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/sched/mm.h>
39 #include <linux/shmem_fs.h>
40 #include <linux/hugetlb.h>
41 #include <linux/pagemap.h>
42 #include <linux/smp.h>
43 #include <linux/page-flags.h>
44 #include <linux/backing-dev.h>
45 #include <linux/bit_spinlock.h>
46 #include <linux/rcupdate.h>
47 #include <linux/limits.h>
48 #include <linux/export.h>
49 #include <linux/mutex.h>
50 #include <linux/rbtree.h>
51 #include <linux/slab.h>
52 #include <linux/swap.h>
53 #include <linux/swapops.h>
54 #include <linux/spinlock.h>
55 #include <linux/eventfd.h>
56 #include <linux/poll.h>
57 #include <linux/sort.h>
59 #include <linux/seq_file.h>
60 #include <linux/vmpressure.h>
61 #include <linux/mm_inline.h>
62 #include <linux/swap_cgroup.h>
63 #include <linux/cpu.h>
64 #include <linux/oom.h>
65 #include <linux/lockdep.h>
66 #include <linux/file.h>
67 #include <linux/tracehook.h>
73 #include <linux/uaccess.h>
75 #include <trace/events/vmscan.h>
77 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
78 EXPORT_SYMBOL(memory_cgrp_subsys
);
80 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
82 #define MEM_CGROUP_RECLAIM_RETRIES 5
84 /* Socket memory accounting disabled? */
85 static bool cgroup_memory_nosocket
;
87 /* Kernel memory accounting disabled? */
88 static bool cgroup_memory_nokmem
;
90 /* Whether the swap controller is active */
91 #ifdef CONFIG_MEMCG_SWAP
92 int do_swap_account __read_mostly
;
94 #define do_swap_account 0
97 /* Whether legacy memory+swap accounting is active */
98 static bool do_memsw_account(void)
100 return !cgroup_subsys_on_dfl(memory_cgrp_subsys
) && do_swap_account
;
103 static const char *const mem_cgroup_lru_names
[] = {
111 #define THRESHOLDS_EVENTS_TARGET 128
112 #define SOFTLIMIT_EVENTS_TARGET 1024
113 #define NUMAINFO_EVENTS_TARGET 1024
116 * Cgroups above their limits are maintained in a RB-Tree, independent of
117 * their hierarchy representation
120 struct mem_cgroup_tree_per_node
{
121 struct rb_root rb_root
;
122 struct rb_node
*rb_rightmost
;
126 struct mem_cgroup_tree
{
127 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
130 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
133 struct mem_cgroup_eventfd_list
{
134 struct list_head list
;
135 struct eventfd_ctx
*eventfd
;
139 * cgroup_event represents events which userspace want to receive.
141 struct mem_cgroup_event
{
143 * memcg which the event belongs to.
145 struct mem_cgroup
*memcg
;
147 * eventfd to signal userspace about the event.
149 struct eventfd_ctx
*eventfd
;
151 * Each of these stored in a list by the cgroup.
153 struct list_head list
;
155 * register_event() callback will be used to add new userspace
156 * waiter for changes related to this event. Use eventfd_signal()
157 * on eventfd to send notification to userspace.
159 int (*register_event
)(struct mem_cgroup
*memcg
,
160 struct eventfd_ctx
*eventfd
, const char *args
);
162 * unregister_event() callback will be called when userspace closes
163 * the eventfd or on cgroup removing. This callback must be set,
164 * if you want provide notification functionality.
166 void (*unregister_event
)(struct mem_cgroup
*memcg
,
167 struct eventfd_ctx
*eventfd
);
169 * All fields below needed to unregister event when
170 * userspace closes eventfd.
173 wait_queue_head_t
*wqh
;
174 wait_queue_entry_t wait
;
175 struct work_struct remove
;
178 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
179 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
181 /* Stuffs for move charges at task migration. */
183 * Types of charges to be moved.
185 #define MOVE_ANON 0x1U
186 #define MOVE_FILE 0x2U
187 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
189 /* "mc" and its members are protected by cgroup_mutex */
190 static struct move_charge_struct
{
191 spinlock_t lock
; /* for from, to */
192 struct mm_struct
*mm
;
193 struct mem_cgroup
*from
;
194 struct mem_cgroup
*to
;
196 unsigned long precharge
;
197 unsigned long moved_charge
;
198 unsigned long moved_swap
;
199 struct task_struct
*moving_task
; /* a task moving charges */
200 wait_queue_head_t waitq
; /* a waitq for other context */
202 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
203 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
207 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
208 * limit reclaim to prevent infinite loops, if they ever occur.
210 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
211 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
214 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
215 MEM_CGROUP_CHARGE_TYPE_ANON
,
216 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
217 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
221 /* for encoding cft->private value on file */
230 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
231 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
232 #define MEMFILE_ATTR(val) ((val) & 0xffff)
233 /* Used for OOM nofiier */
234 #define OOM_CONTROL (0)
236 static inline bool should_force_charge(void)
238 return tsk_is_oom_victim(current
) || fatal_signal_pending(current
) ||
239 (current
->flags
& PF_EXITING
);
242 /* Some nice accessors for the vmpressure. */
243 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
246 memcg
= root_mem_cgroup
;
247 return &memcg
->vmpressure
;
250 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
252 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
255 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
257 return (memcg
== root_mem_cgroup
);
262 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
263 * The main reason for not using cgroup id for this:
264 * this works better in sparse environments, where we have a lot of memcgs,
265 * but only a few kmem-limited. Or also, if we have, for instance, 200
266 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
267 * 200 entry array for that.
269 * The current size of the caches array is stored in memcg_nr_cache_ids. It
270 * will double each time we have to increase it.
272 static DEFINE_IDA(memcg_cache_ida
);
273 int memcg_nr_cache_ids
;
275 /* Protects memcg_nr_cache_ids */
276 static DECLARE_RWSEM(memcg_cache_ids_sem
);
278 void memcg_get_cache_ids(void)
280 down_read(&memcg_cache_ids_sem
);
283 void memcg_put_cache_ids(void)
285 up_read(&memcg_cache_ids_sem
);
289 * MIN_SIZE is different than 1, because we would like to avoid going through
290 * the alloc/free process all the time. In a small machine, 4 kmem-limited
291 * cgroups is a reasonable guess. In the future, it could be a parameter or
292 * tunable, but that is strictly not necessary.
294 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
295 * this constant directly from cgroup, but it is understandable that this is
296 * better kept as an internal representation in cgroup.c. In any case, the
297 * cgrp_id space is not getting any smaller, and we don't have to necessarily
298 * increase ours as well if it increases.
300 #define MEMCG_CACHES_MIN_SIZE 4
301 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
304 * A lot of the calls to the cache allocation functions are expected to be
305 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
306 * conditional to this static branch, we'll have to allow modules that does
307 * kmem_cache_alloc and the such to see this symbol as well
309 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
310 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
312 struct workqueue_struct
*memcg_kmem_cache_wq
;
314 #endif /* !CONFIG_SLOB */
317 * mem_cgroup_css_from_page - css of the memcg associated with a page
318 * @page: page of interest
320 * If memcg is bound to the default hierarchy, css of the memcg associated
321 * with @page is returned. The returned css remains associated with @page
322 * until it is released.
324 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
327 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
329 struct mem_cgroup
*memcg
;
331 memcg
= page
->mem_cgroup
;
333 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
334 memcg
= root_mem_cgroup
;
340 * page_cgroup_ino - return inode number of the memcg a page is charged to
343 * Look up the closest online ancestor of the memory cgroup @page is charged to
344 * and return its inode number or 0 if @page is not charged to any cgroup. It
345 * is safe to call this function without holding a reference to @page.
347 * Note, this function is inherently racy, because there is nothing to prevent
348 * the cgroup inode from getting torn down and potentially reallocated a moment
349 * after page_cgroup_ino() returns, so it only should be used by callers that
350 * do not care (such as procfs interfaces).
352 ino_t
page_cgroup_ino(struct page
*page
)
354 struct mem_cgroup
*memcg
;
355 unsigned long ino
= 0;
358 memcg
= READ_ONCE(page
->mem_cgroup
);
359 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
360 memcg
= parent_mem_cgroup(memcg
);
362 ino
= cgroup_ino(memcg
->css
.cgroup
);
367 static struct mem_cgroup_per_node
*
368 mem_cgroup_page_nodeinfo(struct mem_cgroup
*memcg
, struct page
*page
)
370 int nid
= page_to_nid(page
);
372 return memcg
->nodeinfo
[nid
];
375 static struct mem_cgroup_tree_per_node
*
376 soft_limit_tree_node(int nid
)
378 return soft_limit_tree
.rb_tree_per_node
[nid
];
381 static struct mem_cgroup_tree_per_node
*
382 soft_limit_tree_from_page(struct page
*page
)
384 int nid
= page_to_nid(page
);
386 return soft_limit_tree
.rb_tree_per_node
[nid
];
389 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node
*mz
,
390 struct mem_cgroup_tree_per_node
*mctz
,
391 unsigned long new_usage_in_excess
)
393 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
394 struct rb_node
*parent
= NULL
;
395 struct mem_cgroup_per_node
*mz_node
;
396 bool rightmost
= true;
401 mz
->usage_in_excess
= new_usage_in_excess
;
402 if (!mz
->usage_in_excess
)
406 mz_node
= rb_entry(parent
, struct mem_cgroup_per_node
,
408 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
) {
414 * We can't avoid mem cgroups that are over their soft
415 * limit by the same amount
417 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
422 mctz
->rb_rightmost
= &mz
->tree_node
;
424 rb_link_node(&mz
->tree_node
, parent
, p
);
425 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
429 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
430 struct mem_cgroup_tree_per_node
*mctz
)
435 if (&mz
->tree_node
== mctz
->rb_rightmost
)
436 mctz
->rb_rightmost
= rb_prev(&mz
->tree_node
);
438 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
442 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
443 struct mem_cgroup_tree_per_node
*mctz
)
447 spin_lock_irqsave(&mctz
->lock
, flags
);
448 __mem_cgroup_remove_exceeded(mz
, mctz
);
449 spin_unlock_irqrestore(&mctz
->lock
, flags
);
452 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
454 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
455 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
456 unsigned long excess
= 0;
458 if (nr_pages
> soft_limit
)
459 excess
= nr_pages
- soft_limit
;
464 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
466 unsigned long excess
;
467 struct mem_cgroup_per_node
*mz
;
468 struct mem_cgroup_tree_per_node
*mctz
;
470 mctz
= soft_limit_tree_from_page(page
);
474 * Necessary to update all ancestors when hierarchy is used.
475 * because their event counter is not touched.
477 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
478 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
479 excess
= soft_limit_excess(memcg
);
481 * We have to update the tree if mz is on RB-tree or
482 * mem is over its softlimit.
484 if (excess
|| mz
->on_tree
) {
487 spin_lock_irqsave(&mctz
->lock
, flags
);
488 /* if on-tree, remove it */
490 __mem_cgroup_remove_exceeded(mz
, mctz
);
492 * Insert again. mz->usage_in_excess will be updated.
493 * If excess is 0, no tree ops.
495 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
496 spin_unlock_irqrestore(&mctz
->lock
, flags
);
501 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
503 struct mem_cgroup_tree_per_node
*mctz
;
504 struct mem_cgroup_per_node
*mz
;
508 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
509 mctz
= soft_limit_tree_node(nid
);
511 mem_cgroup_remove_exceeded(mz
, mctz
);
515 static struct mem_cgroup_per_node
*
516 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
518 struct mem_cgroup_per_node
*mz
;
522 if (!mctz
->rb_rightmost
)
523 goto done
; /* Nothing to reclaim from */
525 mz
= rb_entry(mctz
->rb_rightmost
,
526 struct mem_cgroup_per_node
, tree_node
);
528 * Remove the node now but someone else can add it back,
529 * we will to add it back at the end of reclaim to its correct
530 * position in the tree.
532 __mem_cgroup_remove_exceeded(mz
, mctz
);
533 if (!soft_limit_excess(mz
->memcg
) ||
534 !css_tryget_online(&mz
->memcg
->css
))
540 static struct mem_cgroup_per_node
*
541 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
543 struct mem_cgroup_per_node
*mz
;
545 spin_lock_irq(&mctz
->lock
);
546 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
547 spin_unlock_irq(&mctz
->lock
);
552 * Return page count for single (non recursive) @memcg.
554 * Implementation Note: reading percpu statistics for memcg.
556 * Both of vmstat[] and percpu_counter has threshold and do periodic
557 * synchronization to implement "quick" read. There are trade-off between
558 * reading cost and precision of value. Then, we may have a chance to implement
559 * a periodic synchronization of counter in memcg's counter.
561 * But this _read() function is used for user interface now. The user accounts
562 * memory usage by memory cgroup and he _always_ requires exact value because
563 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
564 * have to visit all online cpus and make sum. So, for now, unnecessary
565 * synchronization is not implemented. (just implemented for cpu hotplug)
567 * If there are kernel internal actions which can make use of some not-exact
568 * value, and reading all cpu value can be performance bottleneck in some
569 * common workload, threshold and synchronization as vmstat[] should be
572 * The parameter idx can be of type enum memcg_event_item or vm_event_item.
575 static unsigned long memcg_sum_events(struct mem_cgroup
*memcg
,
578 unsigned long val
= 0;
581 for_each_possible_cpu(cpu
)
582 val
+= per_cpu(memcg
->stat
->events
[event
], cpu
);
586 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
588 bool compound
, int nr_pages
)
591 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
592 * counted as CACHE even if it's on ANON LRU.
595 __this_cpu_add(memcg
->stat
->count
[MEMCG_RSS
], nr_pages
);
597 __this_cpu_add(memcg
->stat
->count
[MEMCG_CACHE
], nr_pages
);
598 if (PageSwapBacked(page
))
599 __this_cpu_add(memcg
->stat
->count
[NR_SHMEM
], nr_pages
);
603 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
604 __this_cpu_add(memcg
->stat
->count
[MEMCG_RSS_HUGE
], nr_pages
);
607 /* pagein of a big page is an event. So, ignore page size */
609 __this_cpu_inc(memcg
->stat
->events
[PGPGIN
]);
611 __this_cpu_inc(memcg
->stat
->events
[PGPGOUT
]);
612 nr_pages
= -nr_pages
; /* for event */
615 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
618 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
619 int nid
, unsigned int lru_mask
)
621 struct lruvec
*lruvec
= mem_cgroup_lruvec(NODE_DATA(nid
), memcg
);
622 unsigned long nr
= 0;
625 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
628 if (!(BIT(lru
) & lru_mask
))
630 nr
+= mem_cgroup_get_lru_size(lruvec
, lru
);
635 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
636 unsigned int lru_mask
)
638 unsigned long nr
= 0;
641 for_each_node_state(nid
, N_MEMORY
)
642 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
646 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
647 enum mem_cgroup_events_target target
)
649 unsigned long val
, next
;
651 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
652 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
653 /* from time_after() in jiffies.h */
654 if ((long)(next
- val
) < 0) {
656 case MEM_CGROUP_TARGET_THRESH
:
657 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
659 case MEM_CGROUP_TARGET_SOFTLIMIT
:
660 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
662 case MEM_CGROUP_TARGET_NUMAINFO
:
663 next
= val
+ NUMAINFO_EVENTS_TARGET
;
668 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
675 * Check events in order.
678 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
680 /* threshold event is triggered in finer grain than soft limit */
681 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
682 MEM_CGROUP_TARGET_THRESH
))) {
684 bool do_numainfo __maybe_unused
;
686 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
687 MEM_CGROUP_TARGET_SOFTLIMIT
);
689 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
690 MEM_CGROUP_TARGET_NUMAINFO
);
692 mem_cgroup_threshold(memcg
);
693 if (unlikely(do_softlimit
))
694 mem_cgroup_update_tree(memcg
, page
);
696 if (unlikely(do_numainfo
))
697 atomic_inc(&memcg
->numainfo_events
);
702 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
705 * mm_update_next_owner() may clear mm->owner to NULL
706 * if it races with swapoff, page migration, etc.
707 * So this can be called with p == NULL.
712 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
714 EXPORT_SYMBOL(mem_cgroup_from_task
);
716 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
718 struct mem_cgroup
*memcg
= NULL
;
723 * Page cache insertions can happen withou an
724 * actual mm context, e.g. during disk probing
725 * on boot, loopback IO, acct() writes etc.
728 memcg
= root_mem_cgroup
;
730 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
731 if (unlikely(!memcg
))
732 memcg
= root_mem_cgroup
;
734 } while (!css_tryget_online(&memcg
->css
));
740 * mem_cgroup_iter - iterate over memory cgroup hierarchy
741 * @root: hierarchy root
742 * @prev: previously returned memcg, NULL on first invocation
743 * @reclaim: cookie for shared reclaim walks, NULL for full walks
745 * Returns references to children of the hierarchy below @root, or
746 * @root itself, or %NULL after a full round-trip.
748 * Caller must pass the return value in @prev on subsequent
749 * invocations for reference counting, or use mem_cgroup_iter_break()
750 * to cancel a hierarchy walk before the round-trip is complete.
752 * Reclaimers can specify a zone and a priority level in @reclaim to
753 * divide up the memcgs in the hierarchy among all concurrent
754 * reclaimers operating on the same zone and priority.
756 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
757 struct mem_cgroup
*prev
,
758 struct mem_cgroup_reclaim_cookie
*reclaim
)
760 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
761 struct cgroup_subsys_state
*css
= NULL
;
762 struct mem_cgroup
*memcg
= NULL
;
763 struct mem_cgroup
*pos
= NULL
;
765 if (mem_cgroup_disabled())
769 root
= root_mem_cgroup
;
771 if (prev
&& !reclaim
)
774 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
783 struct mem_cgroup_per_node
*mz
;
785 mz
= mem_cgroup_nodeinfo(root
, reclaim
->pgdat
->node_id
);
786 iter
= &mz
->iter
[reclaim
->priority
];
788 if (prev
&& reclaim
->generation
!= iter
->generation
)
792 pos
= READ_ONCE(iter
->position
);
793 if (!pos
|| css_tryget(&pos
->css
))
796 * css reference reached zero, so iter->position will
797 * be cleared by ->css_released. However, we should not
798 * rely on this happening soon, because ->css_released
799 * is called from a work queue, and by busy-waiting we
800 * might block it. So we clear iter->position right
803 (void)cmpxchg(&iter
->position
, pos
, NULL
);
811 css
= css_next_descendant_pre(css
, &root
->css
);
814 * Reclaimers share the hierarchy walk, and a
815 * new one might jump in right at the end of
816 * the hierarchy - make sure they see at least
817 * one group and restart from the beginning.
825 * Verify the css and acquire a reference. The root
826 * is provided by the caller, so we know it's alive
827 * and kicking, and don't take an extra reference.
829 memcg
= mem_cgroup_from_css(css
);
831 if (css
== &root
->css
)
842 * The position could have already been updated by a competing
843 * thread, so check that the value hasn't changed since we read
844 * it to avoid reclaiming from the same cgroup twice.
846 (void)cmpxchg(&iter
->position
, pos
, memcg
);
854 reclaim
->generation
= iter
->generation
;
860 if (prev
&& prev
!= root
)
867 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
868 * @root: hierarchy root
869 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
871 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
872 struct mem_cgroup
*prev
)
875 root
= root_mem_cgroup
;
876 if (prev
&& prev
!= root
)
880 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
882 struct mem_cgroup
*memcg
= dead_memcg
;
883 struct mem_cgroup_reclaim_iter
*iter
;
884 struct mem_cgroup_per_node
*mz
;
888 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
890 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
891 for (i
= 0; i
<= DEF_PRIORITY
; i
++) {
893 cmpxchg(&iter
->position
,
901 * Iteration constructs for visiting all cgroups (under a tree). If
902 * loops are exited prematurely (break), mem_cgroup_iter_break() must
903 * be used for reference counting.
905 #define for_each_mem_cgroup_tree(iter, root) \
906 for (iter = mem_cgroup_iter(root, NULL, NULL); \
908 iter = mem_cgroup_iter(root, iter, NULL))
910 #define for_each_mem_cgroup(iter) \
911 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
913 iter = mem_cgroup_iter(NULL, iter, NULL))
916 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
917 * @memcg: hierarchy root
918 * @fn: function to call for each task
919 * @arg: argument passed to @fn
921 * This function iterates over tasks attached to @memcg or to any of its
922 * descendants and calls @fn for each task. If @fn returns a non-zero
923 * value, the function breaks the iteration loop and returns the value.
924 * Otherwise, it will iterate over all tasks and return 0.
926 * This function must not be called for the root memory cgroup.
928 int mem_cgroup_scan_tasks(struct mem_cgroup
*memcg
,
929 int (*fn
)(struct task_struct
*, void *), void *arg
)
931 struct mem_cgroup
*iter
;
934 BUG_ON(memcg
== root_mem_cgroup
);
936 for_each_mem_cgroup_tree(iter
, memcg
) {
937 struct css_task_iter it
;
938 struct task_struct
*task
;
940 css_task_iter_start(&iter
->css
, 0, &it
);
941 while (!ret
&& (task
= css_task_iter_next(&it
)))
943 css_task_iter_end(&it
);
945 mem_cgroup_iter_break(memcg
, iter
);
953 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
955 * @zone: zone of the page
957 * This function is only safe when following the LRU page isolation
958 * and putback protocol: the LRU lock must be held, and the page must
959 * either be PageLRU() or the caller must have isolated/allocated it.
961 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct pglist_data
*pgdat
)
963 struct mem_cgroup_per_node
*mz
;
964 struct mem_cgroup
*memcg
;
965 struct lruvec
*lruvec
;
967 if (mem_cgroup_disabled()) {
968 lruvec
= &pgdat
->lruvec
;
972 memcg
= page
->mem_cgroup
;
974 * Swapcache readahead pages are added to the LRU - and
975 * possibly migrated - before they are charged.
978 memcg
= root_mem_cgroup
;
980 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
981 lruvec
= &mz
->lruvec
;
984 * Since a node can be onlined after the mem_cgroup was created,
985 * we have to be prepared to initialize lruvec->zone here;
986 * and if offlined then reonlined, we need to reinitialize it.
988 if (unlikely(lruvec
->pgdat
!= pgdat
))
989 lruvec
->pgdat
= pgdat
;
994 * mem_cgroup_update_lru_size - account for adding or removing an lru page
995 * @lruvec: mem_cgroup per zone lru vector
996 * @lru: index of lru list the page is sitting on
997 * @zid: zone id of the accounted pages
998 * @nr_pages: positive when adding or negative when removing
1000 * This function must be called under lru_lock, just before a page is added
1001 * to or just after a page is removed from an lru list (that ordering being
1002 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1004 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1005 int zid
, int nr_pages
)
1007 struct mem_cgroup_per_node
*mz
;
1008 unsigned long *lru_size
;
1011 if (mem_cgroup_disabled())
1014 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
1015 lru_size
= &mz
->lru_zone_size
[zid
][lru
];
1018 *lru_size
+= nr_pages
;
1021 if (WARN_ONCE(size
< 0,
1022 "%s(%p, %d, %d): lru_size %ld\n",
1023 __func__
, lruvec
, lru
, nr_pages
, size
)) {
1029 *lru_size
+= nr_pages
;
1032 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1034 struct mem_cgroup
*task_memcg
;
1035 struct task_struct
*p
;
1038 p
= find_lock_task_mm(task
);
1040 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1044 * All threads may have already detached their mm's, but the oom
1045 * killer still needs to detect if they have already been oom
1046 * killed to prevent needlessly killing additional tasks.
1049 task_memcg
= mem_cgroup_from_task(task
);
1050 css_get(&task_memcg
->css
);
1053 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1054 css_put(&task_memcg
->css
);
1059 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1060 * @memcg: the memory cgroup
1062 * Returns the maximum amount of memory @mem can be charged with, in
1065 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1067 unsigned long margin
= 0;
1068 unsigned long count
;
1069 unsigned long limit
;
1071 count
= page_counter_read(&memcg
->memory
);
1072 limit
= READ_ONCE(memcg
->memory
.limit
);
1074 margin
= limit
- count
;
1076 if (do_memsw_account()) {
1077 count
= page_counter_read(&memcg
->memsw
);
1078 limit
= READ_ONCE(memcg
->memsw
.limit
);
1080 margin
= min(margin
, limit
- count
);
1089 * A routine for checking "mem" is under move_account() or not.
1091 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1092 * moving cgroups. This is for waiting at high-memory pressure
1095 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1097 struct mem_cgroup
*from
;
1098 struct mem_cgroup
*to
;
1101 * Unlike task_move routines, we access mc.to, mc.from not under
1102 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1104 spin_lock(&mc
.lock
);
1110 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1111 mem_cgroup_is_descendant(to
, memcg
);
1113 spin_unlock(&mc
.lock
);
1117 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1119 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1120 if (mem_cgroup_under_move(memcg
)) {
1122 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1123 /* moving charge context might have finished. */
1126 finish_wait(&mc
.waitq
, &wait
);
1133 unsigned int memcg1_stats
[] = {
1144 static const char *const memcg1_stat_names
[] = {
1155 #define K(x) ((x) << (PAGE_SHIFT-10))
1157 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1158 * @memcg: The memory cgroup that went over limit
1159 * @p: Task that is going to be killed
1161 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1164 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1166 struct mem_cgroup
*iter
;
1172 pr_info("Task in ");
1173 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1174 pr_cont(" killed as a result of limit of ");
1176 pr_info("Memory limit reached of cgroup ");
1179 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1184 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1185 K((u64
)page_counter_read(&memcg
->memory
)),
1186 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1187 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1188 K((u64
)page_counter_read(&memcg
->memsw
)),
1189 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1190 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1191 K((u64
)page_counter_read(&memcg
->kmem
)),
1192 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1194 for_each_mem_cgroup_tree(iter
, memcg
) {
1195 pr_info("Memory cgroup stats for ");
1196 pr_cont_cgroup_path(iter
->css
.cgroup
);
1199 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
1200 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_swap_account
)
1202 pr_cont(" %s:%luKB", memcg1_stat_names
[i
],
1203 K(memcg_page_state(iter
, memcg1_stats
[i
])));
1206 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1207 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1208 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1215 * This function returns the number of memcg under hierarchy tree. Returns
1216 * 1(self count) if no children.
1218 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1221 struct mem_cgroup
*iter
;
1223 for_each_mem_cgroup_tree(iter
, memcg
)
1229 * Return the memory (and swap, if configured) limit for a memcg.
1231 unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1233 unsigned long limit
;
1235 limit
= memcg
->memory
.limit
;
1236 if (mem_cgroup_swappiness(memcg
)) {
1237 unsigned long memsw_limit
;
1238 unsigned long swap_limit
;
1240 memsw_limit
= memcg
->memsw
.limit
;
1241 swap_limit
= memcg
->swap
.limit
;
1242 swap_limit
= min(swap_limit
, (unsigned long)total_swap_pages
);
1243 limit
= min(limit
+ swap_limit
, memsw_limit
);
1248 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1251 struct oom_control oc
= {
1255 .gfp_mask
= gfp_mask
,
1260 if (mutex_lock_killable(&oom_lock
))
1263 * A few threads which were not waiting at mutex_lock_killable() can
1264 * fail to bail out. Therefore, check again after holding oom_lock.
1266 ret
= should_force_charge() || out_of_memory(&oc
);
1267 mutex_unlock(&oom_lock
);
1271 #if MAX_NUMNODES > 1
1274 * test_mem_cgroup_node_reclaimable
1275 * @memcg: the target memcg
1276 * @nid: the node ID to be checked.
1277 * @noswap : specify true here if the user wants flle only information.
1279 * This function returns whether the specified memcg contains any
1280 * reclaimable pages on a node. Returns true if there are any reclaimable
1281 * pages in the node.
1283 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1284 int nid
, bool noswap
)
1286 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1288 if (noswap
|| !total_swap_pages
)
1290 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1297 * Always updating the nodemask is not very good - even if we have an empty
1298 * list or the wrong list here, we can start from some node and traverse all
1299 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1302 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1306 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1307 * pagein/pageout changes since the last update.
1309 if (!atomic_read(&memcg
->numainfo_events
))
1311 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1314 /* make a nodemask where this memcg uses memory from */
1315 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1317 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1319 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1320 node_clear(nid
, memcg
->scan_nodes
);
1323 atomic_set(&memcg
->numainfo_events
, 0);
1324 atomic_set(&memcg
->numainfo_updating
, 0);
1328 * Selecting a node where we start reclaim from. Because what we need is just
1329 * reducing usage counter, start from anywhere is O,K. Considering
1330 * memory reclaim from current node, there are pros. and cons.
1332 * Freeing memory from current node means freeing memory from a node which
1333 * we'll use or we've used. So, it may make LRU bad. And if several threads
1334 * hit limits, it will see a contention on a node. But freeing from remote
1335 * node means more costs for memory reclaim because of memory latency.
1337 * Now, we use round-robin. Better algorithm is welcomed.
1339 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1343 mem_cgroup_may_update_nodemask(memcg
);
1344 node
= memcg
->last_scanned_node
;
1346 node
= next_node_in(node
, memcg
->scan_nodes
);
1348 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1349 * last time it really checked all the LRUs due to rate limiting.
1350 * Fallback to the current node in that case for simplicity.
1352 if (unlikely(node
== MAX_NUMNODES
))
1353 node
= numa_node_id();
1355 memcg
->last_scanned_node
= node
;
1359 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1365 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1368 unsigned long *total_scanned
)
1370 struct mem_cgroup
*victim
= NULL
;
1373 unsigned long excess
;
1374 unsigned long nr_scanned
;
1375 struct mem_cgroup_reclaim_cookie reclaim
= {
1380 excess
= soft_limit_excess(root_memcg
);
1383 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1388 * If we have not been able to reclaim
1389 * anything, it might because there are
1390 * no reclaimable pages under this hierarchy
1395 * We want to do more targeted reclaim.
1396 * excess >> 2 is not to excessive so as to
1397 * reclaim too much, nor too less that we keep
1398 * coming back to reclaim from this cgroup
1400 if (total
>= (excess
>> 2) ||
1401 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1406 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1407 pgdat
, &nr_scanned
);
1408 *total_scanned
+= nr_scanned
;
1409 if (!soft_limit_excess(root_memcg
))
1412 mem_cgroup_iter_break(root_memcg
, victim
);
1416 #ifdef CONFIG_LOCKDEP
1417 static struct lockdep_map memcg_oom_lock_dep_map
= {
1418 .name
= "memcg_oom_lock",
1422 static DEFINE_SPINLOCK(memcg_oom_lock
);
1425 * Check OOM-Killer is already running under our hierarchy.
1426 * If someone is running, return false.
1428 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1430 struct mem_cgroup
*iter
, *failed
= NULL
;
1432 spin_lock(&memcg_oom_lock
);
1434 for_each_mem_cgroup_tree(iter
, memcg
) {
1435 if (iter
->oom_lock
) {
1437 * this subtree of our hierarchy is already locked
1438 * so we cannot give a lock.
1441 mem_cgroup_iter_break(memcg
, iter
);
1444 iter
->oom_lock
= true;
1449 * OK, we failed to lock the whole subtree so we have
1450 * to clean up what we set up to the failing subtree
1452 for_each_mem_cgroup_tree(iter
, memcg
) {
1453 if (iter
== failed
) {
1454 mem_cgroup_iter_break(memcg
, iter
);
1457 iter
->oom_lock
= false;
1460 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1462 spin_unlock(&memcg_oom_lock
);
1467 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1469 struct mem_cgroup
*iter
;
1471 spin_lock(&memcg_oom_lock
);
1472 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1473 for_each_mem_cgroup_tree(iter
, memcg
)
1474 iter
->oom_lock
= false;
1475 spin_unlock(&memcg_oom_lock
);
1478 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1480 struct mem_cgroup
*iter
;
1482 spin_lock(&memcg_oom_lock
);
1483 for_each_mem_cgroup_tree(iter
, memcg
)
1485 spin_unlock(&memcg_oom_lock
);
1488 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1490 struct mem_cgroup
*iter
;
1493 * When a new child is created while the hierarchy is under oom,
1494 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1496 spin_lock(&memcg_oom_lock
);
1497 for_each_mem_cgroup_tree(iter
, memcg
)
1498 if (iter
->under_oom
> 0)
1500 spin_unlock(&memcg_oom_lock
);
1503 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1505 struct oom_wait_info
{
1506 struct mem_cgroup
*memcg
;
1507 wait_queue_entry_t wait
;
1510 static int memcg_oom_wake_function(wait_queue_entry_t
*wait
,
1511 unsigned mode
, int sync
, void *arg
)
1513 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1514 struct mem_cgroup
*oom_wait_memcg
;
1515 struct oom_wait_info
*oom_wait_info
;
1517 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1518 oom_wait_memcg
= oom_wait_info
->memcg
;
1520 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1521 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1523 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1526 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1529 * For the following lockless ->under_oom test, the only required
1530 * guarantee is that it must see the state asserted by an OOM when
1531 * this function is called as a result of userland actions
1532 * triggered by the notification of the OOM. This is trivially
1533 * achieved by invoking mem_cgroup_mark_under_oom() before
1534 * triggering notification.
1536 if (memcg
&& memcg
->under_oom
)
1537 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1540 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1542 if (!current
->memcg_may_oom
)
1545 * We are in the middle of the charge context here, so we
1546 * don't want to block when potentially sitting on a callstack
1547 * that holds all kinds of filesystem and mm locks.
1549 * Also, the caller may handle a failed allocation gracefully
1550 * (like optional page cache readahead) and so an OOM killer
1551 * invocation might not even be necessary.
1553 * That's why we don't do anything here except remember the
1554 * OOM context and then deal with it at the end of the page
1555 * fault when the stack is unwound, the locks are released,
1556 * and when we know whether the fault was overall successful.
1558 css_get(&memcg
->css
);
1559 current
->memcg_in_oom
= memcg
;
1560 current
->memcg_oom_gfp_mask
= mask
;
1561 current
->memcg_oom_order
= order
;
1565 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1566 * @handle: actually kill/wait or just clean up the OOM state
1568 * This has to be called at the end of a page fault if the memcg OOM
1569 * handler was enabled.
1571 * Memcg supports userspace OOM handling where failed allocations must
1572 * sleep on a waitqueue until the userspace task resolves the
1573 * situation. Sleeping directly in the charge context with all kinds
1574 * of locks held is not a good idea, instead we remember an OOM state
1575 * in the task and mem_cgroup_oom_synchronize() has to be called at
1576 * the end of the page fault to complete the OOM handling.
1578 * Returns %true if an ongoing memcg OOM situation was detected and
1579 * completed, %false otherwise.
1581 bool mem_cgroup_oom_synchronize(bool handle
)
1583 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1584 struct oom_wait_info owait
;
1587 /* OOM is global, do not handle */
1594 owait
.memcg
= memcg
;
1595 owait
.wait
.flags
= 0;
1596 owait
.wait
.func
= memcg_oom_wake_function
;
1597 owait
.wait
.private = current
;
1598 INIT_LIST_HEAD(&owait
.wait
.entry
);
1600 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1601 mem_cgroup_mark_under_oom(memcg
);
1603 locked
= mem_cgroup_oom_trylock(memcg
);
1606 mem_cgroup_oom_notify(memcg
);
1608 if (locked
&& !memcg
->oom_kill_disable
) {
1609 mem_cgroup_unmark_under_oom(memcg
);
1610 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1611 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1612 current
->memcg_oom_order
);
1615 mem_cgroup_unmark_under_oom(memcg
);
1616 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1620 mem_cgroup_oom_unlock(memcg
);
1622 * There is no guarantee that an OOM-lock contender
1623 * sees the wakeups triggered by the OOM kill
1624 * uncharges. Wake any sleepers explicitely.
1626 memcg_oom_recover(memcg
);
1629 current
->memcg_in_oom
= NULL
;
1630 css_put(&memcg
->css
);
1635 * lock_page_memcg - lock a page->mem_cgroup binding
1638 * This function protects unlocked LRU pages from being moved to
1641 * It ensures lifetime of the returned memcg. Caller is responsible
1642 * for the lifetime of the page; __unlock_page_memcg() is available
1643 * when @page might get freed inside the locked section.
1645 struct mem_cgroup
*lock_page_memcg(struct page
*page
)
1647 struct mem_cgroup
*memcg
;
1648 unsigned long flags
;
1651 * The RCU lock is held throughout the transaction. The fast
1652 * path can get away without acquiring the memcg->move_lock
1653 * because page moving starts with an RCU grace period.
1655 * The RCU lock also protects the memcg from being freed when
1656 * the page state that is going to change is the only thing
1657 * preventing the page itself from being freed. E.g. writeback
1658 * doesn't hold a page reference and relies on PG_writeback to
1659 * keep off truncation, migration and so forth.
1663 if (mem_cgroup_disabled())
1666 memcg
= page
->mem_cgroup
;
1667 if (unlikely(!memcg
))
1670 if (atomic_read(&memcg
->moving_account
) <= 0)
1673 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1674 if (memcg
!= page
->mem_cgroup
) {
1675 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1680 * When charge migration first begins, we can have locked and
1681 * unlocked page stat updates happening concurrently. Track
1682 * the task who has the lock for unlock_page_memcg().
1684 memcg
->move_lock_task
= current
;
1685 memcg
->move_lock_flags
= flags
;
1689 EXPORT_SYMBOL(lock_page_memcg
);
1692 * __unlock_page_memcg - unlock and unpin a memcg
1695 * Unlock and unpin a memcg returned by lock_page_memcg().
1697 void __unlock_page_memcg(struct mem_cgroup
*memcg
)
1699 if (memcg
&& memcg
->move_lock_task
== current
) {
1700 unsigned long flags
= memcg
->move_lock_flags
;
1702 memcg
->move_lock_task
= NULL
;
1703 memcg
->move_lock_flags
= 0;
1705 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1712 * unlock_page_memcg - unlock a page->mem_cgroup binding
1715 void unlock_page_memcg(struct page
*page
)
1717 __unlock_page_memcg(page
->mem_cgroup
);
1719 EXPORT_SYMBOL(unlock_page_memcg
);
1722 * size of first charge trial. "32" comes from vmscan.c's magic value.
1723 * TODO: maybe necessary to use big numbers in big irons.
1725 #define CHARGE_BATCH 32U
1726 struct memcg_stock_pcp
{
1727 struct mem_cgroup
*cached
; /* this never be root cgroup */
1728 unsigned int nr_pages
;
1729 struct work_struct work
;
1730 unsigned long flags
;
1731 #define FLUSHING_CACHED_CHARGE 0
1733 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1734 static DEFINE_MUTEX(percpu_charge_mutex
);
1737 * consume_stock: Try to consume stocked charge on this cpu.
1738 * @memcg: memcg to consume from.
1739 * @nr_pages: how many pages to charge.
1741 * The charges will only happen if @memcg matches the current cpu's memcg
1742 * stock, and at least @nr_pages are available in that stock. Failure to
1743 * service an allocation will refill the stock.
1745 * returns true if successful, false otherwise.
1747 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1749 struct memcg_stock_pcp
*stock
;
1750 unsigned long flags
;
1753 if (nr_pages
> CHARGE_BATCH
)
1756 local_irq_save(flags
);
1758 stock
= this_cpu_ptr(&memcg_stock
);
1759 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
1760 stock
->nr_pages
-= nr_pages
;
1764 local_irq_restore(flags
);
1770 * Returns stocks cached in percpu and reset cached information.
1772 static void drain_stock(struct memcg_stock_pcp
*stock
)
1774 struct mem_cgroup
*old
= stock
->cached
;
1776 if (stock
->nr_pages
) {
1777 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
1778 if (do_memsw_account())
1779 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
1780 css_put_many(&old
->css
, stock
->nr_pages
);
1781 stock
->nr_pages
= 0;
1783 stock
->cached
= NULL
;
1786 static void drain_local_stock(struct work_struct
*dummy
)
1788 struct memcg_stock_pcp
*stock
;
1789 unsigned long flags
;
1792 * The only protection from memory hotplug vs. drain_stock races is
1793 * that we always operate on local CPU stock here with IRQ disabled
1795 local_irq_save(flags
);
1797 stock
= this_cpu_ptr(&memcg_stock
);
1799 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
1801 local_irq_restore(flags
);
1805 * Cache charges(val) to local per_cpu area.
1806 * This will be consumed by consume_stock() function, later.
1808 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1810 struct memcg_stock_pcp
*stock
;
1811 unsigned long flags
;
1813 local_irq_save(flags
);
1815 stock
= this_cpu_ptr(&memcg_stock
);
1816 if (stock
->cached
!= memcg
) { /* reset if necessary */
1818 stock
->cached
= memcg
;
1820 stock
->nr_pages
+= nr_pages
;
1822 if (stock
->nr_pages
> CHARGE_BATCH
)
1825 local_irq_restore(flags
);
1829 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1830 * of the hierarchy under it.
1832 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
1836 /* If someone's already draining, avoid adding running more workers. */
1837 if (!mutex_trylock(&percpu_charge_mutex
))
1840 * Notify other cpus that system-wide "drain" is running
1841 * We do not care about races with the cpu hotplug because cpu down
1842 * as well as workers from this path always operate on the local
1843 * per-cpu data. CPU up doesn't touch memcg_stock at all.
1846 for_each_online_cpu(cpu
) {
1847 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1848 struct mem_cgroup
*memcg
;
1850 memcg
= stock
->cached
;
1851 if (!memcg
|| !stock
->nr_pages
|| !css_tryget(&memcg
->css
))
1853 if (!mem_cgroup_is_descendant(memcg
, root_memcg
)) {
1854 css_put(&memcg
->css
);
1857 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
1859 drain_local_stock(&stock
->work
);
1861 schedule_work_on(cpu
, &stock
->work
);
1863 css_put(&memcg
->css
);
1866 mutex_unlock(&percpu_charge_mutex
);
1869 static int memcg_hotplug_cpu_dead(unsigned int cpu
)
1871 struct memcg_stock_pcp
*stock
;
1873 stock
= &per_cpu(memcg_stock
, cpu
);
1878 static void reclaim_high(struct mem_cgroup
*memcg
,
1879 unsigned int nr_pages
,
1883 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
1885 mem_cgroup_event(memcg
, MEMCG_HIGH
);
1886 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
1887 } while ((memcg
= parent_mem_cgroup(memcg
)));
1890 static void high_work_func(struct work_struct
*work
)
1892 struct mem_cgroup
*memcg
;
1894 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
1895 reclaim_high(memcg
, CHARGE_BATCH
, GFP_KERNEL
);
1899 * Scheduled by try_charge() to be executed from the userland return path
1900 * and reclaims memory over the high limit.
1902 void mem_cgroup_handle_over_high(void)
1904 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
1905 struct mem_cgroup
*memcg
;
1907 if (likely(!nr_pages
))
1910 memcg
= get_mem_cgroup_from_mm(current
->mm
);
1911 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
1912 css_put(&memcg
->css
);
1913 current
->memcg_nr_pages_over_high
= 0;
1916 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1917 unsigned int nr_pages
)
1919 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
1920 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1921 struct mem_cgroup
*mem_over_limit
;
1922 struct page_counter
*counter
;
1923 unsigned long nr_reclaimed
;
1924 bool may_swap
= true;
1925 bool drained
= false;
1927 if (mem_cgroup_is_root(memcg
))
1930 if (consume_stock(memcg
, nr_pages
))
1933 if (!do_memsw_account() ||
1934 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
1935 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
1937 if (do_memsw_account())
1938 page_counter_uncharge(&memcg
->memsw
, batch
);
1939 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
1941 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
1945 if (batch
> nr_pages
) {
1951 * Unlike in global OOM situations, memcg is not in a physical
1952 * memory shortage. Allow dying and OOM-killed tasks to
1953 * bypass the last charges so that they can exit quickly and
1954 * free their memory.
1956 if (unlikely(should_force_charge()))
1960 * Prevent unbounded recursion when reclaim operations need to
1961 * allocate memory. This might exceed the limits temporarily,
1962 * but we prefer facilitating memory reclaim and getting back
1963 * under the limit over triggering OOM kills in these cases.
1965 if (unlikely(current
->flags
& PF_MEMALLOC
))
1968 if (unlikely(task_in_memcg_oom(current
)))
1971 if (!gfpflags_allow_blocking(gfp_mask
))
1974 mem_cgroup_event(mem_over_limit
, MEMCG_MAX
);
1976 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
1977 gfp_mask
, may_swap
);
1979 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
1983 drain_all_stock(mem_over_limit
);
1988 if (gfp_mask
& __GFP_NORETRY
)
1991 * Even though the limit is exceeded at this point, reclaim
1992 * may have been able to free some pages. Retry the charge
1993 * before killing the task.
1995 * Only for regular pages, though: huge pages are rather
1996 * unlikely to succeed so close to the limit, and we fall back
1997 * to regular pages anyway in case of failure.
1999 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2002 * At task move, charge accounts can be doubly counted. So, it's
2003 * better to wait until the end of task_move if something is going on.
2005 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2011 if (gfp_mask
& __GFP_NOFAIL
)
2014 if (fatal_signal_pending(current
))
2017 mem_cgroup_event(mem_over_limit
, MEMCG_OOM
);
2019 mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2020 get_order(nr_pages
* PAGE_SIZE
));
2022 if (!(gfp_mask
& __GFP_NOFAIL
))
2026 * The allocation either can't fail or will lead to more memory
2027 * being freed very soon. Allow memory usage go over the limit
2028 * temporarily by force charging it.
2030 page_counter_charge(&memcg
->memory
, nr_pages
);
2031 if (do_memsw_account())
2032 page_counter_charge(&memcg
->memsw
, nr_pages
);
2033 css_get_many(&memcg
->css
, nr_pages
);
2038 css_get_many(&memcg
->css
, batch
);
2039 if (batch
> nr_pages
)
2040 refill_stock(memcg
, batch
- nr_pages
);
2043 * If the hierarchy is above the normal consumption range, schedule
2044 * reclaim on returning to userland. We can perform reclaim here
2045 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2046 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2047 * not recorded as it most likely matches current's and won't
2048 * change in the meantime. As high limit is checked again before
2049 * reclaim, the cost of mismatch is negligible.
2052 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2053 /* Don't bother a random interrupted task */
2054 if (in_interrupt()) {
2055 schedule_work(&memcg
->high_work
);
2058 current
->memcg_nr_pages_over_high
+= batch
;
2059 set_notify_resume(current
);
2062 } while ((memcg
= parent_mem_cgroup(memcg
)));
2067 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2069 if (mem_cgroup_is_root(memcg
))
2072 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2073 if (do_memsw_account())
2074 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2076 css_put_many(&memcg
->css
, nr_pages
);
2079 static void lock_page_lru(struct page
*page
, int *isolated
)
2081 struct zone
*zone
= page_zone(page
);
2083 spin_lock_irq(zone_lru_lock(zone
));
2084 if (PageLRU(page
)) {
2085 struct lruvec
*lruvec
;
2087 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
2089 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2095 static void unlock_page_lru(struct page
*page
, int isolated
)
2097 struct zone
*zone
= page_zone(page
);
2100 struct lruvec
*lruvec
;
2102 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
2103 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2105 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2107 spin_unlock_irq(zone_lru_lock(zone
));
2110 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2115 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2118 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2119 * may already be on some other mem_cgroup's LRU. Take care of it.
2122 lock_page_lru(page
, &isolated
);
2125 * Nobody should be changing or seriously looking at
2126 * page->mem_cgroup at this point:
2128 * - the page is uncharged
2130 * - the page is off-LRU
2132 * - an anonymous fault has exclusive page access, except for
2133 * a locked page table
2135 * - a page cache insertion, a swapin fault, or a migration
2136 * have the page locked
2138 page
->mem_cgroup
= memcg
;
2141 unlock_page_lru(page
, isolated
);
2145 static int memcg_alloc_cache_id(void)
2150 id
= ida_simple_get(&memcg_cache_ida
,
2151 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2155 if (id
< memcg_nr_cache_ids
)
2159 * There's no space for the new id in memcg_caches arrays,
2160 * so we have to grow them.
2162 down_write(&memcg_cache_ids_sem
);
2164 size
= 2 * (id
+ 1);
2165 if (size
< MEMCG_CACHES_MIN_SIZE
)
2166 size
= MEMCG_CACHES_MIN_SIZE
;
2167 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2168 size
= MEMCG_CACHES_MAX_SIZE
;
2170 err
= memcg_update_all_caches(size
);
2172 err
= memcg_update_all_list_lrus(size
);
2174 memcg_nr_cache_ids
= size
;
2176 up_write(&memcg_cache_ids_sem
);
2179 ida_simple_remove(&memcg_cache_ida
, id
);
2185 static void memcg_free_cache_id(int id
)
2187 ida_simple_remove(&memcg_cache_ida
, id
);
2190 struct memcg_kmem_cache_create_work
{
2191 struct mem_cgroup
*memcg
;
2192 struct kmem_cache
*cachep
;
2193 struct work_struct work
;
2196 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2198 struct memcg_kmem_cache_create_work
*cw
=
2199 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2200 struct mem_cgroup
*memcg
= cw
->memcg
;
2201 struct kmem_cache
*cachep
= cw
->cachep
;
2203 memcg_create_kmem_cache(memcg
, cachep
);
2205 css_put(&memcg
->css
);
2210 * Enqueue the creation of a per-memcg kmem_cache.
2212 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2213 struct kmem_cache
*cachep
)
2215 struct memcg_kmem_cache_create_work
*cw
;
2217 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
| __GFP_NOWARN
);
2221 css_get(&memcg
->css
);
2224 cw
->cachep
= cachep
;
2225 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2227 queue_work(memcg_kmem_cache_wq
, &cw
->work
);
2230 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2231 struct kmem_cache
*cachep
)
2234 * We need to stop accounting when we kmalloc, because if the
2235 * corresponding kmalloc cache is not yet created, the first allocation
2236 * in __memcg_schedule_kmem_cache_create will recurse.
2238 * However, it is better to enclose the whole function. Depending on
2239 * the debugging options enabled, INIT_WORK(), for instance, can
2240 * trigger an allocation. This too, will make us recurse. Because at
2241 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2242 * the safest choice is to do it like this, wrapping the whole function.
2244 current
->memcg_kmem_skip_account
= 1;
2245 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2246 current
->memcg_kmem_skip_account
= 0;
2249 static inline bool memcg_kmem_bypass(void)
2251 if (in_interrupt() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
2257 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2258 * @cachep: the original global kmem cache
2260 * Return the kmem_cache we're supposed to use for a slab allocation.
2261 * We try to use the current memcg's version of the cache.
2263 * If the cache does not exist yet, if we are the first user of it, we
2264 * create it asynchronously in a workqueue and let the current allocation
2265 * go through with the original cache.
2267 * This function takes a reference to the cache it returns to assure it
2268 * won't get destroyed while we are working with it. Once the caller is
2269 * done with it, memcg_kmem_put_cache() must be called to release the
2272 struct kmem_cache
*memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2274 struct mem_cgroup
*memcg
;
2275 struct kmem_cache
*memcg_cachep
;
2278 VM_BUG_ON(!is_root_cache(cachep
));
2280 if (memcg_kmem_bypass())
2283 if (current
->memcg_kmem_skip_account
)
2286 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2287 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2291 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2292 if (likely(memcg_cachep
))
2293 return memcg_cachep
;
2296 * If we are in a safe context (can wait, and not in interrupt
2297 * context), we could be be predictable and return right away.
2298 * This would guarantee that the allocation being performed
2299 * already belongs in the new cache.
2301 * However, there are some clashes that can arrive from locking.
2302 * For instance, because we acquire the slab_mutex while doing
2303 * memcg_create_kmem_cache, this means no further allocation
2304 * could happen with the slab_mutex held. So it's better to
2307 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2309 css_put(&memcg
->css
);
2314 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2315 * @cachep: the cache returned by memcg_kmem_get_cache
2317 void memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2319 if (!is_root_cache(cachep
))
2320 css_put(&cachep
->memcg_params
.memcg
->css
);
2324 * memcg_kmem_charge: charge a kmem page
2325 * @page: page to charge
2326 * @gfp: reclaim mode
2327 * @order: allocation order
2328 * @memcg: memory cgroup to charge
2330 * Returns 0 on success, an error code on failure.
2332 int memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2333 struct mem_cgroup
*memcg
)
2335 unsigned int nr_pages
= 1 << order
;
2336 struct page_counter
*counter
;
2339 ret
= try_charge(memcg
, gfp
, nr_pages
);
2343 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2344 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2345 cancel_charge(memcg
, nr_pages
);
2349 page
->mem_cgroup
= memcg
;
2355 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2356 * @page: page to charge
2357 * @gfp: reclaim mode
2358 * @order: allocation order
2360 * Returns 0 on success, an error code on failure.
2362 int memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2364 struct mem_cgroup
*memcg
;
2367 if (memcg_kmem_bypass())
2370 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2371 if (!mem_cgroup_is_root(memcg
)) {
2372 ret
= memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2374 __SetPageKmemcg(page
);
2376 css_put(&memcg
->css
);
2380 * memcg_kmem_uncharge: uncharge a kmem page
2381 * @page: page to uncharge
2382 * @order: allocation order
2384 void memcg_kmem_uncharge(struct page
*page
, int order
)
2386 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2387 unsigned int nr_pages
= 1 << order
;
2392 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2394 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2395 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2397 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2398 if (do_memsw_account())
2399 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2401 page
->mem_cgroup
= NULL
;
2403 /* slab pages do not have PageKmemcg flag set */
2404 if (PageKmemcg(page
))
2405 __ClearPageKmemcg(page
);
2407 css_put_many(&memcg
->css
, nr_pages
);
2409 #endif /* !CONFIG_SLOB */
2411 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2414 * Because tail pages are not marked as "used", set it. We're under
2415 * zone_lru_lock and migration entries setup in all page mappings.
2417 void mem_cgroup_split_huge_fixup(struct page
*head
)
2421 if (mem_cgroup_disabled())
2424 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2425 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2427 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEMCG_RSS_HUGE
],
2430 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2432 #ifdef CONFIG_MEMCG_SWAP
2433 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2436 this_cpu_add(memcg
->stat
->count
[MEMCG_SWAP
], nr_entries
);
2440 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2441 * @entry: swap entry to be moved
2442 * @from: mem_cgroup which the entry is moved from
2443 * @to: mem_cgroup which the entry is moved to
2445 * It succeeds only when the swap_cgroup's record for this entry is the same
2446 * as the mem_cgroup's id of @from.
2448 * Returns 0 on success, -EINVAL on failure.
2450 * The caller must have charged to @to, IOW, called page_counter_charge() about
2451 * both res and memsw, and called css_get().
2453 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2454 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2456 unsigned short old_id
, new_id
;
2458 old_id
= mem_cgroup_id(from
);
2459 new_id
= mem_cgroup_id(to
);
2461 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2462 mem_cgroup_swap_statistics(from
, -1);
2463 mem_cgroup_swap_statistics(to
, 1);
2469 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2470 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2476 static DEFINE_MUTEX(memcg_limit_mutex
);
2478 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2479 unsigned long limit
)
2481 unsigned long curusage
;
2482 unsigned long oldusage
;
2483 bool enlarge
= false;
2488 * For keeping hierarchical_reclaim simple, how long we should retry
2489 * is depends on callers. We set our retry-count to be function
2490 * of # of children which we should visit in this loop.
2492 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2493 mem_cgroup_count_children(memcg
);
2495 oldusage
= page_counter_read(&memcg
->memory
);
2498 if (signal_pending(current
)) {
2503 mutex_lock(&memcg_limit_mutex
);
2504 if (limit
> memcg
->memsw
.limit
) {
2505 mutex_unlock(&memcg_limit_mutex
);
2509 if (limit
> memcg
->memory
.limit
)
2511 ret
= page_counter_limit(&memcg
->memory
, limit
);
2512 mutex_unlock(&memcg_limit_mutex
);
2517 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2519 curusage
= page_counter_read(&memcg
->memory
);
2520 /* Usage is reduced ? */
2521 if (curusage
>= oldusage
)
2524 oldusage
= curusage
;
2525 } while (retry_count
);
2527 if (!ret
&& enlarge
)
2528 memcg_oom_recover(memcg
);
2533 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2534 unsigned long limit
)
2536 unsigned long curusage
;
2537 unsigned long oldusage
;
2538 bool enlarge
= false;
2542 /* see mem_cgroup_resize_res_limit */
2543 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2544 mem_cgroup_count_children(memcg
);
2546 oldusage
= page_counter_read(&memcg
->memsw
);
2549 if (signal_pending(current
)) {
2554 mutex_lock(&memcg_limit_mutex
);
2555 if (limit
< memcg
->memory
.limit
) {
2556 mutex_unlock(&memcg_limit_mutex
);
2560 if (limit
> memcg
->memsw
.limit
)
2562 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2563 mutex_unlock(&memcg_limit_mutex
);
2568 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2570 curusage
= page_counter_read(&memcg
->memsw
);
2571 /* Usage is reduced ? */
2572 if (curusage
>= oldusage
)
2575 oldusage
= curusage
;
2576 } while (retry_count
);
2578 if (!ret
&& enlarge
)
2579 memcg_oom_recover(memcg
);
2584 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
2586 unsigned long *total_scanned
)
2588 unsigned long nr_reclaimed
= 0;
2589 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
2590 unsigned long reclaimed
;
2592 struct mem_cgroup_tree_per_node
*mctz
;
2593 unsigned long excess
;
2594 unsigned long nr_scanned
;
2599 mctz
= soft_limit_tree_node(pgdat
->node_id
);
2602 * Do not even bother to check the largest node if the root
2603 * is empty. Do it lockless to prevent lock bouncing. Races
2604 * are acceptable as soft limit is best effort anyway.
2606 if (!mctz
|| RB_EMPTY_ROOT(&mctz
->rb_root
))
2610 * This loop can run a while, specially if mem_cgroup's continuously
2611 * keep exceeding their soft limit and putting the system under
2618 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2623 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
2624 gfp_mask
, &nr_scanned
);
2625 nr_reclaimed
+= reclaimed
;
2626 *total_scanned
+= nr_scanned
;
2627 spin_lock_irq(&mctz
->lock
);
2628 __mem_cgroup_remove_exceeded(mz
, mctz
);
2631 * If we failed to reclaim anything from this memory cgroup
2632 * it is time to move on to the next cgroup
2636 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2638 excess
= soft_limit_excess(mz
->memcg
);
2640 * One school of thought says that we should not add
2641 * back the node to the tree if reclaim returns 0.
2642 * But our reclaim could return 0, simply because due
2643 * to priority we are exposing a smaller subset of
2644 * memory to reclaim from. Consider this as a longer
2647 /* If excess == 0, no tree ops */
2648 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2649 spin_unlock_irq(&mctz
->lock
);
2650 css_put(&mz
->memcg
->css
);
2653 * Could not reclaim anything and there are no more
2654 * mem cgroups to try or we seem to be looping without
2655 * reclaiming anything.
2657 if (!nr_reclaimed
&&
2659 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2661 } while (!nr_reclaimed
);
2663 css_put(&next_mz
->memcg
->css
);
2664 return nr_reclaimed
;
2668 * Test whether @memcg has children, dead or alive. Note that this
2669 * function doesn't care whether @memcg has use_hierarchy enabled and
2670 * returns %true if there are child csses according to the cgroup
2671 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2673 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2678 ret
= css_next_child(NULL
, &memcg
->css
);
2684 * Reclaims as many pages from the given memcg as possible.
2686 * Caller is responsible for holding css reference for memcg.
2688 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2690 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2692 /* we call try-to-free pages for make this cgroup empty */
2693 lru_add_drain_all();
2694 /* try to free all pages in this cgroup */
2695 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2698 if (signal_pending(current
))
2701 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2705 /* maybe some writeback is necessary */
2706 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2714 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
2715 char *buf
, size_t nbytes
,
2718 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2720 if (mem_cgroup_is_root(memcg
))
2722 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
2725 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
2728 return mem_cgroup_from_css(css
)->use_hierarchy
;
2731 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
2732 struct cftype
*cft
, u64 val
)
2735 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2736 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
2738 if (memcg
->use_hierarchy
== val
)
2742 * If parent's use_hierarchy is set, we can't make any modifications
2743 * in the child subtrees. If it is unset, then the change can
2744 * occur, provided the current cgroup has no children.
2746 * For the root cgroup, parent_mem is NULL, we allow value to be
2747 * set if there are no children.
2749 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
2750 (val
== 1 || val
== 0)) {
2751 if (!memcg_has_children(memcg
))
2752 memcg
->use_hierarchy
= val
;
2761 static void tree_stat(struct mem_cgroup
*memcg
, unsigned long *stat
)
2763 struct mem_cgroup
*iter
;
2766 memset(stat
, 0, sizeof(*stat
) * MEMCG_NR_STAT
);
2768 for_each_mem_cgroup_tree(iter
, memcg
) {
2769 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
2770 stat
[i
] += memcg_page_state(iter
, i
);
2774 static void tree_events(struct mem_cgroup
*memcg
, unsigned long *events
)
2776 struct mem_cgroup
*iter
;
2779 memset(events
, 0, sizeof(*events
) * MEMCG_NR_EVENTS
);
2781 for_each_mem_cgroup_tree(iter
, memcg
) {
2782 for (i
= 0; i
< MEMCG_NR_EVENTS
; i
++)
2783 events
[i
] += memcg_sum_events(iter
, i
);
2787 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
2789 unsigned long val
= 0;
2791 if (mem_cgroup_is_root(memcg
)) {
2792 struct mem_cgroup
*iter
;
2794 for_each_mem_cgroup_tree(iter
, memcg
) {
2795 val
+= memcg_page_state(iter
, MEMCG_CACHE
);
2796 val
+= memcg_page_state(iter
, MEMCG_RSS
);
2798 val
+= memcg_page_state(iter
, MEMCG_SWAP
);
2802 val
= page_counter_read(&memcg
->memory
);
2804 val
= page_counter_read(&memcg
->memsw
);
2817 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
2820 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2821 struct page_counter
*counter
;
2823 switch (MEMFILE_TYPE(cft
->private)) {
2825 counter
= &memcg
->memory
;
2828 counter
= &memcg
->memsw
;
2831 counter
= &memcg
->kmem
;
2834 counter
= &memcg
->tcpmem
;
2840 switch (MEMFILE_ATTR(cft
->private)) {
2842 if (counter
== &memcg
->memory
)
2843 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
2844 if (counter
== &memcg
->memsw
)
2845 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
2846 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
2848 return (u64
)counter
->limit
* PAGE_SIZE
;
2850 return (u64
)counter
->watermark
* PAGE_SIZE
;
2852 return counter
->failcnt
;
2853 case RES_SOFT_LIMIT
:
2854 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
2861 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2865 if (cgroup_memory_nokmem
)
2868 BUG_ON(memcg
->kmemcg_id
>= 0);
2869 BUG_ON(memcg
->kmem_state
);
2871 memcg_id
= memcg_alloc_cache_id();
2875 static_branch_inc(&memcg_kmem_enabled_key
);
2877 * A memory cgroup is considered kmem-online as soon as it gets
2878 * kmemcg_id. Setting the id after enabling static branching will
2879 * guarantee no one starts accounting before all call sites are
2882 memcg
->kmemcg_id
= memcg_id
;
2883 memcg
->kmem_state
= KMEM_ONLINE
;
2884 INIT_LIST_HEAD(&memcg
->kmem_caches
);
2889 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2891 struct cgroup_subsys_state
*css
;
2892 struct mem_cgroup
*parent
, *child
;
2895 if (memcg
->kmem_state
!= KMEM_ONLINE
)
2898 * Clear the online state before clearing memcg_caches array
2899 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2900 * guarantees that no cache will be created for this cgroup
2901 * after we are done (see memcg_create_kmem_cache()).
2903 memcg
->kmem_state
= KMEM_ALLOCATED
;
2905 memcg_deactivate_kmem_caches(memcg
);
2907 kmemcg_id
= memcg
->kmemcg_id
;
2908 BUG_ON(kmemcg_id
< 0);
2910 parent
= parent_mem_cgroup(memcg
);
2912 parent
= root_mem_cgroup
;
2915 * Change kmemcg_id of this cgroup and all its descendants to the
2916 * parent's id, and then move all entries from this cgroup's list_lrus
2917 * to ones of the parent. After we have finished, all list_lrus
2918 * corresponding to this cgroup are guaranteed to remain empty. The
2919 * ordering is imposed by list_lru_node->lock taken by
2920 * memcg_drain_all_list_lrus().
2922 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2923 css_for_each_descendant_pre(css
, &memcg
->css
) {
2924 child
= mem_cgroup_from_css(css
);
2925 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
2926 child
->kmemcg_id
= parent
->kmemcg_id
;
2927 if (!memcg
->use_hierarchy
)
2932 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
2934 memcg_free_cache_id(kmemcg_id
);
2937 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2939 /* css_alloc() failed, offlining didn't happen */
2940 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
2941 memcg_offline_kmem(memcg
);
2943 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
2944 memcg_destroy_kmem_caches(memcg
);
2945 static_branch_dec(&memcg_kmem_enabled_key
);
2946 WARN_ON(page_counter_read(&memcg
->kmem
));
2950 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2954 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2957 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2960 #endif /* !CONFIG_SLOB */
2962 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2963 unsigned long limit
)
2967 mutex_lock(&memcg_limit_mutex
);
2968 ret
= page_counter_limit(&memcg
->kmem
, limit
);
2969 mutex_unlock(&memcg_limit_mutex
);
2973 static int memcg_update_tcp_limit(struct mem_cgroup
*memcg
, unsigned long limit
)
2977 mutex_lock(&memcg_limit_mutex
);
2979 ret
= page_counter_limit(&memcg
->tcpmem
, limit
);
2983 if (!memcg
->tcpmem_active
) {
2985 * The active flag needs to be written after the static_key
2986 * update. This is what guarantees that the socket activation
2987 * function is the last one to run. See mem_cgroup_sk_alloc()
2988 * for details, and note that we don't mark any socket as
2989 * belonging to this memcg until that flag is up.
2991 * We need to do this, because static_keys will span multiple
2992 * sites, but we can't control their order. If we mark a socket
2993 * as accounted, but the accounting functions are not patched in
2994 * yet, we'll lose accounting.
2996 * We never race with the readers in mem_cgroup_sk_alloc(),
2997 * because when this value change, the code to process it is not
3000 static_branch_inc(&memcg_sockets_enabled_key
);
3001 memcg
->tcpmem_active
= true;
3004 mutex_unlock(&memcg_limit_mutex
);
3009 * The user of this function is...
3012 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3013 char *buf
, size_t nbytes
, loff_t off
)
3015 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3016 unsigned long nr_pages
;
3019 buf
= strstrip(buf
);
3020 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3024 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3026 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3030 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3032 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
3035 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
3038 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
3041 ret
= memcg_update_tcp_limit(memcg
, nr_pages
);
3045 case RES_SOFT_LIMIT
:
3046 memcg
->soft_limit
= nr_pages
;
3050 return ret
?: nbytes
;
3053 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3054 size_t nbytes
, loff_t off
)
3056 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3057 struct page_counter
*counter
;
3059 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3061 counter
= &memcg
->memory
;
3064 counter
= &memcg
->memsw
;
3067 counter
= &memcg
->kmem
;
3070 counter
= &memcg
->tcpmem
;
3076 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3078 page_counter_reset_watermark(counter
);
3081 counter
->failcnt
= 0;
3090 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3093 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3097 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3098 struct cftype
*cft
, u64 val
)
3100 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3102 if (val
& ~MOVE_MASK
)
3106 * No kind of locking is needed in here, because ->can_attach() will
3107 * check this value once in the beginning of the process, and then carry
3108 * on with stale data. This means that changes to this value will only
3109 * affect task migrations starting after the change.
3111 memcg
->move_charge_at_immigrate
= val
;
3115 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3116 struct cftype
*cft
, u64 val
)
3123 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3127 unsigned int lru_mask
;
3130 static const struct numa_stat stats
[] = {
3131 { "total", LRU_ALL
},
3132 { "file", LRU_ALL_FILE
},
3133 { "anon", LRU_ALL_ANON
},
3134 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3136 const struct numa_stat
*stat
;
3139 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3141 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3142 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3143 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3144 for_each_node_state(nid
, N_MEMORY
) {
3145 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3147 seq_printf(m
, " N%d=%lu", nid
, nr
);
3152 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3153 struct mem_cgroup
*iter
;
3156 for_each_mem_cgroup_tree(iter
, memcg
)
3157 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3158 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3159 for_each_node_state(nid
, N_MEMORY
) {
3161 for_each_mem_cgroup_tree(iter
, memcg
)
3162 nr
+= mem_cgroup_node_nr_lru_pages(
3163 iter
, nid
, stat
->lru_mask
);
3164 seq_printf(m
, " N%d=%lu", nid
, nr
);
3171 #endif /* CONFIG_NUMA */
3173 /* Universal VM events cgroup1 shows, original sort order */
3174 unsigned int memcg1_events
[] = {
3181 static const char *const memcg1_event_names
[] = {
3188 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3190 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3191 unsigned long memory
, memsw
;
3192 struct mem_cgroup
*mi
;
3195 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names
) != ARRAY_SIZE(memcg1_stats
));
3196 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3198 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
3199 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
3201 seq_printf(m
, "%s %lu\n", memcg1_stat_names
[i
],
3202 memcg_page_state(memcg
, memcg1_stats
[i
]) *
3206 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
3207 seq_printf(m
, "%s %lu\n", memcg1_event_names
[i
],
3208 memcg_sum_events(memcg
, memcg1_events
[i
]));
3210 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3211 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3212 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3214 /* Hierarchical information */
3215 memory
= memsw
= PAGE_COUNTER_MAX
;
3216 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3217 memory
= min(memory
, mi
->memory
.limit
);
3218 memsw
= min(memsw
, mi
->memsw
.limit
);
3220 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3221 (u64
)memory
* PAGE_SIZE
);
3222 if (do_memsw_account())
3223 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3224 (u64
)memsw
* PAGE_SIZE
);
3226 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
3227 unsigned long long val
= 0;
3229 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
3231 for_each_mem_cgroup_tree(mi
, memcg
)
3232 val
+= memcg_page_state(mi
, memcg1_stats
[i
]) *
3234 seq_printf(m
, "total_%s %llu\n", memcg1_stat_names
[i
], val
);
3237 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++) {
3238 unsigned long long val
= 0;
3240 for_each_mem_cgroup_tree(mi
, memcg
)
3241 val
+= memcg_sum_events(mi
, memcg1_events
[i
]);
3242 seq_printf(m
, "total_%s %llu\n", memcg1_event_names
[i
], val
);
3245 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3246 unsigned long long val
= 0;
3248 for_each_mem_cgroup_tree(mi
, memcg
)
3249 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3250 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3253 #ifdef CONFIG_DEBUG_VM
3256 struct mem_cgroup_per_node
*mz
;
3257 struct zone_reclaim_stat
*rstat
;
3258 unsigned long recent_rotated
[2] = {0, 0};
3259 unsigned long recent_scanned
[2] = {0, 0};
3261 for_each_online_pgdat(pgdat
) {
3262 mz
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
3263 rstat
= &mz
->lruvec
.reclaim_stat
;
3265 recent_rotated
[0] += rstat
->recent_rotated
[0];
3266 recent_rotated
[1] += rstat
->recent_rotated
[1];
3267 recent_scanned
[0] += rstat
->recent_scanned
[0];
3268 recent_scanned
[1] += rstat
->recent_scanned
[1];
3270 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3271 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3272 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3273 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3280 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3283 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3285 return mem_cgroup_swappiness(memcg
);
3288 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3289 struct cftype
*cft
, u64 val
)
3291 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3297 memcg
->swappiness
= val
;
3299 vm_swappiness
= val
;
3304 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3306 struct mem_cgroup_threshold_ary
*t
;
3307 unsigned long usage
;
3312 t
= rcu_dereference(memcg
->thresholds
.primary
);
3314 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3319 usage
= mem_cgroup_usage(memcg
, swap
);
3322 * current_threshold points to threshold just below or equal to usage.
3323 * If it's not true, a threshold was crossed after last
3324 * call of __mem_cgroup_threshold().
3326 i
= t
->current_threshold
;
3329 * Iterate backward over array of thresholds starting from
3330 * current_threshold and check if a threshold is crossed.
3331 * If none of thresholds below usage is crossed, we read
3332 * only one element of the array here.
3334 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3335 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3337 /* i = current_threshold + 1 */
3341 * Iterate forward over array of thresholds starting from
3342 * current_threshold+1 and check if a threshold is crossed.
3343 * If none of thresholds above usage is crossed, we read
3344 * only one element of the array here.
3346 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3347 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3349 /* Update current_threshold */
3350 t
->current_threshold
= i
- 1;
3355 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3358 __mem_cgroup_threshold(memcg
, false);
3359 if (do_memsw_account())
3360 __mem_cgroup_threshold(memcg
, true);
3362 memcg
= parent_mem_cgroup(memcg
);
3366 static int compare_thresholds(const void *a
, const void *b
)
3368 const struct mem_cgroup_threshold
*_a
= a
;
3369 const struct mem_cgroup_threshold
*_b
= b
;
3371 if (_a
->threshold
> _b
->threshold
)
3374 if (_a
->threshold
< _b
->threshold
)
3380 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3382 struct mem_cgroup_eventfd_list
*ev
;
3384 spin_lock(&memcg_oom_lock
);
3386 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3387 eventfd_signal(ev
->eventfd
, 1);
3389 spin_unlock(&memcg_oom_lock
);
3393 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3395 struct mem_cgroup
*iter
;
3397 for_each_mem_cgroup_tree(iter
, memcg
)
3398 mem_cgroup_oom_notify_cb(iter
);
3401 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3402 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3404 struct mem_cgroup_thresholds
*thresholds
;
3405 struct mem_cgroup_threshold_ary
*new;
3406 unsigned long threshold
;
3407 unsigned long usage
;
3410 ret
= page_counter_memparse(args
, "-1", &threshold
);
3414 mutex_lock(&memcg
->thresholds_lock
);
3417 thresholds
= &memcg
->thresholds
;
3418 usage
= mem_cgroup_usage(memcg
, false);
3419 } else if (type
== _MEMSWAP
) {
3420 thresholds
= &memcg
->memsw_thresholds
;
3421 usage
= mem_cgroup_usage(memcg
, true);
3425 /* Check if a threshold crossed before adding a new one */
3426 if (thresholds
->primary
)
3427 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3429 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3431 /* Allocate memory for new array of thresholds */
3432 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3440 /* Copy thresholds (if any) to new array */
3441 if (thresholds
->primary
) {
3442 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3443 sizeof(struct mem_cgroup_threshold
));
3446 /* Add new threshold */
3447 new->entries
[size
- 1].eventfd
= eventfd
;
3448 new->entries
[size
- 1].threshold
= threshold
;
3450 /* Sort thresholds. Registering of new threshold isn't time-critical */
3451 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3452 compare_thresholds
, NULL
);
3454 /* Find current threshold */
3455 new->current_threshold
= -1;
3456 for (i
= 0; i
< size
; i
++) {
3457 if (new->entries
[i
].threshold
<= usage
) {
3459 * new->current_threshold will not be used until
3460 * rcu_assign_pointer(), so it's safe to increment
3463 ++new->current_threshold
;
3468 /* Free old spare buffer and save old primary buffer as spare */
3469 kfree(thresholds
->spare
);
3470 thresholds
->spare
= thresholds
->primary
;
3472 rcu_assign_pointer(thresholds
->primary
, new);
3474 /* To be sure that nobody uses thresholds */
3478 mutex_unlock(&memcg
->thresholds_lock
);
3483 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3484 struct eventfd_ctx
*eventfd
, const char *args
)
3486 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3489 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3490 struct eventfd_ctx
*eventfd
, const char *args
)
3492 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3495 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3496 struct eventfd_ctx
*eventfd
, enum res_type type
)
3498 struct mem_cgroup_thresholds
*thresholds
;
3499 struct mem_cgroup_threshold_ary
*new;
3500 unsigned long usage
;
3503 mutex_lock(&memcg
->thresholds_lock
);
3506 thresholds
= &memcg
->thresholds
;
3507 usage
= mem_cgroup_usage(memcg
, false);
3508 } else if (type
== _MEMSWAP
) {
3509 thresholds
= &memcg
->memsw_thresholds
;
3510 usage
= mem_cgroup_usage(memcg
, true);
3514 if (!thresholds
->primary
)
3517 /* Check if a threshold crossed before removing */
3518 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3520 /* Calculate new number of threshold */
3522 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3523 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3527 new = thresholds
->spare
;
3529 /* Set thresholds array to NULL if we don't have thresholds */
3538 /* Copy thresholds and find current threshold */
3539 new->current_threshold
= -1;
3540 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3541 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3544 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3545 if (new->entries
[j
].threshold
<= usage
) {
3547 * new->current_threshold will not be used
3548 * until rcu_assign_pointer(), so it's safe to increment
3551 ++new->current_threshold
;
3557 /* Swap primary and spare array */
3558 thresholds
->spare
= thresholds
->primary
;
3560 rcu_assign_pointer(thresholds
->primary
, new);
3562 /* To be sure that nobody uses thresholds */
3565 /* If all events are unregistered, free the spare array */
3567 kfree(thresholds
->spare
);
3568 thresholds
->spare
= NULL
;
3571 mutex_unlock(&memcg
->thresholds_lock
);
3574 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3575 struct eventfd_ctx
*eventfd
)
3577 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3580 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3581 struct eventfd_ctx
*eventfd
)
3583 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3586 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3587 struct eventfd_ctx
*eventfd
, const char *args
)
3589 struct mem_cgroup_eventfd_list
*event
;
3591 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3595 spin_lock(&memcg_oom_lock
);
3597 event
->eventfd
= eventfd
;
3598 list_add(&event
->list
, &memcg
->oom_notify
);
3600 /* already in OOM ? */
3601 if (memcg
->under_oom
)
3602 eventfd_signal(eventfd
, 1);
3603 spin_unlock(&memcg_oom_lock
);
3608 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3609 struct eventfd_ctx
*eventfd
)
3611 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3613 spin_lock(&memcg_oom_lock
);
3615 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3616 if (ev
->eventfd
== eventfd
) {
3617 list_del(&ev
->list
);
3622 spin_unlock(&memcg_oom_lock
);
3625 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3627 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3629 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3630 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3631 seq_printf(sf
, "oom_kill %lu\n", memcg_sum_events(memcg
, OOM_KILL
));
3635 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3636 struct cftype
*cft
, u64 val
)
3638 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3640 /* cannot set to root cgroup and only 0 and 1 are allowed */
3641 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3644 memcg
->oom_kill_disable
= val
;
3646 memcg_oom_recover(memcg
);
3651 #ifdef CONFIG_CGROUP_WRITEBACK
3653 struct list_head
*mem_cgroup_cgwb_list(struct mem_cgroup
*memcg
)
3655 return &memcg
->cgwb_list
;
3658 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3660 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3663 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3665 wb_domain_exit(&memcg
->cgwb_domain
);
3668 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3670 wb_domain_size_changed(&memcg
->cgwb_domain
);
3673 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3675 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3677 if (!memcg
->css
.parent
)
3680 return &memcg
->cgwb_domain
;
3684 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3685 * @wb: bdi_writeback in question
3686 * @pfilepages: out parameter for number of file pages
3687 * @pheadroom: out parameter for number of allocatable pages according to memcg
3688 * @pdirty: out parameter for number of dirty pages
3689 * @pwriteback: out parameter for number of pages under writeback
3691 * Determine the numbers of file, headroom, dirty, and writeback pages in
3692 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3693 * is a bit more involved.
3695 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3696 * headroom is calculated as the lowest headroom of itself and the
3697 * ancestors. Note that this doesn't consider the actual amount of
3698 * available memory in the system. The caller should further cap
3699 * *@pheadroom accordingly.
3701 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
3702 unsigned long *pheadroom
, unsigned long *pdirty
,
3703 unsigned long *pwriteback
)
3705 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3706 struct mem_cgroup
*parent
;
3708 *pdirty
= memcg_page_state(memcg
, NR_FILE_DIRTY
);
3710 /* this should eventually include NR_UNSTABLE_NFS */
3711 *pwriteback
= memcg_page_state(memcg
, NR_WRITEBACK
);
3712 *pfilepages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
3713 (1 << LRU_ACTIVE_FILE
));
3714 *pheadroom
= PAGE_COUNTER_MAX
;
3716 while ((parent
= parent_mem_cgroup(memcg
))) {
3717 unsigned long ceiling
= min(memcg
->memory
.limit
, memcg
->high
);
3718 unsigned long used
= page_counter_read(&memcg
->memory
);
3720 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
3725 #else /* CONFIG_CGROUP_WRITEBACK */
3727 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3732 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3736 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3740 #endif /* CONFIG_CGROUP_WRITEBACK */
3743 * DO NOT USE IN NEW FILES.
3745 * "cgroup.event_control" implementation.
3747 * This is way over-engineered. It tries to support fully configurable
3748 * events for each user. Such level of flexibility is completely
3749 * unnecessary especially in the light of the planned unified hierarchy.
3751 * Please deprecate this and replace with something simpler if at all
3756 * Unregister event and free resources.
3758 * Gets called from workqueue.
3760 static void memcg_event_remove(struct work_struct
*work
)
3762 struct mem_cgroup_event
*event
=
3763 container_of(work
, struct mem_cgroup_event
, remove
);
3764 struct mem_cgroup
*memcg
= event
->memcg
;
3766 remove_wait_queue(event
->wqh
, &event
->wait
);
3768 event
->unregister_event(memcg
, event
->eventfd
);
3770 /* Notify userspace the event is going away. */
3771 eventfd_signal(event
->eventfd
, 1);
3773 eventfd_ctx_put(event
->eventfd
);
3775 css_put(&memcg
->css
);
3779 * Gets called on POLLHUP on eventfd when user closes it.
3781 * Called with wqh->lock held and interrupts disabled.
3783 static int memcg_event_wake(wait_queue_entry_t
*wait
, unsigned mode
,
3784 int sync
, void *key
)
3786 struct mem_cgroup_event
*event
=
3787 container_of(wait
, struct mem_cgroup_event
, wait
);
3788 struct mem_cgroup
*memcg
= event
->memcg
;
3789 unsigned long flags
= (unsigned long)key
;
3791 if (flags
& POLLHUP
) {
3793 * If the event has been detached at cgroup removal, we
3794 * can simply return knowing the other side will cleanup
3797 * We can't race against event freeing since the other
3798 * side will require wqh->lock via remove_wait_queue(),
3801 spin_lock(&memcg
->event_list_lock
);
3802 if (!list_empty(&event
->list
)) {
3803 list_del_init(&event
->list
);
3805 * We are in atomic context, but cgroup_event_remove()
3806 * may sleep, so we have to call it in workqueue.
3808 schedule_work(&event
->remove
);
3810 spin_unlock(&memcg
->event_list_lock
);
3816 static void memcg_event_ptable_queue_proc(struct file
*file
,
3817 wait_queue_head_t
*wqh
, poll_table
*pt
)
3819 struct mem_cgroup_event
*event
=
3820 container_of(pt
, struct mem_cgroup_event
, pt
);
3823 add_wait_queue(wqh
, &event
->wait
);
3827 * DO NOT USE IN NEW FILES.
3829 * Parse input and register new cgroup event handler.
3831 * Input must be in format '<event_fd> <control_fd> <args>'.
3832 * Interpretation of args is defined by control file implementation.
3834 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
3835 char *buf
, size_t nbytes
, loff_t off
)
3837 struct cgroup_subsys_state
*css
= of_css(of
);
3838 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3839 struct mem_cgroup_event
*event
;
3840 struct cgroup_subsys_state
*cfile_css
;
3841 unsigned int efd
, cfd
;
3848 buf
= strstrip(buf
);
3850 efd
= simple_strtoul(buf
, &endp
, 10);
3855 cfd
= simple_strtoul(buf
, &endp
, 10);
3856 if ((*endp
!= ' ') && (*endp
!= '\0'))
3860 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
3864 event
->memcg
= memcg
;
3865 INIT_LIST_HEAD(&event
->list
);
3866 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
3867 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
3868 INIT_WORK(&event
->remove
, memcg_event_remove
);
3876 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
3877 if (IS_ERR(event
->eventfd
)) {
3878 ret
= PTR_ERR(event
->eventfd
);
3885 goto out_put_eventfd
;
3888 /* the process need read permission on control file */
3889 /* AV: shouldn't we check that it's been opened for read instead? */
3890 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
3895 * Determine the event callbacks and set them in @event. This used
3896 * to be done via struct cftype but cgroup core no longer knows
3897 * about these events. The following is crude but the whole thing
3898 * is for compatibility anyway.
3900 * DO NOT ADD NEW FILES.
3902 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
3904 if (!strcmp(name
, "memory.usage_in_bytes")) {
3905 event
->register_event
= mem_cgroup_usage_register_event
;
3906 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
3907 } else if (!strcmp(name
, "memory.oom_control")) {
3908 event
->register_event
= mem_cgroup_oom_register_event
;
3909 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
3910 } else if (!strcmp(name
, "memory.pressure_level")) {
3911 event
->register_event
= vmpressure_register_event
;
3912 event
->unregister_event
= vmpressure_unregister_event
;
3913 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
3914 event
->register_event
= memsw_cgroup_usage_register_event
;
3915 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
3922 * Verify @cfile should belong to @css. Also, remaining events are
3923 * automatically removed on cgroup destruction but the removal is
3924 * asynchronous, so take an extra ref on @css.
3926 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
3927 &memory_cgrp_subsys
);
3929 if (IS_ERR(cfile_css
))
3931 if (cfile_css
!= css
) {
3936 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
3940 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
3942 spin_lock(&memcg
->event_list_lock
);
3943 list_add(&event
->list
, &memcg
->event_list
);
3944 spin_unlock(&memcg
->event_list_lock
);
3956 eventfd_ctx_put(event
->eventfd
);
3965 static struct cftype mem_cgroup_legacy_files
[] = {
3967 .name
= "usage_in_bytes",
3968 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
3969 .read_u64
= mem_cgroup_read_u64
,
3972 .name
= "max_usage_in_bytes",
3973 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
3974 .write
= mem_cgroup_reset
,
3975 .read_u64
= mem_cgroup_read_u64
,
3978 .name
= "limit_in_bytes",
3979 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
3980 .write
= mem_cgroup_write
,
3981 .read_u64
= mem_cgroup_read_u64
,
3984 .name
= "soft_limit_in_bytes",
3985 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
3986 .write
= mem_cgroup_write
,
3987 .read_u64
= mem_cgroup_read_u64
,
3991 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
3992 .write
= mem_cgroup_reset
,
3993 .read_u64
= mem_cgroup_read_u64
,
3997 .seq_show
= memcg_stat_show
,
4000 .name
= "force_empty",
4001 .write
= mem_cgroup_force_empty_write
,
4004 .name
= "use_hierarchy",
4005 .write_u64
= mem_cgroup_hierarchy_write
,
4006 .read_u64
= mem_cgroup_hierarchy_read
,
4009 .name
= "cgroup.event_control", /* XXX: for compat */
4010 .write
= memcg_write_event_control
,
4011 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
4014 .name
= "swappiness",
4015 .read_u64
= mem_cgroup_swappiness_read
,
4016 .write_u64
= mem_cgroup_swappiness_write
,
4019 .name
= "move_charge_at_immigrate",
4020 .read_u64
= mem_cgroup_move_charge_read
,
4021 .write_u64
= mem_cgroup_move_charge_write
,
4024 .name
= "oom_control",
4025 .seq_show
= mem_cgroup_oom_control_read
,
4026 .write_u64
= mem_cgroup_oom_control_write
,
4027 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4030 .name
= "pressure_level",
4034 .name
= "numa_stat",
4035 .seq_show
= memcg_numa_stat_show
,
4039 .name
= "kmem.limit_in_bytes",
4040 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4041 .write
= mem_cgroup_write
,
4042 .read_u64
= mem_cgroup_read_u64
,
4045 .name
= "kmem.usage_in_bytes",
4046 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4047 .read_u64
= mem_cgroup_read_u64
,
4050 .name
= "kmem.failcnt",
4051 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4052 .write
= mem_cgroup_reset
,
4053 .read_u64
= mem_cgroup_read_u64
,
4056 .name
= "kmem.max_usage_in_bytes",
4057 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4058 .write
= mem_cgroup_reset
,
4059 .read_u64
= mem_cgroup_read_u64
,
4061 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4063 .name
= "kmem.slabinfo",
4064 .seq_start
= memcg_slab_start
,
4065 .seq_next
= memcg_slab_next
,
4066 .seq_stop
= memcg_slab_stop
,
4067 .seq_show
= memcg_slab_show
,
4071 .name
= "kmem.tcp.limit_in_bytes",
4072 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4073 .write
= mem_cgroup_write
,
4074 .read_u64
= mem_cgroup_read_u64
,
4077 .name
= "kmem.tcp.usage_in_bytes",
4078 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4079 .read_u64
= mem_cgroup_read_u64
,
4082 .name
= "kmem.tcp.failcnt",
4083 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4084 .write
= mem_cgroup_reset
,
4085 .read_u64
= mem_cgroup_read_u64
,
4088 .name
= "kmem.tcp.max_usage_in_bytes",
4089 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4090 .write
= mem_cgroup_reset
,
4091 .read_u64
= mem_cgroup_read_u64
,
4093 { }, /* terminate */
4097 * Private memory cgroup IDR
4099 * Swap-out records and page cache shadow entries need to store memcg
4100 * references in constrained space, so we maintain an ID space that is
4101 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4102 * memory-controlled cgroups to 64k.
4104 * However, there usually are many references to the oflline CSS after
4105 * the cgroup has been destroyed, such as page cache or reclaimable
4106 * slab objects, that don't need to hang on to the ID. We want to keep
4107 * those dead CSS from occupying IDs, or we might quickly exhaust the
4108 * relatively small ID space and prevent the creation of new cgroups
4109 * even when there are much fewer than 64k cgroups - possibly none.
4111 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4112 * be freed and recycled when it's no longer needed, which is usually
4113 * when the CSS is offlined.
4115 * The only exception to that are records of swapped out tmpfs/shmem
4116 * pages that need to be attributed to live ancestors on swapin. But
4117 * those references are manageable from userspace.
4120 static DEFINE_IDR(mem_cgroup_idr
);
4122 static void mem_cgroup_id_remove(struct mem_cgroup
*memcg
)
4124 if (memcg
->id
.id
> 0) {
4125 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4130 static void mem_cgroup_id_get_many(struct mem_cgroup
*memcg
, unsigned int n
)
4132 VM_BUG_ON(atomic_read(&memcg
->id
.ref
) <= 0);
4133 atomic_add(n
, &memcg
->id
.ref
);
4136 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
4138 VM_BUG_ON(atomic_read(&memcg
->id
.ref
) < n
);
4139 if (atomic_sub_and_test(n
, &memcg
->id
.ref
)) {
4140 mem_cgroup_id_remove(memcg
);
4142 /* Memcg ID pins CSS */
4143 css_put(&memcg
->css
);
4147 static inline void mem_cgroup_id_get(struct mem_cgroup
*memcg
)
4149 mem_cgroup_id_get_many(memcg
, 1);
4152 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
4154 mem_cgroup_id_put_many(memcg
, 1);
4158 * mem_cgroup_from_id - look up a memcg from a memcg id
4159 * @id: the memcg id to look up
4161 * Caller must hold rcu_read_lock().
4163 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
4165 WARN_ON_ONCE(!rcu_read_lock_held());
4166 return idr_find(&mem_cgroup_idr
, id
);
4169 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4171 struct mem_cgroup_per_node
*pn
;
4174 * This routine is called against possible nodes.
4175 * But it's BUG to call kmalloc() against offline node.
4177 * TODO: this routine can waste much memory for nodes which will
4178 * never be onlined. It's better to use memory hotplug callback
4181 if (!node_state(node
, N_NORMAL_MEMORY
))
4183 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4187 pn
->lruvec_stat
= alloc_percpu(struct lruvec_stat
);
4188 if (!pn
->lruvec_stat
) {
4193 lruvec_init(&pn
->lruvec
);
4194 pn
->usage_in_excess
= 0;
4195 pn
->on_tree
= false;
4198 memcg
->nodeinfo
[node
] = pn
;
4202 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4204 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
4209 free_percpu(pn
->lruvec_stat
);
4213 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
4218 free_mem_cgroup_per_node_info(memcg
, node
);
4219 free_percpu(memcg
->stat
);
4223 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4225 memcg_wb_domain_exit(memcg
);
4226 __mem_cgroup_free(memcg
);
4229 static struct mem_cgroup
*mem_cgroup_alloc(void)
4231 struct mem_cgroup
*memcg
;
4235 size
= sizeof(struct mem_cgroup
);
4236 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4238 memcg
= kzalloc(size
, GFP_KERNEL
);
4242 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
4243 1, MEM_CGROUP_ID_MAX
,
4245 if (memcg
->id
.id
< 0)
4248 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4253 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
4256 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4259 INIT_WORK(&memcg
->high_work
, high_work_func
);
4260 memcg
->last_scanned_node
= MAX_NUMNODES
;
4261 INIT_LIST_HEAD(&memcg
->oom_notify
);
4262 mutex_init(&memcg
->thresholds_lock
);
4263 spin_lock_init(&memcg
->move_lock
);
4264 vmpressure_init(&memcg
->vmpressure
);
4265 INIT_LIST_HEAD(&memcg
->event_list
);
4266 spin_lock_init(&memcg
->event_list_lock
);
4267 memcg
->socket_pressure
= jiffies
;
4269 memcg
->kmemcg_id
= -1;
4271 #ifdef CONFIG_CGROUP_WRITEBACK
4272 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4274 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
4277 mem_cgroup_id_remove(memcg
);
4278 __mem_cgroup_free(memcg
);
4282 static struct cgroup_subsys_state
* __ref
4283 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4285 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
4286 struct mem_cgroup
*memcg
;
4287 long error
= -ENOMEM
;
4289 memcg
= mem_cgroup_alloc();
4291 return ERR_PTR(error
);
4293 memcg
->high
= PAGE_COUNTER_MAX
;
4294 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4296 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4297 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4299 if (parent
&& parent
->use_hierarchy
) {
4300 memcg
->use_hierarchy
= true;
4301 page_counter_init(&memcg
->memory
, &parent
->memory
);
4302 page_counter_init(&memcg
->swap
, &parent
->swap
);
4303 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4304 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4305 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
4307 page_counter_init(&memcg
->memory
, NULL
);
4308 page_counter_init(&memcg
->swap
, NULL
);
4309 page_counter_init(&memcg
->memsw
, NULL
);
4310 page_counter_init(&memcg
->kmem
, NULL
);
4311 page_counter_init(&memcg
->tcpmem
, NULL
);
4313 * Deeper hierachy with use_hierarchy == false doesn't make
4314 * much sense so let cgroup subsystem know about this
4315 * unfortunate state in our controller.
4317 if (parent
!= root_mem_cgroup
)
4318 memory_cgrp_subsys
.broken_hierarchy
= true;
4321 /* The following stuff does not apply to the root */
4323 root_mem_cgroup
= memcg
;
4327 error
= memcg_online_kmem(memcg
);
4331 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4332 static_branch_inc(&memcg_sockets_enabled_key
);
4336 mem_cgroup_id_remove(memcg
);
4337 mem_cgroup_free(memcg
);
4338 return ERR_PTR(-ENOMEM
);
4341 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4343 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4345 /* Online state pins memcg ID, memcg ID pins CSS */
4346 atomic_set(&memcg
->id
.ref
, 1);
4351 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4353 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4354 struct mem_cgroup_event
*event
, *tmp
;
4357 * Unregister events and notify userspace.
4358 * Notify userspace about cgroup removing only after rmdir of cgroup
4359 * directory to avoid race between userspace and kernelspace.
4361 spin_lock(&memcg
->event_list_lock
);
4362 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4363 list_del_init(&event
->list
);
4364 schedule_work(&event
->remove
);
4366 spin_unlock(&memcg
->event_list_lock
);
4370 memcg_offline_kmem(memcg
);
4371 wb_memcg_offline(memcg
);
4373 mem_cgroup_id_put(memcg
);
4376 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4378 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4380 invalidate_reclaim_iterators(memcg
);
4383 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4385 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4387 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4388 static_branch_dec(&memcg_sockets_enabled_key
);
4390 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
4391 static_branch_dec(&memcg_sockets_enabled_key
);
4393 vmpressure_cleanup(&memcg
->vmpressure
);
4394 cancel_work_sync(&memcg
->high_work
);
4395 mem_cgroup_remove_from_trees(memcg
);
4396 memcg_free_kmem(memcg
);
4397 mem_cgroup_free(memcg
);
4401 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4402 * @css: the target css
4404 * Reset the states of the mem_cgroup associated with @css. This is
4405 * invoked when the userland requests disabling on the default hierarchy
4406 * but the memcg is pinned through dependency. The memcg should stop
4407 * applying policies and should revert to the vanilla state as it may be
4408 * made visible again.
4410 * The current implementation only resets the essential configurations.
4411 * This needs to be expanded to cover all the visible parts.
4413 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4415 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4417 page_counter_limit(&memcg
->memory
, PAGE_COUNTER_MAX
);
4418 page_counter_limit(&memcg
->swap
, PAGE_COUNTER_MAX
);
4419 page_counter_limit(&memcg
->memsw
, PAGE_COUNTER_MAX
);
4420 page_counter_limit(&memcg
->kmem
, PAGE_COUNTER_MAX
);
4421 page_counter_limit(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
4423 memcg
->high
= PAGE_COUNTER_MAX
;
4424 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4425 memcg_wb_domain_size_changed(memcg
);
4429 /* Handlers for move charge at task migration. */
4430 static int mem_cgroup_do_precharge(unsigned long count
)
4434 /* Try a single bulk charge without reclaim first, kswapd may wake */
4435 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4437 mc
.precharge
+= count
;
4441 /* Try charges one by one with reclaim, but do not retry */
4443 ret
= try_charge(mc
.to
, GFP_KERNEL
| __GFP_NORETRY
, 1);
4457 enum mc_target_type
{
4464 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4465 unsigned long addr
, pte_t ptent
)
4467 struct page
*page
= _vm_normal_page(vma
, addr
, ptent
, true);
4469 if (!page
|| !page_mapped(page
))
4471 if (PageAnon(page
)) {
4472 if (!(mc
.flags
& MOVE_ANON
))
4475 if (!(mc
.flags
& MOVE_FILE
))
4478 if (!get_page_unless_zero(page
))
4484 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
4485 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4486 pte_t ptent
, swp_entry_t
*entry
)
4488 struct page
*page
= NULL
;
4489 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4491 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4495 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
4496 * a device and because they are not accessible by CPU they are store
4497 * as special swap entry in the CPU page table.
4499 if (is_device_private_entry(ent
)) {
4500 page
= device_private_entry_to_page(ent
);
4502 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
4503 * a refcount of 1 when free (unlike normal page)
4505 if (!page_ref_add_unless(page
, 1, 1))
4511 * Because lookup_swap_cache() updates some statistics counter,
4512 * we call find_get_page() with swapper_space directly.
4514 page
= find_get_page(swap_address_space(ent
), swp_offset(ent
));
4515 if (do_memsw_account())
4516 entry
->val
= ent
.val
;
4521 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4522 pte_t ptent
, swp_entry_t
*entry
)
4528 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4529 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4531 struct page
*page
= NULL
;
4532 struct address_space
*mapping
;
4535 if (!vma
->vm_file
) /* anonymous vma */
4537 if (!(mc
.flags
& MOVE_FILE
))
4540 mapping
= vma
->vm_file
->f_mapping
;
4541 pgoff
= linear_page_index(vma
, addr
);
4543 /* page is moved even if it's not RSS of this task(page-faulted). */
4545 /* shmem/tmpfs may report page out on swap: account for that too. */
4546 if (shmem_mapping(mapping
)) {
4547 page
= find_get_entry(mapping
, pgoff
);
4548 if (radix_tree_exceptional_entry(page
)) {
4549 swp_entry_t swp
= radix_to_swp_entry(page
);
4550 if (do_memsw_account())
4552 page
= find_get_page(swap_address_space(swp
),
4556 page
= find_get_page(mapping
, pgoff
);
4558 page
= find_get_page(mapping
, pgoff
);
4564 * mem_cgroup_move_account - move account of the page
4566 * @compound: charge the page as compound or small page
4567 * @from: mem_cgroup which the page is moved from.
4568 * @to: mem_cgroup which the page is moved to. @from != @to.
4570 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4572 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4575 static int mem_cgroup_move_account(struct page
*page
,
4577 struct mem_cgroup
*from
,
4578 struct mem_cgroup
*to
)
4580 unsigned long flags
;
4581 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
4585 VM_BUG_ON(from
== to
);
4586 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4587 VM_BUG_ON(compound
&& !PageTransHuge(page
));
4590 * Prevent mem_cgroup_migrate() from looking at
4591 * page->mem_cgroup of its source page while we change it.
4594 if (!trylock_page(page
))
4598 if (page
->mem_cgroup
!= from
)
4601 anon
= PageAnon(page
);
4603 spin_lock_irqsave(&from
->move_lock
, flags
);
4605 if (!anon
&& page_mapped(page
)) {
4606 __this_cpu_sub(from
->stat
->count
[NR_FILE_MAPPED
], nr_pages
);
4607 __this_cpu_add(to
->stat
->count
[NR_FILE_MAPPED
], nr_pages
);
4611 * move_lock grabbed above and caller set from->moving_account, so
4612 * mod_memcg_page_state will serialize updates to PageDirty.
4613 * So mapping should be stable for dirty pages.
4615 if (!anon
&& PageDirty(page
)) {
4616 struct address_space
*mapping
= page_mapping(page
);
4618 if (mapping_cap_account_dirty(mapping
)) {
4619 __this_cpu_sub(from
->stat
->count
[NR_FILE_DIRTY
],
4621 __this_cpu_add(to
->stat
->count
[NR_FILE_DIRTY
],
4626 if (PageWriteback(page
)) {
4627 __this_cpu_sub(from
->stat
->count
[NR_WRITEBACK
], nr_pages
);
4628 __this_cpu_add(to
->stat
->count
[NR_WRITEBACK
], nr_pages
);
4632 * It is safe to change page->mem_cgroup here because the page
4633 * is referenced, charged, and isolated - we can't race with
4634 * uncharging, charging, migration, or LRU putback.
4637 /* caller should have done css_get */
4638 page
->mem_cgroup
= to
;
4639 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4643 local_irq_disable();
4644 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
4645 memcg_check_events(to
, page
);
4646 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
4647 memcg_check_events(from
, page
);
4656 * get_mctgt_type - get target type of moving charge
4657 * @vma: the vma the pte to be checked belongs
4658 * @addr: the address corresponding to the pte to be checked
4659 * @ptent: the pte to be checked
4660 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4663 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4664 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4665 * move charge. if @target is not NULL, the page is stored in target->page
4666 * with extra refcnt got(Callers should handle it).
4667 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4668 * target for charge migration. if @target is not NULL, the entry is stored
4670 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PUBLIC
4671 * or MEMORY_DEVICE_PRIVATE (so ZONE_DEVICE page and thus not on the lru).
4672 * For now we such page is charge like a regular page would be as for all
4673 * intent and purposes it is just special memory taking the place of a
4676 * See Documentations/vm/hmm.txt and include/linux/hmm.h
4678 * Called with pte lock held.
4681 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4682 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4684 struct page
*page
= NULL
;
4685 enum mc_target_type ret
= MC_TARGET_NONE
;
4686 swp_entry_t ent
= { .val
= 0 };
4688 if (pte_present(ptent
))
4689 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4690 else if (is_swap_pte(ptent
))
4691 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
4692 else if (pte_none(ptent
))
4693 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4695 if (!page
&& !ent
.val
)
4699 * Do only loose check w/o serialization.
4700 * mem_cgroup_move_account() checks the page is valid or
4701 * not under LRU exclusion.
4703 if (page
->mem_cgroup
== mc
.from
) {
4704 ret
= MC_TARGET_PAGE
;
4705 if (is_device_private_page(page
) ||
4706 is_device_public_page(page
))
4707 ret
= MC_TARGET_DEVICE
;
4709 target
->page
= page
;
4711 if (!ret
|| !target
)
4715 * There is a swap entry and a page doesn't exist or isn't charged.
4716 * But we cannot move a tail-page in a THP.
4718 if (ent
.val
&& !ret
&& (!page
|| !PageTransCompound(page
)) &&
4719 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4720 ret
= MC_TARGET_SWAP
;
4727 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4729 * We don't consider PMD mapped swapping or file mapped pages because THP does
4730 * not support them for now.
4731 * Caller should make sure that pmd_trans_huge(pmd) is true.
4733 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4734 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4736 struct page
*page
= NULL
;
4737 enum mc_target_type ret
= MC_TARGET_NONE
;
4739 if (unlikely(is_swap_pmd(pmd
))) {
4740 VM_BUG_ON(thp_migration_supported() &&
4741 !is_pmd_migration_entry(pmd
));
4744 page
= pmd_page(pmd
);
4745 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4746 if (!(mc
.flags
& MOVE_ANON
))
4748 if (page
->mem_cgroup
== mc
.from
) {
4749 ret
= MC_TARGET_PAGE
;
4752 target
->page
= page
;
4758 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4759 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4761 return MC_TARGET_NONE
;
4765 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4766 unsigned long addr
, unsigned long end
,
4767 struct mm_walk
*walk
)
4769 struct vm_area_struct
*vma
= walk
->vma
;
4773 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4776 * Note their can not be MC_TARGET_DEVICE for now as we do not
4777 * support transparent huge page with MEMORY_DEVICE_PUBLIC or
4778 * MEMORY_DEVICE_PRIVATE but this might change.
4780 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4781 mc
.precharge
+= HPAGE_PMD_NR
;
4786 if (pmd_trans_unstable(pmd
))
4788 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4789 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4790 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4791 mc
.precharge
++; /* increment precharge temporarily */
4792 pte_unmap_unlock(pte
- 1, ptl
);
4798 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4800 unsigned long precharge
;
4802 struct mm_walk mem_cgroup_count_precharge_walk
= {
4803 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4806 down_read(&mm
->mmap_sem
);
4807 walk_page_range(0, mm
->highest_vm_end
,
4808 &mem_cgroup_count_precharge_walk
);
4809 up_read(&mm
->mmap_sem
);
4811 precharge
= mc
.precharge
;
4817 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4819 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4821 VM_BUG_ON(mc
.moving_task
);
4822 mc
.moving_task
= current
;
4823 return mem_cgroup_do_precharge(precharge
);
4826 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4827 static void __mem_cgroup_clear_mc(void)
4829 struct mem_cgroup
*from
= mc
.from
;
4830 struct mem_cgroup
*to
= mc
.to
;
4832 /* we must uncharge all the leftover precharges from mc.to */
4834 cancel_charge(mc
.to
, mc
.precharge
);
4838 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4839 * we must uncharge here.
4841 if (mc
.moved_charge
) {
4842 cancel_charge(mc
.from
, mc
.moved_charge
);
4843 mc
.moved_charge
= 0;
4845 /* we must fixup refcnts and charges */
4846 if (mc
.moved_swap
) {
4847 /* uncharge swap account from the old cgroup */
4848 if (!mem_cgroup_is_root(mc
.from
))
4849 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4851 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
4854 * we charged both to->memory and to->memsw, so we
4855 * should uncharge to->memory.
4857 if (!mem_cgroup_is_root(mc
.to
))
4858 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4860 mem_cgroup_id_get_many(mc
.to
, mc
.moved_swap
);
4861 css_put_many(&mc
.to
->css
, mc
.moved_swap
);
4865 memcg_oom_recover(from
);
4866 memcg_oom_recover(to
);
4867 wake_up_all(&mc
.waitq
);
4870 static void mem_cgroup_clear_mc(void)
4872 struct mm_struct
*mm
= mc
.mm
;
4875 * we must clear moving_task before waking up waiters at the end of
4878 mc
.moving_task
= NULL
;
4879 __mem_cgroup_clear_mc();
4880 spin_lock(&mc
.lock
);
4884 spin_unlock(&mc
.lock
);
4889 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4891 struct cgroup_subsys_state
*css
;
4892 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
4893 struct mem_cgroup
*from
;
4894 struct task_struct
*leader
, *p
;
4895 struct mm_struct
*mm
;
4896 unsigned long move_flags
;
4899 /* charge immigration isn't supported on the default hierarchy */
4900 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4904 * Multi-process migrations only happen on the default hierarchy
4905 * where charge immigration is not used. Perform charge
4906 * immigration if @tset contains a leader and whine if there are
4910 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
4913 memcg
= mem_cgroup_from_css(css
);
4919 * We are now commited to this value whatever it is. Changes in this
4920 * tunable will only affect upcoming migrations, not the current one.
4921 * So we need to save it, and keep it going.
4923 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
4927 from
= mem_cgroup_from_task(p
);
4929 VM_BUG_ON(from
== memcg
);
4931 mm
= get_task_mm(p
);
4934 /* We move charges only when we move a owner of the mm */
4935 if (mm
->owner
== p
) {
4938 VM_BUG_ON(mc
.precharge
);
4939 VM_BUG_ON(mc
.moved_charge
);
4940 VM_BUG_ON(mc
.moved_swap
);
4942 spin_lock(&mc
.lock
);
4946 mc
.flags
= move_flags
;
4947 spin_unlock(&mc
.lock
);
4948 /* We set mc.moving_task later */
4950 ret
= mem_cgroup_precharge_mc(mm
);
4952 mem_cgroup_clear_mc();
4959 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4962 mem_cgroup_clear_mc();
4965 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4966 unsigned long addr
, unsigned long end
,
4967 struct mm_walk
*walk
)
4970 struct vm_area_struct
*vma
= walk
->vma
;
4973 enum mc_target_type target_type
;
4974 union mc_target target
;
4977 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4979 if (mc
.precharge
< HPAGE_PMD_NR
) {
4983 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
4984 if (target_type
== MC_TARGET_PAGE
) {
4986 if (!isolate_lru_page(page
)) {
4987 if (!mem_cgroup_move_account(page
, true,
4989 mc
.precharge
-= HPAGE_PMD_NR
;
4990 mc
.moved_charge
+= HPAGE_PMD_NR
;
4992 putback_lru_page(page
);
4995 } else if (target_type
== MC_TARGET_DEVICE
) {
4997 if (!mem_cgroup_move_account(page
, true,
4999 mc
.precharge
-= HPAGE_PMD_NR
;
5000 mc
.moved_charge
+= HPAGE_PMD_NR
;
5008 if (pmd_trans_unstable(pmd
))
5011 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5012 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
5013 pte_t ptent
= *(pte
++);
5014 bool device
= false;
5020 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
5021 case MC_TARGET_DEVICE
:
5024 case MC_TARGET_PAGE
:
5027 * We can have a part of the split pmd here. Moving it
5028 * can be done but it would be too convoluted so simply
5029 * ignore such a partial THP and keep it in original
5030 * memcg. There should be somebody mapping the head.
5032 if (PageTransCompound(page
))
5034 if (!device
&& isolate_lru_page(page
))
5036 if (!mem_cgroup_move_account(page
, false,
5039 /* we uncharge from mc.from later. */
5043 putback_lru_page(page
);
5044 put
: /* get_mctgt_type() gets the page */
5047 case MC_TARGET_SWAP
:
5049 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
5051 /* we fixup refcnts and charges later. */
5059 pte_unmap_unlock(pte
- 1, ptl
);
5064 * We have consumed all precharges we got in can_attach().
5065 * We try charge one by one, but don't do any additional
5066 * charges to mc.to if we have failed in charge once in attach()
5069 ret
= mem_cgroup_do_precharge(1);
5077 static void mem_cgroup_move_charge(void)
5079 struct mm_walk mem_cgroup_move_charge_walk
= {
5080 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
5084 lru_add_drain_all();
5086 * Signal lock_page_memcg() to take the memcg's move_lock
5087 * while we're moving its pages to another memcg. Then wait
5088 * for already started RCU-only updates to finish.
5090 atomic_inc(&mc
.from
->moving_account
);
5093 if (unlikely(!down_read_trylock(&mc
.mm
->mmap_sem
))) {
5095 * Someone who are holding the mmap_sem might be waiting in
5096 * waitq. So we cancel all extra charges, wake up all waiters,
5097 * and retry. Because we cancel precharges, we might not be able
5098 * to move enough charges, but moving charge is a best-effort
5099 * feature anyway, so it wouldn't be a big problem.
5101 __mem_cgroup_clear_mc();
5106 * When we have consumed all precharges and failed in doing
5107 * additional charge, the page walk just aborts.
5109 walk_page_range(0, mc
.mm
->highest_vm_end
, &mem_cgroup_move_charge_walk
);
5111 up_read(&mc
.mm
->mmap_sem
);
5112 atomic_dec(&mc
.from
->moving_account
);
5115 static void mem_cgroup_move_task(void)
5118 mem_cgroup_move_charge();
5119 mem_cgroup_clear_mc();
5122 #else /* !CONFIG_MMU */
5123 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5127 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5130 static void mem_cgroup_move_task(void)
5136 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5137 * to verify whether we're attached to the default hierarchy on each mount
5140 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5143 * use_hierarchy is forced on the default hierarchy. cgroup core
5144 * guarantees that @root doesn't have any children, so turning it
5145 * on for the root memcg is enough.
5147 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5148 root_mem_cgroup
->use_hierarchy
= true;
5150 root_mem_cgroup
->use_hierarchy
= false;
5153 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5156 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5158 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
5161 static int memory_low_show(struct seq_file
*m
, void *v
)
5163 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5164 unsigned long low
= READ_ONCE(memcg
->low
);
5166 if (low
== PAGE_COUNTER_MAX
)
5167 seq_puts(m
, "max\n");
5169 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
5174 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5175 char *buf
, size_t nbytes
, loff_t off
)
5177 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5181 buf
= strstrip(buf
);
5182 err
= page_counter_memparse(buf
, "max", &low
);
5191 static int memory_high_show(struct seq_file
*m
, void *v
)
5193 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5194 unsigned long high
= READ_ONCE(memcg
->high
);
5196 if (high
== PAGE_COUNTER_MAX
)
5197 seq_puts(m
, "max\n");
5199 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
5204 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5205 char *buf
, size_t nbytes
, loff_t off
)
5207 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5208 unsigned long nr_pages
;
5212 buf
= strstrip(buf
);
5213 err
= page_counter_memparse(buf
, "max", &high
);
5219 nr_pages
= page_counter_read(&memcg
->memory
);
5220 if (nr_pages
> high
)
5221 try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
5224 memcg_wb_domain_size_changed(memcg
);
5228 static int memory_max_show(struct seq_file
*m
, void *v
)
5230 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5231 unsigned long max
= READ_ONCE(memcg
->memory
.limit
);
5233 if (max
== PAGE_COUNTER_MAX
)
5234 seq_puts(m
, "max\n");
5236 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5241 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5242 char *buf
, size_t nbytes
, loff_t off
)
5244 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5245 unsigned int nr_reclaims
= MEM_CGROUP_RECLAIM_RETRIES
;
5246 bool drained
= false;
5250 buf
= strstrip(buf
);
5251 err
= page_counter_memparse(buf
, "max", &max
);
5255 xchg(&memcg
->memory
.limit
, max
);
5258 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
5260 if (nr_pages
<= max
)
5263 if (signal_pending(current
)) {
5269 drain_all_stock(memcg
);
5275 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
5281 mem_cgroup_event(memcg
, MEMCG_OOM
);
5282 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
5286 memcg_wb_domain_size_changed(memcg
);
5290 static int memory_events_show(struct seq_file
*m
, void *v
)
5292 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5294 seq_printf(m
, "low %lu\n", memcg_sum_events(memcg
, MEMCG_LOW
));
5295 seq_printf(m
, "high %lu\n", memcg_sum_events(memcg
, MEMCG_HIGH
));
5296 seq_printf(m
, "max %lu\n", memcg_sum_events(memcg
, MEMCG_MAX
));
5297 seq_printf(m
, "oom %lu\n", memcg_sum_events(memcg
, MEMCG_OOM
));
5298 seq_printf(m
, "oom_kill %lu\n", memcg_sum_events(memcg
, OOM_KILL
));
5303 static int memory_stat_show(struct seq_file
*m
, void *v
)
5305 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5306 unsigned long stat
[MEMCG_NR_STAT
];
5307 unsigned long events
[MEMCG_NR_EVENTS
];
5311 * Provide statistics on the state of the memory subsystem as
5312 * well as cumulative event counters that show past behavior.
5314 * This list is ordered following a combination of these gradients:
5315 * 1) generic big picture -> specifics and details
5316 * 2) reflecting userspace activity -> reflecting kernel heuristics
5318 * Current memory state:
5321 tree_stat(memcg
, stat
);
5322 tree_events(memcg
, events
);
5324 seq_printf(m
, "anon %llu\n",
5325 (u64
)stat
[MEMCG_RSS
] * PAGE_SIZE
);
5326 seq_printf(m
, "file %llu\n",
5327 (u64
)stat
[MEMCG_CACHE
] * PAGE_SIZE
);
5328 seq_printf(m
, "kernel_stack %llu\n",
5329 (u64
)stat
[MEMCG_KERNEL_STACK_KB
] * 1024);
5330 seq_printf(m
, "slab %llu\n",
5331 (u64
)(stat
[NR_SLAB_RECLAIMABLE
] +
5332 stat
[NR_SLAB_UNRECLAIMABLE
]) * PAGE_SIZE
);
5333 seq_printf(m
, "sock %llu\n",
5334 (u64
)stat
[MEMCG_SOCK
] * PAGE_SIZE
);
5336 seq_printf(m
, "shmem %llu\n",
5337 (u64
)stat
[NR_SHMEM
] * PAGE_SIZE
);
5338 seq_printf(m
, "file_mapped %llu\n",
5339 (u64
)stat
[NR_FILE_MAPPED
] * PAGE_SIZE
);
5340 seq_printf(m
, "file_dirty %llu\n",
5341 (u64
)stat
[NR_FILE_DIRTY
] * PAGE_SIZE
);
5342 seq_printf(m
, "file_writeback %llu\n",
5343 (u64
)stat
[NR_WRITEBACK
] * PAGE_SIZE
);
5345 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5346 struct mem_cgroup
*mi
;
5347 unsigned long val
= 0;
5349 for_each_mem_cgroup_tree(mi
, memcg
)
5350 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
));
5351 seq_printf(m
, "%s %llu\n",
5352 mem_cgroup_lru_names
[i
], (u64
)val
* PAGE_SIZE
);
5355 seq_printf(m
, "slab_reclaimable %llu\n",
5356 (u64
)stat
[NR_SLAB_RECLAIMABLE
] * PAGE_SIZE
);
5357 seq_printf(m
, "slab_unreclaimable %llu\n",
5358 (u64
)stat
[NR_SLAB_UNRECLAIMABLE
] * PAGE_SIZE
);
5360 /* Accumulated memory events */
5362 seq_printf(m
, "pgfault %lu\n", events
[PGFAULT
]);
5363 seq_printf(m
, "pgmajfault %lu\n", events
[PGMAJFAULT
]);
5365 seq_printf(m
, "pgrefill %lu\n", events
[PGREFILL
]);
5366 seq_printf(m
, "pgscan %lu\n", events
[PGSCAN_KSWAPD
] +
5367 events
[PGSCAN_DIRECT
]);
5368 seq_printf(m
, "pgsteal %lu\n", events
[PGSTEAL_KSWAPD
] +
5369 events
[PGSTEAL_DIRECT
]);
5370 seq_printf(m
, "pgactivate %lu\n", events
[PGACTIVATE
]);
5371 seq_printf(m
, "pgdeactivate %lu\n", events
[PGDEACTIVATE
]);
5372 seq_printf(m
, "pglazyfree %lu\n", events
[PGLAZYFREE
]);
5373 seq_printf(m
, "pglazyfreed %lu\n", events
[PGLAZYFREED
]);
5375 seq_printf(m
, "workingset_refault %lu\n",
5376 stat
[WORKINGSET_REFAULT
]);
5377 seq_printf(m
, "workingset_activate %lu\n",
5378 stat
[WORKINGSET_ACTIVATE
]);
5379 seq_printf(m
, "workingset_nodereclaim %lu\n",
5380 stat
[WORKINGSET_NODERECLAIM
]);
5385 static struct cftype memory_files
[] = {
5388 .flags
= CFTYPE_NOT_ON_ROOT
,
5389 .read_u64
= memory_current_read
,
5393 .flags
= CFTYPE_NOT_ON_ROOT
,
5394 .seq_show
= memory_low_show
,
5395 .write
= memory_low_write
,
5399 .flags
= CFTYPE_NOT_ON_ROOT
,
5400 .seq_show
= memory_high_show
,
5401 .write
= memory_high_write
,
5405 .flags
= CFTYPE_NOT_ON_ROOT
,
5406 .seq_show
= memory_max_show
,
5407 .write
= memory_max_write
,
5411 .flags
= CFTYPE_NOT_ON_ROOT
,
5412 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
5413 .seq_show
= memory_events_show
,
5417 .flags
= CFTYPE_NOT_ON_ROOT
,
5418 .seq_show
= memory_stat_show
,
5423 struct cgroup_subsys memory_cgrp_subsys
= {
5424 .css_alloc
= mem_cgroup_css_alloc
,
5425 .css_online
= mem_cgroup_css_online
,
5426 .css_offline
= mem_cgroup_css_offline
,
5427 .css_released
= mem_cgroup_css_released
,
5428 .css_free
= mem_cgroup_css_free
,
5429 .css_reset
= mem_cgroup_css_reset
,
5430 .can_attach
= mem_cgroup_can_attach
,
5431 .cancel_attach
= mem_cgroup_cancel_attach
,
5432 .post_attach
= mem_cgroup_move_task
,
5433 .bind
= mem_cgroup_bind
,
5434 .dfl_cftypes
= memory_files
,
5435 .legacy_cftypes
= mem_cgroup_legacy_files
,
5440 * mem_cgroup_low - check if memory consumption is below the normal range
5441 * @root: the top ancestor of the sub-tree being checked
5442 * @memcg: the memory cgroup to check
5444 * Returns %true if memory consumption of @memcg, and that of all
5445 * ancestors up to (but not including) @root, is below the normal range.
5447 * @root is exclusive; it is never low when looked at directly and isn't
5448 * checked when traversing the hierarchy.
5450 * Excluding @root enables using memory.low to prioritize memory usage
5451 * between cgroups within a subtree of the hierarchy that is limited by
5452 * memory.high or memory.max.
5454 * For example, given cgroup A with children B and C:
5462 * 1. A/memory.current > A/memory.high
5463 * 2. A/B/memory.current < A/B/memory.low
5464 * 3. A/C/memory.current >= A/C/memory.low
5466 * As 'A' is high, i.e. triggers reclaim from 'A', and 'B' is low, we
5467 * should reclaim from 'C' until 'A' is no longer high or until we can
5468 * no longer reclaim from 'C'. If 'A', i.e. @root, isn't excluded by
5469 * mem_cgroup_low when reclaming from 'A', then 'B' won't be considered
5470 * low and we will reclaim indiscriminately from both 'B' and 'C'.
5472 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5474 if (mem_cgroup_disabled())
5478 root
= root_mem_cgroup
;
5482 for (; memcg
!= root
; memcg
= parent_mem_cgroup(memcg
)) {
5483 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5491 * mem_cgroup_try_charge - try charging a page
5492 * @page: page to charge
5493 * @mm: mm context of the victim
5494 * @gfp_mask: reclaim mode
5495 * @memcgp: charged memcg return
5496 * @compound: charge the page as compound or small page
5498 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5499 * pages according to @gfp_mask if necessary.
5501 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5502 * Otherwise, an error code is returned.
5504 * After page->mapping has been set up, the caller must finalize the
5505 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5506 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5508 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5509 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
5512 struct mem_cgroup
*memcg
= NULL
;
5513 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5516 if (mem_cgroup_disabled())
5519 if (PageSwapCache(page
)) {
5521 * Every swap fault against a single page tries to charge the
5522 * page, bail as early as possible. shmem_unuse() encounters
5523 * already charged pages, too. The USED bit is protected by
5524 * the page lock, which serializes swap cache removal, which
5525 * in turn serializes uncharging.
5527 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5528 if (compound_head(page
)->mem_cgroup
)
5531 if (do_swap_account
) {
5532 swp_entry_t ent
= { .val
= page_private(page
), };
5533 unsigned short id
= lookup_swap_cgroup_id(ent
);
5536 memcg
= mem_cgroup_from_id(id
);
5537 if (memcg
&& !css_tryget_online(&memcg
->css
))
5544 memcg
= get_mem_cgroup_from_mm(mm
);
5546 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5548 css_put(&memcg
->css
);
5555 * mem_cgroup_commit_charge - commit a page charge
5556 * @page: page to charge
5557 * @memcg: memcg to charge the page to
5558 * @lrucare: page might be on LRU already
5559 * @compound: charge the page as compound or small page
5561 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5562 * after page->mapping has been set up. This must happen atomically
5563 * as part of the page instantiation, i.e. under the page table lock
5564 * for anonymous pages, under the page lock for page and swap cache.
5566 * In addition, the page must not be on the LRU during the commit, to
5567 * prevent racing with task migration. If it might be, use @lrucare.
5569 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5571 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5572 bool lrucare
, bool compound
)
5574 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5576 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5577 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5579 if (mem_cgroup_disabled())
5582 * Swap faults will attempt to charge the same page multiple
5583 * times. But reuse_swap_page() might have removed the page
5584 * from swapcache already, so we can't check PageSwapCache().
5589 commit_charge(page
, memcg
, lrucare
);
5591 local_irq_disable();
5592 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
5593 memcg_check_events(memcg
, page
);
5596 if (do_memsw_account() && PageSwapCache(page
)) {
5597 swp_entry_t entry
= { .val
= page_private(page
) };
5599 * The swap entry might not get freed for a long time,
5600 * let's not wait for it. The page already received a
5601 * memory+swap charge, drop the swap entry duplicate.
5603 mem_cgroup_uncharge_swap(entry
, nr_pages
);
5608 * mem_cgroup_cancel_charge - cancel a page charge
5609 * @page: page to charge
5610 * @memcg: memcg to charge the page to
5611 * @compound: charge the page as compound or small page
5613 * Cancel a charge transaction started by mem_cgroup_try_charge().
5615 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5618 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5620 if (mem_cgroup_disabled())
5623 * Swap faults will attempt to charge the same page multiple
5624 * times. But reuse_swap_page() might have removed the page
5625 * from swapcache already, so we can't check PageSwapCache().
5630 cancel_charge(memcg
, nr_pages
);
5633 struct uncharge_gather
{
5634 struct mem_cgroup
*memcg
;
5635 unsigned long pgpgout
;
5636 unsigned long nr_anon
;
5637 unsigned long nr_file
;
5638 unsigned long nr_kmem
;
5639 unsigned long nr_huge
;
5640 unsigned long nr_shmem
;
5641 struct page
*dummy_page
;
5644 static inline void uncharge_gather_clear(struct uncharge_gather
*ug
)
5646 memset(ug
, 0, sizeof(*ug
));
5649 static void uncharge_batch(const struct uncharge_gather
*ug
)
5651 unsigned long nr_pages
= ug
->nr_anon
+ ug
->nr_file
+ ug
->nr_kmem
;
5652 unsigned long flags
;
5654 if (!mem_cgroup_is_root(ug
->memcg
)) {
5655 page_counter_uncharge(&ug
->memcg
->memory
, nr_pages
);
5656 if (do_memsw_account())
5657 page_counter_uncharge(&ug
->memcg
->memsw
, nr_pages
);
5658 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && ug
->nr_kmem
)
5659 page_counter_uncharge(&ug
->memcg
->kmem
, ug
->nr_kmem
);
5660 memcg_oom_recover(ug
->memcg
);
5663 local_irq_save(flags
);
5664 __this_cpu_sub(ug
->memcg
->stat
->count
[MEMCG_RSS
], ug
->nr_anon
);
5665 __this_cpu_sub(ug
->memcg
->stat
->count
[MEMCG_CACHE
], ug
->nr_file
);
5666 __this_cpu_sub(ug
->memcg
->stat
->count
[MEMCG_RSS_HUGE
], ug
->nr_huge
);
5667 __this_cpu_sub(ug
->memcg
->stat
->count
[NR_SHMEM
], ug
->nr_shmem
);
5668 __this_cpu_add(ug
->memcg
->stat
->events
[PGPGOUT
], ug
->pgpgout
);
5669 __this_cpu_add(ug
->memcg
->stat
->nr_page_events
, nr_pages
);
5670 memcg_check_events(ug
->memcg
, ug
->dummy_page
);
5671 local_irq_restore(flags
);
5673 if (!mem_cgroup_is_root(ug
->memcg
))
5674 css_put_many(&ug
->memcg
->css
, nr_pages
);
5677 static void uncharge_page(struct page
*page
, struct uncharge_gather
*ug
)
5679 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5680 VM_BUG_ON_PAGE(page_count(page
) && !is_zone_device_page(page
) &&
5681 !PageHWPoison(page
) , page
);
5683 if (!page
->mem_cgroup
)
5687 * Nobody should be changing or seriously looking at
5688 * page->mem_cgroup at this point, we have fully
5689 * exclusive access to the page.
5692 if (ug
->memcg
!= page
->mem_cgroup
) {
5695 uncharge_gather_clear(ug
);
5697 ug
->memcg
= page
->mem_cgroup
;
5700 if (!PageKmemcg(page
)) {
5701 unsigned int nr_pages
= 1;
5703 if (PageTransHuge(page
)) {
5704 nr_pages
<<= compound_order(page
);
5705 ug
->nr_huge
+= nr_pages
;
5708 ug
->nr_anon
+= nr_pages
;
5710 ug
->nr_file
+= nr_pages
;
5711 if (PageSwapBacked(page
))
5712 ug
->nr_shmem
+= nr_pages
;
5716 ug
->nr_kmem
+= 1 << compound_order(page
);
5717 __ClearPageKmemcg(page
);
5720 ug
->dummy_page
= page
;
5721 page
->mem_cgroup
= NULL
;
5724 static void uncharge_list(struct list_head
*page_list
)
5726 struct uncharge_gather ug
;
5727 struct list_head
*next
;
5729 uncharge_gather_clear(&ug
);
5732 * Note that the list can be a single page->lru; hence the
5733 * do-while loop instead of a simple list_for_each_entry().
5735 next
= page_list
->next
;
5739 page
= list_entry(next
, struct page
, lru
);
5740 next
= page
->lru
.next
;
5742 uncharge_page(page
, &ug
);
5743 } while (next
!= page_list
);
5746 uncharge_batch(&ug
);
5750 * mem_cgroup_uncharge - uncharge a page
5751 * @page: page to uncharge
5753 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5754 * mem_cgroup_commit_charge().
5756 void mem_cgroup_uncharge(struct page
*page
)
5758 struct uncharge_gather ug
;
5760 if (mem_cgroup_disabled())
5763 /* Don't touch page->lru of any random page, pre-check: */
5764 if (!page
->mem_cgroup
)
5767 uncharge_gather_clear(&ug
);
5768 uncharge_page(page
, &ug
);
5769 uncharge_batch(&ug
);
5773 * mem_cgroup_uncharge_list - uncharge a list of page
5774 * @page_list: list of pages to uncharge
5776 * Uncharge a list of pages previously charged with
5777 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5779 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5781 if (mem_cgroup_disabled())
5784 if (!list_empty(page_list
))
5785 uncharge_list(page_list
);
5789 * mem_cgroup_migrate - charge a page's replacement
5790 * @oldpage: currently circulating page
5791 * @newpage: replacement page
5793 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5794 * be uncharged upon free.
5796 * Both pages must be locked, @newpage->mapping must be set up.
5798 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
5800 struct mem_cgroup
*memcg
;
5801 unsigned int nr_pages
;
5803 unsigned long flags
;
5805 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5806 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5807 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5808 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5811 if (mem_cgroup_disabled())
5814 /* Page cache replacement: new page already charged? */
5815 if (newpage
->mem_cgroup
)
5818 /* Swapcache readahead pages can get replaced before being charged */
5819 memcg
= oldpage
->mem_cgroup
;
5823 /* Force-charge the new page. The old one will be freed soon */
5824 compound
= PageTransHuge(newpage
);
5825 nr_pages
= compound
? hpage_nr_pages(newpage
) : 1;
5827 page_counter_charge(&memcg
->memory
, nr_pages
);
5828 if (do_memsw_account())
5829 page_counter_charge(&memcg
->memsw
, nr_pages
);
5830 css_get_many(&memcg
->css
, nr_pages
);
5832 commit_charge(newpage
, memcg
, false);
5834 local_irq_save(flags
);
5835 mem_cgroup_charge_statistics(memcg
, newpage
, compound
, nr_pages
);
5836 memcg_check_events(memcg
, newpage
);
5837 local_irq_restore(flags
);
5840 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
5841 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
5843 void mem_cgroup_sk_alloc(struct sock
*sk
)
5845 struct mem_cgroup
*memcg
;
5847 if (!mem_cgroup_sockets_enabled
)
5851 * Socket cloning can throw us here with sk_memcg already
5852 * filled. It won't however, necessarily happen from
5853 * process context. So the test for root memcg given
5854 * the current task's memcg won't help us in this case.
5856 * Respecting the original socket's memcg is a better
5857 * decision in this case.
5860 css_get(&sk
->sk_memcg
->css
);
5865 memcg
= mem_cgroup_from_task(current
);
5866 if (memcg
== root_mem_cgroup
)
5868 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
5870 if (css_tryget_online(&memcg
->css
))
5871 sk
->sk_memcg
= memcg
;
5876 void mem_cgroup_sk_free(struct sock
*sk
)
5879 css_put(&sk
->sk_memcg
->css
);
5883 * mem_cgroup_charge_skmem - charge socket memory
5884 * @memcg: memcg to charge
5885 * @nr_pages: number of pages to charge
5887 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5888 * @memcg's configured limit, %false if the charge had to be forced.
5890 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5892 gfp_t gfp_mask
= GFP_KERNEL
;
5894 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5895 struct page_counter
*fail
;
5897 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
5898 memcg
->tcpmem_pressure
= 0;
5901 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
5902 memcg
->tcpmem_pressure
= 1;
5906 /* Don't block in the packet receive path */
5908 gfp_mask
= GFP_NOWAIT
;
5910 this_cpu_add(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5912 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
5915 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
5920 * mem_cgroup_uncharge_skmem - uncharge socket memory
5921 * @memcg - memcg to uncharge
5922 * @nr_pages - number of pages to uncharge
5924 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5926 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5927 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
5931 this_cpu_sub(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5933 refill_stock(memcg
, nr_pages
);
5936 static int __init
cgroup_memory(char *s
)
5940 while ((token
= strsep(&s
, ",")) != NULL
) {
5943 if (!strcmp(token
, "nosocket"))
5944 cgroup_memory_nosocket
= true;
5945 if (!strcmp(token
, "nokmem"))
5946 cgroup_memory_nokmem
= true;
5950 __setup("cgroup.memory=", cgroup_memory
);
5953 * subsys_initcall() for memory controller.
5955 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
5956 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
5957 * basically everything that doesn't depend on a specific mem_cgroup structure
5958 * should be initialized from here.
5960 static int __init
mem_cgroup_init(void)
5966 * Kmem cache creation is mostly done with the slab_mutex held,
5967 * so use a workqueue with limited concurrency to avoid stalling
5968 * all worker threads in case lots of cgroups are created and
5969 * destroyed simultaneously.
5971 memcg_kmem_cache_wq
= alloc_workqueue("memcg_kmem_cache", 0, 1);
5972 BUG_ON(!memcg_kmem_cache_wq
);
5975 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD
, "mm/memctrl:dead", NULL
,
5976 memcg_hotplug_cpu_dead
);
5978 for_each_possible_cpu(cpu
)
5979 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5982 for_each_node(node
) {
5983 struct mem_cgroup_tree_per_node
*rtpn
;
5985 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5986 node_online(node
) ? node
: NUMA_NO_NODE
);
5988 rtpn
->rb_root
= RB_ROOT
;
5989 rtpn
->rb_rightmost
= NULL
;
5990 spin_lock_init(&rtpn
->lock
);
5991 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5996 subsys_initcall(mem_cgroup_init
);
5998 #ifdef CONFIG_MEMCG_SWAP
5999 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
6001 while (!atomic_inc_not_zero(&memcg
->id
.ref
)) {
6003 * The root cgroup cannot be destroyed, so it's refcount must
6006 if (WARN_ON_ONCE(memcg
== root_mem_cgroup
)) {
6010 memcg
= parent_mem_cgroup(memcg
);
6012 memcg
= root_mem_cgroup
;
6018 * mem_cgroup_swapout - transfer a memsw charge to swap
6019 * @page: page whose memsw charge to transfer
6020 * @entry: swap entry to move the charge to
6022 * Transfer the memsw charge of @page to @entry.
6024 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
6026 struct mem_cgroup
*memcg
, *swap_memcg
;
6027 unsigned int nr_entries
;
6028 unsigned short oldid
;
6030 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6031 VM_BUG_ON_PAGE(page_count(page
), page
);
6033 if (!do_memsw_account())
6036 memcg
= page
->mem_cgroup
;
6038 /* Readahead page, never charged */
6043 * In case the memcg owning these pages has been offlined and doesn't
6044 * have an ID allocated to it anymore, charge the closest online
6045 * ancestor for the swap instead and transfer the memory+swap charge.
6047 swap_memcg
= mem_cgroup_id_get_online(memcg
);
6048 nr_entries
= hpage_nr_pages(page
);
6049 /* Get references for the tail pages, too */
6051 mem_cgroup_id_get_many(swap_memcg
, nr_entries
- 1);
6052 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
),
6054 VM_BUG_ON_PAGE(oldid
, page
);
6055 mem_cgroup_swap_statistics(swap_memcg
, nr_entries
);
6057 page
->mem_cgroup
= NULL
;
6059 if (!mem_cgroup_is_root(memcg
))
6060 page_counter_uncharge(&memcg
->memory
, nr_entries
);
6062 if (memcg
!= swap_memcg
) {
6063 if (!mem_cgroup_is_root(swap_memcg
))
6064 page_counter_charge(&swap_memcg
->memsw
, nr_entries
);
6065 page_counter_uncharge(&memcg
->memsw
, nr_entries
);
6069 * Interrupts should be disabled here because the caller holds the
6070 * mapping->tree_lock lock which is taken with interrupts-off. It is
6071 * important here to have the interrupts disabled because it is the
6072 * only synchronisation we have for udpating the per-CPU variables.
6074 VM_BUG_ON(!irqs_disabled());
6075 mem_cgroup_charge_statistics(memcg
, page
, PageTransHuge(page
),
6077 memcg_check_events(memcg
, page
);
6079 if (!mem_cgroup_is_root(memcg
))
6080 css_put_many(&memcg
->css
, nr_entries
);
6084 * mem_cgroup_try_charge_swap - try charging swap space for a page
6085 * @page: page being added to swap
6086 * @entry: swap entry to charge
6088 * Try to charge @page's memcg for the swap space at @entry.
6090 * Returns 0 on success, -ENOMEM on failure.
6092 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
6094 unsigned int nr_pages
= hpage_nr_pages(page
);
6095 struct page_counter
*counter
;
6096 struct mem_cgroup
*memcg
;
6097 unsigned short oldid
;
6099 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
6102 memcg
= page
->mem_cgroup
;
6104 /* Readahead page, never charged */
6108 memcg
= mem_cgroup_id_get_online(memcg
);
6110 if (!mem_cgroup_is_root(memcg
) &&
6111 !page_counter_try_charge(&memcg
->swap
, nr_pages
, &counter
)) {
6112 mem_cgroup_id_put(memcg
);
6116 /* Get references for the tail pages, too */
6118 mem_cgroup_id_get_many(memcg
, nr_pages
- 1);
6119 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
), nr_pages
);
6120 VM_BUG_ON_PAGE(oldid
, page
);
6121 mem_cgroup_swap_statistics(memcg
, nr_pages
);
6127 * mem_cgroup_uncharge_swap - uncharge swap space
6128 * @entry: swap entry to uncharge
6129 * @nr_pages: the amount of swap space to uncharge
6131 void mem_cgroup_uncharge_swap(swp_entry_t entry
, unsigned int nr_pages
)
6133 struct mem_cgroup
*memcg
;
6136 if (!do_swap_account
)
6139 id
= swap_cgroup_record(entry
, 0, nr_pages
);
6141 memcg
= mem_cgroup_from_id(id
);
6143 if (!mem_cgroup_is_root(memcg
)) {
6144 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
6145 page_counter_uncharge(&memcg
->swap
, nr_pages
);
6147 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
6149 mem_cgroup_swap_statistics(memcg
, -nr_pages
);
6150 mem_cgroup_id_put_many(memcg
, nr_pages
);
6155 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
6157 long nr_swap_pages
= get_nr_swap_pages();
6159 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
6160 return nr_swap_pages
;
6161 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
6162 nr_swap_pages
= min_t(long, nr_swap_pages
,
6163 READ_ONCE(memcg
->swap
.limit
) -
6164 page_counter_read(&memcg
->swap
));
6165 return nr_swap_pages
;
6168 bool mem_cgroup_swap_full(struct page
*page
)
6170 struct mem_cgroup
*memcg
;
6172 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
6176 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
6179 memcg
= page
->mem_cgroup
;
6183 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
6184 if (page_counter_read(&memcg
->swap
) * 2 >= memcg
->swap
.limit
)
6190 /* for remember boot option*/
6191 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6192 static int really_do_swap_account __initdata
= 1;
6194 static int really_do_swap_account __initdata
;
6197 static int __init
enable_swap_account(char *s
)
6199 if (!strcmp(s
, "1"))
6200 really_do_swap_account
= 1;
6201 else if (!strcmp(s
, "0"))
6202 really_do_swap_account
= 0;
6205 __setup("swapaccount=", enable_swap_account
);
6207 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
6210 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6212 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
6215 static int swap_max_show(struct seq_file
*m
, void *v
)
6217 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
6218 unsigned long max
= READ_ONCE(memcg
->swap
.limit
);
6220 if (max
== PAGE_COUNTER_MAX
)
6221 seq_puts(m
, "max\n");
6223 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
6228 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
6229 char *buf
, size_t nbytes
, loff_t off
)
6231 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6235 buf
= strstrip(buf
);
6236 err
= page_counter_memparse(buf
, "max", &max
);
6240 mutex_lock(&memcg_limit_mutex
);
6241 err
= page_counter_limit(&memcg
->swap
, max
);
6242 mutex_unlock(&memcg_limit_mutex
);
6249 static struct cftype swap_files
[] = {
6251 .name
= "swap.current",
6252 .flags
= CFTYPE_NOT_ON_ROOT
,
6253 .read_u64
= swap_current_read
,
6257 .flags
= CFTYPE_NOT_ON_ROOT
,
6258 .seq_show
= swap_max_show
,
6259 .write
= swap_max_write
,
6264 static struct cftype memsw_cgroup_files
[] = {
6266 .name
= "memsw.usage_in_bytes",
6267 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6268 .read_u64
= mem_cgroup_read_u64
,
6271 .name
= "memsw.max_usage_in_bytes",
6272 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6273 .write
= mem_cgroup_reset
,
6274 .read_u64
= mem_cgroup_read_u64
,
6277 .name
= "memsw.limit_in_bytes",
6278 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6279 .write
= mem_cgroup_write
,
6280 .read_u64
= mem_cgroup_read_u64
,
6283 .name
= "memsw.failcnt",
6284 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6285 .write
= mem_cgroup_reset
,
6286 .read_u64
= mem_cgroup_read_u64
,
6288 { }, /* terminate */
6291 static int __init
mem_cgroup_swap_init(void)
6293 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6294 do_swap_account
= 1;
6295 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
6297 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
6298 memsw_cgroup_files
));
6302 subsys_initcall(mem_cgroup_swap_init
);
6304 #endif /* CONFIG_MEMCG_SWAP */