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 <linux/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_node
{
136 struct rb_root rb_root
;
140 struct mem_cgroup_tree
{
141 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
144 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
147 struct mem_cgroup_eventfd_list
{
148 struct list_head list
;
149 struct eventfd_ctx
*eventfd
;
153 * cgroup_event represents events which userspace want to receive.
155 struct mem_cgroup_event
{
157 * memcg which the event belongs to.
159 struct mem_cgroup
*memcg
;
161 * eventfd to signal userspace about the event.
163 struct eventfd_ctx
*eventfd
;
165 * Each of these stored in a list by the cgroup.
167 struct list_head list
;
169 * register_event() callback will be used to add new userspace
170 * waiter for changes related to this event. Use eventfd_signal()
171 * on eventfd to send notification to userspace.
173 int (*register_event
)(struct mem_cgroup
*memcg
,
174 struct eventfd_ctx
*eventfd
, const char *args
);
176 * unregister_event() callback will be called when userspace closes
177 * the eventfd or on cgroup removing. This callback must be set,
178 * if you want provide notification functionality.
180 void (*unregister_event
)(struct mem_cgroup
*memcg
,
181 struct eventfd_ctx
*eventfd
);
183 * All fields below needed to unregister event when
184 * userspace closes eventfd.
187 wait_queue_head_t
*wqh
;
189 struct work_struct remove
;
192 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
193 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
195 /* Stuffs for move charges at task migration. */
197 * Types of charges to be moved.
199 #define MOVE_ANON 0x1U
200 #define MOVE_FILE 0x2U
201 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
203 /* "mc" and its members are protected by cgroup_mutex */
204 static struct move_charge_struct
{
205 spinlock_t lock
; /* for from, to */
206 struct mm_struct
*mm
;
207 struct mem_cgroup
*from
;
208 struct mem_cgroup
*to
;
210 unsigned long precharge
;
211 unsigned long moved_charge
;
212 unsigned long moved_swap
;
213 struct task_struct
*moving_task
; /* a task moving charges */
214 wait_queue_head_t waitq
; /* a waitq for other context */
216 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
217 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
221 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
222 * limit reclaim to prevent infinite loops, if they ever occur.
224 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
225 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
228 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
229 MEM_CGROUP_CHARGE_TYPE_ANON
,
230 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
231 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
235 /* for encoding cft->private value on file */
244 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
245 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
246 #define MEMFILE_ATTR(val) ((val) & 0xffff)
247 /* Used for OOM nofiier */
248 #define OOM_CONTROL (0)
250 /* Some nice accessors for the vmpressure. */
251 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
254 memcg
= root_mem_cgroup
;
255 return &memcg
->vmpressure
;
258 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
260 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
263 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
265 return (memcg
== root_mem_cgroup
);
270 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
271 * The main reason for not using cgroup id for this:
272 * this works better in sparse environments, where we have a lot of memcgs,
273 * but only a few kmem-limited. Or also, if we have, for instance, 200
274 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
275 * 200 entry array for that.
277 * The current size of the caches array is stored in memcg_nr_cache_ids. It
278 * will double each time we have to increase it.
280 static DEFINE_IDA(memcg_cache_ida
);
281 int memcg_nr_cache_ids
;
283 /* Protects memcg_nr_cache_ids */
284 static DECLARE_RWSEM(memcg_cache_ids_sem
);
286 void memcg_get_cache_ids(void)
288 down_read(&memcg_cache_ids_sem
);
291 void memcg_put_cache_ids(void)
293 up_read(&memcg_cache_ids_sem
);
297 * MIN_SIZE is different than 1, because we would like to avoid going through
298 * the alloc/free process all the time. In a small machine, 4 kmem-limited
299 * cgroups is a reasonable guess. In the future, it could be a parameter or
300 * tunable, but that is strictly not necessary.
302 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
303 * this constant directly from cgroup, but it is understandable that this is
304 * better kept as an internal representation in cgroup.c. In any case, the
305 * cgrp_id space is not getting any smaller, and we don't have to necessarily
306 * increase ours as well if it increases.
308 #define MEMCG_CACHES_MIN_SIZE 4
309 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
312 * A lot of the calls to the cache allocation functions are expected to be
313 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
314 * conditional to this static branch, we'll have to allow modules that does
315 * kmem_cache_alloc and the such to see this symbol as well
317 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
318 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
320 #endif /* !CONFIG_SLOB */
323 * mem_cgroup_css_from_page - css of the memcg associated with a page
324 * @page: page of interest
326 * If memcg is bound to the default hierarchy, css of the memcg associated
327 * with @page is returned. The returned css remains associated with @page
328 * until it is released.
330 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
333 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
335 struct mem_cgroup
*memcg
;
337 memcg
= page
->mem_cgroup
;
339 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
340 memcg
= root_mem_cgroup
;
346 * page_cgroup_ino - return inode number of the memcg a page is charged to
349 * Look up the closest online ancestor of the memory cgroup @page is charged to
350 * and return its inode number or 0 if @page is not charged to any cgroup. It
351 * is safe to call this function without holding a reference to @page.
353 * Note, this function is inherently racy, because there is nothing to prevent
354 * the cgroup inode from getting torn down and potentially reallocated a moment
355 * after page_cgroup_ino() returns, so it only should be used by callers that
356 * do not care (such as procfs interfaces).
358 ino_t
page_cgroup_ino(struct page
*page
)
360 struct mem_cgroup
*memcg
;
361 unsigned long ino
= 0;
364 memcg
= READ_ONCE(page
->mem_cgroup
);
365 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
366 memcg
= parent_mem_cgroup(memcg
);
368 ino
= cgroup_ino(memcg
->css
.cgroup
);
373 static struct mem_cgroup_per_node
*
374 mem_cgroup_page_nodeinfo(struct mem_cgroup
*memcg
, struct page
*page
)
376 int nid
= page_to_nid(page
);
378 return memcg
->nodeinfo
[nid
];
381 static struct mem_cgroup_tree_per_node
*
382 soft_limit_tree_node(int nid
)
384 return soft_limit_tree
.rb_tree_per_node
[nid
];
387 static struct mem_cgroup_tree_per_node
*
388 soft_limit_tree_from_page(struct page
*page
)
390 int nid
= page_to_nid(page
);
392 return soft_limit_tree
.rb_tree_per_node
[nid
];
395 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node
*mz
,
396 struct mem_cgroup_tree_per_node
*mctz
,
397 unsigned long new_usage_in_excess
)
399 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
400 struct rb_node
*parent
= NULL
;
401 struct mem_cgroup_per_node
*mz_node
;
406 mz
->usage_in_excess
= new_usage_in_excess
;
407 if (!mz
->usage_in_excess
)
411 mz_node
= rb_entry(parent
, struct mem_cgroup_per_node
,
413 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
416 * We can't avoid mem cgroups that are over their soft
417 * limit by the same amount
419 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
422 rb_link_node(&mz
->tree_node
, parent
, p
);
423 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
427 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
428 struct mem_cgroup_tree_per_node
*mctz
)
432 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
436 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
437 struct mem_cgroup_tree_per_node
*mctz
)
441 spin_lock_irqsave(&mctz
->lock
, flags
);
442 __mem_cgroup_remove_exceeded(mz
, mctz
);
443 spin_unlock_irqrestore(&mctz
->lock
, flags
);
446 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
448 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
449 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
450 unsigned long excess
= 0;
452 if (nr_pages
> soft_limit
)
453 excess
= nr_pages
- soft_limit
;
458 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
460 unsigned long excess
;
461 struct mem_cgroup_per_node
*mz
;
462 struct mem_cgroup_tree_per_node
*mctz
;
464 mctz
= soft_limit_tree_from_page(page
);
468 * Necessary to update all ancestors when hierarchy is used.
469 * because their event counter is not touched.
471 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
472 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
473 excess
= soft_limit_excess(memcg
);
475 * We have to update the tree if mz is on RB-tree or
476 * mem is over its softlimit.
478 if (excess
|| mz
->on_tree
) {
481 spin_lock_irqsave(&mctz
->lock
, flags
);
482 /* if on-tree, remove it */
484 __mem_cgroup_remove_exceeded(mz
, mctz
);
486 * Insert again. mz->usage_in_excess will be updated.
487 * If excess is 0, no tree ops.
489 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
490 spin_unlock_irqrestore(&mctz
->lock
, flags
);
495 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
497 struct mem_cgroup_tree_per_node
*mctz
;
498 struct mem_cgroup_per_node
*mz
;
502 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
503 mctz
= soft_limit_tree_node(nid
);
505 mem_cgroup_remove_exceeded(mz
, mctz
);
509 static struct mem_cgroup_per_node
*
510 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
512 struct rb_node
*rightmost
= NULL
;
513 struct mem_cgroup_per_node
*mz
;
517 rightmost
= rb_last(&mctz
->rb_root
);
519 goto done
; /* Nothing to reclaim from */
521 mz
= rb_entry(rightmost
, struct mem_cgroup_per_node
, tree_node
);
523 * Remove the node now but someone else can add it back,
524 * we will to add it back at the end of reclaim to its correct
525 * position in the tree.
527 __mem_cgroup_remove_exceeded(mz
, mctz
);
528 if (!soft_limit_excess(mz
->memcg
) ||
529 !css_tryget_online(&mz
->memcg
->css
))
535 static struct mem_cgroup_per_node
*
536 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
538 struct mem_cgroup_per_node
*mz
;
540 spin_lock_irq(&mctz
->lock
);
541 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
542 spin_unlock_irq(&mctz
->lock
);
547 * Return page count for single (non recursive) @memcg.
549 * Implementation Note: reading percpu statistics for memcg.
551 * Both of vmstat[] and percpu_counter has threshold and do periodic
552 * synchronization to implement "quick" read. There are trade-off between
553 * reading cost and precision of value. Then, we may have a chance to implement
554 * a periodic synchronization of counter in memcg's counter.
556 * But this _read() function is used for user interface now. The user accounts
557 * memory usage by memory cgroup and he _always_ requires exact value because
558 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
559 * have to visit all online cpus and make sum. So, for now, unnecessary
560 * synchronization is not implemented. (just implemented for cpu hotplug)
562 * If there are kernel internal actions which can make use of some not-exact
563 * value, and reading all cpu value can be performance bottleneck in some
564 * common workload, threshold and synchronization as vmstat[] should be
568 mem_cgroup_read_stat(struct mem_cgroup
*memcg
, enum mem_cgroup_stat_index idx
)
573 /* Per-cpu values can be negative, use a signed accumulator */
574 for_each_possible_cpu(cpu
)
575 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
577 * Summing races with updates, so val may be negative. Avoid exposing
578 * transient negative values.
585 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
586 enum mem_cgroup_events_index idx
)
588 unsigned long val
= 0;
591 for_each_possible_cpu(cpu
)
592 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
596 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
598 bool compound
, int nr_pages
)
601 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
602 * counted as CACHE even if it's on ANON LRU.
605 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
608 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
612 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
613 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
617 /* pagein of a big page is an event. So, ignore page size */
619 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
621 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
622 nr_pages
= -nr_pages
; /* for event */
625 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
628 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
629 int nid
, unsigned int lru_mask
)
631 struct lruvec
*lruvec
= mem_cgroup_lruvec(NODE_DATA(nid
), memcg
);
632 unsigned long nr
= 0;
635 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
638 if (!(BIT(lru
) & lru_mask
))
640 nr
+= mem_cgroup_get_lru_size(lruvec
, lru
);
645 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
646 unsigned int lru_mask
)
648 unsigned long nr
= 0;
651 for_each_node_state(nid
, N_MEMORY
)
652 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
656 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
657 enum mem_cgroup_events_target target
)
659 unsigned long val
, next
;
661 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
662 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
663 /* from time_after() in jiffies.h */
664 if ((long)next
- (long)val
< 0) {
666 case MEM_CGROUP_TARGET_THRESH
:
667 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
669 case MEM_CGROUP_TARGET_SOFTLIMIT
:
670 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
672 case MEM_CGROUP_TARGET_NUMAINFO
:
673 next
= val
+ NUMAINFO_EVENTS_TARGET
;
678 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
685 * Check events in order.
688 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
690 /* threshold event is triggered in finer grain than soft limit */
691 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
692 MEM_CGROUP_TARGET_THRESH
))) {
694 bool do_numainfo __maybe_unused
;
696 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
697 MEM_CGROUP_TARGET_SOFTLIMIT
);
699 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
700 MEM_CGROUP_TARGET_NUMAINFO
);
702 mem_cgroup_threshold(memcg
);
703 if (unlikely(do_softlimit
))
704 mem_cgroup_update_tree(memcg
, page
);
706 if (unlikely(do_numainfo
))
707 atomic_inc(&memcg
->numainfo_events
);
712 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
715 * mm_update_next_owner() may clear mm->owner to NULL
716 * if it races with swapoff, page migration, etc.
717 * So this can be called with p == NULL.
722 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
724 EXPORT_SYMBOL(mem_cgroup_from_task
);
726 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
728 struct mem_cgroup
*memcg
= NULL
;
733 * Page cache insertions can happen withou an
734 * actual mm context, e.g. during disk probing
735 * on boot, loopback IO, acct() writes etc.
738 memcg
= root_mem_cgroup
;
740 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
741 if (unlikely(!memcg
))
742 memcg
= root_mem_cgroup
;
744 } while (!css_tryget_online(&memcg
->css
));
750 * mem_cgroup_iter - iterate over memory cgroup hierarchy
751 * @root: hierarchy root
752 * @prev: previously returned memcg, NULL on first invocation
753 * @reclaim: cookie for shared reclaim walks, NULL for full walks
755 * Returns references to children of the hierarchy below @root, or
756 * @root itself, or %NULL after a full round-trip.
758 * Caller must pass the return value in @prev on subsequent
759 * invocations for reference counting, or use mem_cgroup_iter_break()
760 * to cancel a hierarchy walk before the round-trip is complete.
762 * Reclaimers can specify a zone and a priority level in @reclaim to
763 * divide up the memcgs in the hierarchy among all concurrent
764 * reclaimers operating on the same zone and priority.
766 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
767 struct mem_cgroup
*prev
,
768 struct mem_cgroup_reclaim_cookie
*reclaim
)
770 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
771 struct cgroup_subsys_state
*css
= NULL
;
772 struct mem_cgroup
*memcg
= NULL
;
773 struct mem_cgroup
*pos
= NULL
;
775 if (mem_cgroup_disabled())
779 root
= root_mem_cgroup
;
781 if (prev
&& !reclaim
)
784 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
793 struct mem_cgroup_per_node
*mz
;
795 mz
= mem_cgroup_nodeinfo(root
, reclaim
->pgdat
->node_id
);
796 iter
= &mz
->iter
[reclaim
->priority
];
798 if (prev
&& reclaim
->generation
!= iter
->generation
)
802 pos
= READ_ONCE(iter
->position
);
803 if (!pos
|| css_tryget(&pos
->css
))
806 * css reference reached zero, so iter->position will
807 * be cleared by ->css_released. However, we should not
808 * rely on this happening soon, because ->css_released
809 * is called from a work queue, and by busy-waiting we
810 * might block it. So we clear iter->position right
813 (void)cmpxchg(&iter
->position
, pos
, NULL
);
821 css
= css_next_descendant_pre(css
, &root
->css
);
824 * Reclaimers share the hierarchy walk, and a
825 * new one might jump in right at the end of
826 * the hierarchy - make sure they see at least
827 * one group and restart from the beginning.
835 * Verify the css and acquire a reference. The root
836 * is provided by the caller, so we know it's alive
837 * and kicking, and don't take an extra reference.
839 memcg
= mem_cgroup_from_css(css
);
841 if (css
== &root
->css
)
852 * The position could have already been updated by a competing
853 * thread, so check that the value hasn't changed since we read
854 * it to avoid reclaiming from the same cgroup twice.
856 (void)cmpxchg(&iter
->position
, pos
, memcg
);
864 reclaim
->generation
= iter
->generation
;
870 if (prev
&& prev
!= root
)
877 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
878 * @root: hierarchy root
879 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
881 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
882 struct mem_cgroup
*prev
)
885 root
= root_mem_cgroup
;
886 if (prev
&& prev
!= root
)
890 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
892 struct mem_cgroup
*memcg
= dead_memcg
;
893 struct mem_cgroup_reclaim_iter
*iter
;
894 struct mem_cgroup_per_node
*mz
;
898 while ((memcg
= parent_mem_cgroup(memcg
))) {
900 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
901 for (i
= 0; i
<= DEF_PRIORITY
; i
++) {
903 cmpxchg(&iter
->position
,
911 * Iteration constructs for visiting all cgroups (under a tree). If
912 * loops are exited prematurely (break), mem_cgroup_iter_break() must
913 * be used for reference counting.
915 #define for_each_mem_cgroup_tree(iter, root) \
916 for (iter = mem_cgroup_iter(root, NULL, NULL); \
918 iter = mem_cgroup_iter(root, iter, NULL))
920 #define for_each_mem_cgroup(iter) \
921 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
923 iter = mem_cgroup_iter(NULL, iter, NULL))
926 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
927 * @memcg: hierarchy root
928 * @fn: function to call for each task
929 * @arg: argument passed to @fn
931 * This function iterates over tasks attached to @memcg or to any of its
932 * descendants and calls @fn for each task. If @fn returns a non-zero
933 * value, the function breaks the iteration loop and returns the value.
934 * Otherwise, it will iterate over all tasks and return 0.
936 * This function must not be called for the root memory cgroup.
938 int mem_cgroup_scan_tasks(struct mem_cgroup
*memcg
,
939 int (*fn
)(struct task_struct
*, void *), void *arg
)
941 struct mem_cgroup
*iter
;
944 BUG_ON(memcg
== root_mem_cgroup
);
946 for_each_mem_cgroup_tree(iter
, memcg
) {
947 struct css_task_iter it
;
948 struct task_struct
*task
;
950 css_task_iter_start(&iter
->css
, &it
);
951 while (!ret
&& (task
= css_task_iter_next(&it
)))
953 css_task_iter_end(&it
);
955 mem_cgroup_iter_break(memcg
, iter
);
963 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
965 * @zone: zone of the page
967 * This function is only safe when following the LRU page isolation
968 * and putback protocol: the LRU lock must be held, and the page must
969 * either be PageLRU() or the caller must have isolated/allocated it.
971 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct pglist_data
*pgdat
)
973 struct mem_cgroup_per_node
*mz
;
974 struct mem_cgroup
*memcg
;
975 struct lruvec
*lruvec
;
977 if (mem_cgroup_disabled()) {
978 lruvec
= &pgdat
->lruvec
;
982 memcg
= page
->mem_cgroup
;
984 * Swapcache readahead pages are added to the LRU - and
985 * possibly migrated - before they are charged.
988 memcg
= root_mem_cgroup
;
990 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
991 lruvec
= &mz
->lruvec
;
994 * Since a node can be onlined after the mem_cgroup was created,
995 * we have to be prepared to initialize lruvec->zone here;
996 * and if offlined then reonlined, we need to reinitialize it.
998 if (unlikely(lruvec
->pgdat
!= pgdat
))
999 lruvec
->pgdat
= pgdat
;
1004 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1005 * @lruvec: mem_cgroup per zone lru vector
1006 * @lru: index of lru list the page is sitting on
1007 * @zid: zone id of the accounted pages
1008 * @nr_pages: positive when adding or negative when removing
1010 * This function must be called under lru_lock, just before a page is added
1011 * to or just after a page is removed from an lru list (that ordering being
1012 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1014 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1015 int zid
, int nr_pages
)
1017 struct mem_cgroup_per_node
*mz
;
1018 unsigned long *lru_size
;
1021 if (mem_cgroup_disabled())
1024 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
1025 lru_size
= &mz
->lru_zone_size
[zid
][lru
];
1028 *lru_size
+= nr_pages
;
1031 if (WARN_ONCE(size
< 0,
1032 "%s(%p, %d, %d): lru_size %ld\n",
1033 __func__
, lruvec
, lru
, nr_pages
, size
)) {
1039 *lru_size
+= nr_pages
;
1042 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1044 struct mem_cgroup
*task_memcg
;
1045 struct task_struct
*p
;
1048 p
= find_lock_task_mm(task
);
1050 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1054 * All threads may have already detached their mm's, but the oom
1055 * killer still needs to detect if they have already been oom
1056 * killed to prevent needlessly killing additional tasks.
1059 task_memcg
= mem_cgroup_from_task(task
);
1060 css_get(&task_memcg
->css
);
1063 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1064 css_put(&task_memcg
->css
);
1069 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1070 * @memcg: the memory cgroup
1072 * Returns the maximum amount of memory @mem can be charged with, in
1075 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1077 unsigned long margin
= 0;
1078 unsigned long count
;
1079 unsigned long limit
;
1081 count
= page_counter_read(&memcg
->memory
);
1082 limit
= READ_ONCE(memcg
->memory
.limit
);
1084 margin
= limit
- count
;
1086 if (do_memsw_account()) {
1087 count
= page_counter_read(&memcg
->memsw
);
1088 limit
= READ_ONCE(memcg
->memsw
.limit
);
1090 margin
= min(margin
, limit
- count
);
1099 * A routine for checking "mem" is under move_account() or not.
1101 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1102 * moving cgroups. This is for waiting at high-memory pressure
1105 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1107 struct mem_cgroup
*from
;
1108 struct mem_cgroup
*to
;
1111 * Unlike task_move routines, we access mc.to, mc.from not under
1112 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1114 spin_lock(&mc
.lock
);
1120 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1121 mem_cgroup_is_descendant(to
, memcg
);
1123 spin_unlock(&mc
.lock
);
1127 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1129 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1130 if (mem_cgroup_under_move(memcg
)) {
1132 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1133 /* moving charge context might have finished. */
1136 finish_wait(&mc
.waitq
, &wait
);
1143 #define K(x) ((x) << (PAGE_SHIFT-10))
1145 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1146 * @memcg: The memory cgroup that went over limit
1147 * @p: Task that is going to be killed
1149 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1152 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1154 struct mem_cgroup
*iter
;
1160 pr_info("Task in ");
1161 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1162 pr_cont(" killed as a result of limit of ");
1164 pr_info("Memory limit reached of cgroup ");
1167 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1172 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1173 K((u64
)page_counter_read(&memcg
->memory
)),
1174 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1175 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1176 K((u64
)page_counter_read(&memcg
->memsw
)),
1177 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1178 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1179 K((u64
)page_counter_read(&memcg
->kmem
)),
1180 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1182 for_each_mem_cgroup_tree(iter
, memcg
) {
1183 pr_info("Memory cgroup stats for ");
1184 pr_cont_cgroup_path(iter
->css
.cgroup
);
1187 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1188 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1190 pr_cont(" %s:%luKB", mem_cgroup_stat_names
[i
],
1191 K(mem_cgroup_read_stat(iter
, i
)));
1194 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1195 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1196 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1203 * This function returns the number of memcg under hierarchy tree. Returns
1204 * 1(self count) if no children.
1206 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1209 struct mem_cgroup
*iter
;
1211 for_each_mem_cgroup_tree(iter
, memcg
)
1217 * Return the memory (and swap, if configured) limit for a memcg.
1219 unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1221 unsigned long limit
;
1223 limit
= memcg
->memory
.limit
;
1224 if (mem_cgroup_swappiness(memcg
)) {
1225 unsigned long memsw_limit
;
1226 unsigned long swap_limit
;
1228 memsw_limit
= memcg
->memsw
.limit
;
1229 swap_limit
= memcg
->swap
.limit
;
1230 swap_limit
= min(swap_limit
, (unsigned long)total_swap_pages
);
1231 limit
= min(limit
+ swap_limit
, memsw_limit
);
1236 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1239 struct oom_control oc
= {
1243 .gfp_mask
= gfp_mask
,
1248 mutex_lock(&oom_lock
);
1249 ret
= out_of_memory(&oc
);
1250 mutex_unlock(&oom_lock
);
1254 #if MAX_NUMNODES > 1
1257 * test_mem_cgroup_node_reclaimable
1258 * @memcg: the target memcg
1259 * @nid: the node ID to be checked.
1260 * @noswap : specify true here if the user wants flle only information.
1262 * This function returns whether the specified memcg contains any
1263 * reclaimable pages on a node. Returns true if there are any reclaimable
1264 * pages in the node.
1266 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1267 int nid
, bool noswap
)
1269 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1271 if (noswap
|| !total_swap_pages
)
1273 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1280 * Always updating the nodemask is not very good - even if we have an empty
1281 * list or the wrong list here, we can start from some node and traverse all
1282 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1285 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1289 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1290 * pagein/pageout changes since the last update.
1292 if (!atomic_read(&memcg
->numainfo_events
))
1294 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1297 /* make a nodemask where this memcg uses memory from */
1298 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1300 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1302 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1303 node_clear(nid
, memcg
->scan_nodes
);
1306 atomic_set(&memcg
->numainfo_events
, 0);
1307 atomic_set(&memcg
->numainfo_updating
, 0);
1311 * Selecting a node where we start reclaim from. Because what we need is just
1312 * reducing usage counter, start from anywhere is O,K. Considering
1313 * memory reclaim from current node, there are pros. and cons.
1315 * Freeing memory from current node means freeing memory from a node which
1316 * we'll use or we've used. So, it may make LRU bad. And if several threads
1317 * hit limits, it will see a contention on a node. But freeing from remote
1318 * node means more costs for memory reclaim because of memory latency.
1320 * Now, we use round-robin. Better algorithm is welcomed.
1322 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1326 mem_cgroup_may_update_nodemask(memcg
);
1327 node
= memcg
->last_scanned_node
;
1329 node
= next_node_in(node
, memcg
->scan_nodes
);
1331 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1332 * last time it really checked all the LRUs due to rate limiting.
1333 * Fallback to the current node in that case for simplicity.
1335 if (unlikely(node
== MAX_NUMNODES
))
1336 node
= numa_node_id();
1338 memcg
->last_scanned_node
= node
;
1342 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1348 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1351 unsigned long *total_scanned
)
1353 struct mem_cgroup
*victim
= NULL
;
1356 unsigned long excess
;
1357 unsigned long nr_scanned
;
1358 struct mem_cgroup_reclaim_cookie reclaim
= {
1363 excess
= soft_limit_excess(root_memcg
);
1366 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1371 * If we have not been able to reclaim
1372 * anything, it might because there are
1373 * no reclaimable pages under this hierarchy
1378 * We want to do more targeted reclaim.
1379 * excess >> 2 is not to excessive so as to
1380 * reclaim too much, nor too less that we keep
1381 * coming back to reclaim from this cgroup
1383 if (total
>= (excess
>> 2) ||
1384 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1389 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1390 pgdat
, &nr_scanned
);
1391 *total_scanned
+= nr_scanned
;
1392 if (!soft_limit_excess(root_memcg
))
1395 mem_cgroup_iter_break(root_memcg
, victim
);
1399 #ifdef CONFIG_LOCKDEP
1400 static struct lockdep_map memcg_oom_lock_dep_map
= {
1401 .name
= "memcg_oom_lock",
1405 static DEFINE_SPINLOCK(memcg_oom_lock
);
1408 * Check OOM-Killer is already running under our hierarchy.
1409 * If someone is running, return false.
1411 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1413 struct mem_cgroup
*iter
, *failed
= NULL
;
1415 spin_lock(&memcg_oom_lock
);
1417 for_each_mem_cgroup_tree(iter
, memcg
) {
1418 if (iter
->oom_lock
) {
1420 * this subtree of our hierarchy is already locked
1421 * so we cannot give a lock.
1424 mem_cgroup_iter_break(memcg
, iter
);
1427 iter
->oom_lock
= true;
1432 * OK, we failed to lock the whole subtree so we have
1433 * to clean up what we set up to the failing subtree
1435 for_each_mem_cgroup_tree(iter
, memcg
) {
1436 if (iter
== failed
) {
1437 mem_cgroup_iter_break(memcg
, iter
);
1440 iter
->oom_lock
= false;
1443 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1445 spin_unlock(&memcg_oom_lock
);
1450 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1452 struct mem_cgroup
*iter
;
1454 spin_lock(&memcg_oom_lock
);
1455 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1456 for_each_mem_cgroup_tree(iter
, memcg
)
1457 iter
->oom_lock
= false;
1458 spin_unlock(&memcg_oom_lock
);
1461 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1463 struct mem_cgroup
*iter
;
1465 spin_lock(&memcg_oom_lock
);
1466 for_each_mem_cgroup_tree(iter
, memcg
)
1468 spin_unlock(&memcg_oom_lock
);
1471 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1473 struct mem_cgroup
*iter
;
1476 * When a new child is created while the hierarchy is under oom,
1477 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1479 spin_lock(&memcg_oom_lock
);
1480 for_each_mem_cgroup_tree(iter
, memcg
)
1481 if (iter
->under_oom
> 0)
1483 spin_unlock(&memcg_oom_lock
);
1486 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1488 struct oom_wait_info
{
1489 struct mem_cgroup
*memcg
;
1493 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1494 unsigned mode
, int sync
, void *arg
)
1496 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1497 struct mem_cgroup
*oom_wait_memcg
;
1498 struct oom_wait_info
*oom_wait_info
;
1500 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1501 oom_wait_memcg
= oom_wait_info
->memcg
;
1503 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1504 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1506 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1509 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1512 * For the following lockless ->under_oom test, the only required
1513 * guarantee is that it must see the state asserted by an OOM when
1514 * this function is called as a result of userland actions
1515 * triggered by the notification of the OOM. This is trivially
1516 * achieved by invoking mem_cgroup_mark_under_oom() before
1517 * triggering notification.
1519 if (memcg
&& memcg
->under_oom
)
1520 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1523 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1525 if (!current
->memcg_may_oom
)
1528 * We are in the middle of the charge context here, so we
1529 * don't want to block when potentially sitting on a callstack
1530 * that holds all kinds of filesystem and mm locks.
1532 * Also, the caller may handle a failed allocation gracefully
1533 * (like optional page cache readahead) and so an OOM killer
1534 * invocation might not even be necessary.
1536 * That's why we don't do anything here except remember the
1537 * OOM context and then deal with it at the end of the page
1538 * fault when the stack is unwound, the locks are released,
1539 * and when we know whether the fault was overall successful.
1541 css_get(&memcg
->css
);
1542 current
->memcg_in_oom
= memcg
;
1543 current
->memcg_oom_gfp_mask
= mask
;
1544 current
->memcg_oom_order
= order
;
1548 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1549 * @handle: actually kill/wait or just clean up the OOM state
1551 * This has to be called at the end of a page fault if the memcg OOM
1552 * handler was enabled.
1554 * Memcg supports userspace OOM handling where failed allocations must
1555 * sleep on a waitqueue until the userspace task resolves the
1556 * situation. Sleeping directly in the charge context with all kinds
1557 * of locks held is not a good idea, instead we remember an OOM state
1558 * in the task and mem_cgroup_oom_synchronize() has to be called at
1559 * the end of the page fault to complete the OOM handling.
1561 * Returns %true if an ongoing memcg OOM situation was detected and
1562 * completed, %false otherwise.
1564 bool mem_cgroup_oom_synchronize(bool handle
)
1566 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1567 struct oom_wait_info owait
;
1570 /* OOM is global, do not handle */
1577 owait
.memcg
= memcg
;
1578 owait
.wait
.flags
= 0;
1579 owait
.wait
.func
= memcg_oom_wake_function
;
1580 owait
.wait
.private = current
;
1581 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1583 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1584 mem_cgroup_mark_under_oom(memcg
);
1586 locked
= mem_cgroup_oom_trylock(memcg
);
1589 mem_cgroup_oom_notify(memcg
);
1591 if (locked
&& !memcg
->oom_kill_disable
) {
1592 mem_cgroup_unmark_under_oom(memcg
);
1593 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1594 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1595 current
->memcg_oom_order
);
1598 mem_cgroup_unmark_under_oom(memcg
);
1599 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1603 mem_cgroup_oom_unlock(memcg
);
1605 * There is no guarantee that an OOM-lock contender
1606 * sees the wakeups triggered by the OOM kill
1607 * uncharges. Wake any sleepers explicitely.
1609 memcg_oom_recover(memcg
);
1612 current
->memcg_in_oom
= NULL
;
1613 css_put(&memcg
->css
);
1618 * lock_page_memcg - lock a page->mem_cgroup binding
1621 * This function protects unlocked LRU pages from being moved to
1622 * another cgroup and stabilizes their page->mem_cgroup binding.
1624 void lock_page_memcg(struct page
*page
)
1626 struct mem_cgroup
*memcg
;
1627 unsigned long flags
;
1630 * The RCU lock is held throughout the transaction. The fast
1631 * path can get away without acquiring the memcg->move_lock
1632 * because page moving starts with an RCU grace period.
1636 if (mem_cgroup_disabled())
1639 memcg
= page
->mem_cgroup
;
1640 if (unlikely(!memcg
))
1643 if (atomic_read(&memcg
->moving_account
) <= 0)
1646 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1647 if (memcg
!= page
->mem_cgroup
) {
1648 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1653 * When charge migration first begins, we can have locked and
1654 * unlocked page stat updates happening concurrently. Track
1655 * the task who has the lock for unlock_page_memcg().
1657 memcg
->move_lock_task
= current
;
1658 memcg
->move_lock_flags
= flags
;
1662 EXPORT_SYMBOL(lock_page_memcg
);
1665 * unlock_page_memcg - unlock a page->mem_cgroup binding
1668 void unlock_page_memcg(struct page
*page
)
1670 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
1672 if (memcg
&& memcg
->move_lock_task
== current
) {
1673 unsigned long flags
= memcg
->move_lock_flags
;
1675 memcg
->move_lock_task
= NULL
;
1676 memcg
->move_lock_flags
= 0;
1678 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1683 EXPORT_SYMBOL(unlock_page_memcg
);
1686 * size of first charge trial. "32" comes from vmscan.c's magic value.
1687 * TODO: maybe necessary to use big numbers in big irons.
1689 #define CHARGE_BATCH 32U
1690 struct memcg_stock_pcp
{
1691 struct mem_cgroup
*cached
; /* this never be root cgroup */
1692 unsigned int nr_pages
;
1693 struct work_struct work
;
1694 unsigned long flags
;
1695 #define FLUSHING_CACHED_CHARGE 0
1697 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1698 static DEFINE_MUTEX(percpu_charge_mutex
);
1701 * consume_stock: Try to consume stocked charge on this cpu.
1702 * @memcg: memcg to consume from.
1703 * @nr_pages: how many pages to charge.
1705 * The charges will only happen if @memcg matches the current cpu's memcg
1706 * stock, and at least @nr_pages are available in that stock. Failure to
1707 * service an allocation will refill the stock.
1709 * returns true if successful, false otherwise.
1711 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1713 struct memcg_stock_pcp
*stock
;
1714 unsigned long flags
;
1717 if (nr_pages
> CHARGE_BATCH
)
1720 local_irq_save(flags
);
1722 stock
= this_cpu_ptr(&memcg_stock
);
1723 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
1724 stock
->nr_pages
-= nr_pages
;
1728 local_irq_restore(flags
);
1734 * Returns stocks cached in percpu and reset cached information.
1736 static void drain_stock(struct memcg_stock_pcp
*stock
)
1738 struct mem_cgroup
*old
= stock
->cached
;
1740 if (stock
->nr_pages
) {
1741 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
1742 if (do_memsw_account())
1743 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
1744 css_put_many(&old
->css
, stock
->nr_pages
);
1745 stock
->nr_pages
= 0;
1747 stock
->cached
= NULL
;
1750 static void drain_local_stock(struct work_struct
*dummy
)
1752 struct memcg_stock_pcp
*stock
;
1753 unsigned long flags
;
1755 local_irq_save(flags
);
1757 stock
= this_cpu_ptr(&memcg_stock
);
1759 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
1761 local_irq_restore(flags
);
1765 * Cache charges(val) to local per_cpu area.
1766 * This will be consumed by consume_stock() function, later.
1768 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1770 struct memcg_stock_pcp
*stock
;
1771 unsigned long flags
;
1773 local_irq_save(flags
);
1775 stock
= this_cpu_ptr(&memcg_stock
);
1776 if (stock
->cached
!= memcg
) { /* reset if necessary */
1778 stock
->cached
= memcg
;
1780 stock
->nr_pages
+= nr_pages
;
1782 local_irq_restore(flags
);
1786 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1787 * of the hierarchy under it.
1789 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
1793 /* If someone's already draining, avoid adding running more workers. */
1794 if (!mutex_trylock(&percpu_charge_mutex
))
1796 /* Notify other cpus that system-wide "drain" is running */
1799 for_each_online_cpu(cpu
) {
1800 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1801 struct mem_cgroup
*memcg
;
1803 memcg
= stock
->cached
;
1804 if (!memcg
|| !stock
->nr_pages
)
1806 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
1808 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
1810 drain_local_stock(&stock
->work
);
1812 schedule_work_on(cpu
, &stock
->work
);
1817 mutex_unlock(&percpu_charge_mutex
);
1820 static int memcg_hotplug_cpu_dead(unsigned int cpu
)
1822 struct memcg_stock_pcp
*stock
;
1824 stock
= &per_cpu(memcg_stock
, cpu
);
1829 static void reclaim_high(struct mem_cgroup
*memcg
,
1830 unsigned int nr_pages
,
1834 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
1836 mem_cgroup_events(memcg
, MEMCG_HIGH
, 1);
1837 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
1838 } while ((memcg
= parent_mem_cgroup(memcg
)));
1841 static void high_work_func(struct work_struct
*work
)
1843 struct mem_cgroup
*memcg
;
1845 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
1846 reclaim_high(memcg
, CHARGE_BATCH
, GFP_KERNEL
);
1850 * Scheduled by try_charge() to be executed from the userland return path
1851 * and reclaims memory over the high limit.
1853 void mem_cgroup_handle_over_high(void)
1855 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
1856 struct mem_cgroup
*memcg
;
1858 if (likely(!nr_pages
))
1861 memcg
= get_mem_cgroup_from_mm(current
->mm
);
1862 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
1863 css_put(&memcg
->css
);
1864 current
->memcg_nr_pages_over_high
= 0;
1867 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1868 unsigned int nr_pages
)
1870 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
1871 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1872 struct mem_cgroup
*mem_over_limit
;
1873 struct page_counter
*counter
;
1874 unsigned long nr_reclaimed
;
1875 bool may_swap
= true;
1876 bool drained
= false;
1878 if (mem_cgroup_is_root(memcg
))
1881 if (consume_stock(memcg
, nr_pages
))
1884 if (!do_memsw_account() ||
1885 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
1886 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
1888 if (do_memsw_account())
1889 page_counter_uncharge(&memcg
->memsw
, batch
);
1890 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
1892 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
1896 if (batch
> nr_pages
) {
1902 * Unlike in global OOM situations, memcg is not in a physical
1903 * memory shortage. Allow dying and OOM-killed tasks to
1904 * bypass the last charges so that they can exit quickly and
1905 * free their memory.
1907 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
1908 fatal_signal_pending(current
) ||
1909 current
->flags
& PF_EXITING
))
1913 * Prevent unbounded recursion when reclaim operations need to
1914 * allocate memory. This might exceed the limits temporarily,
1915 * but we prefer facilitating memory reclaim and getting back
1916 * under the limit over triggering OOM kills in these cases.
1918 if (unlikely(current
->flags
& PF_MEMALLOC
))
1921 if (unlikely(task_in_memcg_oom(current
)))
1924 if (!gfpflags_allow_blocking(gfp_mask
))
1927 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
1929 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
1930 gfp_mask
, may_swap
);
1932 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
1936 drain_all_stock(mem_over_limit
);
1941 if (gfp_mask
& __GFP_NORETRY
)
1944 * Even though the limit is exceeded at this point, reclaim
1945 * may have been able to free some pages. Retry the charge
1946 * before killing the task.
1948 * Only for regular pages, though: huge pages are rather
1949 * unlikely to succeed so close to the limit, and we fall back
1950 * to regular pages anyway in case of failure.
1952 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
1955 * At task move, charge accounts can be doubly counted. So, it's
1956 * better to wait until the end of task_move if something is going on.
1958 if (mem_cgroup_wait_acct_move(mem_over_limit
))
1964 if (gfp_mask
& __GFP_NOFAIL
)
1967 if (fatal_signal_pending(current
))
1970 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
1972 mem_cgroup_oom(mem_over_limit
, gfp_mask
,
1973 get_order(nr_pages
* PAGE_SIZE
));
1975 if (!(gfp_mask
& __GFP_NOFAIL
))
1979 * The allocation either can't fail or will lead to more memory
1980 * being freed very soon. Allow memory usage go over the limit
1981 * temporarily by force charging it.
1983 page_counter_charge(&memcg
->memory
, nr_pages
);
1984 if (do_memsw_account())
1985 page_counter_charge(&memcg
->memsw
, nr_pages
);
1986 css_get_many(&memcg
->css
, nr_pages
);
1991 css_get_many(&memcg
->css
, batch
);
1992 if (batch
> nr_pages
)
1993 refill_stock(memcg
, batch
- nr_pages
);
1996 * If the hierarchy is above the normal consumption range, schedule
1997 * reclaim on returning to userland. We can perform reclaim here
1998 * if __GFP_RECLAIM but let's always punt for simplicity and so that
1999 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2000 * not recorded as it most likely matches current's and won't
2001 * change in the meantime. As high limit is checked again before
2002 * reclaim, the cost of mismatch is negligible.
2005 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2006 /* Don't bother a random interrupted task */
2007 if (in_interrupt()) {
2008 schedule_work(&memcg
->high_work
);
2011 current
->memcg_nr_pages_over_high
+= batch
;
2012 set_notify_resume(current
);
2015 } while ((memcg
= parent_mem_cgroup(memcg
)));
2020 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2022 if (mem_cgroup_is_root(memcg
))
2025 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2026 if (do_memsw_account())
2027 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2029 css_put_many(&memcg
->css
, nr_pages
);
2032 static void lock_page_lru(struct page
*page
, int *isolated
)
2034 struct zone
*zone
= page_zone(page
);
2036 spin_lock_irq(zone_lru_lock(zone
));
2037 if (PageLRU(page
)) {
2038 struct lruvec
*lruvec
;
2040 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
2042 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2048 static void unlock_page_lru(struct page
*page
, int isolated
)
2050 struct zone
*zone
= page_zone(page
);
2053 struct lruvec
*lruvec
;
2055 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
2056 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2058 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2060 spin_unlock_irq(zone_lru_lock(zone
));
2063 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2068 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2071 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2072 * may already be on some other mem_cgroup's LRU. Take care of it.
2075 lock_page_lru(page
, &isolated
);
2078 * Nobody should be changing or seriously looking at
2079 * page->mem_cgroup at this point:
2081 * - the page is uncharged
2083 * - the page is off-LRU
2085 * - an anonymous fault has exclusive page access, except for
2086 * a locked page table
2088 * - a page cache insertion, a swapin fault, or a migration
2089 * have the page locked
2091 page
->mem_cgroup
= memcg
;
2094 unlock_page_lru(page
, isolated
);
2098 static int memcg_alloc_cache_id(void)
2103 id
= ida_simple_get(&memcg_cache_ida
,
2104 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2108 if (id
< memcg_nr_cache_ids
)
2112 * There's no space for the new id in memcg_caches arrays,
2113 * so we have to grow them.
2115 down_write(&memcg_cache_ids_sem
);
2117 size
= 2 * (id
+ 1);
2118 if (size
< MEMCG_CACHES_MIN_SIZE
)
2119 size
= MEMCG_CACHES_MIN_SIZE
;
2120 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2121 size
= MEMCG_CACHES_MAX_SIZE
;
2123 err
= memcg_update_all_caches(size
);
2125 err
= memcg_update_all_list_lrus(size
);
2127 memcg_nr_cache_ids
= size
;
2129 up_write(&memcg_cache_ids_sem
);
2132 ida_simple_remove(&memcg_cache_ida
, id
);
2138 static void memcg_free_cache_id(int id
)
2140 ida_simple_remove(&memcg_cache_ida
, id
);
2143 struct memcg_kmem_cache_create_work
{
2144 struct mem_cgroup
*memcg
;
2145 struct kmem_cache
*cachep
;
2146 struct work_struct work
;
2149 static struct workqueue_struct
*memcg_kmem_cache_create_wq
;
2151 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2153 struct memcg_kmem_cache_create_work
*cw
=
2154 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2155 struct mem_cgroup
*memcg
= cw
->memcg
;
2156 struct kmem_cache
*cachep
= cw
->cachep
;
2158 memcg_create_kmem_cache(memcg
, cachep
);
2160 css_put(&memcg
->css
);
2165 * Enqueue the creation of a per-memcg kmem_cache.
2167 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2168 struct kmem_cache
*cachep
)
2170 struct memcg_kmem_cache_create_work
*cw
;
2172 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2176 css_get(&memcg
->css
);
2179 cw
->cachep
= cachep
;
2180 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2182 queue_work(memcg_kmem_cache_create_wq
, &cw
->work
);
2185 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2186 struct kmem_cache
*cachep
)
2189 * We need to stop accounting when we kmalloc, because if the
2190 * corresponding kmalloc cache is not yet created, the first allocation
2191 * in __memcg_schedule_kmem_cache_create will recurse.
2193 * However, it is better to enclose the whole function. Depending on
2194 * the debugging options enabled, INIT_WORK(), for instance, can
2195 * trigger an allocation. This too, will make us recurse. Because at
2196 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2197 * the safest choice is to do it like this, wrapping the whole function.
2199 current
->memcg_kmem_skip_account
= 1;
2200 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2201 current
->memcg_kmem_skip_account
= 0;
2204 static inline bool memcg_kmem_bypass(void)
2206 if (in_interrupt() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
2212 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2213 * @cachep: the original global kmem cache
2215 * Return the kmem_cache we're supposed to use for a slab allocation.
2216 * We try to use the current memcg's version of the cache.
2218 * If the cache does not exist yet, if we are the first user of it, we
2219 * create it asynchronously in a workqueue and let the current allocation
2220 * go through with the original cache.
2222 * This function takes a reference to the cache it returns to assure it
2223 * won't get destroyed while we are working with it. Once the caller is
2224 * done with it, memcg_kmem_put_cache() must be called to release the
2227 struct kmem_cache
*memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2229 struct mem_cgroup
*memcg
;
2230 struct kmem_cache
*memcg_cachep
;
2233 VM_BUG_ON(!is_root_cache(cachep
));
2235 if (memcg_kmem_bypass())
2238 if (current
->memcg_kmem_skip_account
)
2241 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2242 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2246 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2247 if (likely(memcg_cachep
))
2248 return memcg_cachep
;
2251 * If we are in a safe context (can wait, and not in interrupt
2252 * context), we could be be predictable and return right away.
2253 * This would guarantee that the allocation being performed
2254 * already belongs in the new cache.
2256 * However, there are some clashes that can arrive from locking.
2257 * For instance, because we acquire the slab_mutex while doing
2258 * memcg_create_kmem_cache, this means no further allocation
2259 * could happen with the slab_mutex held. So it's better to
2262 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2264 css_put(&memcg
->css
);
2269 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2270 * @cachep: the cache returned by memcg_kmem_get_cache
2272 void memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2274 if (!is_root_cache(cachep
))
2275 css_put(&cachep
->memcg_params
.memcg
->css
);
2279 * memcg_kmem_charge: charge a kmem page
2280 * @page: page to charge
2281 * @gfp: reclaim mode
2282 * @order: allocation order
2283 * @memcg: memory cgroup to charge
2285 * Returns 0 on success, an error code on failure.
2287 int memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2288 struct mem_cgroup
*memcg
)
2290 unsigned int nr_pages
= 1 << order
;
2291 struct page_counter
*counter
;
2294 ret
= try_charge(memcg
, gfp
, nr_pages
);
2298 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2299 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2300 cancel_charge(memcg
, nr_pages
);
2304 page
->mem_cgroup
= memcg
;
2310 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2311 * @page: page to charge
2312 * @gfp: reclaim mode
2313 * @order: allocation order
2315 * Returns 0 on success, an error code on failure.
2317 int memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2319 struct mem_cgroup
*memcg
;
2322 if (memcg_kmem_bypass())
2325 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2326 if (!mem_cgroup_is_root(memcg
)) {
2327 ret
= memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2329 __SetPageKmemcg(page
);
2331 css_put(&memcg
->css
);
2335 * memcg_kmem_uncharge: uncharge a kmem page
2336 * @page: page to uncharge
2337 * @order: allocation order
2339 void memcg_kmem_uncharge(struct page
*page
, int order
)
2341 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2342 unsigned int nr_pages
= 1 << order
;
2347 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2349 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2350 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2352 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2353 if (do_memsw_account())
2354 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2356 page
->mem_cgroup
= NULL
;
2358 /* slab pages do not have PageKmemcg flag set */
2359 if (PageKmemcg(page
))
2360 __ClearPageKmemcg(page
);
2362 css_put_many(&memcg
->css
, nr_pages
);
2364 #endif /* !CONFIG_SLOB */
2366 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2369 * Because tail pages are not marked as "used", set it. We're under
2370 * zone_lru_lock and migration entries setup in all page mappings.
2372 void mem_cgroup_split_huge_fixup(struct page
*head
)
2376 if (mem_cgroup_disabled())
2379 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2380 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2382 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2385 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2387 #ifdef CONFIG_MEMCG_SWAP
2388 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2391 int val
= (charge
) ? 1 : -1;
2392 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2396 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2397 * @entry: swap entry to be moved
2398 * @from: mem_cgroup which the entry is moved from
2399 * @to: mem_cgroup which the entry is moved to
2401 * It succeeds only when the swap_cgroup's record for this entry is the same
2402 * as the mem_cgroup's id of @from.
2404 * Returns 0 on success, -EINVAL on failure.
2406 * The caller must have charged to @to, IOW, called page_counter_charge() about
2407 * both res and memsw, and called css_get().
2409 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2410 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2412 unsigned short old_id
, new_id
;
2414 old_id
= mem_cgroup_id(from
);
2415 new_id
= mem_cgroup_id(to
);
2417 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2418 mem_cgroup_swap_statistics(from
, false);
2419 mem_cgroup_swap_statistics(to
, true);
2425 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2426 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2432 static DEFINE_MUTEX(memcg_limit_mutex
);
2434 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2435 unsigned long limit
)
2437 unsigned long curusage
;
2438 unsigned long oldusage
;
2439 bool enlarge
= false;
2444 * For keeping hierarchical_reclaim simple, how long we should retry
2445 * is depends on callers. We set our retry-count to be function
2446 * of # of children which we should visit in this loop.
2448 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2449 mem_cgroup_count_children(memcg
);
2451 oldusage
= page_counter_read(&memcg
->memory
);
2454 if (signal_pending(current
)) {
2459 mutex_lock(&memcg_limit_mutex
);
2460 if (limit
> memcg
->memsw
.limit
) {
2461 mutex_unlock(&memcg_limit_mutex
);
2465 if (limit
> memcg
->memory
.limit
)
2467 ret
= page_counter_limit(&memcg
->memory
, limit
);
2468 mutex_unlock(&memcg_limit_mutex
);
2473 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2475 curusage
= page_counter_read(&memcg
->memory
);
2476 /* Usage is reduced ? */
2477 if (curusage
>= oldusage
)
2480 oldusage
= curusage
;
2481 } while (retry_count
);
2483 if (!ret
&& enlarge
)
2484 memcg_oom_recover(memcg
);
2489 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2490 unsigned long limit
)
2492 unsigned long curusage
;
2493 unsigned long oldusage
;
2494 bool enlarge
= false;
2498 /* see mem_cgroup_resize_res_limit */
2499 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2500 mem_cgroup_count_children(memcg
);
2502 oldusage
= page_counter_read(&memcg
->memsw
);
2505 if (signal_pending(current
)) {
2510 mutex_lock(&memcg_limit_mutex
);
2511 if (limit
< memcg
->memory
.limit
) {
2512 mutex_unlock(&memcg_limit_mutex
);
2516 if (limit
> memcg
->memsw
.limit
)
2518 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2519 mutex_unlock(&memcg_limit_mutex
);
2524 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2526 curusage
= page_counter_read(&memcg
->memsw
);
2527 /* Usage is reduced ? */
2528 if (curusage
>= oldusage
)
2531 oldusage
= curusage
;
2532 } while (retry_count
);
2534 if (!ret
&& enlarge
)
2535 memcg_oom_recover(memcg
);
2540 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
2542 unsigned long *total_scanned
)
2544 unsigned long nr_reclaimed
= 0;
2545 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
2546 unsigned long reclaimed
;
2548 struct mem_cgroup_tree_per_node
*mctz
;
2549 unsigned long excess
;
2550 unsigned long nr_scanned
;
2555 mctz
= soft_limit_tree_node(pgdat
->node_id
);
2558 * Do not even bother to check the largest node if the root
2559 * is empty. Do it lockless to prevent lock bouncing. Races
2560 * are acceptable as soft limit is best effort anyway.
2562 if (!mctz
|| RB_EMPTY_ROOT(&mctz
->rb_root
))
2566 * This loop can run a while, specially if mem_cgroup's continuously
2567 * keep exceeding their soft limit and putting the system under
2574 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2579 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
2580 gfp_mask
, &nr_scanned
);
2581 nr_reclaimed
+= reclaimed
;
2582 *total_scanned
+= nr_scanned
;
2583 spin_lock_irq(&mctz
->lock
);
2584 __mem_cgroup_remove_exceeded(mz
, mctz
);
2587 * If we failed to reclaim anything from this memory cgroup
2588 * it is time to move on to the next cgroup
2592 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2594 excess
= soft_limit_excess(mz
->memcg
);
2596 * One school of thought says that we should not add
2597 * back the node to the tree if reclaim returns 0.
2598 * But our reclaim could return 0, simply because due
2599 * to priority we are exposing a smaller subset of
2600 * memory to reclaim from. Consider this as a longer
2603 /* If excess == 0, no tree ops */
2604 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2605 spin_unlock_irq(&mctz
->lock
);
2606 css_put(&mz
->memcg
->css
);
2609 * Could not reclaim anything and there are no more
2610 * mem cgroups to try or we seem to be looping without
2611 * reclaiming anything.
2613 if (!nr_reclaimed
&&
2615 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2617 } while (!nr_reclaimed
);
2619 css_put(&next_mz
->memcg
->css
);
2620 return nr_reclaimed
;
2624 * Test whether @memcg has children, dead or alive. Note that this
2625 * function doesn't care whether @memcg has use_hierarchy enabled and
2626 * returns %true if there are child csses according to the cgroup
2627 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2629 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2634 ret
= css_next_child(NULL
, &memcg
->css
);
2640 * Reclaims as many pages from the given memcg as possible.
2642 * Caller is responsible for holding css reference for memcg.
2644 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2646 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2648 /* we call try-to-free pages for make this cgroup empty */
2649 lru_add_drain_all();
2650 /* try to free all pages in this cgroup */
2651 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2654 if (signal_pending(current
))
2657 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2661 /* maybe some writeback is necessary */
2662 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2670 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
2671 char *buf
, size_t nbytes
,
2674 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2676 if (mem_cgroup_is_root(memcg
))
2678 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
2681 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
2684 return mem_cgroup_from_css(css
)->use_hierarchy
;
2687 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
2688 struct cftype
*cft
, u64 val
)
2691 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2692 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
2694 if (memcg
->use_hierarchy
== val
)
2698 * If parent's use_hierarchy is set, we can't make any modifications
2699 * in the child subtrees. If it is unset, then the change can
2700 * occur, provided the current cgroup has no children.
2702 * For the root cgroup, parent_mem is NULL, we allow value to be
2703 * set if there are no children.
2705 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
2706 (val
== 1 || val
== 0)) {
2707 if (!memcg_has_children(memcg
))
2708 memcg
->use_hierarchy
= val
;
2717 static void tree_stat(struct mem_cgroup
*memcg
, unsigned long *stat
)
2719 struct mem_cgroup
*iter
;
2722 memset(stat
, 0, sizeof(*stat
) * MEMCG_NR_STAT
);
2724 for_each_mem_cgroup_tree(iter
, memcg
) {
2725 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
2726 stat
[i
] += mem_cgroup_read_stat(iter
, i
);
2730 static void tree_events(struct mem_cgroup
*memcg
, unsigned long *events
)
2732 struct mem_cgroup
*iter
;
2735 memset(events
, 0, sizeof(*events
) * MEMCG_NR_EVENTS
);
2737 for_each_mem_cgroup_tree(iter
, memcg
) {
2738 for (i
= 0; i
< MEMCG_NR_EVENTS
; i
++)
2739 events
[i
] += mem_cgroup_read_events(iter
, i
);
2743 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
2745 unsigned long val
= 0;
2747 if (mem_cgroup_is_root(memcg
)) {
2748 struct mem_cgroup
*iter
;
2750 for_each_mem_cgroup_tree(iter
, memcg
) {
2751 val
+= mem_cgroup_read_stat(iter
,
2752 MEM_CGROUP_STAT_CACHE
);
2753 val
+= mem_cgroup_read_stat(iter
,
2754 MEM_CGROUP_STAT_RSS
);
2756 val
+= mem_cgroup_read_stat(iter
,
2757 MEM_CGROUP_STAT_SWAP
);
2761 val
= page_counter_read(&memcg
->memory
);
2763 val
= page_counter_read(&memcg
->memsw
);
2776 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
2779 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2780 struct page_counter
*counter
;
2782 switch (MEMFILE_TYPE(cft
->private)) {
2784 counter
= &memcg
->memory
;
2787 counter
= &memcg
->memsw
;
2790 counter
= &memcg
->kmem
;
2793 counter
= &memcg
->tcpmem
;
2799 switch (MEMFILE_ATTR(cft
->private)) {
2801 if (counter
== &memcg
->memory
)
2802 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
2803 if (counter
== &memcg
->memsw
)
2804 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
2805 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
2807 return (u64
)counter
->limit
* PAGE_SIZE
;
2809 return (u64
)counter
->watermark
* PAGE_SIZE
;
2811 return counter
->failcnt
;
2812 case RES_SOFT_LIMIT
:
2813 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
2820 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2824 if (cgroup_memory_nokmem
)
2827 BUG_ON(memcg
->kmemcg_id
>= 0);
2828 BUG_ON(memcg
->kmem_state
);
2830 memcg_id
= memcg_alloc_cache_id();
2834 static_branch_inc(&memcg_kmem_enabled_key
);
2836 * A memory cgroup is considered kmem-online as soon as it gets
2837 * kmemcg_id. Setting the id after enabling static branching will
2838 * guarantee no one starts accounting before all call sites are
2841 memcg
->kmemcg_id
= memcg_id
;
2842 memcg
->kmem_state
= KMEM_ONLINE
;
2847 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2849 struct cgroup_subsys_state
*css
;
2850 struct mem_cgroup
*parent
, *child
;
2853 if (memcg
->kmem_state
!= KMEM_ONLINE
)
2856 * Clear the online state before clearing memcg_caches array
2857 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2858 * guarantees that no cache will be created for this cgroup
2859 * after we are done (see memcg_create_kmem_cache()).
2861 memcg
->kmem_state
= KMEM_ALLOCATED
;
2863 memcg_deactivate_kmem_caches(memcg
);
2865 kmemcg_id
= memcg
->kmemcg_id
;
2866 BUG_ON(kmemcg_id
< 0);
2868 parent
= parent_mem_cgroup(memcg
);
2870 parent
= root_mem_cgroup
;
2873 * Change kmemcg_id of this cgroup and all its descendants to the
2874 * parent's id, and then move all entries from this cgroup's list_lrus
2875 * to ones of the parent. After we have finished, all list_lrus
2876 * corresponding to this cgroup are guaranteed to remain empty. The
2877 * ordering is imposed by list_lru_node->lock taken by
2878 * memcg_drain_all_list_lrus().
2880 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2881 css_for_each_descendant_pre(css
, &memcg
->css
) {
2882 child
= mem_cgroup_from_css(css
);
2883 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
2884 child
->kmemcg_id
= parent
->kmemcg_id
;
2885 if (!memcg
->use_hierarchy
)
2890 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
2892 memcg_free_cache_id(kmemcg_id
);
2895 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2897 /* css_alloc() failed, offlining didn't happen */
2898 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
2899 memcg_offline_kmem(memcg
);
2901 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
2902 memcg_destroy_kmem_caches(memcg
);
2903 static_branch_dec(&memcg_kmem_enabled_key
);
2904 WARN_ON(page_counter_read(&memcg
->kmem
));
2908 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2912 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2915 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2918 #endif /* !CONFIG_SLOB */
2920 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2921 unsigned long limit
)
2925 mutex_lock(&memcg_limit_mutex
);
2926 ret
= page_counter_limit(&memcg
->kmem
, limit
);
2927 mutex_unlock(&memcg_limit_mutex
);
2931 static int memcg_update_tcp_limit(struct mem_cgroup
*memcg
, unsigned long limit
)
2935 mutex_lock(&memcg_limit_mutex
);
2937 ret
= page_counter_limit(&memcg
->tcpmem
, limit
);
2941 if (!memcg
->tcpmem_active
) {
2943 * The active flag needs to be written after the static_key
2944 * update. This is what guarantees that the socket activation
2945 * function is the last one to run. See mem_cgroup_sk_alloc()
2946 * for details, and note that we don't mark any socket as
2947 * belonging to this memcg until that flag is up.
2949 * We need to do this, because static_keys will span multiple
2950 * sites, but we can't control their order. If we mark a socket
2951 * as accounted, but the accounting functions are not patched in
2952 * yet, we'll lose accounting.
2954 * We never race with the readers in mem_cgroup_sk_alloc(),
2955 * because when this value change, the code to process it is not
2958 static_branch_inc(&memcg_sockets_enabled_key
);
2959 memcg
->tcpmem_active
= true;
2962 mutex_unlock(&memcg_limit_mutex
);
2967 * The user of this function is...
2970 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
2971 char *buf
, size_t nbytes
, loff_t off
)
2973 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2974 unsigned long nr_pages
;
2977 buf
= strstrip(buf
);
2978 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
2982 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
2984 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
2988 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
2990 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
2993 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
2996 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
2999 ret
= memcg_update_tcp_limit(memcg
, nr_pages
);
3003 case RES_SOFT_LIMIT
:
3004 memcg
->soft_limit
= nr_pages
;
3008 return ret
?: nbytes
;
3011 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3012 size_t nbytes
, loff_t off
)
3014 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3015 struct page_counter
*counter
;
3017 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3019 counter
= &memcg
->memory
;
3022 counter
= &memcg
->memsw
;
3025 counter
= &memcg
->kmem
;
3028 counter
= &memcg
->tcpmem
;
3034 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3036 page_counter_reset_watermark(counter
);
3039 counter
->failcnt
= 0;
3048 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3051 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3055 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3056 struct cftype
*cft
, u64 val
)
3058 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3060 if (val
& ~MOVE_MASK
)
3064 * No kind of locking is needed in here, because ->can_attach() will
3065 * check this value once in the beginning of the process, and then carry
3066 * on with stale data. This means that changes to this value will only
3067 * affect task migrations starting after the change.
3069 memcg
->move_charge_at_immigrate
= val
;
3073 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3074 struct cftype
*cft
, u64 val
)
3081 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3085 unsigned int lru_mask
;
3088 static const struct numa_stat stats
[] = {
3089 { "total", LRU_ALL
},
3090 { "file", LRU_ALL_FILE
},
3091 { "anon", LRU_ALL_ANON
},
3092 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3094 const struct numa_stat
*stat
;
3097 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3099 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3100 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3101 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3102 for_each_node_state(nid
, N_MEMORY
) {
3103 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3105 seq_printf(m
, " N%d=%lu", nid
, nr
);
3110 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3111 struct mem_cgroup
*iter
;
3114 for_each_mem_cgroup_tree(iter
, memcg
)
3115 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3116 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3117 for_each_node_state(nid
, N_MEMORY
) {
3119 for_each_mem_cgroup_tree(iter
, memcg
)
3120 nr
+= mem_cgroup_node_nr_lru_pages(
3121 iter
, nid
, stat
->lru_mask
);
3122 seq_printf(m
, " N%d=%lu", nid
, nr
);
3129 #endif /* CONFIG_NUMA */
3131 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3133 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3134 unsigned long memory
, memsw
;
3135 struct mem_cgroup
*mi
;
3138 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3139 MEM_CGROUP_STAT_NSTATS
);
3140 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3141 MEM_CGROUP_EVENTS_NSTATS
);
3142 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3144 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3145 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3147 seq_printf(m
, "%s %lu\n", mem_cgroup_stat_names
[i
],
3148 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3151 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3152 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3153 mem_cgroup_read_events(memcg
, i
));
3155 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3156 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3157 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3159 /* Hierarchical information */
3160 memory
= memsw
= PAGE_COUNTER_MAX
;
3161 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3162 memory
= min(memory
, mi
->memory
.limit
);
3163 memsw
= min(memsw
, mi
->memsw
.limit
);
3165 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3166 (u64
)memory
* PAGE_SIZE
);
3167 if (do_memsw_account())
3168 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3169 (u64
)memsw
* PAGE_SIZE
);
3171 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3172 unsigned long long val
= 0;
3174 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3176 for_each_mem_cgroup_tree(mi
, memcg
)
3177 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3178 seq_printf(m
, "total_%s %llu\n", mem_cgroup_stat_names
[i
], val
);
3181 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3182 unsigned long long val
= 0;
3184 for_each_mem_cgroup_tree(mi
, memcg
)
3185 val
+= mem_cgroup_read_events(mi
, i
);
3186 seq_printf(m
, "total_%s %llu\n",
3187 mem_cgroup_events_names
[i
], val
);
3190 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3191 unsigned long long val
= 0;
3193 for_each_mem_cgroup_tree(mi
, memcg
)
3194 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3195 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3198 #ifdef CONFIG_DEBUG_VM
3201 struct mem_cgroup_per_node
*mz
;
3202 struct zone_reclaim_stat
*rstat
;
3203 unsigned long recent_rotated
[2] = {0, 0};
3204 unsigned long recent_scanned
[2] = {0, 0};
3206 for_each_online_pgdat(pgdat
) {
3207 mz
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
3208 rstat
= &mz
->lruvec
.reclaim_stat
;
3210 recent_rotated
[0] += rstat
->recent_rotated
[0];
3211 recent_rotated
[1] += rstat
->recent_rotated
[1];
3212 recent_scanned
[0] += rstat
->recent_scanned
[0];
3213 recent_scanned
[1] += rstat
->recent_scanned
[1];
3215 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3216 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3217 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3218 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3225 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3228 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3230 return mem_cgroup_swappiness(memcg
);
3233 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3234 struct cftype
*cft
, u64 val
)
3236 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3242 memcg
->swappiness
= val
;
3244 vm_swappiness
= val
;
3249 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3251 struct mem_cgroup_threshold_ary
*t
;
3252 unsigned long usage
;
3257 t
= rcu_dereference(memcg
->thresholds
.primary
);
3259 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3264 usage
= mem_cgroup_usage(memcg
, swap
);
3267 * current_threshold points to threshold just below or equal to usage.
3268 * If it's not true, a threshold was crossed after last
3269 * call of __mem_cgroup_threshold().
3271 i
= t
->current_threshold
;
3274 * Iterate backward over array of thresholds starting from
3275 * current_threshold and check if a threshold is crossed.
3276 * If none of thresholds below usage is crossed, we read
3277 * only one element of the array here.
3279 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3280 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3282 /* i = current_threshold + 1 */
3286 * Iterate forward over array of thresholds starting from
3287 * current_threshold+1 and check if a threshold is crossed.
3288 * If none of thresholds above usage is crossed, we read
3289 * only one element of the array here.
3291 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3292 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3294 /* Update current_threshold */
3295 t
->current_threshold
= i
- 1;
3300 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3303 __mem_cgroup_threshold(memcg
, false);
3304 if (do_memsw_account())
3305 __mem_cgroup_threshold(memcg
, true);
3307 memcg
= parent_mem_cgroup(memcg
);
3311 static int compare_thresholds(const void *a
, const void *b
)
3313 const struct mem_cgroup_threshold
*_a
= a
;
3314 const struct mem_cgroup_threshold
*_b
= b
;
3316 if (_a
->threshold
> _b
->threshold
)
3319 if (_a
->threshold
< _b
->threshold
)
3325 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3327 struct mem_cgroup_eventfd_list
*ev
;
3329 spin_lock(&memcg_oom_lock
);
3331 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3332 eventfd_signal(ev
->eventfd
, 1);
3334 spin_unlock(&memcg_oom_lock
);
3338 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3340 struct mem_cgroup
*iter
;
3342 for_each_mem_cgroup_tree(iter
, memcg
)
3343 mem_cgroup_oom_notify_cb(iter
);
3346 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3347 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3349 struct mem_cgroup_thresholds
*thresholds
;
3350 struct mem_cgroup_threshold_ary
*new;
3351 unsigned long threshold
;
3352 unsigned long usage
;
3355 ret
= page_counter_memparse(args
, "-1", &threshold
);
3359 mutex_lock(&memcg
->thresholds_lock
);
3362 thresholds
= &memcg
->thresholds
;
3363 usage
= mem_cgroup_usage(memcg
, false);
3364 } else if (type
== _MEMSWAP
) {
3365 thresholds
= &memcg
->memsw_thresholds
;
3366 usage
= mem_cgroup_usage(memcg
, true);
3370 /* Check if a threshold crossed before adding a new one */
3371 if (thresholds
->primary
)
3372 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3374 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3376 /* Allocate memory for new array of thresholds */
3377 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3385 /* Copy thresholds (if any) to new array */
3386 if (thresholds
->primary
) {
3387 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3388 sizeof(struct mem_cgroup_threshold
));
3391 /* Add new threshold */
3392 new->entries
[size
- 1].eventfd
= eventfd
;
3393 new->entries
[size
- 1].threshold
= threshold
;
3395 /* Sort thresholds. Registering of new threshold isn't time-critical */
3396 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3397 compare_thresholds
, NULL
);
3399 /* Find current threshold */
3400 new->current_threshold
= -1;
3401 for (i
= 0; i
< size
; i
++) {
3402 if (new->entries
[i
].threshold
<= usage
) {
3404 * new->current_threshold will not be used until
3405 * rcu_assign_pointer(), so it's safe to increment
3408 ++new->current_threshold
;
3413 /* Free old spare buffer and save old primary buffer as spare */
3414 kfree(thresholds
->spare
);
3415 thresholds
->spare
= thresholds
->primary
;
3417 rcu_assign_pointer(thresholds
->primary
, new);
3419 /* To be sure that nobody uses thresholds */
3423 mutex_unlock(&memcg
->thresholds_lock
);
3428 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3429 struct eventfd_ctx
*eventfd
, const char *args
)
3431 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3434 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3435 struct eventfd_ctx
*eventfd
, const char *args
)
3437 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3440 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3441 struct eventfd_ctx
*eventfd
, enum res_type type
)
3443 struct mem_cgroup_thresholds
*thresholds
;
3444 struct mem_cgroup_threshold_ary
*new;
3445 unsigned long usage
;
3448 mutex_lock(&memcg
->thresholds_lock
);
3451 thresholds
= &memcg
->thresholds
;
3452 usage
= mem_cgroup_usage(memcg
, false);
3453 } else if (type
== _MEMSWAP
) {
3454 thresholds
= &memcg
->memsw_thresholds
;
3455 usage
= mem_cgroup_usage(memcg
, true);
3459 if (!thresholds
->primary
)
3462 /* Check if a threshold crossed before removing */
3463 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3465 /* Calculate new number of threshold */
3467 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3468 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3472 new = thresholds
->spare
;
3474 /* Set thresholds array to NULL if we don't have thresholds */
3483 /* Copy thresholds and find current threshold */
3484 new->current_threshold
= -1;
3485 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3486 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3489 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3490 if (new->entries
[j
].threshold
<= usage
) {
3492 * new->current_threshold will not be used
3493 * until rcu_assign_pointer(), so it's safe to increment
3496 ++new->current_threshold
;
3502 /* Swap primary and spare array */
3503 thresholds
->spare
= thresholds
->primary
;
3505 rcu_assign_pointer(thresholds
->primary
, new);
3507 /* To be sure that nobody uses thresholds */
3510 /* If all events are unregistered, free the spare array */
3512 kfree(thresholds
->spare
);
3513 thresholds
->spare
= NULL
;
3516 mutex_unlock(&memcg
->thresholds_lock
);
3519 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3520 struct eventfd_ctx
*eventfd
)
3522 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3525 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3526 struct eventfd_ctx
*eventfd
)
3528 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3531 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3532 struct eventfd_ctx
*eventfd
, const char *args
)
3534 struct mem_cgroup_eventfd_list
*event
;
3536 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3540 spin_lock(&memcg_oom_lock
);
3542 event
->eventfd
= eventfd
;
3543 list_add(&event
->list
, &memcg
->oom_notify
);
3545 /* already in OOM ? */
3546 if (memcg
->under_oom
)
3547 eventfd_signal(eventfd
, 1);
3548 spin_unlock(&memcg_oom_lock
);
3553 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3554 struct eventfd_ctx
*eventfd
)
3556 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3558 spin_lock(&memcg_oom_lock
);
3560 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3561 if (ev
->eventfd
== eventfd
) {
3562 list_del(&ev
->list
);
3567 spin_unlock(&memcg_oom_lock
);
3570 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3572 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3574 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3575 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3579 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3580 struct cftype
*cft
, u64 val
)
3582 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3584 /* cannot set to root cgroup and only 0 and 1 are allowed */
3585 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3588 memcg
->oom_kill_disable
= val
;
3590 memcg_oom_recover(memcg
);
3595 #ifdef CONFIG_CGROUP_WRITEBACK
3597 struct list_head
*mem_cgroup_cgwb_list(struct mem_cgroup
*memcg
)
3599 return &memcg
->cgwb_list
;
3602 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3604 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3607 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3609 wb_domain_exit(&memcg
->cgwb_domain
);
3612 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3614 wb_domain_size_changed(&memcg
->cgwb_domain
);
3617 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3619 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3621 if (!memcg
->css
.parent
)
3624 return &memcg
->cgwb_domain
;
3628 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3629 * @wb: bdi_writeback in question
3630 * @pfilepages: out parameter for number of file pages
3631 * @pheadroom: out parameter for number of allocatable pages according to memcg
3632 * @pdirty: out parameter for number of dirty pages
3633 * @pwriteback: out parameter for number of pages under writeback
3635 * Determine the numbers of file, headroom, dirty, and writeback pages in
3636 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3637 * is a bit more involved.
3639 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3640 * headroom is calculated as the lowest headroom of itself and the
3641 * ancestors. Note that this doesn't consider the actual amount of
3642 * available memory in the system. The caller should further cap
3643 * *@pheadroom accordingly.
3645 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
3646 unsigned long *pheadroom
, unsigned long *pdirty
,
3647 unsigned long *pwriteback
)
3649 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3650 struct mem_cgroup
*parent
;
3652 *pdirty
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
3654 /* this should eventually include NR_UNSTABLE_NFS */
3655 *pwriteback
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
3656 *pfilepages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
3657 (1 << LRU_ACTIVE_FILE
));
3658 *pheadroom
= PAGE_COUNTER_MAX
;
3660 while ((parent
= parent_mem_cgroup(memcg
))) {
3661 unsigned long ceiling
= min(memcg
->memory
.limit
, memcg
->high
);
3662 unsigned long used
= page_counter_read(&memcg
->memory
);
3664 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
3669 #else /* CONFIG_CGROUP_WRITEBACK */
3671 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3676 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3680 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3684 #endif /* CONFIG_CGROUP_WRITEBACK */
3687 * DO NOT USE IN NEW FILES.
3689 * "cgroup.event_control" implementation.
3691 * This is way over-engineered. It tries to support fully configurable
3692 * events for each user. Such level of flexibility is completely
3693 * unnecessary especially in the light of the planned unified hierarchy.
3695 * Please deprecate this and replace with something simpler if at all
3700 * Unregister event and free resources.
3702 * Gets called from workqueue.
3704 static void memcg_event_remove(struct work_struct
*work
)
3706 struct mem_cgroup_event
*event
=
3707 container_of(work
, struct mem_cgroup_event
, remove
);
3708 struct mem_cgroup
*memcg
= event
->memcg
;
3710 remove_wait_queue(event
->wqh
, &event
->wait
);
3712 event
->unregister_event(memcg
, event
->eventfd
);
3714 /* Notify userspace the event is going away. */
3715 eventfd_signal(event
->eventfd
, 1);
3717 eventfd_ctx_put(event
->eventfd
);
3719 css_put(&memcg
->css
);
3723 * Gets called on POLLHUP on eventfd when user closes it.
3725 * Called with wqh->lock held and interrupts disabled.
3727 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
3728 int sync
, void *key
)
3730 struct mem_cgroup_event
*event
=
3731 container_of(wait
, struct mem_cgroup_event
, wait
);
3732 struct mem_cgroup
*memcg
= event
->memcg
;
3733 unsigned long flags
= (unsigned long)key
;
3735 if (flags
& POLLHUP
) {
3737 * If the event has been detached at cgroup removal, we
3738 * can simply return knowing the other side will cleanup
3741 * We can't race against event freeing since the other
3742 * side will require wqh->lock via remove_wait_queue(),
3745 spin_lock(&memcg
->event_list_lock
);
3746 if (!list_empty(&event
->list
)) {
3747 list_del_init(&event
->list
);
3749 * We are in atomic context, but cgroup_event_remove()
3750 * may sleep, so we have to call it in workqueue.
3752 schedule_work(&event
->remove
);
3754 spin_unlock(&memcg
->event_list_lock
);
3760 static void memcg_event_ptable_queue_proc(struct file
*file
,
3761 wait_queue_head_t
*wqh
, poll_table
*pt
)
3763 struct mem_cgroup_event
*event
=
3764 container_of(pt
, struct mem_cgroup_event
, pt
);
3767 add_wait_queue(wqh
, &event
->wait
);
3771 * DO NOT USE IN NEW FILES.
3773 * Parse input and register new cgroup event handler.
3775 * Input must be in format '<event_fd> <control_fd> <args>'.
3776 * Interpretation of args is defined by control file implementation.
3778 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
3779 char *buf
, size_t nbytes
, loff_t off
)
3781 struct cgroup_subsys_state
*css
= of_css(of
);
3782 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3783 struct mem_cgroup_event
*event
;
3784 struct cgroup_subsys_state
*cfile_css
;
3785 unsigned int efd
, cfd
;
3792 buf
= strstrip(buf
);
3794 efd
= simple_strtoul(buf
, &endp
, 10);
3799 cfd
= simple_strtoul(buf
, &endp
, 10);
3800 if ((*endp
!= ' ') && (*endp
!= '\0'))
3804 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
3808 event
->memcg
= memcg
;
3809 INIT_LIST_HEAD(&event
->list
);
3810 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
3811 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
3812 INIT_WORK(&event
->remove
, memcg_event_remove
);
3820 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
3821 if (IS_ERR(event
->eventfd
)) {
3822 ret
= PTR_ERR(event
->eventfd
);
3829 goto out_put_eventfd
;
3832 /* the process need read permission on control file */
3833 /* AV: shouldn't we check that it's been opened for read instead? */
3834 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
3839 * Determine the event callbacks and set them in @event. This used
3840 * to be done via struct cftype but cgroup core no longer knows
3841 * about these events. The following is crude but the whole thing
3842 * is for compatibility anyway.
3844 * DO NOT ADD NEW FILES.
3846 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
3848 if (!strcmp(name
, "memory.usage_in_bytes")) {
3849 event
->register_event
= mem_cgroup_usage_register_event
;
3850 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
3851 } else if (!strcmp(name
, "memory.oom_control")) {
3852 event
->register_event
= mem_cgroup_oom_register_event
;
3853 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
3854 } else if (!strcmp(name
, "memory.pressure_level")) {
3855 event
->register_event
= vmpressure_register_event
;
3856 event
->unregister_event
= vmpressure_unregister_event
;
3857 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
3858 event
->register_event
= memsw_cgroup_usage_register_event
;
3859 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
3866 * Verify @cfile should belong to @css. Also, remaining events are
3867 * automatically removed on cgroup destruction but the removal is
3868 * asynchronous, so take an extra ref on @css.
3870 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
3871 &memory_cgrp_subsys
);
3873 if (IS_ERR(cfile_css
))
3875 if (cfile_css
!= css
) {
3880 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
3884 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
3886 spin_lock(&memcg
->event_list_lock
);
3887 list_add(&event
->list
, &memcg
->event_list
);
3888 spin_unlock(&memcg
->event_list_lock
);
3900 eventfd_ctx_put(event
->eventfd
);
3909 static struct cftype mem_cgroup_legacy_files
[] = {
3911 .name
= "usage_in_bytes",
3912 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
3913 .read_u64
= mem_cgroup_read_u64
,
3916 .name
= "max_usage_in_bytes",
3917 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
3918 .write
= mem_cgroup_reset
,
3919 .read_u64
= mem_cgroup_read_u64
,
3922 .name
= "limit_in_bytes",
3923 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
3924 .write
= mem_cgroup_write
,
3925 .read_u64
= mem_cgroup_read_u64
,
3928 .name
= "soft_limit_in_bytes",
3929 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
3930 .write
= mem_cgroup_write
,
3931 .read_u64
= mem_cgroup_read_u64
,
3935 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
3936 .write
= mem_cgroup_reset
,
3937 .read_u64
= mem_cgroup_read_u64
,
3941 .seq_show
= memcg_stat_show
,
3944 .name
= "force_empty",
3945 .write
= mem_cgroup_force_empty_write
,
3948 .name
= "use_hierarchy",
3949 .write_u64
= mem_cgroup_hierarchy_write
,
3950 .read_u64
= mem_cgroup_hierarchy_read
,
3953 .name
= "cgroup.event_control", /* XXX: for compat */
3954 .write
= memcg_write_event_control
,
3955 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
3958 .name
= "swappiness",
3959 .read_u64
= mem_cgroup_swappiness_read
,
3960 .write_u64
= mem_cgroup_swappiness_write
,
3963 .name
= "move_charge_at_immigrate",
3964 .read_u64
= mem_cgroup_move_charge_read
,
3965 .write_u64
= mem_cgroup_move_charge_write
,
3968 .name
= "oom_control",
3969 .seq_show
= mem_cgroup_oom_control_read
,
3970 .write_u64
= mem_cgroup_oom_control_write
,
3971 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
3974 .name
= "pressure_level",
3978 .name
= "numa_stat",
3979 .seq_show
= memcg_numa_stat_show
,
3983 .name
= "kmem.limit_in_bytes",
3984 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
3985 .write
= mem_cgroup_write
,
3986 .read_u64
= mem_cgroup_read_u64
,
3989 .name
= "kmem.usage_in_bytes",
3990 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
3991 .read_u64
= mem_cgroup_read_u64
,
3994 .name
= "kmem.failcnt",
3995 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
3996 .write
= mem_cgroup_reset
,
3997 .read_u64
= mem_cgroup_read_u64
,
4000 .name
= "kmem.max_usage_in_bytes",
4001 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4002 .write
= mem_cgroup_reset
,
4003 .read_u64
= mem_cgroup_read_u64
,
4005 #ifdef CONFIG_SLABINFO
4007 .name
= "kmem.slabinfo",
4008 .seq_start
= slab_start
,
4009 .seq_next
= slab_next
,
4010 .seq_stop
= slab_stop
,
4011 .seq_show
= memcg_slab_show
,
4015 .name
= "kmem.tcp.limit_in_bytes",
4016 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4017 .write
= mem_cgroup_write
,
4018 .read_u64
= mem_cgroup_read_u64
,
4021 .name
= "kmem.tcp.usage_in_bytes",
4022 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4023 .read_u64
= mem_cgroup_read_u64
,
4026 .name
= "kmem.tcp.failcnt",
4027 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4028 .write
= mem_cgroup_reset
,
4029 .read_u64
= mem_cgroup_read_u64
,
4032 .name
= "kmem.tcp.max_usage_in_bytes",
4033 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4034 .write
= mem_cgroup_reset
,
4035 .read_u64
= mem_cgroup_read_u64
,
4037 { }, /* terminate */
4041 * Private memory cgroup IDR
4043 * Swap-out records and page cache shadow entries need to store memcg
4044 * references in constrained space, so we maintain an ID space that is
4045 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4046 * memory-controlled cgroups to 64k.
4048 * However, there usually are many references to the oflline CSS after
4049 * the cgroup has been destroyed, such as page cache or reclaimable
4050 * slab objects, that don't need to hang on to the ID. We want to keep
4051 * those dead CSS from occupying IDs, or we might quickly exhaust the
4052 * relatively small ID space and prevent the creation of new cgroups
4053 * even when there are much fewer than 64k cgroups - possibly none.
4055 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4056 * be freed and recycled when it's no longer needed, which is usually
4057 * when the CSS is offlined.
4059 * The only exception to that are records of swapped out tmpfs/shmem
4060 * pages that need to be attributed to live ancestors on swapin. But
4061 * those references are manageable from userspace.
4064 static DEFINE_IDR(mem_cgroup_idr
);
4066 static void mem_cgroup_id_get_many(struct mem_cgroup
*memcg
, unsigned int n
)
4068 VM_BUG_ON(atomic_read(&memcg
->id
.ref
) <= 0);
4069 atomic_add(n
, &memcg
->id
.ref
);
4072 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
4074 VM_BUG_ON(atomic_read(&memcg
->id
.ref
) < n
);
4075 if (atomic_sub_and_test(n
, &memcg
->id
.ref
)) {
4076 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4079 /* Memcg ID pins CSS */
4080 css_put(&memcg
->css
);
4084 static inline void mem_cgroup_id_get(struct mem_cgroup
*memcg
)
4086 mem_cgroup_id_get_many(memcg
, 1);
4089 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
4091 mem_cgroup_id_put_many(memcg
, 1);
4095 * mem_cgroup_from_id - look up a memcg from a memcg id
4096 * @id: the memcg id to look up
4098 * Caller must hold rcu_read_lock().
4100 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
4102 WARN_ON_ONCE(!rcu_read_lock_held());
4103 return idr_find(&mem_cgroup_idr
, id
);
4106 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4108 struct mem_cgroup_per_node
*pn
;
4111 * This routine is called against possible nodes.
4112 * But it's BUG to call kmalloc() against offline node.
4114 * TODO: this routine can waste much memory for nodes which will
4115 * never be onlined. It's better to use memory hotplug callback
4118 if (!node_state(node
, N_NORMAL_MEMORY
))
4120 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4124 lruvec_init(&pn
->lruvec
);
4125 pn
->usage_in_excess
= 0;
4126 pn
->on_tree
= false;
4129 memcg
->nodeinfo
[node
] = pn
;
4133 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4135 kfree(memcg
->nodeinfo
[node
]);
4138 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
4143 free_mem_cgroup_per_node_info(memcg
, node
);
4144 free_percpu(memcg
->stat
);
4148 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4150 memcg_wb_domain_exit(memcg
);
4151 __mem_cgroup_free(memcg
);
4154 static struct mem_cgroup
*mem_cgroup_alloc(void)
4156 struct mem_cgroup
*memcg
;
4160 size
= sizeof(struct mem_cgroup
);
4161 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4163 memcg
= kzalloc(size
, GFP_KERNEL
);
4167 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
4168 1, MEM_CGROUP_ID_MAX
,
4170 if (memcg
->id
.id
< 0)
4173 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4178 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
4181 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4184 INIT_WORK(&memcg
->high_work
, high_work_func
);
4185 memcg
->last_scanned_node
= MAX_NUMNODES
;
4186 INIT_LIST_HEAD(&memcg
->oom_notify
);
4187 mutex_init(&memcg
->thresholds_lock
);
4188 spin_lock_init(&memcg
->move_lock
);
4189 vmpressure_init(&memcg
->vmpressure
);
4190 INIT_LIST_HEAD(&memcg
->event_list
);
4191 spin_lock_init(&memcg
->event_list_lock
);
4192 memcg
->socket_pressure
= jiffies
;
4194 memcg
->kmemcg_id
= -1;
4196 #ifdef CONFIG_CGROUP_WRITEBACK
4197 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4199 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
4202 if (memcg
->id
.id
> 0)
4203 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4204 __mem_cgroup_free(memcg
);
4208 static struct cgroup_subsys_state
* __ref
4209 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4211 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
4212 struct mem_cgroup
*memcg
;
4213 long error
= -ENOMEM
;
4215 memcg
= mem_cgroup_alloc();
4217 return ERR_PTR(error
);
4219 memcg
->high
= PAGE_COUNTER_MAX
;
4220 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4222 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4223 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4225 if (parent
&& parent
->use_hierarchy
) {
4226 memcg
->use_hierarchy
= true;
4227 page_counter_init(&memcg
->memory
, &parent
->memory
);
4228 page_counter_init(&memcg
->swap
, &parent
->swap
);
4229 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4230 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4231 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
4233 page_counter_init(&memcg
->memory
, NULL
);
4234 page_counter_init(&memcg
->swap
, NULL
);
4235 page_counter_init(&memcg
->memsw
, NULL
);
4236 page_counter_init(&memcg
->kmem
, NULL
);
4237 page_counter_init(&memcg
->tcpmem
, NULL
);
4239 * Deeper hierachy with use_hierarchy == false doesn't make
4240 * much sense so let cgroup subsystem know about this
4241 * unfortunate state in our controller.
4243 if (parent
!= root_mem_cgroup
)
4244 memory_cgrp_subsys
.broken_hierarchy
= true;
4247 /* The following stuff does not apply to the root */
4249 root_mem_cgroup
= memcg
;
4253 error
= memcg_online_kmem(memcg
);
4257 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4258 static_branch_inc(&memcg_sockets_enabled_key
);
4262 mem_cgroup_free(memcg
);
4263 return ERR_PTR(-ENOMEM
);
4266 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4268 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4270 /* Online state pins memcg ID, memcg ID pins CSS */
4271 atomic_set(&memcg
->id
.ref
, 1);
4276 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4278 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4279 struct mem_cgroup_event
*event
, *tmp
;
4282 * Unregister events and notify userspace.
4283 * Notify userspace about cgroup removing only after rmdir of cgroup
4284 * directory to avoid race between userspace and kernelspace.
4286 spin_lock(&memcg
->event_list_lock
);
4287 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4288 list_del_init(&event
->list
);
4289 schedule_work(&event
->remove
);
4291 spin_unlock(&memcg
->event_list_lock
);
4293 memcg_offline_kmem(memcg
);
4294 wb_memcg_offline(memcg
);
4296 mem_cgroup_id_put(memcg
);
4299 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4301 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4303 invalidate_reclaim_iterators(memcg
);
4306 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4308 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4310 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4311 static_branch_dec(&memcg_sockets_enabled_key
);
4313 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
4314 static_branch_dec(&memcg_sockets_enabled_key
);
4316 vmpressure_cleanup(&memcg
->vmpressure
);
4317 cancel_work_sync(&memcg
->high_work
);
4318 mem_cgroup_remove_from_trees(memcg
);
4319 memcg_free_kmem(memcg
);
4320 mem_cgroup_free(memcg
);
4324 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4325 * @css: the target css
4327 * Reset the states of the mem_cgroup associated with @css. This is
4328 * invoked when the userland requests disabling on the default hierarchy
4329 * but the memcg is pinned through dependency. The memcg should stop
4330 * applying policies and should revert to the vanilla state as it may be
4331 * made visible again.
4333 * The current implementation only resets the essential configurations.
4334 * This needs to be expanded to cover all the visible parts.
4336 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4338 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4340 page_counter_limit(&memcg
->memory
, PAGE_COUNTER_MAX
);
4341 page_counter_limit(&memcg
->swap
, PAGE_COUNTER_MAX
);
4342 page_counter_limit(&memcg
->memsw
, PAGE_COUNTER_MAX
);
4343 page_counter_limit(&memcg
->kmem
, PAGE_COUNTER_MAX
);
4344 page_counter_limit(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
4346 memcg
->high
= PAGE_COUNTER_MAX
;
4347 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4348 memcg_wb_domain_size_changed(memcg
);
4352 /* Handlers for move charge at task migration. */
4353 static int mem_cgroup_do_precharge(unsigned long count
)
4357 /* Try a single bulk charge without reclaim first, kswapd may wake */
4358 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4360 mc
.precharge
+= count
;
4364 /* Try charges one by one with reclaim, but do not retry */
4366 ret
= try_charge(mc
.to
, GFP_KERNEL
| __GFP_NORETRY
, 1);
4380 enum mc_target_type
{
4386 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4387 unsigned long addr
, pte_t ptent
)
4389 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4391 if (!page
|| !page_mapped(page
))
4393 if (PageAnon(page
)) {
4394 if (!(mc
.flags
& MOVE_ANON
))
4397 if (!(mc
.flags
& MOVE_FILE
))
4400 if (!get_page_unless_zero(page
))
4407 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4408 pte_t ptent
, swp_entry_t
*entry
)
4410 struct page
*page
= NULL
;
4411 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4413 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4416 * Because lookup_swap_cache() updates some statistics counter,
4417 * we call find_get_page() with swapper_space directly.
4419 page
= find_get_page(swap_address_space(ent
), swp_offset(ent
));
4420 if (do_memsw_account())
4421 entry
->val
= ent
.val
;
4426 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4427 pte_t ptent
, swp_entry_t
*entry
)
4433 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4434 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4436 struct page
*page
= NULL
;
4437 struct address_space
*mapping
;
4440 if (!vma
->vm_file
) /* anonymous vma */
4442 if (!(mc
.flags
& MOVE_FILE
))
4445 mapping
= vma
->vm_file
->f_mapping
;
4446 pgoff
= linear_page_index(vma
, addr
);
4448 /* page is moved even if it's not RSS of this task(page-faulted). */
4450 /* shmem/tmpfs may report page out on swap: account for that too. */
4451 if (shmem_mapping(mapping
)) {
4452 page
= find_get_entry(mapping
, pgoff
);
4453 if (radix_tree_exceptional_entry(page
)) {
4454 swp_entry_t swp
= radix_to_swp_entry(page
);
4455 if (do_memsw_account())
4457 page
= find_get_page(swap_address_space(swp
),
4461 page
= find_get_page(mapping
, pgoff
);
4463 page
= find_get_page(mapping
, pgoff
);
4469 * mem_cgroup_move_account - move account of the page
4471 * @compound: charge the page as compound or small page
4472 * @from: mem_cgroup which the page is moved from.
4473 * @to: mem_cgroup which the page is moved to. @from != @to.
4475 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4477 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4480 static int mem_cgroup_move_account(struct page
*page
,
4482 struct mem_cgroup
*from
,
4483 struct mem_cgroup
*to
)
4485 unsigned long flags
;
4486 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
4490 VM_BUG_ON(from
== to
);
4491 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4492 VM_BUG_ON(compound
&& !PageTransHuge(page
));
4495 * Prevent mem_cgroup_migrate() from looking at
4496 * page->mem_cgroup of its source page while we change it.
4499 if (!trylock_page(page
))
4503 if (page
->mem_cgroup
!= from
)
4506 anon
= PageAnon(page
);
4508 spin_lock_irqsave(&from
->move_lock
, flags
);
4510 if (!anon
&& page_mapped(page
)) {
4511 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4513 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4518 * move_lock grabbed above and caller set from->moving_account, so
4519 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4520 * So mapping should be stable for dirty pages.
4522 if (!anon
&& PageDirty(page
)) {
4523 struct address_space
*mapping
= page_mapping(page
);
4525 if (mapping_cap_account_dirty(mapping
)) {
4526 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4528 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4533 if (PageWriteback(page
)) {
4534 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4536 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4541 * It is safe to change page->mem_cgroup here because the page
4542 * is referenced, charged, and isolated - we can't race with
4543 * uncharging, charging, migration, or LRU putback.
4546 /* caller should have done css_get */
4547 page
->mem_cgroup
= to
;
4548 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4552 local_irq_disable();
4553 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
4554 memcg_check_events(to
, page
);
4555 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
4556 memcg_check_events(from
, page
);
4565 * get_mctgt_type - get target type of moving charge
4566 * @vma: the vma the pte to be checked belongs
4567 * @addr: the address corresponding to the pte to be checked
4568 * @ptent: the pte to be checked
4569 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4572 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4573 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4574 * move charge. if @target is not NULL, the page is stored in target->page
4575 * with extra refcnt got(Callers should handle it).
4576 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4577 * target for charge migration. if @target is not NULL, the entry is stored
4580 * Called with pte lock held.
4583 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4584 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4586 struct page
*page
= NULL
;
4587 enum mc_target_type ret
= MC_TARGET_NONE
;
4588 swp_entry_t ent
= { .val
= 0 };
4590 if (pte_present(ptent
))
4591 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4592 else if (is_swap_pte(ptent
))
4593 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
4594 else if (pte_none(ptent
))
4595 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4597 if (!page
&& !ent
.val
)
4601 * Do only loose check w/o serialization.
4602 * mem_cgroup_move_account() checks the page is valid or
4603 * not under LRU exclusion.
4605 if (page
->mem_cgroup
== mc
.from
) {
4606 ret
= MC_TARGET_PAGE
;
4608 target
->page
= page
;
4610 if (!ret
|| !target
)
4613 /* There is a swap entry and a page doesn't exist or isn't charged */
4614 if (ent
.val
&& !ret
&&
4615 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4616 ret
= MC_TARGET_SWAP
;
4623 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4625 * We don't consider swapping or file mapped pages because THP does not
4626 * support them for now.
4627 * Caller should make sure that pmd_trans_huge(pmd) is true.
4629 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4630 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4632 struct page
*page
= NULL
;
4633 enum mc_target_type ret
= MC_TARGET_NONE
;
4635 page
= pmd_page(pmd
);
4636 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4637 if (!(mc
.flags
& MOVE_ANON
))
4639 if (page
->mem_cgroup
== mc
.from
) {
4640 ret
= MC_TARGET_PAGE
;
4643 target
->page
= page
;
4649 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4650 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4652 return MC_TARGET_NONE
;
4656 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4657 unsigned long addr
, unsigned long end
,
4658 struct mm_walk
*walk
)
4660 struct vm_area_struct
*vma
= walk
->vma
;
4664 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4666 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4667 mc
.precharge
+= HPAGE_PMD_NR
;
4672 if (pmd_trans_unstable(pmd
))
4674 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4675 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4676 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4677 mc
.precharge
++; /* increment precharge temporarily */
4678 pte_unmap_unlock(pte
- 1, ptl
);
4684 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4686 unsigned long precharge
;
4688 struct mm_walk mem_cgroup_count_precharge_walk
= {
4689 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4692 down_read(&mm
->mmap_sem
);
4693 walk_page_range(0, mm
->highest_vm_end
,
4694 &mem_cgroup_count_precharge_walk
);
4695 up_read(&mm
->mmap_sem
);
4697 precharge
= mc
.precharge
;
4703 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4705 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4707 VM_BUG_ON(mc
.moving_task
);
4708 mc
.moving_task
= current
;
4709 return mem_cgroup_do_precharge(precharge
);
4712 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4713 static void __mem_cgroup_clear_mc(void)
4715 struct mem_cgroup
*from
= mc
.from
;
4716 struct mem_cgroup
*to
= mc
.to
;
4718 /* we must uncharge all the leftover precharges from mc.to */
4720 cancel_charge(mc
.to
, mc
.precharge
);
4724 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4725 * we must uncharge here.
4727 if (mc
.moved_charge
) {
4728 cancel_charge(mc
.from
, mc
.moved_charge
);
4729 mc
.moved_charge
= 0;
4731 /* we must fixup refcnts and charges */
4732 if (mc
.moved_swap
) {
4733 /* uncharge swap account from the old cgroup */
4734 if (!mem_cgroup_is_root(mc
.from
))
4735 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4737 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
4740 * we charged both to->memory and to->memsw, so we
4741 * should uncharge to->memory.
4743 if (!mem_cgroup_is_root(mc
.to
))
4744 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4746 mem_cgroup_id_get_many(mc
.to
, mc
.moved_swap
);
4747 css_put_many(&mc
.to
->css
, mc
.moved_swap
);
4751 memcg_oom_recover(from
);
4752 memcg_oom_recover(to
);
4753 wake_up_all(&mc
.waitq
);
4756 static void mem_cgroup_clear_mc(void)
4758 struct mm_struct
*mm
= mc
.mm
;
4761 * we must clear moving_task before waking up waiters at the end of
4764 mc
.moving_task
= NULL
;
4765 __mem_cgroup_clear_mc();
4766 spin_lock(&mc
.lock
);
4770 spin_unlock(&mc
.lock
);
4775 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4777 struct cgroup_subsys_state
*css
;
4778 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
4779 struct mem_cgroup
*from
;
4780 struct task_struct
*leader
, *p
;
4781 struct mm_struct
*mm
;
4782 unsigned long move_flags
;
4785 /* charge immigration isn't supported on the default hierarchy */
4786 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4790 * Multi-process migrations only happen on the default hierarchy
4791 * where charge immigration is not used. Perform charge
4792 * immigration if @tset contains a leader and whine if there are
4796 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
4799 memcg
= mem_cgroup_from_css(css
);
4805 * We are now commited to this value whatever it is. Changes in this
4806 * tunable will only affect upcoming migrations, not the current one.
4807 * So we need to save it, and keep it going.
4809 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
4813 from
= mem_cgroup_from_task(p
);
4815 VM_BUG_ON(from
== memcg
);
4817 mm
= get_task_mm(p
);
4820 /* We move charges only when we move a owner of the mm */
4821 if (mm
->owner
== p
) {
4824 VM_BUG_ON(mc
.precharge
);
4825 VM_BUG_ON(mc
.moved_charge
);
4826 VM_BUG_ON(mc
.moved_swap
);
4828 spin_lock(&mc
.lock
);
4832 mc
.flags
= move_flags
;
4833 spin_unlock(&mc
.lock
);
4834 /* We set mc.moving_task later */
4836 ret
= mem_cgroup_precharge_mc(mm
);
4838 mem_cgroup_clear_mc();
4845 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4848 mem_cgroup_clear_mc();
4851 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4852 unsigned long addr
, unsigned long end
,
4853 struct mm_walk
*walk
)
4856 struct vm_area_struct
*vma
= walk
->vma
;
4859 enum mc_target_type target_type
;
4860 union mc_target target
;
4863 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4865 if (mc
.precharge
< HPAGE_PMD_NR
) {
4869 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
4870 if (target_type
== MC_TARGET_PAGE
) {
4872 if (!isolate_lru_page(page
)) {
4873 if (!mem_cgroup_move_account(page
, true,
4875 mc
.precharge
-= HPAGE_PMD_NR
;
4876 mc
.moved_charge
+= HPAGE_PMD_NR
;
4878 putback_lru_page(page
);
4886 if (pmd_trans_unstable(pmd
))
4889 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4890 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4891 pte_t ptent
= *(pte
++);
4897 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
4898 case MC_TARGET_PAGE
:
4901 * We can have a part of the split pmd here. Moving it
4902 * can be done but it would be too convoluted so simply
4903 * ignore such a partial THP and keep it in original
4904 * memcg. There should be somebody mapping the head.
4906 if (PageTransCompound(page
))
4908 if (isolate_lru_page(page
))
4910 if (!mem_cgroup_move_account(page
, false,
4913 /* we uncharge from mc.from later. */
4916 putback_lru_page(page
);
4917 put
: /* get_mctgt_type() gets the page */
4920 case MC_TARGET_SWAP
:
4922 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
4924 /* we fixup refcnts and charges later. */
4932 pte_unmap_unlock(pte
- 1, ptl
);
4937 * We have consumed all precharges we got in can_attach().
4938 * We try charge one by one, but don't do any additional
4939 * charges to mc.to if we have failed in charge once in attach()
4942 ret
= mem_cgroup_do_precharge(1);
4950 static void mem_cgroup_move_charge(void)
4952 struct mm_walk mem_cgroup_move_charge_walk
= {
4953 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4957 lru_add_drain_all();
4959 * Signal lock_page_memcg() to take the memcg's move_lock
4960 * while we're moving its pages to another memcg. Then wait
4961 * for already started RCU-only updates to finish.
4963 atomic_inc(&mc
.from
->moving_account
);
4966 if (unlikely(!down_read_trylock(&mc
.mm
->mmap_sem
))) {
4968 * Someone who are holding the mmap_sem might be waiting in
4969 * waitq. So we cancel all extra charges, wake up all waiters,
4970 * and retry. Because we cancel precharges, we might not be able
4971 * to move enough charges, but moving charge is a best-effort
4972 * feature anyway, so it wouldn't be a big problem.
4974 __mem_cgroup_clear_mc();
4979 * When we have consumed all precharges and failed in doing
4980 * additional charge, the page walk just aborts.
4982 walk_page_range(0, mc
.mm
->highest_vm_end
, &mem_cgroup_move_charge_walk
);
4984 up_read(&mc
.mm
->mmap_sem
);
4985 atomic_dec(&mc
.from
->moving_account
);
4988 static void mem_cgroup_move_task(void)
4991 mem_cgroup_move_charge();
4992 mem_cgroup_clear_mc();
4995 #else /* !CONFIG_MMU */
4996 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5000 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5003 static void mem_cgroup_move_task(void)
5009 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5010 * to verify whether we're attached to the default hierarchy on each mount
5013 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5016 * use_hierarchy is forced on the default hierarchy. cgroup core
5017 * guarantees that @root doesn't have any children, so turning it
5018 * on for the root memcg is enough.
5020 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5021 root_mem_cgroup
->use_hierarchy
= true;
5023 root_mem_cgroup
->use_hierarchy
= false;
5026 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5029 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5031 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
5034 static int memory_low_show(struct seq_file
*m
, void *v
)
5036 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5037 unsigned long low
= READ_ONCE(memcg
->low
);
5039 if (low
== PAGE_COUNTER_MAX
)
5040 seq_puts(m
, "max\n");
5042 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
5047 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5048 char *buf
, size_t nbytes
, loff_t off
)
5050 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5054 buf
= strstrip(buf
);
5055 err
= page_counter_memparse(buf
, "max", &low
);
5064 static int memory_high_show(struct seq_file
*m
, void *v
)
5066 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5067 unsigned long high
= READ_ONCE(memcg
->high
);
5069 if (high
== PAGE_COUNTER_MAX
)
5070 seq_puts(m
, "max\n");
5072 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
5077 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5078 char *buf
, size_t nbytes
, loff_t off
)
5080 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5081 unsigned long nr_pages
;
5085 buf
= strstrip(buf
);
5086 err
= page_counter_memparse(buf
, "max", &high
);
5092 nr_pages
= page_counter_read(&memcg
->memory
);
5093 if (nr_pages
> high
)
5094 try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
5097 memcg_wb_domain_size_changed(memcg
);
5101 static int memory_max_show(struct seq_file
*m
, void *v
)
5103 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5104 unsigned long max
= READ_ONCE(memcg
->memory
.limit
);
5106 if (max
== PAGE_COUNTER_MAX
)
5107 seq_puts(m
, "max\n");
5109 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5114 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5115 char *buf
, size_t nbytes
, loff_t off
)
5117 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5118 unsigned int nr_reclaims
= MEM_CGROUP_RECLAIM_RETRIES
;
5119 bool drained
= false;
5123 buf
= strstrip(buf
);
5124 err
= page_counter_memparse(buf
, "max", &max
);
5128 xchg(&memcg
->memory
.limit
, max
);
5131 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
5133 if (nr_pages
<= max
)
5136 if (signal_pending(current
)) {
5142 drain_all_stock(memcg
);
5148 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
5154 mem_cgroup_events(memcg
, MEMCG_OOM
, 1);
5155 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
5159 memcg_wb_domain_size_changed(memcg
);
5163 static int memory_events_show(struct seq_file
*m
, void *v
)
5165 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5167 seq_printf(m
, "low %lu\n", mem_cgroup_read_events(memcg
, MEMCG_LOW
));
5168 seq_printf(m
, "high %lu\n", mem_cgroup_read_events(memcg
, MEMCG_HIGH
));
5169 seq_printf(m
, "max %lu\n", mem_cgroup_read_events(memcg
, MEMCG_MAX
));
5170 seq_printf(m
, "oom %lu\n", mem_cgroup_read_events(memcg
, MEMCG_OOM
));
5175 static int memory_stat_show(struct seq_file
*m
, void *v
)
5177 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5178 unsigned long stat
[MEMCG_NR_STAT
];
5179 unsigned long events
[MEMCG_NR_EVENTS
];
5183 * Provide statistics on the state of the memory subsystem as
5184 * well as cumulative event counters that show past behavior.
5186 * This list is ordered following a combination of these gradients:
5187 * 1) generic big picture -> specifics and details
5188 * 2) reflecting userspace activity -> reflecting kernel heuristics
5190 * Current memory state:
5193 tree_stat(memcg
, stat
);
5194 tree_events(memcg
, events
);
5196 seq_printf(m
, "anon %llu\n",
5197 (u64
)stat
[MEM_CGROUP_STAT_RSS
] * PAGE_SIZE
);
5198 seq_printf(m
, "file %llu\n",
5199 (u64
)stat
[MEM_CGROUP_STAT_CACHE
] * PAGE_SIZE
);
5200 seq_printf(m
, "kernel_stack %llu\n",
5201 (u64
)stat
[MEMCG_KERNEL_STACK_KB
] * 1024);
5202 seq_printf(m
, "slab %llu\n",
5203 (u64
)(stat
[MEMCG_SLAB_RECLAIMABLE
] +
5204 stat
[MEMCG_SLAB_UNRECLAIMABLE
]) * PAGE_SIZE
);
5205 seq_printf(m
, "sock %llu\n",
5206 (u64
)stat
[MEMCG_SOCK
] * PAGE_SIZE
);
5208 seq_printf(m
, "file_mapped %llu\n",
5209 (u64
)stat
[MEM_CGROUP_STAT_FILE_MAPPED
] * PAGE_SIZE
);
5210 seq_printf(m
, "file_dirty %llu\n",
5211 (u64
)stat
[MEM_CGROUP_STAT_DIRTY
] * PAGE_SIZE
);
5212 seq_printf(m
, "file_writeback %llu\n",
5213 (u64
)stat
[MEM_CGROUP_STAT_WRITEBACK
] * PAGE_SIZE
);
5215 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5216 struct mem_cgroup
*mi
;
5217 unsigned long val
= 0;
5219 for_each_mem_cgroup_tree(mi
, memcg
)
5220 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
));
5221 seq_printf(m
, "%s %llu\n",
5222 mem_cgroup_lru_names
[i
], (u64
)val
* PAGE_SIZE
);
5225 seq_printf(m
, "slab_reclaimable %llu\n",
5226 (u64
)stat
[MEMCG_SLAB_RECLAIMABLE
] * PAGE_SIZE
);
5227 seq_printf(m
, "slab_unreclaimable %llu\n",
5228 (u64
)stat
[MEMCG_SLAB_UNRECLAIMABLE
] * PAGE_SIZE
);
5230 /* Accumulated memory events */
5232 seq_printf(m
, "pgfault %lu\n",
5233 events
[MEM_CGROUP_EVENTS_PGFAULT
]);
5234 seq_printf(m
, "pgmajfault %lu\n",
5235 events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
5240 static struct cftype memory_files
[] = {
5243 .flags
= CFTYPE_NOT_ON_ROOT
,
5244 .read_u64
= memory_current_read
,
5248 .flags
= CFTYPE_NOT_ON_ROOT
,
5249 .seq_show
= memory_low_show
,
5250 .write
= memory_low_write
,
5254 .flags
= CFTYPE_NOT_ON_ROOT
,
5255 .seq_show
= memory_high_show
,
5256 .write
= memory_high_write
,
5260 .flags
= CFTYPE_NOT_ON_ROOT
,
5261 .seq_show
= memory_max_show
,
5262 .write
= memory_max_write
,
5266 .flags
= CFTYPE_NOT_ON_ROOT
,
5267 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
5268 .seq_show
= memory_events_show
,
5272 .flags
= CFTYPE_NOT_ON_ROOT
,
5273 .seq_show
= memory_stat_show
,
5278 struct cgroup_subsys memory_cgrp_subsys
= {
5279 .css_alloc
= mem_cgroup_css_alloc
,
5280 .css_online
= mem_cgroup_css_online
,
5281 .css_offline
= mem_cgroup_css_offline
,
5282 .css_released
= mem_cgroup_css_released
,
5283 .css_free
= mem_cgroup_css_free
,
5284 .css_reset
= mem_cgroup_css_reset
,
5285 .can_attach
= mem_cgroup_can_attach
,
5286 .cancel_attach
= mem_cgroup_cancel_attach
,
5287 .post_attach
= mem_cgroup_move_task
,
5288 .bind
= mem_cgroup_bind
,
5289 .dfl_cftypes
= memory_files
,
5290 .legacy_cftypes
= mem_cgroup_legacy_files
,
5295 * mem_cgroup_low - check if memory consumption is below the normal range
5296 * @root: the highest ancestor to consider
5297 * @memcg: the memory cgroup to check
5299 * Returns %true if memory consumption of @memcg, and that of all
5300 * configurable ancestors up to @root, is below the normal range.
5302 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5304 if (mem_cgroup_disabled())
5308 * The toplevel group doesn't have a configurable range, so
5309 * it's never low when looked at directly, and it is not
5310 * considered an ancestor when assessing the hierarchy.
5313 if (memcg
== root_mem_cgroup
)
5316 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5319 while (memcg
!= root
) {
5320 memcg
= parent_mem_cgroup(memcg
);
5322 if (memcg
== root_mem_cgroup
)
5325 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5332 * mem_cgroup_try_charge - try charging a page
5333 * @page: page to charge
5334 * @mm: mm context of the victim
5335 * @gfp_mask: reclaim mode
5336 * @memcgp: charged memcg return
5337 * @compound: charge the page as compound or small page
5339 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5340 * pages according to @gfp_mask if necessary.
5342 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5343 * Otherwise, an error code is returned.
5345 * After page->mapping has been set up, the caller must finalize the
5346 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5347 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5349 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5350 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
5353 struct mem_cgroup
*memcg
= NULL
;
5354 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5357 if (mem_cgroup_disabled())
5360 if (PageSwapCache(page
)) {
5362 * Every swap fault against a single page tries to charge the
5363 * page, bail as early as possible. shmem_unuse() encounters
5364 * already charged pages, too. The USED bit is protected by
5365 * the page lock, which serializes swap cache removal, which
5366 * in turn serializes uncharging.
5368 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5369 if (page
->mem_cgroup
)
5372 if (do_swap_account
) {
5373 swp_entry_t ent
= { .val
= page_private(page
), };
5374 unsigned short id
= lookup_swap_cgroup_id(ent
);
5377 memcg
= mem_cgroup_from_id(id
);
5378 if (memcg
&& !css_tryget_online(&memcg
->css
))
5385 memcg
= get_mem_cgroup_from_mm(mm
);
5387 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5389 css_put(&memcg
->css
);
5396 * mem_cgroup_commit_charge - commit a page charge
5397 * @page: page to charge
5398 * @memcg: memcg to charge the page to
5399 * @lrucare: page might be on LRU already
5400 * @compound: charge the page as compound or small page
5402 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5403 * after page->mapping has been set up. This must happen atomically
5404 * as part of the page instantiation, i.e. under the page table lock
5405 * for anonymous pages, under the page lock for page and swap cache.
5407 * In addition, the page must not be on the LRU during the commit, to
5408 * prevent racing with task migration. If it might be, use @lrucare.
5410 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5412 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5413 bool lrucare
, bool compound
)
5415 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5417 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5418 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5420 if (mem_cgroup_disabled())
5423 * Swap faults will attempt to charge the same page multiple
5424 * times. But reuse_swap_page() might have removed the page
5425 * from swapcache already, so we can't check PageSwapCache().
5430 commit_charge(page
, memcg
, lrucare
);
5432 local_irq_disable();
5433 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
5434 memcg_check_events(memcg
, page
);
5437 if (do_memsw_account() && PageSwapCache(page
)) {
5438 swp_entry_t entry
= { .val
= page_private(page
) };
5440 * The swap entry might not get freed for a long time,
5441 * let's not wait for it. The page already received a
5442 * memory+swap charge, drop the swap entry duplicate.
5444 mem_cgroup_uncharge_swap(entry
);
5449 * mem_cgroup_cancel_charge - cancel a page charge
5450 * @page: page to charge
5451 * @memcg: memcg to charge the page to
5452 * @compound: charge the page as compound or small page
5454 * Cancel a charge transaction started by mem_cgroup_try_charge().
5456 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5459 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5461 if (mem_cgroup_disabled())
5464 * Swap faults will attempt to charge the same page multiple
5465 * times. But reuse_swap_page() might have removed the page
5466 * from swapcache already, so we can't check PageSwapCache().
5471 cancel_charge(memcg
, nr_pages
);
5474 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5475 unsigned long nr_anon
, unsigned long nr_file
,
5476 unsigned long nr_huge
, unsigned long nr_kmem
,
5477 struct page
*dummy_page
)
5479 unsigned long nr_pages
= nr_anon
+ nr_file
+ nr_kmem
;
5480 unsigned long flags
;
5482 if (!mem_cgroup_is_root(memcg
)) {
5483 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5484 if (do_memsw_account())
5485 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5486 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && nr_kmem
)
5487 page_counter_uncharge(&memcg
->kmem
, nr_kmem
);
5488 memcg_oom_recover(memcg
);
5491 local_irq_save(flags
);
5492 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5493 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5494 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5495 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5496 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5497 memcg_check_events(memcg
, dummy_page
);
5498 local_irq_restore(flags
);
5500 if (!mem_cgroup_is_root(memcg
))
5501 css_put_many(&memcg
->css
, nr_pages
);
5504 static void uncharge_list(struct list_head
*page_list
)
5506 struct mem_cgroup
*memcg
= NULL
;
5507 unsigned long nr_anon
= 0;
5508 unsigned long nr_file
= 0;
5509 unsigned long nr_huge
= 0;
5510 unsigned long nr_kmem
= 0;
5511 unsigned long pgpgout
= 0;
5512 struct list_head
*next
;
5516 * Note that the list can be a single page->lru; hence the
5517 * do-while loop instead of a simple list_for_each_entry().
5519 next
= page_list
->next
;
5521 page
= list_entry(next
, struct page
, lru
);
5522 next
= page
->lru
.next
;
5524 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5525 VM_BUG_ON_PAGE(page_count(page
), page
);
5527 if (!page
->mem_cgroup
)
5531 * Nobody should be changing or seriously looking at
5532 * page->mem_cgroup at this point, we have fully
5533 * exclusive access to the page.
5536 if (memcg
!= page
->mem_cgroup
) {
5538 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5539 nr_huge
, nr_kmem
, page
);
5540 pgpgout
= nr_anon
= nr_file
=
5541 nr_huge
= nr_kmem
= 0;
5543 memcg
= page
->mem_cgroup
;
5546 if (!PageKmemcg(page
)) {
5547 unsigned int nr_pages
= 1;
5549 if (PageTransHuge(page
)) {
5550 nr_pages
<<= compound_order(page
);
5551 nr_huge
+= nr_pages
;
5554 nr_anon
+= nr_pages
;
5556 nr_file
+= nr_pages
;
5559 nr_kmem
+= 1 << compound_order(page
);
5560 __ClearPageKmemcg(page
);
5563 page
->mem_cgroup
= NULL
;
5564 } while (next
!= page_list
);
5567 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5568 nr_huge
, nr_kmem
, page
);
5572 * mem_cgroup_uncharge - uncharge a page
5573 * @page: page to uncharge
5575 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5576 * mem_cgroup_commit_charge().
5578 void mem_cgroup_uncharge(struct page
*page
)
5580 if (mem_cgroup_disabled())
5583 /* Don't touch page->lru of any random page, pre-check: */
5584 if (!page
->mem_cgroup
)
5587 INIT_LIST_HEAD(&page
->lru
);
5588 uncharge_list(&page
->lru
);
5592 * mem_cgroup_uncharge_list - uncharge a list of page
5593 * @page_list: list of pages to uncharge
5595 * Uncharge a list of pages previously charged with
5596 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5598 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5600 if (mem_cgroup_disabled())
5603 if (!list_empty(page_list
))
5604 uncharge_list(page_list
);
5608 * mem_cgroup_migrate - charge a page's replacement
5609 * @oldpage: currently circulating page
5610 * @newpage: replacement page
5612 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5613 * be uncharged upon free.
5615 * Both pages must be locked, @newpage->mapping must be set up.
5617 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
5619 struct mem_cgroup
*memcg
;
5620 unsigned int nr_pages
;
5622 unsigned long flags
;
5624 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5625 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5626 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5627 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5630 if (mem_cgroup_disabled())
5633 /* Page cache replacement: new page already charged? */
5634 if (newpage
->mem_cgroup
)
5637 /* Swapcache readahead pages can get replaced before being charged */
5638 memcg
= oldpage
->mem_cgroup
;
5642 /* Force-charge the new page. The old one will be freed soon */
5643 compound
= PageTransHuge(newpage
);
5644 nr_pages
= compound
? hpage_nr_pages(newpage
) : 1;
5646 page_counter_charge(&memcg
->memory
, nr_pages
);
5647 if (do_memsw_account())
5648 page_counter_charge(&memcg
->memsw
, nr_pages
);
5649 css_get_many(&memcg
->css
, nr_pages
);
5651 commit_charge(newpage
, memcg
, false);
5653 local_irq_save(flags
);
5654 mem_cgroup_charge_statistics(memcg
, newpage
, compound
, nr_pages
);
5655 memcg_check_events(memcg
, newpage
);
5656 local_irq_restore(flags
);
5659 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
5660 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
5662 void mem_cgroup_sk_alloc(struct sock
*sk
)
5664 struct mem_cgroup
*memcg
;
5666 if (!mem_cgroup_sockets_enabled
)
5670 * Socket cloning can throw us here with sk_memcg already
5671 * filled. It won't however, necessarily happen from
5672 * process context. So the test for root memcg given
5673 * the current task's memcg won't help us in this case.
5675 * Respecting the original socket's memcg is a better
5676 * decision in this case.
5679 BUG_ON(mem_cgroup_is_root(sk
->sk_memcg
));
5680 css_get(&sk
->sk_memcg
->css
);
5685 memcg
= mem_cgroup_from_task(current
);
5686 if (memcg
== root_mem_cgroup
)
5688 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
5690 if (css_tryget_online(&memcg
->css
))
5691 sk
->sk_memcg
= memcg
;
5696 void mem_cgroup_sk_free(struct sock
*sk
)
5699 css_put(&sk
->sk_memcg
->css
);
5703 * mem_cgroup_charge_skmem - charge socket memory
5704 * @memcg: memcg to charge
5705 * @nr_pages: number of pages to charge
5707 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5708 * @memcg's configured limit, %false if the charge had to be forced.
5710 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5712 gfp_t gfp_mask
= GFP_KERNEL
;
5714 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5715 struct page_counter
*fail
;
5717 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
5718 memcg
->tcpmem_pressure
= 0;
5721 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
5722 memcg
->tcpmem_pressure
= 1;
5726 /* Don't block in the packet receive path */
5728 gfp_mask
= GFP_NOWAIT
;
5730 this_cpu_add(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5732 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
5735 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
5740 * mem_cgroup_uncharge_skmem - uncharge socket memory
5741 * @memcg - memcg to uncharge
5742 * @nr_pages - number of pages to uncharge
5744 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5746 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5747 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
5751 this_cpu_sub(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5753 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5754 css_put_many(&memcg
->css
, nr_pages
);
5757 static int __init
cgroup_memory(char *s
)
5761 while ((token
= strsep(&s
, ",")) != NULL
) {
5764 if (!strcmp(token
, "nosocket"))
5765 cgroup_memory_nosocket
= true;
5766 if (!strcmp(token
, "nokmem"))
5767 cgroup_memory_nokmem
= true;
5771 __setup("cgroup.memory=", cgroup_memory
);
5774 * subsys_initcall() for memory controller.
5776 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
5777 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
5778 * basically everything that doesn't depend on a specific mem_cgroup structure
5779 * should be initialized from here.
5781 static int __init
mem_cgroup_init(void)
5787 * Kmem cache creation is mostly done with the slab_mutex held,
5788 * so use a special workqueue to avoid stalling all worker
5789 * threads in case lots of cgroups are created simultaneously.
5791 memcg_kmem_cache_create_wq
=
5792 alloc_ordered_workqueue("memcg_kmem_cache_create", 0);
5793 BUG_ON(!memcg_kmem_cache_create_wq
);
5796 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD
, "mm/memctrl:dead", NULL
,
5797 memcg_hotplug_cpu_dead
);
5799 for_each_possible_cpu(cpu
)
5800 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5803 for_each_node(node
) {
5804 struct mem_cgroup_tree_per_node
*rtpn
;
5806 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5807 node_online(node
) ? node
: NUMA_NO_NODE
);
5809 rtpn
->rb_root
= RB_ROOT
;
5810 spin_lock_init(&rtpn
->lock
);
5811 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5816 subsys_initcall(mem_cgroup_init
);
5818 #ifdef CONFIG_MEMCG_SWAP
5819 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
5821 while (!atomic_inc_not_zero(&memcg
->id
.ref
)) {
5823 * The root cgroup cannot be destroyed, so it's refcount must
5826 if (WARN_ON_ONCE(memcg
== root_mem_cgroup
)) {
5830 memcg
= parent_mem_cgroup(memcg
);
5832 memcg
= root_mem_cgroup
;
5838 * mem_cgroup_swapout - transfer a memsw charge to swap
5839 * @page: page whose memsw charge to transfer
5840 * @entry: swap entry to move the charge to
5842 * Transfer the memsw charge of @page to @entry.
5844 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5846 struct mem_cgroup
*memcg
, *swap_memcg
;
5847 unsigned short oldid
;
5849 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5850 VM_BUG_ON_PAGE(page_count(page
), page
);
5852 if (!do_memsw_account())
5855 memcg
= page
->mem_cgroup
;
5857 /* Readahead page, never charged */
5862 * In case the memcg owning these pages has been offlined and doesn't
5863 * have an ID allocated to it anymore, charge the closest online
5864 * ancestor for the swap instead and transfer the memory+swap charge.
5866 swap_memcg
= mem_cgroup_id_get_online(memcg
);
5867 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
));
5868 VM_BUG_ON_PAGE(oldid
, page
);
5869 mem_cgroup_swap_statistics(swap_memcg
, true);
5871 page
->mem_cgroup
= NULL
;
5873 if (!mem_cgroup_is_root(memcg
))
5874 page_counter_uncharge(&memcg
->memory
, 1);
5876 if (memcg
!= swap_memcg
) {
5877 if (!mem_cgroup_is_root(swap_memcg
))
5878 page_counter_charge(&swap_memcg
->memsw
, 1);
5879 page_counter_uncharge(&memcg
->memsw
, 1);
5883 * Interrupts should be disabled here because the caller holds the
5884 * mapping->tree_lock lock which is taken with interrupts-off. It is
5885 * important here to have the interrupts disabled because it is the
5886 * only synchronisation we have for udpating the per-CPU variables.
5888 VM_BUG_ON(!irqs_disabled());
5889 mem_cgroup_charge_statistics(memcg
, page
, false, -1);
5890 memcg_check_events(memcg
, page
);
5892 if (!mem_cgroup_is_root(memcg
))
5893 css_put(&memcg
->css
);
5897 * mem_cgroup_try_charge_swap - try charging a swap entry
5898 * @page: page being added to swap
5899 * @entry: swap entry to charge
5901 * Try to charge @entry to the memcg that @page belongs to.
5903 * Returns 0 on success, -ENOMEM on failure.
5905 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
5907 struct mem_cgroup
*memcg
;
5908 struct page_counter
*counter
;
5909 unsigned short oldid
;
5911 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
5914 memcg
= page
->mem_cgroup
;
5916 /* Readahead page, never charged */
5920 memcg
= mem_cgroup_id_get_online(memcg
);
5922 if (!mem_cgroup_is_root(memcg
) &&
5923 !page_counter_try_charge(&memcg
->swap
, 1, &counter
)) {
5924 mem_cgroup_id_put(memcg
);
5928 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5929 VM_BUG_ON_PAGE(oldid
, page
);
5930 mem_cgroup_swap_statistics(memcg
, true);
5936 * mem_cgroup_uncharge_swap - uncharge a swap entry
5937 * @entry: swap entry to uncharge
5939 * Drop the swap charge associated with @entry.
5941 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5943 struct mem_cgroup
*memcg
;
5946 if (!do_swap_account
)
5949 id
= swap_cgroup_record(entry
, 0);
5951 memcg
= mem_cgroup_from_id(id
);
5953 if (!mem_cgroup_is_root(memcg
)) {
5954 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5955 page_counter_uncharge(&memcg
->swap
, 1);
5957 page_counter_uncharge(&memcg
->memsw
, 1);
5959 mem_cgroup_swap_statistics(memcg
, false);
5960 mem_cgroup_id_put(memcg
);
5965 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
5967 long nr_swap_pages
= get_nr_swap_pages();
5969 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5970 return nr_swap_pages
;
5971 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5972 nr_swap_pages
= min_t(long, nr_swap_pages
,
5973 READ_ONCE(memcg
->swap
.limit
) -
5974 page_counter_read(&memcg
->swap
));
5975 return nr_swap_pages
;
5978 bool mem_cgroup_swap_full(struct page
*page
)
5980 struct mem_cgroup
*memcg
;
5982 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5986 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5989 memcg
= page
->mem_cgroup
;
5993 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5994 if (page_counter_read(&memcg
->swap
) * 2 >= memcg
->swap
.limit
)
6000 /* for remember boot option*/
6001 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6002 static int really_do_swap_account __initdata
= 1;
6004 static int really_do_swap_account __initdata
;
6007 static int __init
enable_swap_account(char *s
)
6009 if (!strcmp(s
, "1"))
6010 really_do_swap_account
= 1;
6011 else if (!strcmp(s
, "0"))
6012 really_do_swap_account
= 0;
6015 __setup("swapaccount=", enable_swap_account
);
6017 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
6020 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6022 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
6025 static int swap_max_show(struct seq_file
*m
, void *v
)
6027 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
6028 unsigned long max
= READ_ONCE(memcg
->swap
.limit
);
6030 if (max
== PAGE_COUNTER_MAX
)
6031 seq_puts(m
, "max\n");
6033 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
6038 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
6039 char *buf
, size_t nbytes
, loff_t off
)
6041 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6045 buf
= strstrip(buf
);
6046 err
= page_counter_memparse(buf
, "max", &max
);
6050 mutex_lock(&memcg_limit_mutex
);
6051 err
= page_counter_limit(&memcg
->swap
, max
);
6052 mutex_unlock(&memcg_limit_mutex
);
6059 static struct cftype swap_files
[] = {
6061 .name
= "swap.current",
6062 .flags
= CFTYPE_NOT_ON_ROOT
,
6063 .read_u64
= swap_current_read
,
6067 .flags
= CFTYPE_NOT_ON_ROOT
,
6068 .seq_show
= swap_max_show
,
6069 .write
= swap_max_write
,
6074 static struct cftype memsw_cgroup_files
[] = {
6076 .name
= "memsw.usage_in_bytes",
6077 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6078 .read_u64
= mem_cgroup_read_u64
,
6081 .name
= "memsw.max_usage_in_bytes",
6082 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6083 .write
= mem_cgroup_reset
,
6084 .read_u64
= mem_cgroup_read_u64
,
6087 .name
= "memsw.limit_in_bytes",
6088 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6089 .write
= mem_cgroup_write
,
6090 .read_u64
= mem_cgroup_read_u64
,
6093 .name
= "memsw.failcnt",
6094 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6095 .write
= mem_cgroup_reset
,
6096 .read_u64
= mem_cgroup_read_u64
,
6098 { }, /* terminate */
6101 static int __init
mem_cgroup_swap_init(void)
6103 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6104 do_swap_account
= 1;
6105 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
6107 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
6108 memsw_cgroup_files
));
6112 subsys_initcall(mem_cgroup_swap_init
);
6114 #endif /* CONFIG_MEMCG_SWAP */