronization to implement "quick" read. There are trade-off between
566 * reading cost and precision of value. Then, we may have a chance to implement
567 * a periodic synchronization of counter in memcg's counter.
569 * But this _read() function is used for user interface now. The user accounts
570 * memory usage by memory cgroup and he _always_ requires exact value because
571 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
572 * have to visit all online cpus and make sum. So, for now, unnecessary
573 * synchronization is not implemented. (just implemented for cpu hotplug)
575 * If there are kernel internal actions which can make use of some not-exact
576 * value, and reading all cpu value can be performance bottleneck in some
577 * common workload, threshold and synchronization as vmstat[] should be
581 mem_cgroup_read_stat(struct mem_cgroup
*memcg
, enum mem_cgroup_stat_index idx
)
586 /* Per-cpu values can be negative, use a signed accumulator */
587 for_each_possible_cpu(cpu
)
588 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
590 * Summing races with updates, so val may be negative. Avoid exposing
591 * transient negative values.
598 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
599 enum mem_cgroup_events_index idx
)
601 unsigned long val
= 0;
604 for_each_possible_cpu(cpu
)
605 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
609 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
611 bool compound
, int nr_pages
)
614 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
615 * counted as CACHE even if it's on ANON LRU.
618 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
621 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
625 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
626 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
630 /* pagein of a big page is an event. So, ignore page size */
632 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
634 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
635 nr_pages
= -nr_pages
; /* for event */
638 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
641 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
643 unsigned int lru_mask
)
645 unsigned long nr
= 0;
648 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
650 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
651 struct mem_cgroup_per_zone
*mz
;
655 if (!(BIT(lru
) & lru_mask
))
657 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
658 nr
+= mz
->lru_size
[lru
];
664 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
665 unsigned int lru_mask
)
667 unsigned long nr
= 0;
670 for_each_node_state(nid
, N_MEMORY
)
671 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
675 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
676 enum mem_cgroup_events_target target
)
678 unsigned long val
, next
;
680 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
681 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
682 /* from time_after() in jiffies.h */
683 if ((long)next
- (long)val
< 0) {
685 case MEM_CGROUP_TARGET_THRESH
:
686 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
688 case MEM_CGROUP_TARGET_SOFTLIMIT
:
689 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
691 case MEM_CGROUP_TARGET_NUMAINFO
:
692 next
= val
+ NUMAINFO_EVENTS_TARGET
;
697 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
704 * Check events in order.
707 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
709 /* threshold event is triggered in finer grain than soft limit */
710 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
711 MEM_CGROUP_TARGET_THRESH
))) {
713 bool do_numainfo __maybe_unused
;
715 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
716 MEM_CGROUP_TARGET_SOFTLIMIT
);
718 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
719 MEM_CGROUP_TARGET_NUMAINFO
);
721 mem_cgroup_threshold(memcg
);
722 if (unlikely(do_softlimit
))
723 mem_cgroup_update_tree(memcg
, page
);
725 if (unlikely(do_numainfo
))
726 atomic_inc(&memcg
->numainfo_events
);
731 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
734 * mm_update_next_owner() may clear mm->owner to NULL
735 * if it races with swapoff, page migration, etc.
736 * So this can be called with p == NULL.
741 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
743 EXPORT_SYMBOL(mem_cgroup_from_task
);
745 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
747 struct mem_cgroup
*memcg
= NULL
;
752 * Page cache insertions can happen withou an
753 * actual mm context, e.g. during disk probing
754 * on boot, loopback IO, acct() writes etc.
757 memcg
= root_mem_cgroup
;
759 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
760 if (unlikely(!memcg
))
761 memcg
= root_mem_cgroup
;
763 } while (!css_tryget_online(&memcg
->css
));
769 * mem_cgroup_iter - iterate over memory cgroup hierarchy
770 * @root: hierarchy root
771 * @prev: previously returned memcg, NULL on first invocation
772 * @reclaim: cookie for shared reclaim walks, NULL for full walks
774 * Returns references to children of the hierarchy below @root, or
775 * @root itself, or %NULL after a full round-trip.
777 * Caller must pass the return value in @prev on subsequent
778 * invocations for reference counting, or use mem_cgroup_iter_break()
779 * to cancel a hierarchy walk before the round-trip is complete.
781 * Reclaimers can specify a zone and a priority level in @reclaim to
782 * divide up the memcgs in the hierarchy among all concurrent
783 * reclaimers operating on the same zone and priority.
785 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
786 struct mem_cgroup
*prev
,
787 struct mem_cgroup_reclaim_cookie
*reclaim
)
789 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
790 struct cgroup_subsys_state
*css
= NULL
;
791 struct mem_cgroup
*memcg
= NULL
;
792 struct mem_cgroup
*pos
= NULL
;
794 if (mem_cgroup_disabled())
798 root
= root_mem_cgroup
;
800 if (prev
&& !reclaim
)
803 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
812 struct mem_cgroup_per_zone
*mz
;
814 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
815 iter
= &mz
->iter
[reclaim
->priority
];
817 if (prev
&& reclaim
->generation
!= iter
->generation
)
821 pos
= READ_ONCE(iter
->position
);
822 if (!pos
|| css_tryget(&pos
->css
))
825 * css reference reached zero, so iter->position will
826 * be cleared by ->css_released. However, we should not
827 * rely on this happening soon, because ->css_released
828 * is called from a work queue, and by busy-waiting we
829 * might block it. So we clear iter->position right
832 (void)cmpxchg(&iter
->position
, pos
, NULL
);
840 css
= css_next_descendant_pre(css
, &root
->css
);
843 * Reclaimers share the hierarchy walk, and a
844 * new one might jump in right at the end of
845 * the hierarchy - make sure they see at least
846 * one group and restart from the beginning.
854 * Verify the css and acquire a reference. The root
855 * is provided by the caller, so we know it's alive
856 * and kicking, and don't take an extra reference.
858 memcg
= mem_cgroup_from_css(css
);
860 if (css
== &root
->css
)
871 * The position could have already been updated by a competing
872 * thread, so check that the value hasn't changed since we read
873 * it to avoid reclaiming from the same cgroup twice.
875 (void)cmpxchg(&iter
->position
, pos
, memcg
);
883 reclaim
->generation
= iter
->generation
;
889 if (prev
&& prev
!= root
)
896 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
897 * @root: hierarchy root
898 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
900 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
901 struct mem_cgroup
*prev
)
904 root
= root_mem_cgroup
;
905 if (prev
&& prev
!= root
)
909 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
911 struct mem_cgroup
*memcg
= dead_memcg
;
912 struct mem_cgroup_reclaim_iter
*iter
;
913 struct mem_cgroup_per_zone
*mz
;
917 while ((memcg
= parent_mem_cgroup(memcg
))) {
919 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
920 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
921 for (i
= 0; i
<= DEF_PRIORITY
; i
++) {
923 cmpxchg(&iter
->position
,
932 * Iteration constructs for visiting all cgroups (under a tree). If
933 * loops are exited prematurely (break), mem_cgroup_iter_break() must
934 * be used for reference counting.
936 #define for_each_mem_cgroup_tree(iter, root) \
937 for (iter = mem_cgroup_iter(root, NULL, NULL); \
939 iter = mem_cgroup_iter(root, iter, NULL))
941 #define for_each_mem_cgroup(iter) \
942 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
944 iter = mem_cgroup_iter(NULL, iter, NULL))
947 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
948 * @zone: zone of the wanted lruvec
949 * @memcg: memcg of the wanted lruvec
951 * Returns the lru list vector holding pages for the given @zone and
952 * @mem. This can be the global zone lruvec, if the memory controller
955 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
956 struct mem_cgroup
*memcg
)
958 struct mem_cgroup_per_zone
*mz
;
959 struct lruvec
*lruvec
;
961 if (mem_cgroup_disabled()) {
962 lruvec
= &zone
->lruvec
;
966 mz
= mem_cgroup_zone_zoneinfo(memcg
, zone
);
967 lruvec
= &mz
->lruvec
;
970 * Since a node can be onlined after the mem_cgroup was created,
971 * we have to be prepared to initialize lruvec->zone here;
972 * and if offlined then reonlined, we need to reinitialize it.
974 if (unlikely(lruvec
->zone
!= zone
))
980 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
982 * @zone: zone of the page
984 * This function is only safe when following the LRU page isolation
985 * and putback protocol: the LRU lock must be held, and the page must
986 * either be PageLRU() or the caller must have isolated/allocated it.
988 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
990 struct mem_cgroup_per_zone
*mz
;
991 struct mem_cgroup
*memcg
;
992 struct lruvec
*lruvec
;
994 if (mem_cgroup_disabled()) {
995 lruvec
= &zone
->lruvec
;
999 memcg
= page
->mem_cgroup
;
1001 * Swapcache readahead pages are added to the LRU - and
1002 * possibly migrated - before they are charged.
1005 memcg
= root_mem_cgroup
;
1007 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1008 lruvec
= &mz
->lruvec
;
1011 * Since a node can be onlined after the mem_cgroup was created,
1012 * we have to be prepared to initialize lruvec->zone here;
1013 * and if offlined then reonlined, we need to reinitialize it.
1015 if (unlikely(lruvec
->zone
!= zone
))
1016 lruvec
->zone
= zone
;
1021 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1022 * @lruvec: mem_cgroup per zone lru vector
1023 * @lru: index of lru list the page is sitting on
1024 * @nr_pages: positive when adding or negative when removing
1026 * This function must be called when a page is added to or removed from an
1029 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1032 struct mem_cgroup_per_zone
*mz
;
1033 unsigned long *lru_size
;
1035 if (mem_cgroup_disabled())
1038 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1039 lru_size
= mz
->lru_size
+ lru
;
1040 *lru_size
+= nr_pages
;
1041 VM_BUG_ON((long)(*lru_size
) < 0);
1044 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1046 struct mem_cgroup
*task_memcg
;
1047 struct task_struct
*p
;
1050 p
= find_lock_task_mm(task
);
1052 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1056 * All threads may have already detached their mm's, but the oom
1057 * killer still needs to detect if they have already been oom
1058 * killed to prevent needlessly killing additional tasks.
1061 task_memcg
= mem_cgroup_from_task(task
);
1062 css_get(&task_memcg
->css
);
1065 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1066 css_put(&task_memcg
->css
);
1071 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1072 * @memcg: the memory cgroup
1074 * Returns the maximum amount of memory @mem can be charged with, in
1077 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1079 unsigned long margin
= 0;
1080 unsigned long count
;
1081 unsigned long limit
;
1083 count
= page_counter_read(&memcg
->memory
);
1084 limit
= READ_ONCE(memcg
->memory
.limit
);
1086 margin
= limit
- count
;
1088 if (do_memsw_account()) {
1089 count
= page_counter_read(&memcg
->memsw
);
1090 limit
= READ_ONCE(memcg
->memsw
.limit
);
1092 margin
= min(margin
, limit
- count
);
1099 * A routine for checking "mem" is under move_account() or not.
1101 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1102 * moving cgroups. This is for waiting at high-memory pressure
1105 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1107 struct mem_cgroup
*from
;
1108 struct mem_cgroup
*to
;
1111 * Unlike task_move routines, we access mc.to, mc.from not under
1112 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1114 spin_lock(&mc
.lock
);
1120 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1121 mem_cgroup_is_descendant(to
, memcg
);
1123 spin_unlock(&mc
.lock
);
1127 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1129 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1130 if (mem_cgroup_under_move(memcg
)) {
1132 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1133 /* moving charge context might have finished. */
1136 finish_wait(&mc
.waitq
, &wait
);
1143 #define K(x) ((x) << (PAGE_SHIFT-10))
1145 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1146 * @memcg: The memory cgroup that went over limit
1147 * @p: Task that is going to be killed
1149 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1152 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1154 /* oom_info_lock ensures that parallel ooms do not interleave */
1155 static DEFINE_MUTEX(oom_info_lock
);
1156 struct mem_cgroup
*iter
;
1159 mutex_lock(&oom_info_lock
);
1163 pr_info("Task in ");
1164 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1165 pr_cont(" killed as a result of limit of ");
1167 pr_info("Memory limit reached of cgroup ");
1170 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1175 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1176 K((u64
)page_counter_read(&memcg
->memory
)),
1177 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1178 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1179 K((u64
)page_counter_read(&memcg
->memsw
)),
1180 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1181 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1182 K((u64
)page_counter_read(&memcg
->kmem
)),
1183 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1185 for_each_mem_cgroup_tree(iter
, memcg
) {
1186 pr_info("Memory cgroup stats for ");
1187 pr_cont_cgroup_path(iter
->css
.cgroup
);
1190 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1191 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1193 pr_cont(" %s:%luKB", mem_cgroup_stat_names
[i
],
1194 K(mem_cgroup_read_stat(iter
, i
)));
1197 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1198 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1199 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1203 mutex_unlock(&oom_info_lock
);
1207 * This function returns the number of memcg under hierarchy tree. Returns
1208 * 1(self count) if no children.
1210 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1213 struct mem_cgroup
*iter
;
1215 for_each_mem_cgroup_tree(iter
, memcg
)
1221 * Return the memory (and swap, if configured) limit for a memcg.
1223 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1225 unsigned long limit
;
1227 limit
= memcg
->memory
.limit
;
1228 if (mem_cgroup_swappiness(memcg
)) {
1229 unsigned long memsw_limit
;
1230 unsigned long swap_limit
;
1232 memsw_limit
= memcg
->memsw
.limit
;
1233 swap_limit
= memcg
->swap
.limit
;
1234 swap_limit
= min(swap_limit
, (unsigned long)total_swap_pages
);
1235 limit
= min(limit
+ swap_limit
, memsw_limit
);
1240 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1243 struct oom_control oc
= {
1246 .gfp_mask
= gfp_mask
,
1249 struct mem_cgroup
*iter
;
1250 unsigned long chosen_points
= 0;
1251 unsigned long totalpages
;
1252 unsigned int points
= 0;
1253 struct task_struct
*chosen
= NULL
;
1255 mutex_lock(&oom_lock
);
1258 * If current has a pending SIGKILL or is exiting, then automatically
1259 * select it. The goal is to allow it to allocate so that it may
1260 * quickly exit and free its memory.
1262 if (fatal_signal_pending(current
) || task_will_free_mem(current
)) {
1263 mark_oom_victim(current
);
1267 check_panic_on_oom(&oc
, CONSTRAINT_MEMCG
, memcg
);
1268 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1269 for_each_mem_cgroup_tree(iter
, memcg
) {
1270 struct css_task_iter it
;
1271 struct task_struct
*task
;
1273 css_task_iter_start(&iter
->css
, &it
);
1274 while ((task
= css_task_iter_next(&it
))) {
1275 switch (oom_scan_process_thread(&oc
, task
, totalpages
)) {
1276 case OOM_SCAN_SELECT
:
1278 put_task_struct(chosen
);
1280 chosen_points
= ULONG_MAX
;
1281 get_task_struct(chosen
);
1283 case OOM_SCAN_CONTINUE
:
1285 case OOM_SCAN_ABORT
:
1286 css_task_iter_end(&it
);
1287 mem_cgroup_iter_break(memcg
, iter
);
1289 put_task_struct(chosen
);
1294 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1295 if (!points
|| points
< chosen_points
)
1297 /* Prefer thread group leaders for display purposes */
1298 if (points
== chosen_points
&&
1299 thread_group_leader(chosen
))
1303 put_task_struct(chosen
);
1305 chosen_points
= points
;
1306 get_task_struct(chosen
);
1308 css_task_iter_end(&it
);
1312 points
= chosen_points
* 1000 / totalpages
;
1313 oom_kill_process(&oc
, chosen
, points
, totalpages
, memcg
,
1314 "Memory cgroup out of memory");
1317 mutex_unlock(&oom_lock
);
1320 #if MAX_NUMNODES > 1
1323 * test_mem_cgroup_node_reclaimable
1324 * @memcg: the target memcg
1325 * @nid: the node ID to be checked.
1326 * @noswap : specify true here if the user wants flle only information.
1328 * This function returns whether the specified memcg contains any
1329 * reclaimable pages on a node. Returns true if there are any reclaimable
1330 * pages in the node.
1332 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1333 int nid
, bool noswap
)
1335 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1337 if (noswap
|| !total_swap_pages
)
1339 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1346 * Always updating the nodemask is not very good - even if we have an empty
1347 * list or the wrong list here, we can start from some node and traverse all
1348 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1351 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1355 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1356 * pagein/pageout changes since the last update.
1358 if (!atomic_read(&memcg
->numainfo_events
))
1360 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1363 /* make a nodemask where this memcg uses memory from */
1364 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1366 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1368 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1369 node_clear(nid
, memcg
->scan_nodes
);
1372 atomic_set(&memcg
->numainfo_events
, 0);
1373 atomic_set(&memcg
->numainfo_updating
, 0);
1377 * Selecting a node where we start reclaim from. Because what we need is just
1378 * reducing usage counter, start from anywhere is O,K. Considering
1379 * memory reclaim from current node, there are pros. and cons.
1381 * Freeing memory from current node means freeing memory from a node which
1382 * we'll use or we've used. So, it may make LRU bad. And if several threads
1383 * hit limits, it will see a contention on a node. But freeing from remote
1384 * node means more costs for memory reclaim because of memory latency.
1386 * Now, we use round-robin. Better algorithm is welcomed.
1388 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1392 mem_cgroup_may_update_nodemask(memcg
);
1393 node
= memcg
->last_scanned_node
;
1395 node
= next_node(node
, memcg
->scan_nodes
);
1396 if (node
== MAX_NUMNODES
)
1397 node
= first_node(memcg
->scan_nodes
);
1399 * We call this when we hit limit, not when pages are added to LRU.
1400 * No LRU may hold pages because all pages are UNEVICTABLE or
1401 * memcg is too small and all pages are not on LRU. In that case,
1402 * we use curret node.
1404 if (unlikely(node
== MAX_NUMNODES
))
1405 node
= numa_node_id();
1407 memcg
->last_scanned_node
= node
;
1411 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1417 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1420 unsigned long *total_scanned
)
1422 struct mem_cgroup
*victim
= NULL
;
1425 unsigned long excess
;
1426 unsigned long nr_scanned
;
1427 struct mem_cgroup_reclaim_cookie reclaim
= {
1432 excess
= soft_limit_excess(root_memcg
);
1435 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1440 * If we have not been able to reclaim
1441 * anything, it might because there are
1442 * no reclaimable pages under this hierarchy
1447 * We want to do more targeted reclaim.
1448 * excess >> 2 is not to excessive so as to
1449 * reclaim too much, nor too less that we keep
1450 * coming back to reclaim from this cgroup
1452 if (total
>= (excess
>> 2) ||
1453 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1458 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1460 *total_scanned
+= nr_scanned
;
1461 if (!soft_limit_excess(root_memcg
))
1464 mem_cgroup_iter_break(root_memcg
, victim
);
1468 #ifdef CONFIG_LOCKDEP
1469 static struct lockdep_map memcg_oom_lock_dep_map
= {
1470 .name
= "memcg_oom_lock",
1474 static DEFINE_SPINLOCK(memcg_oom_lock
);
1477 * Check OOM-Killer is already running under our hierarchy.
1478 * If someone is running, return false.
1480 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1482 struct mem_cgroup
*iter
, *failed
= NULL
;
1484 spin_lock(&memcg_oom_lock
);
1486 for_each_mem_cgroup_tree(iter
, memcg
) {
1487 if (iter
->oom_lock
) {
1489 * this subtree of our hierarchy is already locked
1490 * so we cannot give a lock.
1493 mem_cgroup_iter_break(memcg
, iter
);
1496 iter
->oom_lock
= true;
1501 * OK, we failed to lock the whole subtree so we have
1502 * to clean up what we set up to the failing subtree
1504 for_each_mem_cgroup_tree(iter
, memcg
) {
1505 if (iter
== failed
) {
1506 mem_cgroup_iter_break(memcg
, iter
);
1509 iter
->oom_lock
= false;
1512 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1514 spin_unlock(&memcg_oom_lock
);
1519 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1521 struct mem_cgroup
*iter
;
1523 spin_lock(&memcg_oom_lock
);
1524 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1525 for_each_mem_cgroup_tree(iter
, memcg
)
1526 iter
->oom_lock
= false;
1527 spin_unlock(&memcg_oom_lock
);
1530 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1532 struct mem_cgroup
*iter
;
1534 spin_lock(&memcg_oom_lock
);
1535 for_each_mem_cgroup_tree(iter
, memcg
)
1537 spin_unlock(&memcg_oom_lock
);
1540 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1542 struct mem_cgroup
*iter
;
1545 * When a new child is created while the hierarchy is under oom,
1546 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1548 spin_lock(&memcg_oom_lock
);
1549 for_each_mem_cgroup_tree(iter
, memcg
)
1550 if (iter
->under_oom
> 0)
1552 spin_unlock(&memcg_oom_lock
);
1555 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1557 struct oom_wait_info
{
1558 struct mem_cgroup
*memcg
;
1562 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1563 unsigned mode
, int sync
, void *arg
)
1565 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1566 struct mem_cgroup
*oom_wait_memcg
;
1567 struct oom_wait_info
*oom_wait_info
;
1569 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1570 oom_wait_memcg
= oom_wait_info
->memcg
;
1572 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1573 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1575 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1578 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1581 * For the following lockless ->under_oom test, the only required
1582 * guarantee is that it must see the state asserted by an OOM when
1583 * this function is called as a result of userland actions
1584 * triggered by the notification of the OOM. This is trivially
1585 * achieved by invoking mem_cgroup_mark_under_oom() before
1586 * triggering notification.
1588 if (memcg
&& memcg
->under_oom
)
1589 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1592 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1594 if (!current
->memcg_may_oom
)
1597 * We are in the middle of the charge context here, so we
1598 * don't want to block when potentially sitting on a callstack
1599 * that holds all kinds of filesystem and mm locks.
1601 * Also, the caller may handle a failed allocation gracefully
1602 * (like optional page cache readahead) and so an OOM killer
1603 * invocation might not even be necessary.
1605 * That's why we don't do anything here except remember the
1606 * OOM context and then deal with it at the end of the page
1607 * fault when the stack is unwound, the locks are released,
1608 * and when we know whether the fault was overall successful.
1610 css_get(&memcg
->css
);
1611 current
->memcg_in_oom
= memcg
;
1612 current
->memcg_oom_gfp_mask
= mask
;
1613 current
->memcg_oom_order
= order
;
1617 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1618 * @handle: actually kill/wait or just clean up the OOM state
1620 * This has to be called at the end of a page fault if the memcg OOM
1621 * handler was enabled.
1623 * Memcg supports userspace OOM handling where failed allocations must
1624 * sleep on a waitqueue until the userspace task resolves the
1625 * situation. Sleeping directly in the charge context with all kinds
1626 * of locks held is not a good idea, instead we remember an OOM state
1627 * in the task and mem_cgroup_oom_synchronize() has to be called at
1628 * the end of the page fault to complete the OOM handling.
1630 * Returns %true if an ongoing memcg OOM situation was detected and
1631 * completed, %false otherwise.
1633 bool mem_cgroup_oom_synchronize(bool handle
)
1635 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1636 struct oom_wait_info owait
;
1639 /* OOM is global, do not handle */
1643 if (!handle
|| oom_killer_disabled
)
1646 owait
.memcg
= memcg
;
1647 owait
.wait
.flags
= 0;
1648 owait
.wait
.func
= memcg_oom_wake_function
;
1649 owait
.wait
.private = current
;
1650 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1652 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1653 mem_cgroup_mark_under_oom(memcg
);
1655 locked
= mem_cgroup_oom_trylock(memcg
);
1658 mem_cgroup_oom_notify(memcg
);
1660 if (locked
&& !memcg
->oom_kill_disable
) {
1661 mem_cgroup_unmark_under_oom(memcg
);
1662 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1663 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1664 current
->memcg_oom_order
);
1667 mem_cgroup_unmark_under_oom(memcg
);
1668 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1672 mem_cgroup_oom_unlock(memcg
);
1674 * There is no guarantee that an OOM-lock contender
1675 * sees the wakeups triggered by the OOM kill
1676 * uncharges. Wake any sleepers explicitely.
1678 memcg_oom_recover(memcg
);
1681 current
->memcg_in_oom
= NULL
;
1682 css_put(&memcg
->css
);
1687 * lock_page_memcg - lock a page->mem_cgroup binding
1690 * This function protects unlocked LRU pages from being moved to
1691 * another cgroup and stabilizes their page->mem_cgroup binding.
1693 struct mem_cgroup
*lock_page_memcg(struct page
*page
)
1695 struct mem_cgroup
*memcg
;
1696 unsigned long flags
;
1699 * The RCU lock is held throughout the transaction. The fast
1700 * path can get away without acquiring the memcg->move_lock
1701 * because page moving starts with an RCU grace period.
1703 * The RCU lock also protects the memcg from being freed when
1704 * the page state that is going to change is the only thing
1705 * preventing the page from being uncharged.
1706 * E.g. end-writeback clearing PageWriteback(), which allows
1707 * migration to go ahead and uncharge the page before the
1708 * account transaction might be complete.
1712 if (mem_cgroup_disabled())
1715 memcg
= page
->mem_cgroup
;
1716 if (unlikely(!memcg
))
1719 if (atomic_read(&memcg
->moving_account
) <= 0)
1722 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1723 if (memcg
!= page
->mem_cgroup
) {
1724 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1729 * When charge migration first begins, we can have locked and
1730 * unlocked page stat updates happening concurrently. Track
1731 * the task who has the lock for unlock_page_memcg().
1733 memcg
->move_lock_task
= current
;
1734 memcg
->move_lock_flags
= flags
;
1738 EXPORT_SYMBOL(lock_page_memcg
);
1741 * unlock_page_memcg - unlock a page->mem_cgroup binding
1742 * @memcg: the memcg returned by lock_page_memcg()
1744 void unlock_page_memcg(struct mem_cgroup
*memcg
)
1746 if (memcg
&& memcg
->move_lock_task
== current
) {
1747 unsigned long flags
= memcg
->move_lock_flags
;
1749 memcg
->move_lock_task
= NULL
;
1750 memcg
->move_lock_flags
= 0;
1752 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1757 EXPORT_SYMBOL(unlock_page_memcg
);
1760 * size of first charge trial. "32" comes from vmscan.c's magic value.
1761 * TODO: maybe necessary to use big numbers in big irons.
1763 #define CHARGE_BATCH 32U
1764 struct memcg_stock_pcp
{
1765 struct mem_cgroup
*cached
; /* this never be root cgroup */
1766 unsigned int nr_pages
;
1767 struct work_struct work
;
1768 unsigned long flags
;
1769 #define FLUSHING_CACHED_CHARGE 0
1771 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1772 static DEFINE_MUTEX(percpu_charge_mutex
);
1775 * consume_stock: Try to consume stocked charge on this cpu.
1776 * @memcg: memcg to consume from.
1777 * @nr_pages: how many pages to charge.
1779 * The charges will only happen if @memcg matches the current cpu's memcg
1780 * stock, and at least @nr_pages are available in that stock. Failure to
1781 * service an allocation will refill the stock.
1783 * returns true if successful, false otherwise.
1785 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1787 struct memcg_stock_pcp
*stock
;
1790 if (nr_pages
> CHARGE_BATCH
)
1793 stock
= &get_cpu_var(memcg_stock
);
1794 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
1795 stock
->nr_pages
-= nr_pages
;
1798 put_cpu_var(memcg_stock
);
1803 * Returns stocks cached in percpu and reset cached information.
1805 static void drain_stock(struct memcg_stock_pcp
*stock
)
1807 struct mem_cgroup
*old
= stock
->cached
;
1809 if (stock
->nr_pages
) {
1810 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
1811 if (do_memsw_account())
1812 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
1813 css_put_many(&old
->css
, stock
->nr_pages
);
1814 stock
->nr_pages
= 0;
1816 stock
->cached
= NULL
;
1820 * This must be called under preempt disabled or must be called by
1821 * a thread which is pinned to local cpu.
1823 static void drain_local_stock(struct work_struct
*dummy
)
1825 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
1827 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
1831 * Cache charges(val) to local per_cpu area.
1832 * This will be consumed by consume_stock() function, later.
1834 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1836 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
1838 if (stock
->cached
!= memcg
) { /* reset if necessary */
1840 stock
->cached
= memcg
;
1842 stock
->nr_pages
+= nr_pages
;
1843 put_cpu_var(memcg_stock
);
1847 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1848 * of the hierarchy under it.
1850 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
1854 /* If someone's already draining, avoid adding running more workers. */
1855 if (!mutex_trylock(&percpu_charge_mutex
))
1857 /* Notify other cpus that system-wide "drain" is running */
1860 for_each_online_cpu(cpu
) {
1861 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1862 struct mem_cgroup
*memcg
;
1864 memcg
= stock
->cached
;
1865 if (!memcg
|| !stock
->nr_pages
)
1867 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
1869 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
1871 drain_local_stock(&stock
->work
);
1873 schedule_work_on(cpu
, &stock
->work
);
1878 mutex_unlock(&percpu_charge_mutex
);
1881 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
1882 unsigned long action
,
1885 int cpu
= (unsigned long)hcpu
;
1886 struct memcg_stock_pcp
*stock
;
1888 if (action
== CPU_ONLINE
)
1891 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
1894 stock
= &per_cpu(memcg_stock
, cpu
);
1899 static void reclaim_high(struct mem_cgroup
*memcg
,
1900 unsigned int nr_pages
,
1904 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
1906 mem_cgroup_events(memcg
, MEMCG_HIGH
, 1);
1907 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
1908 } while ((memcg
= parent_mem_cgroup(memcg
)));
1911 static void high_work_func(struct work_struct
*work
)
1913 struct mem_cgroup
*memcg
;
1915 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
1916 reclaim_high(memcg
, CHARGE_BATCH
, GFP_KERNEL
);
1920 * Scheduled by try_charge() to be executed from the userland return path
1921 * and reclaims memory over the high limit.
1923 void mem_cgroup_handle_over_high(void)
1925 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
1926 struct mem_cgroup
*memcg
;
1928 if (likely(!nr_pages
))
1931 memcg
= get_mem_cgroup_from_mm(current
->mm
);
1932 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
1933 css_put(&memcg
->css
);
1934 current
->memcg_nr_pages_over_high
= 0;
1937 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1938 unsigned int nr_pages
)
1940 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
1941 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1942 struct mem_cgroup
*mem_over_limit
;
1943 struct page_counter
*counter
;
1944 unsigned long nr_reclaimed
;
1945 bool may_swap
= true;
1946 bool drained
= false;
1948 if (mem_cgroup_is_root(memcg
))
1951 if (consume_stock(memcg
, nr_pages
))
1954 if (!do_memsw_account() ||
1955 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
1956 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
1958 if (do_memsw_account())
1959 page_counter_uncharge(&memcg
->memsw
, batch
);
1960 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
1962 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
1966 if (batch
> nr_pages
) {
1972 * Unlike in global OOM situations, memcg is not in a physical
1973 * memory shortage. Allow dying and OOM-killed tasks to
1974 * bypass the last charges so that they can exit quickly and
1975 * free their memory.
1977 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
1978 fatal_signal_pending(current
) ||
1979 current
->flags
& PF_EXITING
))
1982 if (unlikely(task_in_memcg_oom(current
)))
1985 if (!gfpflags_allow_blocking(gfp_mask
))
1988 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
1990 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
1991 gfp_mask
, may_swap
);
1993 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
1997 drain_all_stock(mem_over_limit
);
2002 if (gfp_mask
& __GFP_NORETRY
)
2005 * Even though the limit is exceeded at this point, reclaim
2006 * may have been able to free some pages. Retry the charge
2007 * before killing the task.
2009 * Only for regular pages, though: huge pages are rather
2010 * unlikely to succeed so close to the limit, and we fall back
2011 * to regular pages anyway in case of failure.
2013 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2016 * At task move, charge accounts can be doubly counted. So, it's
2017 * better to wait until the end of task_move if something is going on.
2019 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2025 if (gfp_mask
& __GFP_NOFAIL
)
2028 if (fatal_signal_pending(current
))
2031 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
2033 mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2034 get_order(nr_pages
* PAGE_SIZE
));
2036 if (!(gfp_mask
& __GFP_NOFAIL
))
2040 * The allocation either can't fail or will lead to more memory
2041 * being freed very soon. Allow memory usage go over the limit
2042 * temporarily by force charging it.
2044 page_counter_charge(&memcg
->memory
, nr_pages
);
2045 if (do_memsw_account())
2046 page_counter_charge(&memcg
->memsw
, nr_pages
);
2047 css_get_many(&memcg
->css
, nr_pages
);
2052 css_get_many(&memcg
->css
, batch
);
2053 if (batch
> nr_pages
)
2054 refill_stock(memcg
, batch
- nr_pages
);
2057 * If the hierarchy is above the normal consumption range, schedule
2058 * reclaim on returning to userland. We can perform reclaim here
2059 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2060 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2061 * not recorded as it most likely matches current's and won't
2062 * change in the meantime. As high limit is checked again before
2063 * reclaim, the cost of mismatch is negligible.
2066 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2067 /* Don't bother a random interrupted task */
2068 if (in_interrupt()) {
2069 schedule_work(&memcg
->high_work
);
2072 current
->memcg_nr_pages_over_high
+= batch
;
2073 set_notify_resume(current
);
2076 } while ((memcg
= parent_mem_cgroup(memcg
)));
2081 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2083 if (mem_cgroup_is_root(memcg
))
2086 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2087 if (do_memsw_account())
2088 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2090 css_put_many(&memcg
->css
, nr_pages
);
2093 static void lock_page_lru(struct page
*page
, int *isolated
)
2095 struct zone
*zone
= page_zone(page
);
2097 spin_lock_irq(&zone
->lru_lock
);
2098 if (PageLRU(page
)) {
2099 struct lruvec
*lruvec
;
2101 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2103 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2109 static void unlock_page_lru(struct page
*page
, int isolated
)
2111 struct zone
*zone
= page_zone(page
);
2114 struct lruvec
*lruvec
;
2116 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2117 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2119 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2121 spin_unlock_irq(&zone
->lru_lock
);
2124 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2129 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2132 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2133 * may already be on some other mem_cgroup's LRU. Take care of it.
2136 lock_page_lru(page
, &isolated
);
2139 * Nobody should be changing or seriously looking at
2140 * page->mem_cgroup at this point:
2142 * - the page is uncharged
2144 * - the page is off-LRU
2146 * - an anonymous fault has exclusive page access, except for
2147 * a locked page table
2149 * - a page cache insertion, a swapin fault, or a migration
2150 * have the page locked
2152 page
->mem_cgroup
= memcg
;
2155 unlock_page_lru(page
, isolated
);
2159 static int memcg_alloc_cache_id(void)
2164 id
= ida_simple_get(&memcg_cache_ida
,
2165 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2169 if (id
< memcg_nr_cache_ids
)
2173 * There's no space for the new id in memcg_caches arrays,
2174 * so we have to grow them.
2176 down_write(&memcg_cache_ids_sem
);
2178 size
= 2 * (id
+ 1);
2179 if (size
< MEMCG_CACHES_MIN_SIZE
)
2180 size
= MEMCG_CACHES_MIN_SIZE
;
2181 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2182 size
= MEMCG_CACHES_MAX_SIZE
;
2184 err
= memcg_update_all_caches(size
);
2186 err
= memcg_update_all_list_lrus(size
);
2188 memcg_nr_cache_ids
= size
;
2190 up_write(&memcg_cache_ids_sem
);
2193 ida_simple_remove(&memcg_cache_ida
, id
);
2199 static void memcg_free_cache_id(int id
)
2201 ida_simple_remove(&memcg_cache_ida
, id
);
2204 struct memcg_kmem_cache_create_work
{
2205 struct mem_cgroup
*memcg
;
2206 struct kmem_cache
*cachep
;
2207 struct work_struct work
;
2210 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2212 struct memcg_kmem_cache_create_work
*cw
=
2213 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2214 struct mem_cgroup
*memcg
= cw
->memcg
;
2215 struct kmem_cache
*cachep
= cw
->cachep
;
2217 memcg_create_kmem_cache(memcg
, cachep
);
2219 css_put(&memcg
->css
);
2224 * Enqueue the creation of a per-memcg kmem_cache.
2226 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2227 struct kmem_cache
*cachep
)
2229 struct memcg_kmem_cache_create_work
*cw
;
2231 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2235 css_get(&memcg
->css
);
2238 cw
->cachep
= cachep
;
2239 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2241 schedule_work(&cw
->work
);
2244 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2245 struct kmem_cache
*cachep
)
2248 * We need to stop accounting when we kmalloc, because if the
2249 * corresponding kmalloc cache is not yet created, the first allocation
2250 * in __memcg_schedule_kmem_cache_create will recurse.
2252 * However, it is better to enclose the whole function. Depending on
2253 * the debugging options enabled, INIT_WORK(), for instance, can
2254 * trigger an allocation. This too, will make us recurse. Because at
2255 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2256 * the safest choice is to do it like this, wrapping the whole function.
2258 current
->memcg_kmem_skip_account
= 1;
2259 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2260 current
->memcg_kmem_skip_account
= 0;
2264 * Return the kmem_cache we're supposed to use for a slab allocation.
2265 * We try to use the current memcg's version of the cache.
2267 * If the cache does not exist yet, if we are the first user of it,
2268 * we either create it immediately, if possible, or create it asynchronously
2270 * In the latter case, we will let the current allocation go through with
2271 * the original cache.
2273 * Can't be called in interrupt context or from kernel threads.
2274 * This function needs to be called with rcu_read_lock() held.
2276 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2278 struct mem_cgroup
*memcg
;
2279 struct kmem_cache
*memcg_cachep
;
2282 VM_BUG_ON(!is_root_cache(cachep
));
2284 if (cachep
->flags
& SLAB_ACCOUNT
)
2285 gfp
|= __GFP_ACCOUNT
;
2287 if (!(gfp
& __GFP_ACCOUNT
))
2290 if (current
->memcg_kmem_skip_account
)
2293 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2294 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2298 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2299 if (likely(memcg_cachep
))
2300 return memcg_cachep
;
2303 * If we are in a safe context (can wait, and not in interrupt
2304 * context), we could be be predictable and return right away.
2305 * This would guarantee that the allocation being performed
2306 * already belongs in the new cache.
2308 * However, there are some clashes that can arrive from locking.
2309 * For instance, because we acquire the slab_mutex while doing
2310 * memcg_create_kmem_cache, this means no further allocation
2311 * could happen with the slab_mutex held. So it's better to
2314 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2316 css_put(&memcg
->css
);
2320 void __memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2322 if (!is_root_cache(cachep
))
2323 css_put(&cachep
->memcg_params
.memcg
->css
);
2326 int __memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2327 struct mem_cgroup
*memcg
)
2329 unsigned int nr_pages
= 1 << order
;
2330 struct page_counter
*counter
;
2333 if (!memcg_kmem_online(memcg
))
2336 ret
= try_charge(memcg
, gfp
, nr_pages
);
2340 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2341 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2342 cancel_charge(memcg
, nr_pages
);
2346 page
->mem_cgroup
= memcg
;
2351 int __memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2353 struct mem_cgroup
*memcg
;
2356 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2357 ret
= __memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2358 css_put(&memcg
->css
);
2362 void __memcg_kmem_uncharge(struct page
*page
, int order
)
2364 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2365 unsigned int nr_pages
= 1 << order
;
2370 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2372 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2373 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2375 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2376 if (do_memsw_account())
2377 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2379 page
->mem_cgroup
= NULL
;
2380 css_put_many(&memcg
->css
, nr_pages
);
2382 #endif /* !CONFIG_SLOB */
2384 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2387 * Because tail pages are not marked as "used", set it. We're under
2388 * zone->lru_lock and migration entries setup in all page mappings.
2390 void mem_cgroup_split_huge_fixup(struct page
*head
)
2394 if (mem_cgroup_disabled())
2397 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2398 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2400 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2403 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2405 #ifdef CONFIG_MEMCG_SWAP
2406 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2409 int val
= (charge
) ? 1 : -1;
2410 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2414 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2415 * @entry: swap entry to be moved
2416 * @from: mem_cgroup which the entry is moved from
2417 * @to: mem_cgroup which the entry is moved to
2419 * It succeeds only when the swap_cgroup's record for this entry is the same
2420 * as the mem_cgroup's id of @from.
2422 * Returns 0 on success, -EINVAL on failure.
2424 * The caller must have charged to @to, IOW, called page_counter_charge() about
2425 * both res and memsw, and called css_get().
2427 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2428 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2430 unsigned short old_id
, new_id
;
2432 old_id
= mem_cgroup_id(from
);
2433 new_id
= mem_cgroup_id(to
);
2435 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2436 mem_cgroup_swap_statistics(from
, false);
2437 mem_cgroup_swap_statistics(to
, true);
2443 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2444 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2450 static DEFINE_MUTEX(memcg_limit_mutex
);
2452 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2453 unsigned long limit
)
2455 unsigned long curusage
;
2456 unsigned long oldusage
;
2457 bool enlarge
= false;
2462 * For keeping hierarchical_reclaim simple, how long we should retry
2463 * is depends on callers. We set our retry-count to be function
2464 * of # of children which we should visit in this loop.
2466 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2467 mem_cgroup_count_children(memcg
);
2469 oldusage
= page_counter_read(&memcg
->memory
);
2472 if (signal_pending(current
)) {
2477 mutex_lock(&memcg_limit_mutex
);
2478 if (limit
> memcg
->memsw
.limit
) {
2479 mutex_unlock(&memcg_limit_mutex
);
2483 if (limit
> memcg
->memory
.limit
)
2485 ret
= page_counter_limit(&memcg
->memory
, limit
);
2486 mutex_unlock(&memcg_limit_mutex
);
2491 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2493 curusage
= page_counter_read(&memcg
->memory
);
2494 /* Usage is reduced ? */
2495 if (curusage
>= oldusage
)
2498 oldusage
= curusage
;
2499 } while (retry_count
);
2501 if (!ret
&& enlarge
)
2502 memcg_oom_recover(memcg
);
2507 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2508 unsigned long limit
)
2510 unsigned long curusage
;
2511 unsigned long oldusage
;
2512 bool enlarge
= false;
2516 /* see mem_cgroup_resize_res_limit */
2517 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2518 mem_cgroup_count_children(memcg
);
2520 oldusage
= page_counter_read(&memcg
->memsw
);
2523 if (signal_pending(current
)) {
2528 mutex_lock(&memcg_limit_mutex
);
2529 if (limit
< memcg
->memory
.limit
) {
2530 mutex_unlock(&memcg_limit_mutex
);
2534 if (limit
> memcg
->memsw
.limit
)
2536 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2537 mutex_unlock(&memcg_limit_mutex
);
2542 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2544 curusage
= page_counter_read(&memcg
->memsw
);
2545 /* Usage is reduced ? */
2546 if (curusage
>= oldusage
)
2549 oldusage
= curusage
;
2550 } while (retry_count
);
2552 if (!ret
&& enlarge
)
2553 memcg_oom_recover(memcg
);
2558 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
2560 unsigned long *total_scanned
)
2562 unsigned long nr_reclaimed
= 0;
2563 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
2564 unsigned long reclaimed
;
2566 struct mem_cgroup_tree_per_zone
*mctz
;
2567 unsigned long excess
;
2568 unsigned long nr_scanned
;
2573 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
2575 * This loop can run a while, specially if mem_cgroup's continuously
2576 * keep exceeding their soft limit and putting the system under
2583 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2588 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
2589 gfp_mask
, &nr_scanned
);
2590 nr_reclaimed
+= reclaimed
;
2591 *total_scanned
+= nr_scanned
;
2592 spin_lock_irq(&mctz
->lock
);
2593 __mem_cgroup_remove_exceeded(mz
, mctz
);
2596 * If we failed to reclaim anything from this memory cgroup
2597 * it is time to move on to the next cgroup
2601 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2603 excess
= soft_limit_excess(mz
->memcg
);
2605 * One school of thought says that we should not add
2606 * back the node to the tree if reclaim returns 0.
2607 * But our reclaim could return 0, simply because due
2608 * to priority we are exposing a smaller subset of
2609 * memory to reclaim from. Consider this as a longer
2612 /* If excess == 0, no tree ops */
2613 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2614 spin_unlock_irq(&mctz
->lock
);
2615 css_put(&mz
->memcg
->css
);
2618 * Could not reclaim anything and there are no more
2619 * mem cgroups to try or we seem to be looping without
2620 * reclaiming anything.
2622 if (!nr_reclaimed
&&
2624 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2626 } while (!nr_reclaimed
);
2628 css_put(&next_mz
->memcg
->css
);
2629 return nr_reclaimed
;
2633 * Test whether @memcg has children, dead or alive. Note that this
2634 * function doesn't care whether @memcg has use_hierarchy enabled and
2635 * returns %true if there are child csses according to the cgroup
2636 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2638 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2643 ret
= css_next_child(NULL
, &memcg
->css
);
2649 * Reclaims as many pages from the given memcg as possible and moves
2650 * the rest to the parent.
2652 * Caller is responsible for holding css reference for memcg.
2654 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2656 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2658 /* we call try-to-free pages for make this cgroup empty */
2659 lru_add_drain_all();
2660 /* try to free all pages in this cgroup */
2661 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2664 if (signal_pending(current
))
2667 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2671 /* maybe some writeback is necessary */
2672 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2680 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
2681 char *buf
, size_t nbytes
,
2684 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2686 if (mem_cgroup_is_root(memcg
))
2688 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
2691 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
2694 return mem_cgroup_from_css(css
)->use_hierarchy
;
2697 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
2698 struct cftype
*cft
, u64 val
)
2701 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2702 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
2704 if (memcg
->use_hierarchy
== val
)
2708 * If parent's use_hierarchy is set, we can't make any modifications
2709 * in the child subtrees. If it is unset, then the change can
2710 * occur, provided the current cgroup has no children.
2712 * For the root cgroup, parent_mem is NULL, we allow value to be
2713 * set if there are no children.
2715 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
2716 (val
== 1 || val
== 0)) {
2717 if (!memcg_has_children(memcg
))
2718 memcg
->use_hierarchy
= val
;
2727 static unsigned long tree_stat(struct mem_cgroup
*memcg
,
2728 enum mem_cgroup_stat_index idx
)
2730 struct mem_cgroup
*iter
;
2731 unsigned long val
= 0;
2733 for_each_mem_cgroup_tree(iter
, memcg
)
2734 val
+= mem_cgroup_read_stat(iter
, idx
);
2739 static unsigned long tree_events(struct mem_cgroup
*memcg
,
2740 enum mem_cgroup_events_index idx
)
2742 struct mem_cgroup
*iter
;
2743 unsigned long val
= 0;
2745 for_each_mem_cgroup_tree(iter
, memcg
)
2746 val
+= mem_cgroup_read_events(iter
, idx
);
2751 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
2755 if (mem_cgroup_is_root(memcg
)) {
2756 val
= tree_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
2757 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_RSS
);
2759 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
2762 val
= page_counter_read(&memcg
->memory
);
2764 val
= page_counter_read(&memcg
->memsw
);
2777 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
2780 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2781 struct page_counter
*counter
;
2783 switch (MEMFILE_TYPE(cft
->private)) {
2785 counter
= &memcg
->memory
;
2788 counter
= &memcg
->memsw
;
2791 counter
= &memcg
->kmem
;
2794 counter
= &memcg
->tcpmem
;
2800 switch (MEMFILE_ATTR(cft
->private)) {
2802 if (counter
== &memcg
->memory
)
2803 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
2804 if (counter
== &memcg
->memsw
)
2805 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
2806 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
2808 return (u64
)counter
->limit
* PAGE_SIZE
;
2810 return (u64
)counter
->watermark
* PAGE_SIZE
;
2812 return counter
->failcnt
;
2813 case RES_SOFT_LIMIT
:
2814 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
2821 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2825 BUG_ON(memcg
->kmemcg_id
>= 0);
2826 BUG_ON(memcg
->kmem_state
);
2828 memcg_id
= memcg_alloc_cache_id();
2832 static_branch_inc(&memcg_kmem_enabled_key
);
2834 * A memory cgroup is considered kmem-online as soon as it gets
2835 * kmemcg_id. Setting the id after enabling static branching will
2836 * guarantee no one starts accounting before all call sites are
2839 memcg
->kmemcg_id
= memcg_id
;
2840 memcg
->kmem_state
= KMEM_ONLINE
;
2845 static int memcg_propagate_kmem(struct mem_cgroup
*parent
,
2846 struct mem_cgroup
*memcg
)
2850 mutex_lock(&memcg_limit_mutex
);
2852 * If the parent cgroup is not kmem-online now, it cannot be
2853 * onlined after this point, because it has at least one child
2856 if (memcg_kmem_online(parent
) ||
2857 (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nokmem
))
2858 ret
= memcg_online_kmem(memcg
);
2859 mutex_unlock(&memcg_limit_mutex
);
2863 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2865 struct cgroup_subsys_state
*css
;
2866 struct mem_cgroup
*parent
, *child
;
2869 if (memcg
->kmem_state
!= KMEM_ONLINE
)
2872 * Clear the online state before clearing memcg_caches array
2873 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2874 * guarantees that no cache will be created for this cgroup
2875 * after we are done (see memcg_create_kmem_cache()).
2877 memcg
->kmem_state
= KMEM_ALLOCATED
;
2879 memcg_deactivate_kmem_caches(memcg
);
2881 kmemcg_id
= memcg
->kmemcg_id
;
2882 BUG_ON(kmemcg_id
< 0);
2884 parent
= parent_mem_cgroup(memcg
);
2886 parent
= root_mem_cgroup
;
2889 * Change kmemcg_id of this cgroup and all its descendants to the
2890 * parent's id, and then move all entries from this cgroup's list_lrus
2891 * to ones of the parent. After we have finished, all list_lrus
2892 * corresponding to this cgroup are guaranteed to remain empty. The
2893 * ordering is imposed by list_lru_node->lock taken by
2894 * memcg_drain_all_list_lrus().
2896 css_for_each_descendant_pre(css
, &memcg
->css
) {
2897 child
= mem_cgroup_from_css(css
);
2898 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
2899 child
->kmemcg_id
= parent
->kmemcg_id
;
2900 if (!memcg
->use_hierarchy
)
2903 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
2905 memcg_free_cache_id(kmemcg_id
);
2908 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2910 /* css_alloc() failed, offlining didn't happen */
2911 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
2912 memcg_offline_kmem(memcg
);
2914 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
2915 memcg_destroy_kmem_caches(memcg
);
2916 static_branch_dec(&memcg_kmem_enabled_key
);
2917 WARN_ON(page_counter_read(&memcg
->kmem
));
2921 static int memcg_propagate_kmem(struct mem_cgroup
*parent
, struct mem_cgroup
*memcg
)
2925 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2929 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2932 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2935 #endif /* !CONFIG_SLOB */
2937 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2938 unsigned long limit
)
2942 mutex_lock(&memcg_limit_mutex
);
2943 /* Top-level cgroup doesn't propagate from root */
2944 if (!memcg_kmem_online(memcg
)) {
2945 if (cgroup_is_populated(memcg
->css
.cgroup
) ||
2946 (memcg
->use_hierarchy
&& memcg_has_children(memcg
)))
2950 ret
= memcg_online_kmem(memcg
);
2954 ret
= page_counter_limit(&memcg
->kmem
, limit
);
2956 mutex_unlock(&memcg_limit_mutex
);
2960 static int memcg_update_tcp_limit(struct mem_cgroup
*memcg
, unsigned long limit
)
2964 mutex_lock(&memcg_limit_mutex
);
2966 ret
= page_counter_limit(&memcg
->tcpmem
, limit
);
2970 if (!memcg
->tcpmem_active
) {
2972 * The active flag needs to be written after the static_key
2973 * update. This is what guarantees that the socket activation
2974 * function is the last one to run. See sock_update_memcg() for
2975 * details, and note that we don't mark any socket as belonging
2976 * to this memcg until that flag is up.
2978 * We need to do this, because static_keys will span multiple
2979 * sites, but we can't control their order. If we mark a socket
2980 * as accounted, but the accounting functions are not patched in
2981 * yet, we'll lose accounting.
2983 * We never race with the readers in sock_update_memcg(),
2984 * because when this value change, the code to process it is not
2987 static_branch_inc(&memcg_sockets_enabled_key
);
2988 memcg
->tcpmem_active
= true;
2991 mutex_unlock(&memcg_limit_mutex
);
2996 * The user of this function is...
2999 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3000 char *buf
, size_t nbytes
, loff_t off
)
3002 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3003 unsigned long nr_pages
;
3006 buf
= strstrip(buf
);
3007 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3011 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3013 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3017 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3019 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
3022 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
3025 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
3028 ret
= memcg_update_tcp_limit(memcg
, nr_pages
);
3032 case RES_SOFT_LIMIT
:
3033 memcg
->soft_limit
= nr_pages
;
3037 return ret
?: nbytes
;
3040 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3041 size_t nbytes
, loff_t off
)
3043 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3044 struct page_counter
*counter
;
3046 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3048 counter
= &memcg
->memory
;
3051 counter
= &memcg
->memsw
;
3054 counter
= &memcg
->kmem
;
3057 counter
= &memcg
->tcpmem
;
3063 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3065 page_counter_reset_watermark(counter
);
3068 counter
->failcnt
= 0;
3077 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3080 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3084 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3085 struct cftype
*cft
, u64 val
)
3087 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3089 if (val
& ~MOVE_MASK
)
3093 * No kind of locking is needed in here, because ->can_attach() will
3094 * check this value once in the beginning of the process, and then carry
3095 * on with stale data. This means that changes to this value will only
3096 * affect task migrations starting after the change.
3098 memcg
->move_charge_at_immigrate
= val
;
3102 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3103 struct cftype
*cft
, u64 val
)
3110 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3114 unsigned int lru_mask
;
3117 static const struct numa_stat stats
[] = {
3118 { "total", LRU_ALL
},
3119 { "file", LRU_ALL_FILE
},
3120 { "anon", LRU_ALL_ANON
},
3121 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3123 const struct numa_stat
*stat
;
3126 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3128 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3129 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3130 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3131 for_each_node_state(nid
, N_MEMORY
) {
3132 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3134 seq_printf(m
, " N%d=%lu", nid
, nr
);
3139 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3140 struct mem_cgroup
*iter
;
3143 for_each_mem_cgroup_tree(iter
, memcg
)
3144 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3145 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3146 for_each_node_state(nid
, N_MEMORY
) {
3148 for_each_mem_cgroup_tree(iter
, memcg
)
3149 nr
+= mem_cgroup_node_nr_lru_pages(
3150 iter
, nid
, stat
->lru_mask
);
3151 seq_printf(m
, " N%d=%lu", nid
, nr
);
3158 #endif /* CONFIG_NUMA */
3160 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3162 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3163 unsigned long memory
, memsw
;
3164 struct mem_cgroup
*mi
;
3167 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3168 MEM_CGROUP_STAT_NSTATS
);
3169 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3170 MEM_CGROUP_EVENTS_NSTATS
);
3171 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3173 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3174 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3176 seq_printf(m
, "%s %lu\n", mem_cgroup_stat_names
[i
],
3177 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3180 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3181 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3182 mem_cgroup_read_events(memcg
, i
));
3184 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3185 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3186 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3188 /* Hierarchical information */
3189 memory
= memsw
= PAGE_COUNTER_MAX
;
3190 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3191 memory
= min(memory
, mi
->memory
.limit
);
3192 memsw
= min(memsw
, mi
->memsw
.limit
);
3194 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3195 (u64
)memory
* PAGE_SIZE
);
3196 if (do_memsw_account())
3197 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3198 (u64
)memsw
* PAGE_SIZE
);
3200 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3201 unsigned long long val
= 0;
3203 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3205 for_each_mem_cgroup_tree(mi
, memcg
)
3206 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3207 seq_printf(m
, "total_%s %llu\n", mem_cgroup_stat_names
[i
], val
);
3210 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3211 unsigned long long val
= 0;
3213 for_each_mem_cgroup_tree(mi
, memcg
)
3214 val
+= mem_cgroup_read_events(mi
, i
);
3215 seq_printf(m
, "total_%s %llu\n",
3216 mem_cgroup_events_names
[i
], val
);
3219 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3220 unsigned long long val
= 0;
3222 for_each_mem_cgroup_tree(mi
, memcg
)
3223 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3224 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3227 #ifdef CONFIG_DEBUG_VM
3230 struct mem_cgroup_per_zone
*mz
;
3231 struct zone_reclaim_stat
*rstat
;
3232 unsigned long recent_rotated
[2] = {0, 0};
3233 unsigned long recent_scanned
[2] = {0, 0};
3235 for_each_online_node(nid
)
3236 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3237 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
3238 rstat
= &mz
->lruvec
.reclaim_stat
;
3240 recent_rotated
[0] += rstat
->recent_rotated
[0];
3241 recent_rotated
[1] += rstat
->recent_rotated
[1];
3242 recent_scanned
[0] += rstat
->recent_scanned
[0];
3243 recent_scanned
[1] += rstat
->recent_scanned
[1];
3245 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3246 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3247 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3248 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3255 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3258 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3260 return mem_cgroup_swappiness(memcg
);
3263 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3264 struct cftype
*cft
, u64 val
)
3266 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3272 memcg
->swappiness
= val
;
3274 vm_swappiness
= val
;
3279 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3281 struct mem_cgroup_threshold_ary
*t
;
3282 unsigned long usage
;
3287 t
= rcu_dereference(memcg
->thresholds
.primary
);
3289 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3294 usage
= mem_cgroup_usage(memcg
, swap
);
3297 * current_threshold points to threshold just below or equal to usage.
3298 * If it's not true, a threshold was crossed after last
3299 * call of __mem_cgroup_threshold().
3301 i
= t
->current_threshold
;
3304 * Iterate backward over array of thresholds starting from
3305 * current_threshold and check if a threshold is crossed.
3306 * If none of thresholds below usage is crossed, we read
3307 * only one element of the array here.
3309 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3310 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3312 /* i = current_threshold + 1 */
3316 * Iterate forward over array of thresholds starting from
3317 * current_threshold+1 and check if a threshold is crossed.
3318 * If none of thresholds above usage is crossed, we read
3319 * only one element of the array here.
3321 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3322 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3324 /* Update current_threshold */
3325 t
->current_threshold
= i
- 1;
3330 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3333 __mem_cgroup_threshold(memcg
, false);
3334 if (do_memsw_account())
3335 __mem_cgroup_threshold(memcg
, true);
3337 memcg
= parent_mem_cgroup(memcg
);
3341 static int compare_thresholds(const void *a
, const void *b
)
3343 const struct mem_cgroup_threshold
*_a
= a
;
3344 const struct mem_cgroup_threshold
*_b
= b
;
3346 if (_a
->threshold
> _b
->threshold
)
3349 if (_a
->threshold
< _b
->threshold
)
3355 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3357 struct mem_cgroup_eventfd_list
*ev
;
3359 spin_lock(&memcg_oom_lock
);
3361 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3362 eventfd_signal(ev
->eventfd
, 1);
3364 spin_unlock(&memcg_oom_lock
);
3368 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3370 struct mem_cgroup
*iter
;
3372 for_each_mem_cgroup_tree(iter
, memcg
)
3373 mem_cgroup_oom_notify_cb(iter
);
3376 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3377 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3379 struct mem_cgroup_thresholds
*thresholds
;
3380 struct mem_cgroup_threshold_ary
*new;
3381 unsigned long threshold
;
3382 unsigned long usage
;
3385 ret
= page_counter_memparse(args
, "-1", &threshold
);
3389 mutex_lock(&memcg
->thresholds_lock
);
3392 thresholds
= &memcg
->thresholds
;
3393 usage
= mem_cgroup_usage(memcg
, false);
3394 } else if (type
== _MEMSWAP
) {
3395 thresholds
= &memcg
->memsw_thresholds
;
3396 usage
= mem_cgroup_usage(memcg
, true);
3400 /* Check if a threshold crossed before adding a new one */
3401 if (thresholds
->primary
)
3402 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3404 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3406 /* Allocate memory for new array of thresholds */
3407 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3415 /* Copy thresholds (if any) to new array */
3416 if (thresholds
->primary
) {
3417 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3418 sizeof(struct mem_cgroup_threshold
));
3421 /* Add new threshold */
3422 new->entries
[size
- 1].eventfd
= eventfd
;
3423 new->entries
[size
- 1].threshold
= threshold
;
3425 /* Sort thresholds. Registering of new threshold isn't time-critical */
3426 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3427 compare_thresholds
, NULL
);
3429 /* Find current threshold */
3430 new->current_threshold
= -1;
3431 for (i
= 0; i
< size
; i
++) {
3432 if (new->entries
[i
].threshold
<= usage
) {
3434 * new->current_threshold will not be used until
3435 * rcu_assign_pointer(), so it's safe to increment
3438 ++new->current_threshold
;
3443 /* Free old spare buffer and save old primary buffer as spare */
3444 kfree(thresholds
->spare
);
3445 thresholds
->spare
= thresholds
->primary
;
3447 rcu_assign_pointer(thresholds
->primary
, new);
3449 /* To be sure that nobody uses thresholds */
3453 mutex_unlock(&memcg
->thresholds_lock
);
3458 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3459 struct eventfd_ctx
*eventfd
, const char *args
)
3461 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3464 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3465 struct eventfd_ctx
*eventfd
, const char *args
)
3467 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3470 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3471 struct eventfd_ctx
*eventfd
, enum res_type type
)
3473 struct mem_cgroup_thresholds
*thresholds
;
3474 struct mem_cgroup_threshold_ary
*new;
3475 unsigned long usage
;
3478 mutex_lock(&memcg
->thresholds_lock
);
3481 thresholds
= &memcg
->thresholds
;
3482 usage
= mem_cgroup_usage(memcg
, false);
3483 } else if (type
== _MEMSWAP
) {
3484 thresholds
= &memcg
->memsw_thresholds
;
3485 usage
= mem_cgroup_usage(memcg
, true);
3489 if (!thresholds
->primary
)
3492 /* Check if a threshold crossed before removing */
3493 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3495 /* Calculate new number of threshold */
3497 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3498 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3502 new = thresholds
->spare
;
3504 /* Set thresholds array to NULL if we don't have thresholds */
3513 /* Copy thresholds and find current threshold */
3514 new->current_threshold
= -1;
3515 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3516 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3519 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3520 if (new->entries
[j
].threshold
<= usage
) {
3522 * new->current_threshold will not be used
3523 * until rcu_assign_pointer(), so it's safe to increment
3526 ++new->current_threshold
;
3532 /* Swap primary and spare array */
3533 thresholds
->spare
= thresholds
->primary
;
3535 rcu_assign_pointer(thresholds
->primary
, new);
3537 /* To be sure that nobody uses thresholds */
3540 /* If all events are unregistered, free the spare array */
3542 kfree(thresholds
->spare
);
3543 thresholds
->spare
= NULL
;
3546 mutex_unlock(&memcg
->thresholds_lock
);
3549 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3550 struct eventfd_ctx
*eventfd
)
3552 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3555 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3556 struct eventfd_ctx
*eventfd
)
3558 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3561 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3562 struct eventfd_ctx
*eventfd
, const char *args
)
3564 struct mem_cgroup_eventfd_list
*event
;
3566 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3570 spin_lock(&memcg_oom_lock
);
3572 event
->eventfd
= eventfd
;
3573 list_add(&event
->list
, &memcg
->oom_notify
);
3575 /* already in OOM ? */
3576 if (memcg
->under_oom
)
3577 eventfd_signal(eventfd
, 1);
3578 spin_unlock(&memcg_oom_lock
);
3583 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3584 struct eventfd_ctx
*eventfd
)
3586 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3588 spin_lock(&memcg_oom_lock
);
3590 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3591 if (ev
->eventfd
== eventfd
) {
3592 list_del(&ev
->list
);
3597 spin_unlock(&memcg_oom_lock
);
3600 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3602 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3604 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3605 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3609 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3610 struct cftype
*cft
, u64 val
)
3612 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3614 /* cannot set to root cgroup and only 0 and 1 are allowed */
3615 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3618 memcg
->oom_kill_disable
= val
;
3620 memcg_oom_recover(memcg
);
3625 #ifdef CONFIG_CGROUP_WRITEBACK
3627 struct list_head
*mem_cgroup_cgwb_list(struct mem_cgroup
*memcg
)
3629 return &memcg
->cgwb_list
;
3632 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3634 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3637 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3639 wb_domain_exit(&memcg
->cgwb_domain
);
3642 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3644 wb_domain_size_changed(&memcg
->cgwb_domain
);
3647 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3649 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3651 if (!memcg
->css
.parent
)
3654 return &memcg
->cgwb_domain
;
3658 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3659 * @wb: bdi_writeback in question
3660 * @pfilepages: out parameter for number of file pages
3661 * @pheadroom: out parameter for number of allocatable pages according to memcg
3662 * @pdirty: out parameter for number of dirty pages
3663 * @pwriteback: out parameter for number of pages under writeback
3665 * Determine the numbers of file, headroom, dirty, and writeback pages in
3666 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3667 * is a bit more involved.
3669 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3670 * headroom is calculated as the lowest headroom of itself and the
3671 * ancestors. Note that this doesn't consider the actual amount of
3672 * available memory in the system. The caller should further cap
3673 * *@pheadroom accordingly.
3675 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
3676 unsigned long *pheadroom
, unsigned long *pdirty
,
3677 unsigned long *pwriteback
)
3679 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3680 struct mem_cgroup
*parent
;
3682 *pdirty
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
3684 /* this should eventually include NR_UNSTABLE_NFS */
3685 *pwriteback
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
3686 *pfilepages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
3687 (1 << LRU_ACTIVE_FILE
));
3688 *pheadroom
= PAGE_COUNTER_MAX
;
3690 while ((parent
= parent_mem_cgroup(memcg
))) {
3691 unsigned long ceiling
= min(memcg
->memory
.limit
, memcg
->high
);
3692 unsigned long used
= page_counter_read(&memcg
->memory
);
3694 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
3699 #else /* CONFIG_CGROUP_WRITEBACK */
3701 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3706 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3710 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3714 #endif /* CONFIG_CGROUP_WRITEBACK */
3717 * DO NOT USE IN NEW FILES.
3719 * "cgroup.event_control" implementation.
3721 * This is way over-engineered. It tries to support fully configurable
3722 * events for each user. Such level of flexibility is completely
3723 * unnecessary especially in the light of the planned unified hierarchy.
3725 * Please deprecate this and replace with something simpler if at all
3730 * Unregister event and free resources.
3732 * Gets called from workqueue.
3734 static void memcg_event_remove(struct work_struct
*work
)
3736 struct mem_cgroup_event
*event
=
3737 container_of(work
, struct mem_cgroup_event
, remove
);
3738 struct mem_cgroup
*memcg
= event
->memcg
;
3740 remove_wait_queue(event
->wqh
, &event
->wait
);
3742 event
->unregister_event(memcg
, event
->eventfd
);
3744 /* Notify userspace the event is going away. */
3745 eventfd_signal(event
->eventfd
, 1);
3747 eventfd_ctx_put(event
->eventfd
);
3749 css_put(&memcg
->css
);
3753 * Gets called on POLLHUP on eventfd when user closes it.
3755 * Called with wqh->lock held and interrupts disabled.
3757 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
3758 int sync
, void *key
)
3760 struct mem_cgroup_event
*event
=
3761 container_of(wait
, struct mem_cgroup_event
, wait
);
3762 struct mem_cgroup
*memcg
= event
->memcg
;
3763 unsigned long flags
= (unsigned long)key
;
3765 if (flags
& POLLHUP
) {
3767 * If the event has been detached at cgroup removal, we
3768 * can simply return knowing the other side will cleanup
3771 * We can't race against event freeing since the other
3772 * side will require wqh->lock via remove_wait_queue(),
3775 spin_lock(&memcg
->event_list_lock
);
3776 if (!list_empty(&event
->list
)) {
3777 list_del_init(&event
->list
);
3779 * We are in atomic context, but cgroup_event_remove()
3780 * may sleep, so we have to call it in workqueue.
3782 schedule_work(&event
->remove
);
3784 spin_unlock(&memcg
->event_list_lock
);
3790 static void memcg_event_ptable_queue_proc(struct file
*file
,
3791 wait_queue_head_t
*wqh
, poll_table
*pt
)
3793 struct mem_cgroup_event
*event
=
3794 container_of(pt
, struct mem_cgroup_event
, pt
);
3797 add_wait_queue(wqh
, &event
->wait
);
3801 * DO NOT USE IN NEW FILES.
3803 * Parse input and register new cgroup event handler.
3805 * Input must be in format '<event_fd> <control_fd> <args>'.
3806 * Interpretation of args is defined by control file implementation.
3808 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
3809 char *buf
, size_t nbytes
, loff_t off
)
3811 struct cgroup_subsys_state
*css
= of_css(of
);
3812 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3813 struct mem_cgroup_event
*event
;
3814 struct cgroup_subsys_state
*cfile_css
;
3815 unsigned int efd
, cfd
;
3822 buf
= strstrip(buf
);
3824 efd
= simple_strtoul(buf
, &endp
, 10);
3829 cfd
= simple_strtoul(buf
, &endp
, 10);
3830 if ((*endp
!= ' ') && (*endp
!= '\0'))
3834 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
3838 event
->memcg
= memcg
;
3839 INIT_LIST_HEAD(&event
->list
);
3840 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
3841 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
3842 INIT_WORK(&event
->remove
, memcg_event_remove
);
3850 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
3851 if (IS_ERR(event
->eventfd
)) {
3852 ret
= PTR_ERR(event
->eventfd
);
3859 goto out_put_eventfd
;
3862 /* the process need read permission on control file */
3863 /* AV: shouldn't we check that it's been opened for read instead? */
3864 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
3869 * Determine the event callbacks and set them in @event. This used
3870 * to be done via struct cftype but cgroup core no longer knows
3871 * about these events. The following is crude but the whole thing
3872 * is for compatibility anyway.
3874 * DO NOT ADD NEW FILES.
3876 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
3878 if (!strcmp(name
, "memory.usage_in_bytes")) {
3879 event
->register_event
= mem_cgroup_usage_register_event
;
3880 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
3881 } else if (!strcmp(name
, "memory.oom_control")) {
3882 event
->register_event
= mem_cgroup_oom_register_event
;
3883 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
3884 } else if (!strcmp(name
, "memory.pressure_level")) {
3885 event
->register_event
= vmpressure_register_event
;
3886 event
->unregister_event
= vmpressure_unregister_event
;
3887 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
3888 event
->register_event
= memsw_cgroup_usage_register_event
;
3889 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
3896 * Verify @cfile should belong to @css. Also, remaining events are
3897 * automatically removed on cgroup destruction but the removal is
3898 * asynchronous, so take an extra ref on @css.
3900 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
3901 &memory_cgrp_subsys
);
3903 if (IS_ERR(cfile_css
))
3905 if (cfile_css
!= css
) {
3910 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
3914 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
3916 spin_lock(&memcg
->event_list_lock
);
3917 list_add(&event
->list
, &memcg
->event_list
);
3918 spin_unlock(&memcg
->event_list_lock
);
3930 eventfd_ctx_put(event
->eventfd
);
3939 static struct cftype mem_cgroup_legacy_files
[] = {
3941 .name
= "usage_in_bytes",
3942 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
3943 .read_u64
= mem_cgroup_read_u64
,
3946 .name
= "max_usage_in_bytes",
3947 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
3948 .write
= mem_cgroup_reset
,
3949 .read_u64
= mem_cgroup_read_u64
,
3952 .name
= "limit_in_bytes",
3953 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
3954 .write
= mem_cgroup_write
,
3955 .read_u64
= mem_cgroup_read_u64
,
3958 .name
= "soft_limit_in_bytes",
3959 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
3960 .write
= mem_cgroup_write
,
3961 .read_u64
= mem_cgroup_read_u64
,
3965 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
3966 .write
= mem_cgroup_reset
,
3967 .read_u64
= mem_cgroup_read_u64
,
3971 .seq_show
= memcg_stat_show
,
3974 .name
= "force_empty",
3975 .write
= mem_cgroup_force_empty_write
,
3978 .name
= "use_hierarchy",
3979 .write_u64
= mem_cgroup_hierarchy_write
,
3980 .read_u64
= mem_cgroup_hierarchy_read
,
3983 .name
= "cgroup.event_control", /* XXX: for compat */
3984 .write
= memcg_write_event_control
,
3985 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
3988 .name
= "swappiness",
3989 .read_u64
= mem_cgroup_swappiness_read
,
3990 .write_u64
= mem_cgroup_swappiness_write
,
3993 .name
= "move_charge_at_immigrate",
3994 .read_u64
= mem_cgroup_move_charge_read
,
3995 .write_u64
= mem_cgroup_move_charge_write
,
3998 .name
= "oom_control",
3999 .seq_show
= mem_cgroup_oom_control_read
,
4000 .write_u64
= mem_cgroup_oom_control_write
,
4001 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4004 .name
= "pressure_level",
4008 .name
= "numa_stat",
4009 .seq_show
= memcg_numa_stat_show
,
4013 .name
= "kmem.limit_in_bytes",
4014 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4015 .write
= mem_cgroup_write
,
4016 .read_u64
= mem_cgroup_read_u64
,
4019 .name
= "kmem.usage_in_bytes",
4020 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4021 .read_u64
= mem_cgroup_read_u64
,
4024 .name
= "kmem.failcnt",
4025 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4026 .write
= mem_cgroup_reset
,
4027 .read_u64
= mem_cgroup_read_u64
,
4030 .name
= "kmem.max_usage_in_bytes",
4031 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4032 .write
= mem_cgroup_reset
,
4033 .read_u64
= mem_cgroup_read_u64
,
4035 #ifdef CONFIG_SLABINFO
4037 .name
= "kmem.slabinfo",
4038 .seq_start
= slab_start
,
4039 .seq_next
= slab_next
,
4040 .seq_stop
= slab_stop
,
4041 .seq_show
= memcg_slab_show
,
4045 .name
= "kmem.tcp.limit_in_bytes",
4046 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4047 .write
= mem_cgroup_write
,
4048 .read_u64
= mem_cgroup_read_u64
,
4051 .name
= "kmem.tcp.usage_in_bytes",
4052 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4053 .read_u64
= mem_cgroup_read_u64
,
4056 .name
= "kmem.tcp.failcnt",
4057 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4058 .write
= mem_cgroup_reset
,
4059 .read_u64
= mem_cgroup_read_u64
,
4062 .name
= "kmem.tcp.max_usage_in_bytes",
4063 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4064 .write
= mem_cgroup_reset
,
4065 .read_u64
= mem_cgroup_read_u64
,
4067 { }, /* terminate */
4070 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4072 struct mem_cgroup_per_node
*pn
;
4073 struct mem_cgroup_per_zone
*mz
;
4074 int zone
, tmp
= node
;
4076 * This routine is called against possible nodes.
4077 * But it's BUG to call kmalloc() against offline node.
4079 * TODO: this routine can waste much memory for nodes which will
4080 * never be onlined. It's better to use memory hotplug callback
4083 if (!node_state(node
, N_NORMAL_MEMORY
))
4085 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4089 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4090 mz
= &pn
->zoneinfo
[zone
];
4091 lruvec_init(&mz
->lruvec
);
4092 mz
->usage_in_excess
= 0;
4093 mz
->on_tree
= false;
4096 memcg
->nodeinfo
[node
] = pn
;
4100 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4102 kfree(memcg
->nodeinfo
[node
]);
4105 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4109 memcg_wb_domain_exit(memcg
);
4111 free_mem_cgroup_per_zone_info(memcg
, node
);
4112 free_percpu(memcg
->stat
);
4116 static struct mem_cgroup
*mem_cgroup_alloc(void)
4118 struct mem_cgroup
*memcg
;
4122 size
= sizeof(struct mem_cgroup
);
4123 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4125 memcg
= kzalloc(size
, GFP_KERNEL
);
4129 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4134 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
4137 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4140 INIT_WORK(&memcg
->high_work
, high_work_func
);
4141 memcg
->last_scanned_node
= MAX_NUMNODES
;
4142 INIT_LIST_HEAD(&memcg
->oom_notify
);
4143 mutex_init(&memcg
->thresholds_lock
);
4144 spin_lock_init(&memcg
->move_lock
);
4145 vmpressure_init(&memcg
->vmpressure
);
4146 INIT_LIST_HEAD(&memcg
->event_list
);
4147 spin_lock_init(&memcg
->event_list_lock
);
4148 memcg
->socket_pressure
= jiffies
;
4150 memcg
->kmemcg_id
= -1;
4152 #ifdef CONFIG_CGROUP_WRITEBACK
4153 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4157 mem_cgroup_free(memcg
);
4161 static struct cgroup_subsys_state
* __ref
4162 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4164 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
4165 struct mem_cgroup
*memcg
;
4166 long error
= -ENOMEM
;
4168 memcg
= mem_cgroup_alloc();
4170 return ERR_PTR(error
);
4172 memcg
->high
= PAGE_COUNTER_MAX
;
4173 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4175 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4176 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4178 if (parent
&& parent
->use_hierarchy
) {
4179 memcg
->use_hierarchy
= true;
4180 page_counter_init(&memcg
->memory
, &parent
->memory
);
4181 page_counter_init(&memcg
->swap
, &parent
->swap
);
4182 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4183 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4184 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
4186 page_counter_init(&memcg
->memory
, NULL
);
4187 page_counter_init(&memcg
->swap
, NULL
);
4188 page_counter_init(&memcg
->memsw
, NULL
);
4189 page_counter_init(&memcg
->kmem
, NULL
);
4190 page_counter_init(&memcg
->tcpmem
, NULL
);
4192 * Deeper hierachy with use_hierarchy == false doesn't make
4193 * much sense so let cgroup subsystem know about this
4194 * unfortunate state in our controller.
4196 if (parent
!= root_mem_cgroup
)
4197 memory_cgrp_subsys
.broken_hierarchy
= true;
4200 /* The following stuff does not apply to the root */
4202 root_mem_cgroup
= memcg
;
4206 error
= memcg_propagate_kmem(parent
, memcg
);
4210 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4211 static_branch_inc(&memcg_sockets_enabled_key
);
4215 mem_cgroup_free(memcg
);
4220 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4222 if (css
->id
> MEM_CGROUP_ID_MAX
)
4228 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4230 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4231 struct mem_cgroup_event
*event
, *tmp
;
4234 * Unregister events and notify userspace.
4235 * Notify userspace about cgroup removing only after rmdir of cgroup
4236 * directory to avoid race between userspace and kernelspace.
4238 spin_lock(&memcg
->event_list_lock
);
4239 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4240 list_del_init(&event
->list
);
4241 schedule_work(&event
->remove
);
4243 spin_unlock(&memcg
->event_list_lock
);
4245 memcg_offline_kmem(memcg
);
4246 wb_memcg_offline(memcg
);
4249 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4251 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4253 invalidate_reclaim_iterators(memcg
);
4256 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4258 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4260 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4261 static_branch_dec(&memcg_sockets_enabled_key
);
4263 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
4264 static_branch_dec(&memcg_sockets_enabled_key
);
4266 vmpressure_cleanup(&memcg
->vmpressure
);
4267 cancel_work_sync(&memcg
->high_work
);
4268 mem_cgroup_remove_from_trees(memcg
);
4269 memcg_free_kmem(memcg
);
4270 mem_cgroup_free(memcg
);
4274 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4275 * @css: the target css
4277 * Reset the states of the mem_cgroup associated with @css. This is
4278 * invoked when the userland requests disabling on the default hierarchy
4279 * but the memcg is pinned through dependency. The memcg should stop
4280 * applying policies and should revert to the vanilla state as it may be
4281 * made visible again.
4283 * The current implementation only resets the essential configurations.
4284 * This needs to be expanded to cover all the visible parts.
4286 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4288 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4290 mem_cgroup_resize_limit(memcg
, PAGE_COUNTER_MAX
);
4291 mem_cgroup_resize_memsw_limit(memcg
, PAGE_COUNTER_MAX
);
4292 memcg_update_kmem_limit(memcg
, PAGE_COUNTER_MAX
);
4294 memcg
->high
= PAGE_COUNTER_MAX
;
4295 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4296 memcg_wb_domain_size_changed(memcg
);
4300 /* Handlers for move charge at task migration. */
4301 static int mem_cgroup_do_precharge(unsigned long count
)
4305 /* Try a single bulk charge without reclaim first, kswapd may wake */
4306 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4308 mc
.precharge
+= count
;
4312 /* Try charges one by one with reclaim */
4314 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4324 * get_mctgt_type - get target type of moving charge
4325 * @vma: the vma the pte to be checked belongs
4326 * @addr: the address corresponding to the pte to be checked
4327 * @ptent: the pte to be checked
4328 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4331 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4332 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4333 * move charge. if @target is not NULL, the page is stored in target->page
4334 * with extra refcnt got(Callers should handle it).
4335 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4336 * target for charge migration. if @target is not NULL, the entry is stored
4339 * Called with pte lock held.
4346 enum mc_target_type
{
4352 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4353 unsigned long addr
, pte_t ptent
)
4355 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4357 if (!page
|| !page_mapped(page
))
4359 if (PageAnon(page
)) {
4360 if (!(mc
.flags
& MOVE_ANON
))
4363 if (!(mc
.flags
& MOVE_FILE
))
4366 if (!get_page_unless_zero(page
))
4373 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4374 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4376 struct page
*page
= NULL
;
4377 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4379 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4382 * Because lookup_swap_cache() updates some statistics counter,
4383 * we call find_get_page() with swapper_space directly.
4385 page
= find_get_page(swap_address_space(ent
), ent
.val
);
4386 if (do_memsw_account())
4387 entry
->val
= ent
.val
;
4392 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4393 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4399 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4400 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4402 struct page
*page
= NULL
;
4403 struct address_space
*mapping
;
4406 if (!vma
->vm_file
) /* anonymous vma */
4408 if (!(mc
.flags
& MOVE_FILE
))
4411 mapping
= vma
->vm_file
->f_mapping
;
4412 pgoff
= linear_page_index(vma
, addr
);
4414 /* page is moved even if it's not RSS of this task(page-faulted). */
4416 /* shmem/tmpfs may report page out on swap: account for that too. */
4417 if (shmem_mapping(mapping
)) {
4418 page
= find_get_entry(mapping
, pgoff
);
4419 if (radix_tree_exceptional_entry(page
)) {
4420 swp_entry_t swp
= radix_to_swp_entry(page
);
4421 if (do_memsw_account())
4423 page
= find_get_page(swap_address_space(swp
), swp
.val
);
4426 page
= find_get_page(mapping
, pgoff
);
4428 page
= find_get_page(mapping
, pgoff
);
4434 * mem_cgroup_move_account - move account of the page
4436 * @nr_pages: number of regular pages (>1 for huge pages)
4437 * @from: mem_cgroup which the page is moved from.
4438 * @to: mem_cgroup which the page is moved to. @from != @to.
4440 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4442 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4445 static int mem_cgroup_move_account(struct page
*page
,
4447 struct mem_cgroup
*from
,
4448 struct mem_cgroup
*to
)
4450 unsigned long flags
;
4451 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
4455 VM_BUG_ON(from
== to
);
4456 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4457 VM_BUG_ON(compound
&& !PageTransHuge(page
));
4460 * Prevent mem_cgroup_replace_page() from looking at
4461 * page->mem_cgroup of its source page while we change it.
4464 if (!trylock_page(page
))
4468 if (page
->mem_cgroup
!= from
)
4471 anon
= PageAnon(page
);
4473 spin_lock_irqsave(&from
->move_lock
, flags
);
4475 if (!anon
&& page_mapped(page
)) {
4476 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4478 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4483 * move_lock grabbed above and caller set from->moving_account, so
4484 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4485 * So mapping should be stable for dirty pages.
4487 if (!anon
&& PageDirty(page
)) {
4488 struct address_space
*mapping
= page_mapping(page
);
4490 if (mapping_cap_account_dirty(mapping
)) {
4491 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4493 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4498 if (PageWriteback(page
)) {
4499 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4501 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4506 * It is safe to change page->mem_cgroup here because the page
4507 * is referenced, charged, and isolated - we can't race with
4508 * uncharging, charging, migration, or LRU putback.
4511 /* caller should have done css_get */
4512 page
->mem_cgroup
= to
;
4513 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4517 local_irq_disable();
4518 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
4519 memcg_check_events(to
, page
);
4520 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
4521 memcg_check_events(from
, page
);
4529 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4530 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4532 struct page
*page
= NULL
;
4533 enum mc_target_type ret
= MC_TARGET_NONE
;
4534 swp_entry_t ent
= { .val
= 0 };
4536 if (pte_present(ptent
))
4537 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4538 else if (is_swap_pte(ptent
))
4539 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
4540 else if (pte_none(ptent
))
4541 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4543 if (!page
&& !ent
.val
)
4547 * Do only loose check w/o serialization.
4548 * mem_cgroup_move_account() checks the page is valid or
4549 * not under LRU exclusion.
4551 if (page
->mem_cgroup
== mc
.from
) {
4552 ret
= MC_TARGET_PAGE
;
4554 target
->page
= page
;
4556 if (!ret
|| !target
)
4559 /* There is a swap entry and a page doesn't exist or isn't charged */
4560 if (ent
.val
&& !ret
&&
4561 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4562 ret
= MC_TARGET_SWAP
;
4569 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4571 * We don't consider swapping or file mapped pages because THP does not
4572 * support them for now.
4573 * Caller should make sure that pmd_trans_huge(pmd) is true.
4575 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4576 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4578 struct page
*page
= NULL
;
4579 enum mc_target_type ret
= MC_TARGET_NONE
;
4581 page
= pmd_page(pmd
);
4582 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4583 if (!(mc
.flags
& MOVE_ANON
))
4585 if (page
->mem_cgroup
== mc
.from
) {
4586 ret
= MC_TARGET_PAGE
;
4589 target
->page
= page
;
4595 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4596 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4598 return MC_TARGET_NONE
;
4602 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4603 unsigned long addr
, unsigned long end
,
4604 struct mm_walk
*walk
)
4606 struct vm_area_struct
*vma
= walk
->vma
;
4610 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4612 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4613 mc
.precharge
+= HPAGE_PMD_NR
;
4618 if (pmd_trans_unstable(pmd
))
4620 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4621 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4622 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4623 mc
.precharge
++; /* increment precharge temporarily */
4624 pte_unmap_unlock(pte
- 1, ptl
);
4630 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4632 unsigned long precharge
;
4634 struct mm_walk mem_cgroup_count_precharge_walk
= {
4635 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4638 down_read(&mm
->mmap_sem
);
4639 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk
);
4640 up_read(&mm
->mmap_sem
);
4642 precharge
= mc
.precharge
;
4648 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4650 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4652 VM_BUG_ON(mc
.moving_task
);
4653 mc
.moving_task
= current
;
4654 return mem_cgroup_do_precharge(precharge
);
4657 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4658 static void __mem_cgroup_clear_mc(void)
4660 struct mem_cgroup
*from
= mc
.from
;
4661 struct mem_cgroup
*to
= mc
.to
;
4663 /* we must uncharge all the leftover precharges from mc.to */
4665 cancel_charge(mc
.to
, mc
.precharge
);
4669 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4670 * we must uncharge here.
4672 if (mc
.moved_charge
) {
4673 cancel_charge(mc
.from
, mc
.moved_charge
);
4674 mc
.moved_charge
= 0;
4676 /* we must fixup refcnts and charges */
4677 if (mc
.moved_swap
) {
4678 /* uncharge swap account from the old cgroup */
4679 if (!mem_cgroup_is_root(mc
.from
))
4680 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4683 * we charged both to->memory and to->memsw, so we
4684 * should uncharge to->memory.
4686 if (!mem_cgroup_is_root(mc
.to
))
4687 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4689 css_put_many(&mc
.from
->css
, mc
.moved_swap
);
4691 /* we've already done css_get(mc.to) */
4694 memcg_oom_recover(from
);
4695 memcg_oom_recover(to
);
4696 wake_up_all(&mc
.waitq
);
4699 static void mem_cgroup_clear_mc(void)
4702 * we must clear moving_task before waking up waiters at the end of
4705 mc
.moving_task
= NULL
;
4706 __mem_cgroup_clear_mc();
4707 spin_lock(&mc
.lock
);
4710 spin_unlock(&mc
.lock
);
4713 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4715 struct cgroup_subsys_state
*css
;
4716 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
4717 struct mem_cgroup
*from
;
4718 struct task_struct
*leader
, *p
;
4719 struct mm_struct
*mm
;
4720 unsigned long move_flags
;
4723 /* charge immigration isn't supported on the default hierarchy */
4724 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4728 * Multi-process migrations only happen on the default hierarchy
4729 * where charge immigration is not used. Perform charge
4730 * immigration if @tset contains a leader and whine if there are
4734 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
4737 memcg
= mem_cgroup_from_css(css
);
4743 * We are now commited to this value whatever it is. Changes in this
4744 * tunable will only affect upcoming migrations, not the current one.
4745 * So we need to save it, and keep it going.
4747 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
4751 from
= mem_cgroup_from_task(p
);
4753 VM_BUG_ON(from
== memcg
);
4755 mm
= get_task_mm(p
);
4758 /* We move charges only when we move a owner of the mm */
4759 if (mm
->owner
== p
) {
4762 VM_BUG_ON(mc
.precharge
);
4763 VM_BUG_ON(mc
.moved_charge
);
4764 VM_BUG_ON(mc
.moved_swap
);
4766 spin_lock(&mc
.lock
);
4769 mc
.flags
= move_flags
;
4770 spin_unlock(&mc
.lock
);
4771 /* We set mc.moving_task later */
4773 ret
= mem_cgroup_precharge_mc(mm
);
4775 mem_cgroup_clear_mc();
4781 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4784 mem_cgroup_clear_mc();
4787 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4788 unsigned long addr
, unsigned long end
,
4789 struct mm_walk
*walk
)
4792 struct vm_area_struct
*vma
= walk
->vma
;
4795 enum mc_target_type target_type
;
4796 union mc_target target
;
4799 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4801 if (mc
.precharge
< HPAGE_PMD_NR
) {
4805 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
4806 if (target_type
== MC_TARGET_PAGE
) {
4808 if (!isolate_lru_page(page
)) {
4809 if (!mem_cgroup_move_account(page
, true,
4811 mc
.precharge
-= HPAGE_PMD_NR
;
4812 mc
.moved_charge
+= HPAGE_PMD_NR
;
4814 putback_lru_page(page
);
4822 if (pmd_trans_unstable(pmd
))
4825 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4826 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4827 pte_t ptent
= *(pte
++);
4833 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
4834 case MC_TARGET_PAGE
:
4837 * We can have a part of the split pmd here. Moving it
4838 * can be done but it would be too convoluted so simply
4839 * ignore such a partial THP and keep it in original
4840 * memcg. There should be somebody mapping the head.
4842 if (PageTransCompound(page
))
4844 if (isolate_lru_page(page
))
4846 if (!mem_cgroup_move_account(page
, false,
4849 /* we uncharge from mc.from later. */
4852 putback_lru_page(page
);
4853 put
: /* get_mctgt_type() gets the page */
4856 case MC_TARGET_SWAP
:
4858 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
4860 /* we fixup refcnts and charges later. */
4868 pte_unmap_unlock(pte
- 1, ptl
);
4873 * We have consumed all precharges we got in can_attach().
4874 * We try charge one by one, but don't do any additional
4875 * charges to mc.to if we have failed in charge once in attach()
4878 ret
= mem_cgroup_do_precharge(1);
4886 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
4888 struct mm_walk mem_cgroup_move_charge_walk
= {
4889 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4893 lru_add_drain_all();
4895 * Signal lock_page_memcg() to take the memcg's move_lock
4896 * while we're moving its pages to another memcg. Then wait
4897 * for already started RCU-only updates to finish.
4899 atomic_inc(&mc
.from
->moving_account
);
4902 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
4904 * Someone who are holding the mmap_sem might be waiting in
4905 * waitq. So we cancel all extra charges, wake up all waiters,
4906 * and retry. Because we cancel precharges, we might not be able
4907 * to move enough charges, but moving charge is a best-effort
4908 * feature anyway, so it wouldn't be a big problem.
4910 __mem_cgroup_clear_mc();
4915 * When we have consumed all precharges and failed in doing
4916 * additional charge, the page walk just aborts.
4918 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk
);
4919 up_read(&mm
->mmap_sem
);
4920 atomic_dec(&mc
.from
->moving_account
);
4923 static void mem_cgroup_move_task(struct cgroup_taskset
*tset
)
4925 struct cgroup_subsys_state
*css
;
4926 struct task_struct
*p
= cgroup_taskset_first(tset
, &css
);
4927 struct mm_struct
*mm
= get_task_mm(p
);
4931 mem_cgroup_move_charge(mm
);
4935 mem_cgroup_clear_mc();
4937 #else /* !CONFIG_MMU */
4938 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4942 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4945 static void mem_cgroup_move_task(struct cgroup_taskset
*tset
)
4951 * Cgroup retains root cgroups across [un]mount cycles making it necessary
4952 * to verify whether we're attached to the default hierarchy on each mount
4955 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
4958 * use_hierarchy is forced on the default hierarchy. cgroup core
4959 * guarantees that @root doesn't have any children, so turning it
4960 * on for the root memcg is enough.
4962 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4963 root_mem_cgroup
->use_hierarchy
= true;
4965 root_mem_cgroup
->use_hierarchy
= false;
4968 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
4971 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4973 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
4976 static int memory_low_show(struct seq_file
*m
, void *v
)
4978 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
4979 unsigned long low
= READ_ONCE(memcg
->low
);
4981 if (low
== PAGE_COUNTER_MAX
)
4982 seq_puts(m
, "max\n");
4984 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
4989 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
4990 char *buf
, size_t nbytes
, loff_t off
)
4992 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
4996 buf
= strstrip(buf
);
4997 err
= page_counter_memparse(buf
, "max", &low
);
5006 static int memory_high_show(struct seq_file
*m
, void *v
)
5008 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5009 unsigned long high
= READ_ONCE(memcg
->high
);
5011 if (high
== PAGE_COUNTER_MAX
)
5012 seq_puts(m
, "max\n");
5014 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
5019 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5020 char *buf
, size_t nbytes
, loff_t off
)
5022 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5026 buf
= strstrip(buf
);
5027 err
= page_counter_memparse(buf
, "max", &high
);
5033 memcg_wb_domain_size_changed(memcg
);
5037 static int memory_max_show(struct seq_file
*m
, void *v
)
5039 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5040 unsigned long max
= READ_ONCE(memcg
->memory
.limit
);
5042 if (max
== PAGE_COUNTER_MAX
)
5043 seq_puts(m
, "max\n");
5045 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5050 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5051 char *buf
, size_t nbytes
, loff_t off
)
5053 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5057 buf
= strstrip(buf
);
5058 err
= page_counter_memparse(buf
, "max", &max
);
5062 err
= mem_cgroup_resize_limit(memcg
, max
);
5066 memcg_wb_domain_size_changed(memcg
);
5070 static int memory_events_show(struct seq_file
*m
, void *v
)
5072 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5074 seq_printf(m
, "low %lu\n", mem_cgroup_read_events(memcg
, MEMCG_LOW
));
5075 seq_printf(m
, "high %lu\n", mem_cgroup_read_events(memcg
, MEMCG_HIGH
));
5076 seq_printf(m
, "max %lu\n", mem_cgroup_read_events(memcg
, MEMCG_MAX
));
5077 seq_printf(m
, "oom %lu\n", mem_cgroup_read_events(memcg
, MEMCG_OOM
));
5082 static int memory_stat_show(struct seq_file
*m
, void *v
)
5084 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5088 * Provide statistics on the state of the memory subsystem as
5089 * well as cumulative event counters that show past behavior.
5091 * This list is ordered following a combination of these gradients:
5092 * 1) generic big picture -> specifics and details
5093 * 2) reflecting userspace activity -> reflecting kernel heuristics
5095 * Current memory state:
5098 seq_printf(m
, "anon %llu\n",
5099 (u64
)tree_stat(memcg
, MEM_CGROUP_STAT_RSS
) * PAGE_SIZE
);
5100 seq_printf(m
, "file %llu\n",
5101 (u64
)tree_stat(memcg
, MEM_CGROUP_STAT_CACHE
) * PAGE_SIZE
);
5102 seq_printf(m
, "sock %llu\n",
5103 (u64
)tree_stat(memcg
, MEMCG_SOCK
) * PAGE_SIZE
);
5105 seq_printf(m
, "file_mapped %llu\n",
5106 (u64
)tree_stat(memcg
, MEM_CGROUP_STAT_FILE_MAPPED
) *
5108 seq_printf(m
, "file_dirty %llu\n",
5109 (u64
)tree_stat(memcg
, MEM_CGROUP_STAT_DIRTY
) *
5111 seq_printf(m
, "file_writeback %llu\n",
5112 (u64
)tree_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
) *
5115 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5116 struct mem_cgroup
*mi
;
5117 unsigned long val
= 0;
5119 for_each_mem_cgroup_tree(mi
, memcg
)
5120 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
));
5121 seq_printf(m
, "%s %llu\n",
5122 mem_cgroup_lru_names
[i
], (u64
)val
* PAGE_SIZE
);
5125 /* Accumulated memory events */
5127 seq_printf(m
, "pgfault %lu\n",
5128 tree_events(memcg
, MEM_CGROUP_EVENTS_PGFAULT
));
5129 seq_printf(m
, "pgmajfault %lu\n",
5130 tree_events(memcg
, MEM_CGROUP_EVENTS_PGMAJFAULT
));
5135 static struct cftype memory_files
[] = {
5138 .flags
= CFTYPE_NOT_ON_ROOT
,
5139 .read_u64
= memory_current_read
,
5143 .flags
= CFTYPE_NOT_ON_ROOT
,
5144 .seq_show
= memory_low_show
,
5145 .write
= memory_low_write
,
5149 .flags
= CFTYPE_NOT_ON_ROOT
,
5150 .seq_show
= memory_high_show
,
5151 .write
= memory_high_write
,
5155 .flags
= CFTYPE_NOT_ON_ROOT
,
5156 .seq_show
= memory_max_show
,
5157 .write
= memory_max_write
,
5161 .flags
= CFTYPE_NOT_ON_ROOT
,
5162 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
5163 .seq_show
= memory_events_show
,
5167 .flags
= CFTYPE_NOT_ON_ROOT
,
5168 .seq_show
= memory_stat_show
,
5173 struct cgroup_subsys memory_cgrp_subsys
= {
5174 .css_alloc
= mem_cgroup_css_alloc
,
5175 .css_online
= mem_cgroup_css_online
,
5176 .css_offline
= mem_cgroup_css_offline
,
5177 .css_released
= mem_cgroup_css_released
,
5178 .css_free
= mem_cgroup_css_free
,
5179 .css_reset
= mem_cgroup_css_reset
,
5180 .can_attach
= mem_cgroup_can_attach
,
5181 .cancel_attach
= mem_cgroup_cancel_attach
,
5182 .attach
= mem_cgroup_move_task
,
5183 .bind
= mem_cgroup_bind
,
5184 .dfl_cftypes
= memory_files
,
5185 .legacy_cftypes
= mem_cgroup_legacy_files
,
5190 * mem_cgroup_low - check if memory consumption is below the normal range
5191 * @root: the highest ancestor to consider
5192 * @memcg: the memory cgroup to check
5194 * Returns %true if memory consumption of @memcg, and that of all
5195 * configurable ancestors up to @root, is below the normal range.
5197 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5199 if (mem_cgroup_disabled())
5203 * The toplevel group doesn't have a configurable range, so
5204 * it's never low when looked at directly, and it is not
5205 * considered an ancestor when assessing the hierarchy.
5208 if (memcg
== root_mem_cgroup
)
5211 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5214 while (memcg
!= root
) {
5215 memcg
= parent_mem_cgroup(memcg
);
5217 if (memcg
== root_mem_cgroup
)
5220 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5227 * mem_cgroup_try_charge - try charging a page
5228 * @page: page to charge
5229 * @mm: mm context of the victim
5230 * @gfp_mask: reclaim mode
5231 * @memcgp: charged memcg return
5233 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5234 * pages according to @gfp_mask if necessary.
5236 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5237 * Otherwise, an error code is returned.
5239 * After page->mapping has been set up, the caller must finalize the
5240 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5241 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5243 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5244 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
5247 struct mem_cgroup
*memcg
= NULL
;
5248 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5251 if (mem_cgroup_disabled())
5254 if (PageSwapCache(page
)) {
5256 * Every swap fault against a single page tries to charge the
5257 * page, bail as early as possible. shmem_unuse() encounters
5258 * already charged pages, too. The USED bit is protected by
5259 * the page lock, which serializes swap cache removal, which
5260 * in turn serializes uncharging.
5262 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5263 if (page
->mem_cgroup
)
5266 if (do_swap_account
) {
5267 swp_entry_t ent
= { .val
= page_private(page
), };
5268 unsigned short id
= lookup_swap_cgroup_id(ent
);
5271 memcg
= mem_cgroup_from_id(id
);
5272 if (memcg
&& !css_tryget_online(&memcg
->css
))
5279 memcg
= get_mem_cgroup_from_mm(mm
);
5281 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5283 css_put(&memcg
->css
);
5290 * mem_cgroup_commit_charge - commit a page charge
5291 * @page: page to charge
5292 * @memcg: memcg to charge the page to
5293 * @lrucare: page might be on LRU already
5295 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5296 * after page->mapping has been set up. This must happen atomically
5297 * as part of the page instantiation, i.e. under the page table lock
5298 * for anonymous pages, under the page lock for page and swap cache.
5300 * In addition, the page must not be on the LRU during the commit, to
5301 * prevent racing with task migration. If it might be, use @lrucare.
5303 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5305 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5306 bool lrucare
, bool compound
)
5308 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5310 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5311 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5313 if (mem_cgroup_disabled())
5316 * Swap faults will attempt to charge the same page multiple
5317 * times. But reuse_swap_page() might have removed the page
5318 * from swapcache already, so we can't check PageSwapCache().
5323 commit_charge(page
, memcg
, lrucare
);
5325 local_irq_disable();
5326 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
5327 memcg_check_events(memcg
, page
);
5330 if (do_memsw_account() && PageSwapCache(page
)) {
5331 swp_entry_t entry
= { .val
= page_private(page
) };
5333 * The swap entry might not get freed for a long time,
5334 * let's not wait for it. The page already received a
5335 * memory+swap charge, drop the swap entry duplicate.
5337 mem_cgroup_uncharge_swap(entry
);
5342 * mem_cgroup_cancel_charge - cancel a page charge
5343 * @page: page to charge
5344 * @memcg: memcg to charge the page to
5346 * Cancel a charge transaction started by mem_cgroup_try_charge().
5348 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5351 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5353 if (mem_cgroup_disabled())
5356 * Swap faults will attempt to charge the same page multiple
5357 * times. But reuse_swap_page() might have removed the page
5358 * from swapcache already, so we can't check PageSwapCache().
5363 cancel_charge(memcg
, nr_pages
);
5366 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5367 unsigned long nr_anon
, unsigned long nr_file
,
5368 unsigned long nr_huge
, struct page
*dummy_page
)
5370 unsigned long nr_pages
= nr_anon
+ nr_file
;
5371 unsigned long flags
;
5373 if (!mem_cgroup_is_root(memcg
)) {
5374 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5375 if (do_memsw_account())
5376 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5377 memcg_oom_recover(memcg
);
5380 local_irq_save(flags
);
5381 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5382 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5383 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5384 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5385 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5386 memcg_check_events(memcg
, dummy_page
);
5387 local_irq_restore(flags
);
5389 if (!mem_cgroup_is_root(memcg
))
5390 css_put_many(&memcg
->css
, nr_pages
);
5393 static void uncharge_list(struct list_head
*page_list
)
5395 struct mem_cgroup
*memcg
= NULL
;
5396 unsigned long nr_anon
= 0;
5397 unsigned long nr_file
= 0;
5398 unsigned long nr_huge
= 0;
5399 unsigned long pgpgout
= 0;
5400 struct list_head
*next
;
5403 next
= page_list
->next
;
5405 unsigned int nr_pages
= 1;
5407 page
= list_entry(next
, struct page
, lru
);
5408 next
= page
->lru
.next
;
5410 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5411 VM_BUG_ON_PAGE(page_count(page
), page
);
5413 if (!page
->mem_cgroup
)
5417 * Nobody should be changing or seriously looking at
5418 * page->mem_cgroup at this point, we have fully
5419 * exclusive access to the page.
5422 if (memcg
!= page
->mem_cgroup
) {
5424 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5426 pgpgout
= nr_anon
= nr_file
= nr_huge
= 0;
5428 memcg
= page
->mem_cgroup
;
5431 if (PageTransHuge(page
)) {
5432 nr_pages
<<= compound_order(page
);
5433 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5434 nr_huge
+= nr_pages
;
5438 nr_anon
+= nr_pages
;
5440 nr_file
+= nr_pages
;
5442 page
->mem_cgroup
= NULL
;
5445 } while (next
!= page_list
);
5448 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5453 * mem_cgroup_uncharge - uncharge a page
5454 * @page: page to uncharge
5456 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5457 * mem_cgroup_commit_charge().
5459 void mem_cgroup_uncharge(struct page
*page
)
5461 if (mem_cgroup_disabled())
5464 /* Don't touch page->lru of any random page, pre-check: */
5465 if (!page
->mem_cgroup
)
5468 INIT_LIST_HEAD(&page
->lru
);
5469 uncharge_list(&page
->lru
);
5473 * mem_cgroup_uncharge_list - uncharge a list of page
5474 * @page_list: list of pages to uncharge
5476 * Uncharge a list of pages previously charged with
5477 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5479 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5481 if (mem_cgroup_disabled())
5484 if (!list_empty(page_list
))
5485 uncharge_list(page_list
);
5489 * mem_cgroup_replace_page - migrate a charge to another page
5490 * @oldpage: currently charged page
5491 * @newpage: page to transfer the charge to
5493 * Migrate the charge from @oldpage to @newpage.
5495 * Both pages must be locked, @newpage->mapping must be set up.
5496 * Either or both pages might be on the LRU already.
5498 void mem_cgroup_replace_page(struct page
*oldpage
, struct page
*newpage
)
5500 struct mem_cgroup
*memcg
;
5501 unsigned int nr_pages
;
5504 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5505 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5506 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5507 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5510 if (mem_cgroup_disabled())
5513 /* Page cache replacement: new page already charged? */
5514 if (newpage
->mem_cgroup
)
5517 /* Swapcache readahead pages can get replaced before being charged */
5518 memcg
= oldpage
->mem_cgroup
;
5522 /* Force-charge the new page. The old one will be freed soon */
5523 compound
= PageTransHuge(newpage
);
5524 nr_pages
= compound
? hpage_nr_pages(newpage
) : 1;
5526 page_counter_charge(&memcg
->memory
, nr_pages
);
5527 if (do_memsw_account())
5528 page_counter_charge(&memcg
->memsw
, nr_pages
);
5529 css_get_many(&memcg
->css
, nr_pages
);
5531 commit_charge(newpage
, memcg
, true);
5533 local_irq_disable();
5534 mem_cgroup_charge_statistics(memcg
, newpage
, compound
, nr_pages
);
5535 memcg_check_events(memcg
, newpage
);
5539 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
5540 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
5542 void sock_update_memcg(struct sock
*sk
)
5544 struct mem_cgroup
*memcg
;
5546 /* Socket cloning can throw us here with sk_cgrp already
5547 * filled. It won't however, necessarily happen from
5548 * process context. So the test for root memcg given
5549 * the current task's memcg won't help us in this case.
5551 * Respecting the original socket's memcg is a better
5552 * decision in this case.
5555 BUG_ON(mem_cgroup_is_root(sk
->sk_memcg
));
5556 css_get(&sk
->sk_memcg
->css
);
5561 memcg
= mem_cgroup_from_task(current
);
5562 if (memcg
== root_mem_cgroup
)
5564 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
5566 if (css_tryget_online(&memcg
->css
))
5567 sk
->sk_memcg
= memcg
;
5571 EXPORT_SYMBOL(sock_update_memcg
);
5573 void sock_release_memcg(struct sock
*sk
)
5575 WARN_ON(!sk
->sk_memcg
);
5576 css_put(&sk
->sk_memcg
->css
);
5580 * mem_cgroup_charge_skmem - charge socket memory
5581 * @memcg: memcg to charge
5582 * @nr_pages: number of pages to charge
5584 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5585 * @memcg's configured limit, %false if the charge had to be forced.
5587 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5589 gfp_t gfp_mask
= GFP_KERNEL
;
5591 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5592 struct page_counter
*fail
;
5594 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
5595 memcg
->tcpmem_pressure
= 0;
5598 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
5599 memcg
->tcpmem_pressure
= 1;
5603 /* Don't block in the packet receive path */
5605 gfp_mask
= GFP_NOWAIT
;
5607 this_cpu_add(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5609 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
5612 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
5617 * mem_cgroup_uncharge_skmem - uncharge socket memory
5618 * @memcg - memcg to uncharge
5619 * @nr_pages - number of pages to uncharge
5621 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5623 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5624 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
5628 this_cpu_sub(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5630 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5631 css_put_many(&memcg
->css
, nr_pages
);
5634 static int __init
cgroup_memory(char *s
)
5638 while ((token
= strsep(&s
, ",")) != NULL
) {
5641 if (!strcmp(token
, "nosocket"))
5642 cgroup_memory_nosocket
= true;
5643 if (!strcmp(token
, "nokmem"))
5644 cgroup_memory_nokmem
= true;
5648 __setup("cgroup.memory=", cgroup_memory
);
5651 * subsys_initcall() for memory controller.
5653 * Some parts like hotcpu_notifier() have to be initialized from this context
5654 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5655 * everything that doesn't depend on a specific mem_cgroup structure should
5656 * be initialized from here.
5658 static int __init
mem_cgroup_init(void)
5662 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5664 for_each_possible_cpu(cpu
)
5665 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5668 for_each_node(node
) {
5669 struct mem_cgroup_tree_per_node
*rtpn
;
5672 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5673 node_online(node
) ? node
: NUMA_NO_NODE
);
5675 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5676 struct mem_cgroup_tree_per_zone
*rtpz
;
5678 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
5679 rtpz
->rb_root
= RB_ROOT
;
5680 spin_lock_init(&rtpz
->lock
);
5682 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5687 subsys_initcall(mem_cgroup_init
);
5689 #ifdef CONFIG_MEMCG_SWAP
5691 * mem_cgroup_swapout - transfer a memsw charge to swap
5692 * @page: page whose memsw charge to transfer
5693 * @entry: swap entry to move the charge to
5695 * Transfer the memsw charge of @page to @entry.
5697 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5699 struct mem_cgroup
*memcg
;
5700 unsigned short oldid
;
5702 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5703 VM_BUG_ON_PAGE(page_count(page
), page
);
5705 if (!do_memsw_account())
5708 memcg
= page
->mem_cgroup
;
5710 /* Readahead page, never charged */
5714 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5715 VM_BUG_ON_PAGE(oldid
, page
);
5716 mem_cgroup_swap_statistics(memcg
, true);
5718 page
->mem_cgroup
= NULL
;
5720 if (!mem_cgroup_is_root(memcg
))
5721 page_counter_uncharge(&memcg
->memory
, 1);
5724 * Interrupts should be disabled here because the caller holds the
5725 * mapping->tree_lock lock which is taken with interrupts-off. It is
5726 * important here to have the interrupts disabled because it is the
5727 * only synchronisation we have for udpating the per-CPU variables.
5729 VM_BUG_ON(!irqs_disabled());
5730 mem_cgroup_charge_statistics(memcg
, page
, false, -1);
5731 memcg_check_events(memcg
, page
);
5735 * mem_cgroup_try_charge_swap - try charging a swap entry
5736 * @page: page being added to swap
5737 * @entry: swap entry to charge
5739 * Try to charge @entry to the memcg that @page belongs to.
5741 * Returns 0 on success, -ENOMEM on failure.
5743 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
5745 struct mem_cgroup
*memcg
;
5746 struct page_counter
*counter
;
5747 unsigned short oldid
;
5749 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
5752 memcg
= page
->mem_cgroup
;
5754 /* Readahead page, never charged */
5758 if (!mem_cgroup_is_root(memcg
) &&
5759 !page_counter_try_charge(&memcg
->swap
, 1, &counter
))
5762 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5763 VM_BUG_ON_PAGE(oldid
, page
);
5764 mem_cgroup_swap_statistics(memcg
, true);
5766 css_get(&memcg
->css
);
5771 * mem_cgroup_uncharge_swap - uncharge a swap entry
5772 * @entry: swap entry to uncharge
5774 * Drop the swap charge associated with @entry.
5776 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5778 struct mem_cgroup
*memcg
;
5781 if (!do_swap_account
)
5784 id
= swap_cgroup_record(entry
, 0);
5786 memcg
= mem_cgroup_from_id(id
);
5788 if (!mem_cgroup_is_root(memcg
)) {
5789 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5790 page_counter_uncharge(&memcg
->swap
, 1);
5792 page_counter_uncharge(&memcg
->memsw
, 1);
5794 mem_cgroup_swap_statistics(memcg
, false);
5795 css_put(&memcg
->css
);
5800 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
5802 long nr_swap_pages
= get_nr_swap_pages();
5804 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5805 return nr_swap_pages
;
5806 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5807 nr_swap_pages
= min_t(long, nr_swap_pages
,
5808 READ_ONCE(memcg
->swap
.limit
) -
5809 page_counter_read(&memcg
->swap
));
5810 return nr_swap_pages
;
5813 bool mem_cgroup_swap_full(struct page
*page
)
5815 struct mem_cgroup
*memcg
;
5817 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5821 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5824 memcg
= page
->mem_cgroup
;
5828 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5829 if (page_counter_read(&memcg
->swap
) * 2 >= memcg
->swap
.limit
)
5835 /* for remember boot option*/
5836 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5837 static int really_do_swap_account __initdata
= 1;
5839 static int really_do_swap_account __initdata
;
5842 static int __init
enable_swap_account(char *s
)
5844 if (!strcmp(s
, "1"))
5845 really_do_swap_account
= 1;
5846 else if (!strcmp(s
, "0"))
5847 really_do_swap_account
= 0;
5850 __setup("swapaccount=", enable_swap_account
);
5852 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
5855 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5857 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
5860 static int swap_max_show(struct seq_file
*m
, void *v
)
5862 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5863 unsigned long max
= READ_ONCE(memcg
->swap
.limit
);
5865 if (max
== PAGE_COUNTER_MAX
)
5866 seq_puts(m
, "max\n");
5868 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5873 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
5874 char *buf
, size_t nbytes
, loff_t off
)
5876 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5880 buf
= strstrip(buf
);
5881 err
= page_counter_memparse(buf
, "max", &max
);
5885 mutex_lock(&memcg_limit_mutex
);
5886 err
= page_counter_limit(&memcg
->swap
, max
);
5887 mutex_unlock(&memcg_limit_mutex
);
5894 static struct cftype swap_files
[] = {
5896 .name
= "swap.current",
5897 .flags
= CFTYPE_NOT_ON_ROOT
,
5898 .read_u64
= swap_current_read
,
5902 .flags
= CFTYPE_NOT_ON_ROOT
,
5903 .seq_show
= swap_max_show
,
5904 .write
= swap_max_write
,
5909 static struct cftype memsw_cgroup_files
[] = {
5911 .name
= "memsw.usage_in_bytes",
5912 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5913 .read_u64
= mem_cgroup_read_u64
,
5916 .name
= "memsw.max_usage_in_bytes",
5917 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5918 .write
= mem_cgroup_reset
,
5919 .read_u64
= mem_cgroup_read_u64
,
5922 .name
= "memsw.limit_in_bytes",
5923 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5924 .write
= mem_cgroup_write
,
5925 .read_u64
= mem_cgroup_read_u64
,
5928 .name
= "memsw.failcnt",
5929 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5930 .write
= mem_cgroup_reset
,
5931 .read_u64
= mem_cgroup_read_u64
,
5933 { }, /* terminate */
5936 static int __init
mem_cgroup_swap_init(void)
5938 if (!mem_cgroup_disabled() && really_do_swap_account
) {
5939 do_swap_account
= 1;
5940 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
5942 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
5943 memsw_cgroup_files
));
5947 subsys_initcall(mem_cgroup_swap_init
);
5949 #endif /* CONFIG_MEMCG_SWAP */