1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
25 #include <linux/page_counter.h>
26 #include <linux/memcontrol.h>
27 #include <linux/cgroup.h>
29 #include <linux/sched/mm.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/hugetlb.h>
32 #include <linux/pagemap.h>
33 #include <linux/vm_event_item.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
59 #include <linux/tracehook.h>
65 #include <linux/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
70 EXPORT_SYMBOL(memory_cgrp_subsys
);
72 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
74 #define MEM_CGROUP_RECLAIM_RETRIES 5
76 /* Socket memory accounting disabled? */
77 static bool cgroup_memory_nosocket
;
79 /* Kernel memory accounting disabled? */
80 static bool cgroup_memory_nokmem
;
82 /* Whether the swap controller is active */
83 #ifdef CONFIG_MEMCG_SWAP
84 int do_swap_account __read_mostly
;
86 #define do_swap_account 0
89 /* Whether legacy memory+swap accounting is active */
90 static bool do_memsw_account(void)
92 return !cgroup_subsys_on_dfl(memory_cgrp_subsys
) && do_swap_account
;
95 static const char *const mem_cgroup_lru_names
[] = {
103 #define THRESHOLDS_EVENTS_TARGET 128
104 #define SOFTLIMIT_EVENTS_TARGET 1024
105 #define NUMAINFO_EVENTS_TARGET 1024
108 * Cgroups above their limits are maintained in a RB-Tree, independent of
109 * their hierarchy representation
112 struct mem_cgroup_tree_per_node
{
113 struct rb_root rb_root
;
114 struct rb_node
*rb_rightmost
;
118 struct mem_cgroup_tree
{
119 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
122 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
125 struct mem_cgroup_eventfd_list
{
126 struct list_head list
;
127 struct eventfd_ctx
*eventfd
;
131 * cgroup_event represents events which userspace want to receive.
133 struct mem_cgroup_event
{
135 * memcg which the event belongs to.
137 struct mem_cgroup
*memcg
;
139 * eventfd to signal userspace about the event.
141 struct eventfd_ctx
*eventfd
;
143 * Each of these stored in a list by the cgroup.
145 struct list_head list
;
147 * register_event() callback will be used to add new userspace
148 * waiter for changes related to this event. Use eventfd_signal()
149 * on eventfd to send notification to userspace.
151 int (*register_event
)(struct mem_cgroup
*memcg
,
152 struct eventfd_ctx
*eventfd
, const char *args
);
154 * unregister_event() callback will be called when userspace closes
155 * the eventfd or on cgroup removing. This callback must be set,
156 * if you want provide notification functionality.
158 void (*unregister_event
)(struct mem_cgroup
*memcg
,
159 struct eventfd_ctx
*eventfd
);
161 * All fields below needed to unregister event when
162 * userspace closes eventfd.
165 wait_queue_head_t
*wqh
;
166 wait_queue_entry_t wait
;
167 struct work_struct remove
;
170 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
171 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
173 /* Stuffs for move charges at task migration. */
175 * Types of charges to be moved.
177 #define MOVE_ANON 0x1U
178 #define MOVE_FILE 0x2U
179 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
181 /* "mc" and its members are protected by cgroup_mutex */
182 static struct move_charge_struct
{
183 spinlock_t lock
; /* for from, to */
184 struct mm_struct
*mm
;
185 struct mem_cgroup
*from
;
186 struct mem_cgroup
*to
;
188 unsigned long precharge
;
189 unsigned long moved_charge
;
190 unsigned long moved_swap
;
191 struct task_struct
*moving_task
; /* a task moving charges */
192 wait_queue_head_t waitq
; /* a waitq for other context */
194 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
195 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
199 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
200 * limit reclaim to prevent infinite loops, if they ever occur.
202 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
203 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
206 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
207 MEM_CGROUP_CHARGE_TYPE_ANON
,
208 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
209 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
213 /* for encoding cft->private value on file */
222 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
223 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
224 #define MEMFILE_ATTR(val) ((val) & 0xffff)
225 /* Used for OOM nofiier */
226 #define OOM_CONTROL (0)
229 * Iteration constructs for visiting all cgroups (under a tree). If
230 * loops are exited prematurely (break), mem_cgroup_iter_break() must
231 * be used for reference counting.
233 #define for_each_mem_cgroup_tree(iter, root) \
234 for (iter = mem_cgroup_iter(root, NULL, NULL); \
236 iter = mem_cgroup_iter(root, iter, NULL))
238 #define for_each_mem_cgroup(iter) \
239 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
241 iter = mem_cgroup_iter(NULL, iter, NULL))
243 static inline bool should_force_charge(void)
245 return tsk_is_oom_victim(current
) || fatal_signal_pending(current
) ||
246 (current
->flags
& PF_EXITING
);
249 /* Some nice accessors for the vmpressure. */
250 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
253 memcg
= root_mem_cgroup
;
254 return &memcg
->vmpressure
;
257 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
259 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
262 #ifdef CONFIG_MEMCG_KMEM
264 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
265 * The main reason for not using cgroup id for this:
266 * this works better in sparse environments, where we have a lot of memcgs,
267 * but only a few kmem-limited. Or also, if we have, for instance, 200
268 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
269 * 200 entry array for that.
271 * The current size of the caches array is stored in memcg_nr_cache_ids. It
272 * will double each time we have to increase it.
274 static DEFINE_IDA(memcg_cache_ida
);
275 int memcg_nr_cache_ids
;
277 /* Protects memcg_nr_cache_ids */
278 static DECLARE_RWSEM(memcg_cache_ids_sem
);
280 void memcg_get_cache_ids(void)
282 down_read(&memcg_cache_ids_sem
);
285 void memcg_put_cache_ids(void)
287 up_read(&memcg_cache_ids_sem
);
291 * MIN_SIZE is different than 1, because we would like to avoid going through
292 * the alloc/free process all the time. In a small machine, 4 kmem-limited
293 * cgroups is a reasonable guess. In the future, it could be a parameter or
294 * tunable, but that is strictly not necessary.
296 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
297 * this constant directly from cgroup, but it is understandable that this is
298 * better kept as an internal representation in cgroup.c. In any case, the
299 * cgrp_id space is not getting any smaller, and we don't have to necessarily
300 * increase ours as well if it increases.
302 #define MEMCG_CACHES_MIN_SIZE 4
303 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
306 * A lot of the calls to the cache allocation functions are expected to be
307 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
308 * conditional to this static branch, we'll have to allow modules that does
309 * kmem_cache_alloc and the such to see this symbol as well
311 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
312 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
314 struct workqueue_struct
*memcg_kmem_cache_wq
;
316 static int memcg_shrinker_map_size
;
317 static DEFINE_MUTEX(memcg_shrinker_map_mutex
);
319 static void memcg_free_shrinker_map_rcu(struct rcu_head
*head
)
321 kvfree(container_of(head
, struct memcg_shrinker_map
, rcu
));
324 static int memcg_expand_one_shrinker_map(struct mem_cgroup
*memcg
,
325 int size
, int old_size
)
327 struct memcg_shrinker_map
*new, *old
;
330 lockdep_assert_held(&memcg_shrinker_map_mutex
);
333 old
= rcu_dereference_protected(
334 mem_cgroup_nodeinfo(memcg
, nid
)->shrinker_map
, true);
335 /* Not yet online memcg */
339 new = kvmalloc(sizeof(*new) + size
, GFP_KERNEL
);
343 /* Set all old bits, clear all new bits */
344 memset(new->map
, (int)0xff, old_size
);
345 memset((void *)new->map
+ old_size
, 0, size
- old_size
);
347 rcu_assign_pointer(memcg
->nodeinfo
[nid
]->shrinker_map
, new);
348 call_rcu(&old
->rcu
, memcg_free_shrinker_map_rcu
);
354 static void memcg_free_shrinker_maps(struct mem_cgroup
*memcg
)
356 struct mem_cgroup_per_node
*pn
;
357 struct memcg_shrinker_map
*map
;
360 if (mem_cgroup_is_root(memcg
))
364 pn
= mem_cgroup_nodeinfo(memcg
, nid
);
365 map
= rcu_dereference_protected(pn
->shrinker_map
, true);
368 rcu_assign_pointer(pn
->shrinker_map
, NULL
);
372 static int memcg_alloc_shrinker_maps(struct mem_cgroup
*memcg
)
374 struct memcg_shrinker_map
*map
;
375 int nid
, size
, ret
= 0;
377 if (mem_cgroup_is_root(memcg
))
380 mutex_lock(&memcg_shrinker_map_mutex
);
381 size
= memcg_shrinker_map_size
;
383 map
= kvzalloc(sizeof(*map
) + size
, GFP_KERNEL
);
385 memcg_free_shrinker_maps(memcg
);
389 rcu_assign_pointer(memcg
->nodeinfo
[nid
]->shrinker_map
, map
);
391 mutex_unlock(&memcg_shrinker_map_mutex
);
396 int memcg_expand_shrinker_maps(int new_id
)
398 int size
, old_size
, ret
= 0;
399 struct mem_cgroup
*memcg
;
401 size
= DIV_ROUND_UP(new_id
+ 1, BITS_PER_LONG
) * sizeof(unsigned long);
402 old_size
= memcg_shrinker_map_size
;
403 if (size
<= old_size
)
406 mutex_lock(&memcg_shrinker_map_mutex
);
407 if (!root_mem_cgroup
)
410 for_each_mem_cgroup(memcg
) {
411 if (mem_cgroup_is_root(memcg
))
413 ret
= memcg_expand_one_shrinker_map(memcg
, size
, old_size
);
419 memcg_shrinker_map_size
= size
;
420 mutex_unlock(&memcg_shrinker_map_mutex
);
424 void memcg_set_shrinker_bit(struct mem_cgroup
*memcg
, int nid
, int shrinker_id
)
426 if (shrinker_id
>= 0 && memcg
&& !mem_cgroup_is_root(memcg
)) {
427 struct memcg_shrinker_map
*map
;
430 map
= rcu_dereference(memcg
->nodeinfo
[nid
]->shrinker_map
);
431 /* Pairs with smp mb in shrink_slab() */
432 smp_mb__before_atomic();
433 set_bit(shrinker_id
, map
->map
);
438 #else /* CONFIG_MEMCG_KMEM */
439 static int memcg_alloc_shrinker_maps(struct mem_cgroup
*memcg
)
443 static void memcg_free_shrinker_maps(struct mem_cgroup
*memcg
) { }
444 #endif /* CONFIG_MEMCG_KMEM */
447 * mem_cgroup_css_from_page - css of the memcg associated with a page
448 * @page: page of interest
450 * If memcg is bound to the default hierarchy, css of the memcg associated
451 * with @page is returned. The returned css remains associated with @page
452 * until it is released.
454 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
457 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
459 struct mem_cgroup
*memcg
;
461 memcg
= page
->mem_cgroup
;
463 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
464 memcg
= root_mem_cgroup
;
470 * page_cgroup_ino - return inode number of the memcg a page is charged to
473 * Look up the closest online ancestor of the memory cgroup @page is charged to
474 * and return its inode number or 0 if @page is not charged to any cgroup. It
475 * is safe to call this function without holding a reference to @page.
477 * Note, this function is inherently racy, because there is nothing to prevent
478 * the cgroup inode from getting torn down and potentially reallocated a moment
479 * after page_cgroup_ino() returns, so it only should be used by callers that
480 * do not care (such as procfs interfaces).
482 ino_t
page_cgroup_ino(struct page
*page
)
484 struct mem_cgroup
*memcg
;
485 unsigned long ino
= 0;
488 memcg
= READ_ONCE(page
->mem_cgroup
);
489 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
490 memcg
= parent_mem_cgroup(memcg
);
492 ino
= cgroup_ino(memcg
->css
.cgroup
);
497 static struct mem_cgroup_per_node
*
498 mem_cgroup_page_nodeinfo(struct mem_cgroup
*memcg
, struct page
*page
)
500 int nid
= page_to_nid(page
);
502 return memcg
->nodeinfo
[nid
];
505 static struct mem_cgroup_tree_per_node
*
506 soft_limit_tree_node(int nid
)
508 return soft_limit_tree
.rb_tree_per_node
[nid
];
511 static struct mem_cgroup_tree_per_node
*
512 soft_limit_tree_from_page(struct page
*page
)
514 int nid
= page_to_nid(page
);
516 return soft_limit_tree
.rb_tree_per_node
[nid
];
519 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node
*mz
,
520 struct mem_cgroup_tree_per_node
*mctz
,
521 unsigned long new_usage_in_excess
)
523 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
524 struct rb_node
*parent
= NULL
;
525 struct mem_cgroup_per_node
*mz_node
;
526 bool rightmost
= true;
531 mz
->usage_in_excess
= new_usage_in_excess
;
532 if (!mz
->usage_in_excess
)
536 mz_node
= rb_entry(parent
, struct mem_cgroup_per_node
,
538 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
) {
544 * We can't avoid mem cgroups that are over their soft
545 * limit by the same amount
547 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
552 mctz
->rb_rightmost
= &mz
->tree_node
;
554 rb_link_node(&mz
->tree_node
, parent
, p
);
555 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
559 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
560 struct mem_cgroup_tree_per_node
*mctz
)
565 if (&mz
->tree_node
== mctz
->rb_rightmost
)
566 mctz
->rb_rightmost
= rb_prev(&mz
->tree_node
);
568 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
572 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
573 struct mem_cgroup_tree_per_node
*mctz
)
577 spin_lock_irqsave(&mctz
->lock
, flags
);
578 __mem_cgroup_remove_exceeded(mz
, mctz
);
579 spin_unlock_irqrestore(&mctz
->lock
, flags
);
582 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
584 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
585 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
586 unsigned long excess
= 0;
588 if (nr_pages
> soft_limit
)
589 excess
= nr_pages
- soft_limit
;
594 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
596 unsigned long excess
;
597 struct mem_cgroup_per_node
*mz
;
598 struct mem_cgroup_tree_per_node
*mctz
;
600 mctz
= soft_limit_tree_from_page(page
);
604 * Necessary to update all ancestors when hierarchy is used.
605 * because their event counter is not touched.
607 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
608 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
609 excess
= soft_limit_excess(memcg
);
611 * We have to update the tree if mz is on RB-tree or
612 * mem is over its softlimit.
614 if (excess
|| mz
->on_tree
) {
617 spin_lock_irqsave(&mctz
->lock
, flags
);
618 /* if on-tree, remove it */
620 __mem_cgroup_remove_exceeded(mz
, mctz
);
622 * Insert again. mz->usage_in_excess will be updated.
623 * If excess is 0, no tree ops.
625 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
626 spin_unlock_irqrestore(&mctz
->lock
, flags
);
631 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
633 struct mem_cgroup_tree_per_node
*mctz
;
634 struct mem_cgroup_per_node
*mz
;
638 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
639 mctz
= soft_limit_tree_node(nid
);
641 mem_cgroup_remove_exceeded(mz
, mctz
);
645 static struct mem_cgroup_per_node
*
646 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
648 struct mem_cgroup_per_node
*mz
;
652 if (!mctz
->rb_rightmost
)
653 goto done
; /* Nothing to reclaim from */
655 mz
= rb_entry(mctz
->rb_rightmost
,
656 struct mem_cgroup_per_node
, tree_node
);
658 * Remove the node now but someone else can add it back,
659 * we will to add it back at the end of reclaim to its correct
660 * position in the tree.
662 __mem_cgroup_remove_exceeded(mz
, mctz
);
663 if (!soft_limit_excess(mz
->memcg
) ||
664 !css_tryget_online(&mz
->memcg
->css
))
670 static struct mem_cgroup_per_node
*
671 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
673 struct mem_cgroup_per_node
*mz
;
675 spin_lock_irq(&mctz
->lock
);
676 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
677 spin_unlock_irq(&mctz
->lock
);
682 * __mod_memcg_state - update cgroup memory statistics
683 * @memcg: the memory cgroup
684 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
685 * @val: delta to add to the counter, can be negative
687 void __mod_memcg_state(struct mem_cgroup
*memcg
, int idx
, int val
)
691 if (mem_cgroup_disabled())
694 __this_cpu_add(memcg
->vmstats_local
->stat
[idx
], val
);
696 x
= val
+ __this_cpu_read(memcg
->vmstats_percpu
->stat
[idx
]);
697 if (unlikely(abs(x
) > MEMCG_CHARGE_BATCH
)) {
698 struct mem_cgroup
*mi
;
700 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
701 atomic_long_add(x
, &mi
->vmstats
[idx
]);
704 __this_cpu_write(memcg
->vmstats_percpu
->stat
[idx
], x
);
707 static struct mem_cgroup_per_node
*
708 parent_nodeinfo(struct mem_cgroup_per_node
*pn
, int nid
)
710 struct mem_cgroup
*parent
;
712 parent
= parent_mem_cgroup(pn
->memcg
);
715 return mem_cgroup_nodeinfo(parent
, nid
);
719 * __mod_lruvec_state - update lruvec memory statistics
720 * @lruvec: the lruvec
721 * @idx: the stat item
722 * @val: delta to add to the counter, can be negative
724 * The lruvec is the intersection of the NUMA node and a cgroup. This
725 * function updates the all three counters that are affected by a
726 * change of state at this level: per-node, per-cgroup, per-lruvec.
728 void __mod_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
731 pg_data_t
*pgdat
= lruvec_pgdat(lruvec
);
732 struct mem_cgroup_per_node
*pn
;
733 struct mem_cgroup
*memcg
;
737 __mod_node_page_state(pgdat
, idx
, val
);
739 if (mem_cgroup_disabled())
742 pn
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
746 __mod_memcg_state(memcg
, idx
, val
);
749 __this_cpu_add(pn
->lruvec_stat_local
->count
[idx
], val
);
751 x
= val
+ __this_cpu_read(pn
->lruvec_stat_cpu
->count
[idx
]);
752 if (unlikely(abs(x
) > MEMCG_CHARGE_BATCH
)) {
753 struct mem_cgroup_per_node
*pi
;
755 for (pi
= pn
; pi
; pi
= parent_nodeinfo(pi
, pgdat
->node_id
))
756 atomic_long_add(x
, &pi
->lruvec_stat
[idx
]);
759 __this_cpu_write(pn
->lruvec_stat_cpu
->count
[idx
], x
);
763 * __count_memcg_events - account VM events in a cgroup
764 * @memcg: the memory cgroup
765 * @idx: the event item
766 * @count: the number of events that occured
768 void __count_memcg_events(struct mem_cgroup
*memcg
, enum vm_event_item idx
,
773 if (mem_cgroup_disabled())
776 __this_cpu_add(memcg
->vmstats_local
->events
[idx
], count
);
778 x
= count
+ __this_cpu_read(memcg
->vmstats_percpu
->events
[idx
]);
779 if (unlikely(x
> MEMCG_CHARGE_BATCH
)) {
780 struct mem_cgroup
*mi
;
782 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
783 atomic_long_add(x
, &mi
->vmevents
[idx
]);
786 __this_cpu_write(memcg
->vmstats_percpu
->events
[idx
], x
);
789 static unsigned long memcg_events(struct mem_cgroup
*memcg
, int event
)
791 return atomic_long_read(&memcg
->vmevents
[event
]);
794 static unsigned long memcg_events_local(struct mem_cgroup
*memcg
, int event
)
799 for_each_possible_cpu(cpu
)
800 x
+= per_cpu(memcg
->vmstats_local
->events
[event
], cpu
);
804 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
806 bool compound
, int nr_pages
)
809 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
810 * counted as CACHE even if it's on ANON LRU.
813 __mod_memcg_state(memcg
, MEMCG_RSS
, nr_pages
);
815 __mod_memcg_state(memcg
, MEMCG_CACHE
, nr_pages
);
816 if (PageSwapBacked(page
))
817 __mod_memcg_state(memcg
, NR_SHMEM
, nr_pages
);
821 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
822 __mod_memcg_state(memcg
, MEMCG_RSS_HUGE
, nr_pages
);
825 /* pagein of a big page is an event. So, ignore page size */
827 __count_memcg_events(memcg
, PGPGIN
, 1);
829 __count_memcg_events(memcg
, PGPGOUT
, 1);
830 nr_pages
= -nr_pages
; /* for event */
833 __this_cpu_add(memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
836 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
837 enum mem_cgroup_events_target target
)
839 unsigned long val
, next
;
841 val
= __this_cpu_read(memcg
->vmstats_percpu
->nr_page_events
);
842 next
= __this_cpu_read(memcg
->vmstats_percpu
->targets
[target
]);
843 /* from time_after() in jiffies.h */
844 if ((long)(next
- val
) < 0) {
846 case MEM_CGROUP_TARGET_THRESH
:
847 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
849 case MEM_CGROUP_TARGET_SOFTLIMIT
:
850 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
852 case MEM_CGROUP_TARGET_NUMAINFO
:
853 next
= val
+ NUMAINFO_EVENTS_TARGET
;
858 __this_cpu_write(memcg
->vmstats_percpu
->targets
[target
], next
);
865 * Check events in order.
868 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
870 /* threshold event is triggered in finer grain than soft limit */
871 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
872 MEM_CGROUP_TARGET_THRESH
))) {
874 bool do_numainfo __maybe_unused
;
876 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
877 MEM_CGROUP_TARGET_SOFTLIMIT
);
879 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
880 MEM_CGROUP_TARGET_NUMAINFO
);
882 mem_cgroup_threshold(memcg
);
883 if (unlikely(do_softlimit
))
884 mem_cgroup_update_tree(memcg
, page
);
886 if (unlikely(do_numainfo
))
887 atomic_inc(&memcg
->numainfo_events
);
892 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
895 * mm_update_next_owner() may clear mm->owner to NULL
896 * if it races with swapoff, page migration, etc.
897 * So this can be called with p == NULL.
902 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
904 EXPORT_SYMBOL(mem_cgroup_from_task
);
907 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
908 * @mm: mm from which memcg should be extracted. It can be NULL.
910 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
911 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
914 struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
916 struct mem_cgroup
*memcg
;
918 if (mem_cgroup_disabled())
924 * Page cache insertions can happen withou an
925 * actual mm context, e.g. during disk probing
926 * on boot, loopback IO, acct() writes etc.
929 memcg
= root_mem_cgroup
;
931 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
932 if (unlikely(!memcg
))
933 memcg
= root_mem_cgroup
;
935 } while (!css_tryget_online(&memcg
->css
));
939 EXPORT_SYMBOL(get_mem_cgroup_from_mm
);
942 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
943 * @page: page from which memcg should be extracted.
945 * Obtain a reference on page->memcg and returns it if successful. Otherwise
946 * root_mem_cgroup is returned.
948 struct mem_cgroup
*get_mem_cgroup_from_page(struct page
*page
)
950 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
952 if (mem_cgroup_disabled())
956 if (!memcg
|| !css_tryget_online(&memcg
->css
))
957 memcg
= root_mem_cgroup
;
961 EXPORT_SYMBOL(get_mem_cgroup_from_page
);
964 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
966 static __always_inline
struct mem_cgroup
*get_mem_cgroup_from_current(void)
968 if (unlikely(current
->active_memcg
)) {
969 struct mem_cgroup
*memcg
= root_mem_cgroup
;
972 if (css_tryget_online(¤t
->active_memcg
->css
))
973 memcg
= current
->active_memcg
;
977 return get_mem_cgroup_from_mm(current
->mm
);
981 * mem_cgroup_iter - iterate over memory cgroup hierarchy
982 * @root: hierarchy root
983 * @prev: previously returned memcg, NULL on first invocation
984 * @reclaim: cookie for shared reclaim walks, NULL for full walks
986 * Returns references to children of the hierarchy below @root, or
987 * @root itself, or %NULL after a full round-trip.
989 * Caller must pass the return value in @prev on subsequent
990 * invocations for reference counting, or use mem_cgroup_iter_break()
991 * to cancel a hierarchy walk before the round-trip is complete.
993 * Reclaimers can specify a node and a priority level in @reclaim to
994 * divide up the memcgs in the hierarchy among all concurrent
995 * reclaimers operating on the same node and priority.
997 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
998 struct mem_cgroup
*prev
,
999 struct mem_cgroup_reclaim_cookie
*reclaim
)
1001 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1002 struct cgroup_subsys_state
*css
= NULL
;
1003 struct mem_cgroup
*memcg
= NULL
;
1004 struct mem_cgroup
*pos
= NULL
;
1006 if (mem_cgroup_disabled())
1010 root
= root_mem_cgroup
;
1012 if (prev
&& !reclaim
)
1015 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1024 struct mem_cgroup_per_node
*mz
;
1026 mz
= mem_cgroup_nodeinfo(root
, reclaim
->pgdat
->node_id
);
1027 iter
= &mz
->iter
[reclaim
->priority
];
1029 if (prev
&& reclaim
->generation
!= iter
->generation
)
1033 pos
= READ_ONCE(iter
->position
);
1034 if (!pos
|| css_tryget(&pos
->css
))
1037 * css reference reached zero, so iter->position will
1038 * be cleared by ->css_released. However, we should not
1039 * rely on this happening soon, because ->css_released
1040 * is called from a work queue, and by busy-waiting we
1041 * might block it. So we clear iter->position right
1044 (void)cmpxchg(&iter
->position
, pos
, NULL
);
1052 css
= css_next_descendant_pre(css
, &root
->css
);
1055 * Reclaimers share the hierarchy walk, and a
1056 * new one might jump in right at the end of
1057 * the hierarchy - make sure they see at least
1058 * one group and restart from the beginning.
1066 * Verify the css and acquire a reference. The root
1067 * is provided by the caller, so we know it's alive
1068 * and kicking, and don't take an extra reference.
1070 memcg
= mem_cgroup_from_css(css
);
1072 if (css
== &root
->css
)
1075 if (css_tryget(css
))
1083 * The position could have already been updated by a competing
1084 * thread, so check that the value hasn't changed since we read
1085 * it to avoid reclaiming from the same cgroup twice.
1087 (void)cmpxchg(&iter
->position
, pos
, memcg
);
1095 reclaim
->generation
= iter
->generation
;
1101 if (prev
&& prev
!= root
)
1102 css_put(&prev
->css
);
1108 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1109 * @root: hierarchy root
1110 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1112 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1113 struct mem_cgroup
*prev
)
1116 root
= root_mem_cgroup
;
1117 if (prev
&& prev
!= root
)
1118 css_put(&prev
->css
);
1121 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
1123 struct mem_cgroup
*memcg
= dead_memcg
;
1124 struct mem_cgroup_reclaim_iter
*iter
;
1125 struct mem_cgroup_per_node
*mz
;
1129 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
1130 for_each_node(nid
) {
1131 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
1132 for (i
= 0; i
<= DEF_PRIORITY
; i
++) {
1133 iter
= &mz
->iter
[i
];
1134 cmpxchg(&iter
->position
,
1142 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1143 * @memcg: hierarchy root
1144 * @fn: function to call for each task
1145 * @arg: argument passed to @fn
1147 * This function iterates over tasks attached to @memcg or to any of its
1148 * descendants and calls @fn for each task. If @fn returns a non-zero
1149 * value, the function breaks the iteration loop and returns the value.
1150 * Otherwise, it will iterate over all tasks and return 0.
1152 * This function must not be called for the root memory cgroup.
1154 int mem_cgroup_scan_tasks(struct mem_cgroup
*memcg
,
1155 int (*fn
)(struct task_struct
*, void *), void *arg
)
1157 struct mem_cgroup
*iter
;
1160 BUG_ON(memcg
== root_mem_cgroup
);
1162 for_each_mem_cgroup_tree(iter
, memcg
) {
1163 struct css_task_iter it
;
1164 struct task_struct
*task
;
1166 css_task_iter_start(&iter
->css
, 0, &it
);
1167 while (!ret
&& (task
= css_task_iter_next(&it
)))
1168 ret
= fn(task
, arg
);
1169 css_task_iter_end(&it
);
1171 mem_cgroup_iter_break(memcg
, iter
);
1179 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1181 * @pgdat: pgdat of the page
1183 * This function is only safe when following the LRU page isolation
1184 * and putback protocol: the LRU lock must be held, and the page must
1185 * either be PageLRU() or the caller must have isolated/allocated it.
1187 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct pglist_data
*pgdat
)
1189 struct mem_cgroup_per_node
*mz
;
1190 struct mem_cgroup
*memcg
;
1191 struct lruvec
*lruvec
;
1193 if (mem_cgroup_disabled()) {
1194 lruvec
= &pgdat
->lruvec
;
1198 memcg
= page
->mem_cgroup
;
1200 * Swapcache readahead pages are added to the LRU - and
1201 * possibly migrated - before they are charged.
1204 memcg
= root_mem_cgroup
;
1206 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
1207 lruvec
= &mz
->lruvec
;
1210 * Since a node can be onlined after the mem_cgroup was created,
1211 * we have to be prepared to initialize lruvec->zone here;
1212 * and if offlined then reonlined, we need to reinitialize it.
1214 if (unlikely(lruvec
->pgdat
!= pgdat
))
1215 lruvec
->pgdat
= pgdat
;
1220 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1221 * @lruvec: mem_cgroup per zone lru vector
1222 * @lru: index of lru list the page is sitting on
1223 * @zid: zone id of the accounted pages
1224 * @nr_pages: positive when adding or negative when removing
1226 * This function must be called under lru_lock, just before a page is added
1227 * to or just after a page is removed from an lru list (that ordering being
1228 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1230 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1231 int zid
, int nr_pages
)
1233 struct mem_cgroup_per_node
*mz
;
1234 unsigned long *lru_size
;
1237 if (mem_cgroup_disabled())
1240 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
1241 lru_size
= &mz
->lru_zone_size
[zid
][lru
];
1244 *lru_size
+= nr_pages
;
1247 if (WARN_ONCE(size
< 0,
1248 "%s(%p, %d, %d): lru_size %ld\n",
1249 __func__
, lruvec
, lru
, nr_pages
, size
)) {
1255 *lru_size
+= nr_pages
;
1258 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1260 struct mem_cgroup
*task_memcg
;
1261 struct task_struct
*p
;
1264 p
= find_lock_task_mm(task
);
1266 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1270 * All threads may have already detached their mm's, but the oom
1271 * killer still needs to detect if they have already been oom
1272 * killed to prevent needlessly killing additional tasks.
1275 task_memcg
= mem_cgroup_from_task(task
);
1276 css_get(&task_memcg
->css
);
1279 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1280 css_put(&task_memcg
->css
);
1285 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1286 * @memcg: the memory cgroup
1288 * Returns the maximum amount of memory @mem can be charged with, in
1291 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1293 unsigned long margin
= 0;
1294 unsigned long count
;
1295 unsigned long limit
;
1297 count
= page_counter_read(&memcg
->memory
);
1298 limit
= READ_ONCE(memcg
->memory
.max
);
1300 margin
= limit
- count
;
1302 if (do_memsw_account()) {
1303 count
= page_counter_read(&memcg
->memsw
);
1304 limit
= READ_ONCE(memcg
->memsw
.max
);
1306 margin
= min(margin
, limit
- count
);
1315 * A routine for checking "mem" is under move_account() or not.
1317 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1318 * moving cgroups. This is for waiting at high-memory pressure
1321 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1323 struct mem_cgroup
*from
;
1324 struct mem_cgroup
*to
;
1327 * Unlike task_move routines, we access mc.to, mc.from not under
1328 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1330 spin_lock(&mc
.lock
);
1336 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1337 mem_cgroup_is_descendant(to
, memcg
);
1339 spin_unlock(&mc
.lock
);
1343 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1345 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1346 if (mem_cgroup_under_move(memcg
)) {
1348 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1349 /* moving charge context might have finished. */
1352 finish_wait(&mc
.waitq
, &wait
);
1359 static const unsigned int memcg1_stats
[] = {
1370 static const char *const memcg1_stat_names
[] = {
1381 #define K(x) ((x) << (PAGE_SHIFT-10))
1383 * mem_cgroup_print_oom_context: Print OOM information relevant to
1384 * memory controller.
1385 * @memcg: The memory cgroup that went over limit
1386 * @p: Task that is going to be killed
1388 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1391 void mem_cgroup_print_oom_context(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1396 pr_cont(",oom_memcg=");
1397 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1399 pr_cont(",global_oom");
1401 pr_cont(",task_memcg=");
1402 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1408 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1409 * memory controller.
1410 * @memcg: The memory cgroup that went over limit
1412 void mem_cgroup_print_oom_meminfo(struct mem_cgroup
*memcg
)
1414 struct mem_cgroup
*iter
;
1417 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1418 K((u64
)page_counter_read(&memcg
->memory
)),
1419 K((u64
)memcg
->memory
.max
), memcg
->memory
.failcnt
);
1420 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1421 K((u64
)page_counter_read(&memcg
->memsw
)),
1422 K((u64
)memcg
->memsw
.max
), memcg
->memsw
.failcnt
);
1423 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1424 K((u64
)page_counter_read(&memcg
->kmem
)),
1425 K((u64
)memcg
->kmem
.max
), memcg
->kmem
.failcnt
);
1427 for_each_mem_cgroup_tree(iter
, memcg
) {
1428 pr_info("Memory cgroup stats for ");
1429 pr_cont_cgroup_path(iter
->css
.cgroup
);
1432 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
1433 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_swap_account
)
1435 pr_cont(" %s:%luKB", memcg1_stat_names
[i
],
1436 K(memcg_page_state_local(iter
,
1440 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1441 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1442 K(memcg_page_state_local(iter
,
1450 * Return the memory (and swap, if configured) limit for a memcg.
1452 unsigned long mem_cgroup_get_max(struct mem_cgroup
*memcg
)
1456 max
= memcg
->memory
.max
;
1457 if (mem_cgroup_swappiness(memcg
)) {
1458 unsigned long memsw_max
;
1459 unsigned long swap_max
;
1461 memsw_max
= memcg
->memsw
.max
;
1462 swap_max
= memcg
->swap
.max
;
1463 swap_max
= min(swap_max
, (unsigned long)total_swap_pages
);
1464 max
= min(max
+ swap_max
, memsw_max
);
1469 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1472 struct oom_control oc
= {
1476 .gfp_mask
= gfp_mask
,
1481 if (mutex_lock_killable(&oom_lock
))
1484 * A few threads which were not waiting at mutex_lock_killable() can
1485 * fail to bail out. Therefore, check again after holding oom_lock.
1487 ret
= should_force_charge() || out_of_memory(&oc
);
1488 mutex_unlock(&oom_lock
);
1492 #if MAX_NUMNODES > 1
1495 * test_mem_cgroup_node_reclaimable
1496 * @memcg: the target memcg
1497 * @nid: the node ID to be checked.
1498 * @noswap : specify true here if the user wants flle only information.
1500 * This function returns whether the specified memcg contains any
1501 * reclaimable pages on a node. Returns true if there are any reclaimable
1502 * pages in the node.
1504 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1505 int nid
, bool noswap
)
1507 struct lruvec
*lruvec
= mem_cgroup_lruvec(NODE_DATA(nid
), memcg
);
1509 if (lruvec_page_state(lruvec
, NR_INACTIVE_FILE
) ||
1510 lruvec_page_state(lruvec
, NR_ACTIVE_FILE
))
1512 if (noswap
|| !total_swap_pages
)
1514 if (lruvec_page_state(lruvec
, NR_INACTIVE_ANON
) ||
1515 lruvec_page_state(lruvec
, NR_ACTIVE_ANON
))
1522 * Always updating the nodemask is not very good - even if we have an empty
1523 * list or the wrong list here, we can start from some node and traverse all
1524 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1527 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1531 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1532 * pagein/pageout changes since the last update.
1534 if (!atomic_read(&memcg
->numainfo_events
))
1536 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1539 /* make a nodemask where this memcg uses memory from */
1540 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1542 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1544 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1545 node_clear(nid
, memcg
->scan_nodes
);
1548 atomic_set(&memcg
->numainfo_events
, 0);
1549 atomic_set(&memcg
->numainfo_updating
, 0);
1553 * Selecting a node where we start reclaim from. Because what we need is just
1554 * reducing usage counter, start from anywhere is O,K. Considering
1555 * memory reclaim from current node, there are pros. and cons.
1557 * Freeing memory from current node means freeing memory from a node which
1558 * we'll use or we've used. So, it may make LRU bad. And if several threads
1559 * hit limits, it will see a contention on a node. But freeing from remote
1560 * node means more costs for memory reclaim because of memory latency.
1562 * Now, we use round-robin. Better algorithm is welcomed.
1564 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1568 mem_cgroup_may_update_nodemask(memcg
);
1569 node
= memcg
->last_scanned_node
;
1571 node
= next_node_in(node
, memcg
->scan_nodes
);
1573 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1574 * last time it really checked all the LRUs due to rate limiting.
1575 * Fallback to the current node in that case for simplicity.
1577 if (unlikely(node
== MAX_NUMNODES
))
1578 node
= numa_node_id();
1580 memcg
->last_scanned_node
= node
;
1584 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1590 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1593 unsigned long *total_scanned
)
1595 struct mem_cgroup
*victim
= NULL
;
1598 unsigned long excess
;
1599 unsigned long nr_scanned
;
1600 struct mem_cgroup_reclaim_cookie reclaim
= {
1605 excess
= soft_limit_excess(root_memcg
);
1608 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1613 * If we have not been able to reclaim
1614 * anything, it might because there are
1615 * no reclaimable pages under this hierarchy
1620 * We want to do more targeted reclaim.
1621 * excess >> 2 is not to excessive so as to
1622 * reclaim too much, nor too less that we keep
1623 * coming back to reclaim from this cgroup
1625 if (total
>= (excess
>> 2) ||
1626 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1631 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1632 pgdat
, &nr_scanned
);
1633 *total_scanned
+= nr_scanned
;
1634 if (!soft_limit_excess(root_memcg
))
1637 mem_cgroup_iter_break(root_memcg
, victim
);
1641 #ifdef CONFIG_LOCKDEP
1642 static struct lockdep_map memcg_oom_lock_dep_map
= {
1643 .name
= "memcg_oom_lock",
1647 static DEFINE_SPINLOCK(memcg_oom_lock
);
1650 * Check OOM-Killer is already running under our hierarchy.
1651 * If someone is running, return false.
1653 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1655 struct mem_cgroup
*iter
, *failed
= NULL
;
1657 spin_lock(&memcg_oom_lock
);
1659 for_each_mem_cgroup_tree(iter
, memcg
) {
1660 if (iter
->oom_lock
) {
1662 * this subtree of our hierarchy is already locked
1663 * so we cannot give a lock.
1666 mem_cgroup_iter_break(memcg
, iter
);
1669 iter
->oom_lock
= true;
1674 * OK, we failed to lock the whole subtree so we have
1675 * to clean up what we set up to the failing subtree
1677 for_each_mem_cgroup_tree(iter
, memcg
) {
1678 if (iter
== failed
) {
1679 mem_cgroup_iter_break(memcg
, iter
);
1682 iter
->oom_lock
= false;
1685 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1687 spin_unlock(&memcg_oom_lock
);
1692 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1694 struct mem_cgroup
*iter
;
1696 spin_lock(&memcg_oom_lock
);
1697 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1698 for_each_mem_cgroup_tree(iter
, memcg
)
1699 iter
->oom_lock
= false;
1700 spin_unlock(&memcg_oom_lock
);
1703 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1705 struct mem_cgroup
*iter
;
1707 spin_lock(&memcg_oom_lock
);
1708 for_each_mem_cgroup_tree(iter
, memcg
)
1710 spin_unlock(&memcg_oom_lock
);
1713 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1715 struct mem_cgroup
*iter
;
1718 * When a new child is created while the hierarchy is under oom,
1719 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1721 spin_lock(&memcg_oom_lock
);
1722 for_each_mem_cgroup_tree(iter
, memcg
)
1723 if (iter
->under_oom
> 0)
1725 spin_unlock(&memcg_oom_lock
);
1728 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1730 struct oom_wait_info
{
1731 struct mem_cgroup
*memcg
;
1732 wait_queue_entry_t wait
;
1735 static int memcg_oom_wake_function(wait_queue_entry_t
*wait
,
1736 unsigned mode
, int sync
, void *arg
)
1738 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1739 struct mem_cgroup
*oom_wait_memcg
;
1740 struct oom_wait_info
*oom_wait_info
;
1742 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1743 oom_wait_memcg
= oom_wait_info
->memcg
;
1745 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1746 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1748 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1751 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1754 * For the following lockless ->under_oom test, the only required
1755 * guarantee is that it must see the state asserted by an OOM when
1756 * this function is called as a result of userland actions
1757 * triggered by the notification of the OOM. This is trivially
1758 * achieved by invoking mem_cgroup_mark_under_oom() before
1759 * triggering notification.
1761 if (memcg
&& memcg
->under_oom
)
1762 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1772 static enum oom_status
mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1774 enum oom_status ret
;
1777 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1780 memcg_memory_event(memcg
, MEMCG_OOM
);
1783 * We are in the middle of the charge context here, so we
1784 * don't want to block when potentially sitting on a callstack
1785 * that holds all kinds of filesystem and mm locks.
1787 * cgroup1 allows disabling the OOM killer and waiting for outside
1788 * handling until the charge can succeed; remember the context and put
1789 * the task to sleep at the end of the page fault when all locks are
1792 * On the other hand, in-kernel OOM killer allows for an async victim
1793 * memory reclaim (oom_reaper) and that means that we are not solely
1794 * relying on the oom victim to make a forward progress and we can
1795 * invoke the oom killer here.
1797 * Please note that mem_cgroup_out_of_memory might fail to find a
1798 * victim and then we have to bail out from the charge path.
1800 if (memcg
->oom_kill_disable
) {
1801 if (!current
->in_user_fault
)
1803 css_get(&memcg
->css
);
1804 current
->memcg_in_oom
= memcg
;
1805 current
->memcg_oom_gfp_mask
= mask
;
1806 current
->memcg_oom_order
= order
;
1811 mem_cgroup_mark_under_oom(memcg
);
1813 locked
= mem_cgroup_oom_trylock(memcg
);
1816 mem_cgroup_oom_notify(memcg
);
1818 mem_cgroup_unmark_under_oom(memcg
);
1819 if (mem_cgroup_out_of_memory(memcg
, mask
, order
))
1825 mem_cgroup_oom_unlock(memcg
);
1831 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1832 * @handle: actually kill/wait or just clean up the OOM state
1834 * This has to be called at the end of a page fault if the memcg OOM
1835 * handler was enabled.
1837 * Memcg supports userspace OOM handling where failed allocations must
1838 * sleep on a waitqueue until the userspace task resolves the
1839 * situation. Sleeping directly in the charge context with all kinds
1840 * of locks held is not a good idea, instead we remember an OOM state
1841 * in the task and mem_cgroup_oom_synchronize() has to be called at
1842 * the end of the page fault to complete the OOM handling.
1844 * Returns %true if an ongoing memcg OOM situation was detected and
1845 * completed, %false otherwise.
1847 bool mem_cgroup_oom_synchronize(bool handle
)
1849 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1850 struct oom_wait_info owait
;
1853 /* OOM is global, do not handle */
1860 owait
.memcg
= memcg
;
1861 owait
.wait
.flags
= 0;
1862 owait
.wait
.func
= memcg_oom_wake_function
;
1863 owait
.wait
.private = current
;
1864 INIT_LIST_HEAD(&owait
.wait
.entry
);
1866 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1867 mem_cgroup_mark_under_oom(memcg
);
1869 locked
= mem_cgroup_oom_trylock(memcg
);
1872 mem_cgroup_oom_notify(memcg
);
1874 if (locked
&& !memcg
->oom_kill_disable
) {
1875 mem_cgroup_unmark_under_oom(memcg
);
1876 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1877 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1878 current
->memcg_oom_order
);
1881 mem_cgroup_unmark_under_oom(memcg
);
1882 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1886 mem_cgroup_oom_unlock(memcg
);
1888 * There is no guarantee that an OOM-lock contender
1889 * sees the wakeups triggered by the OOM kill
1890 * uncharges. Wake any sleepers explicitely.
1892 memcg_oom_recover(memcg
);
1895 current
->memcg_in_oom
= NULL
;
1896 css_put(&memcg
->css
);
1901 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1902 * @victim: task to be killed by the OOM killer
1903 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1905 * Returns a pointer to a memory cgroup, which has to be cleaned up
1906 * by killing all belonging OOM-killable tasks.
1908 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1910 struct mem_cgroup
*mem_cgroup_get_oom_group(struct task_struct
*victim
,
1911 struct mem_cgroup
*oom_domain
)
1913 struct mem_cgroup
*oom_group
= NULL
;
1914 struct mem_cgroup
*memcg
;
1916 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1920 oom_domain
= root_mem_cgroup
;
1924 memcg
= mem_cgroup_from_task(victim
);
1925 if (memcg
== root_mem_cgroup
)
1929 * Traverse the memory cgroup hierarchy from the victim task's
1930 * cgroup up to the OOMing cgroup (or root) to find the
1931 * highest-level memory cgroup with oom.group set.
1933 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
1934 if (memcg
->oom_group
)
1937 if (memcg
== oom_domain
)
1942 css_get(&oom_group
->css
);
1949 void mem_cgroup_print_oom_group(struct mem_cgroup
*memcg
)
1951 pr_info("Tasks in ");
1952 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1953 pr_cont(" are going to be killed due to memory.oom.group set\n");
1957 * lock_page_memcg - lock a page->mem_cgroup binding
1960 * This function protects unlocked LRU pages from being moved to
1963 * It ensures lifetime of the returned memcg. Caller is responsible
1964 * for the lifetime of the page; __unlock_page_memcg() is available
1965 * when @page might get freed inside the locked section.
1967 struct mem_cgroup
*lock_page_memcg(struct page
*page
)
1969 struct mem_cgroup
*memcg
;
1970 unsigned long flags
;
1973 * The RCU lock is held throughout the transaction. The fast
1974 * path can get away without acquiring the memcg->move_lock
1975 * because page moving starts with an RCU grace period.
1977 * The RCU lock also protects the memcg from being freed when
1978 * the page state that is going to change is the only thing
1979 * preventing the page itself from being freed. E.g. writeback
1980 * doesn't hold a page reference and relies on PG_writeback to
1981 * keep off truncation, migration and so forth.
1985 if (mem_cgroup_disabled())
1988 memcg
= page
->mem_cgroup
;
1989 if (unlikely(!memcg
))
1992 if (atomic_read(&memcg
->moving_account
) <= 0)
1995 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1996 if (memcg
!= page
->mem_cgroup
) {
1997 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2002 * When charge migration first begins, we can have locked and
2003 * unlocked page stat updates happening concurrently. Track
2004 * the task who has the lock for unlock_page_memcg().
2006 memcg
->move_lock_task
= current
;
2007 memcg
->move_lock_flags
= flags
;
2011 EXPORT_SYMBOL(lock_page_memcg
);
2014 * __unlock_page_memcg - unlock and unpin a memcg
2017 * Unlock and unpin a memcg returned by lock_page_memcg().
2019 void __unlock_page_memcg(struct mem_cgroup
*memcg
)
2021 if (memcg
&& memcg
->move_lock_task
== current
) {
2022 unsigned long flags
= memcg
->move_lock_flags
;
2024 memcg
->move_lock_task
= NULL
;
2025 memcg
->move_lock_flags
= 0;
2027 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2034 * unlock_page_memcg - unlock a page->mem_cgroup binding
2037 void unlock_page_memcg(struct page
*page
)
2039 __unlock_page_memcg(page
->mem_cgroup
);
2041 EXPORT_SYMBOL(unlock_page_memcg
);
2043 struct memcg_stock_pcp
{
2044 struct mem_cgroup
*cached
; /* this never be root cgroup */
2045 unsigned int nr_pages
;
2046 struct work_struct work
;
2047 unsigned long flags
;
2048 #define FLUSHING_CACHED_CHARGE 0
2050 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2051 static DEFINE_MUTEX(percpu_charge_mutex
);
2054 * consume_stock: Try to consume stocked charge on this cpu.
2055 * @memcg: memcg to consume from.
2056 * @nr_pages: how many pages to charge.
2058 * The charges will only happen if @memcg matches the current cpu's memcg
2059 * stock, and at least @nr_pages are available in that stock. Failure to
2060 * service an allocation will refill the stock.
2062 * returns true if successful, false otherwise.
2064 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2066 struct memcg_stock_pcp
*stock
;
2067 unsigned long flags
;
2070 if (nr_pages
> MEMCG_CHARGE_BATCH
)
2073 local_irq_save(flags
);
2075 stock
= this_cpu_ptr(&memcg_stock
);
2076 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
2077 stock
->nr_pages
-= nr_pages
;
2081 local_irq_restore(flags
);
2087 * Returns stocks cached in percpu and reset cached information.
2089 static void drain_stock(struct memcg_stock_pcp
*stock
)
2091 struct mem_cgroup
*old
= stock
->cached
;
2093 if (stock
->nr_pages
) {
2094 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2095 if (do_memsw_account())
2096 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2097 css_put_many(&old
->css
, stock
->nr_pages
);
2098 stock
->nr_pages
= 0;
2100 stock
->cached
= NULL
;
2103 static void drain_local_stock(struct work_struct
*dummy
)
2105 struct memcg_stock_pcp
*stock
;
2106 unsigned long flags
;
2109 * The only protection from memory hotplug vs. drain_stock races is
2110 * that we always operate on local CPU stock here with IRQ disabled
2112 local_irq_save(flags
);
2114 stock
= this_cpu_ptr(&memcg_stock
);
2116 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2118 local_irq_restore(flags
);
2122 * Cache charges(val) to local per_cpu area.
2123 * This will be consumed by consume_stock() function, later.
2125 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2127 struct memcg_stock_pcp
*stock
;
2128 unsigned long flags
;
2130 local_irq_save(flags
);
2132 stock
= this_cpu_ptr(&memcg_stock
);
2133 if (stock
->cached
!= memcg
) { /* reset if necessary */
2135 stock
->cached
= memcg
;
2137 stock
->nr_pages
+= nr_pages
;
2139 if (stock
->nr_pages
> MEMCG_CHARGE_BATCH
)
2142 local_irq_restore(flags
);
2146 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2147 * of the hierarchy under it.
2149 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2153 /* If someone's already draining, avoid adding running more workers. */
2154 if (!mutex_trylock(&percpu_charge_mutex
))
2157 * Notify other cpus that system-wide "drain" is running
2158 * We do not care about races with the cpu hotplug because cpu down
2159 * as well as workers from this path always operate on the local
2160 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2163 for_each_online_cpu(cpu
) {
2164 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2165 struct mem_cgroup
*memcg
;
2167 memcg
= stock
->cached
;
2168 if (!memcg
|| !stock
->nr_pages
|| !css_tryget(&memcg
->css
))
2170 if (!mem_cgroup_is_descendant(memcg
, root_memcg
)) {
2171 css_put(&memcg
->css
);
2174 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2176 drain_local_stock(&stock
->work
);
2178 schedule_work_on(cpu
, &stock
->work
);
2180 css_put(&memcg
->css
);
2183 mutex_unlock(&percpu_charge_mutex
);
2186 static int memcg_hotplug_cpu_dead(unsigned int cpu
)
2188 struct memcg_stock_pcp
*stock
;
2189 struct mem_cgroup
*memcg
, *mi
;
2191 stock
= &per_cpu(memcg_stock
, cpu
);
2194 for_each_mem_cgroup(memcg
) {
2197 for (i
= 0; i
< MEMCG_NR_STAT
; i
++) {
2201 x
= this_cpu_xchg(memcg
->vmstats_percpu
->stat
[i
], 0);
2203 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2204 atomic_long_add(x
, &memcg
->vmstats
[i
]);
2206 if (i
>= NR_VM_NODE_STAT_ITEMS
)
2209 for_each_node(nid
) {
2210 struct mem_cgroup_per_node
*pn
;
2212 pn
= mem_cgroup_nodeinfo(memcg
, nid
);
2213 x
= this_cpu_xchg(pn
->lruvec_stat_cpu
->count
[i
], 0);
2216 atomic_long_add(x
, &pn
->lruvec_stat
[i
]);
2217 } while ((pn
= parent_nodeinfo(pn
, nid
)));
2221 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++) {
2224 x
= this_cpu_xchg(memcg
->vmstats_percpu
->events
[i
], 0);
2226 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2227 atomic_long_add(x
, &memcg
->vmevents
[i
]);
2234 static void reclaim_high(struct mem_cgroup
*memcg
,
2235 unsigned int nr_pages
,
2239 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
2241 memcg_memory_event(memcg
, MEMCG_HIGH
);
2242 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
2243 } while ((memcg
= parent_mem_cgroup(memcg
)));
2246 static void high_work_func(struct work_struct
*work
)
2248 struct mem_cgroup
*memcg
;
2250 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
2251 reclaim_high(memcg
, MEMCG_CHARGE_BATCH
, GFP_KERNEL
);
2255 * Scheduled by try_charge() to be executed from the userland return path
2256 * and reclaims memory over the high limit.
2258 void mem_cgroup_handle_over_high(void)
2260 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
2261 struct mem_cgroup
*memcg
;
2263 if (likely(!nr_pages
))
2266 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2267 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
2268 css_put(&memcg
->css
);
2269 current
->memcg_nr_pages_over_high
= 0;
2272 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2273 unsigned int nr_pages
)
2275 unsigned int batch
= max(MEMCG_CHARGE_BATCH
, nr_pages
);
2276 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2277 struct mem_cgroup
*mem_over_limit
;
2278 struct page_counter
*counter
;
2279 unsigned long nr_reclaimed
;
2280 bool may_swap
= true;
2281 bool drained
= false;
2283 enum oom_status oom_status
;
2285 if (mem_cgroup_is_root(memcg
))
2288 if (consume_stock(memcg
, nr_pages
))
2291 if (!do_memsw_account() ||
2292 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2293 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2295 if (do_memsw_account())
2296 page_counter_uncharge(&memcg
->memsw
, batch
);
2297 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2299 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2303 if (batch
> nr_pages
) {
2309 * Unlike in global OOM situations, memcg is not in a physical
2310 * memory shortage. Allow dying and OOM-killed tasks to
2311 * bypass the last charges so that they can exit quickly and
2312 * free their memory.
2314 if (unlikely(should_force_charge()))
2318 * Prevent unbounded recursion when reclaim operations need to
2319 * allocate memory. This might exceed the limits temporarily,
2320 * but we prefer facilitating memory reclaim and getting back
2321 * under the limit over triggering OOM kills in these cases.
2323 if (unlikely(current
->flags
& PF_MEMALLOC
))
2326 if (unlikely(task_in_memcg_oom(current
)))
2329 if (!gfpflags_allow_blocking(gfp_mask
))
2332 memcg_memory_event(mem_over_limit
, MEMCG_MAX
);
2334 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2335 gfp_mask
, may_swap
);
2337 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2341 drain_all_stock(mem_over_limit
);
2346 if (gfp_mask
& __GFP_NORETRY
)
2349 * Even though the limit is exceeded at this point, reclaim
2350 * may have been able to free some pages. Retry the charge
2351 * before killing the task.
2353 * Only for regular pages, though: huge pages are rather
2354 * unlikely to succeed so close to the limit, and we fall back
2355 * to regular pages anyway in case of failure.
2357 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2360 * At task move, charge accounts can be doubly counted. So, it's
2361 * better to wait until the end of task_move if something is going on.
2363 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2369 if (gfp_mask
& __GFP_RETRY_MAYFAIL
&& oomed
)
2372 if (gfp_mask
& __GFP_NOFAIL
)
2375 if (fatal_signal_pending(current
))
2379 * keep retrying as long as the memcg oom killer is able to make
2380 * a forward progress or bypass the charge if the oom killer
2381 * couldn't make any progress.
2383 oom_status
= mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2384 get_order(nr_pages
* PAGE_SIZE
));
2385 switch (oom_status
) {
2387 nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2396 if (!(gfp_mask
& __GFP_NOFAIL
))
2400 * The allocation either can't fail or will lead to more memory
2401 * being freed very soon. Allow memory usage go over the limit
2402 * temporarily by force charging it.
2404 page_counter_charge(&memcg
->memory
, nr_pages
);
2405 if (do_memsw_account())
2406 page_counter_charge(&memcg
->memsw
, nr_pages
);
2407 css_get_many(&memcg
->css
, nr_pages
);
2412 css_get_many(&memcg
->css
, batch
);
2413 if (batch
> nr_pages
)
2414 refill_stock(memcg
, batch
- nr_pages
);
2417 * If the hierarchy is above the normal consumption range, schedule
2418 * reclaim on returning to userland. We can perform reclaim here
2419 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2420 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2421 * not recorded as it most likely matches current's and won't
2422 * change in the meantime. As high limit is checked again before
2423 * reclaim, the cost of mismatch is negligible.
2426 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2427 /* Don't bother a random interrupted task */
2428 if (in_interrupt()) {
2429 schedule_work(&memcg
->high_work
);
2432 current
->memcg_nr_pages_over_high
+= batch
;
2433 set_notify_resume(current
);
2436 } while ((memcg
= parent_mem_cgroup(memcg
)));
2441 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2443 if (mem_cgroup_is_root(memcg
))
2446 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2447 if (do_memsw_account())
2448 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2450 css_put_many(&memcg
->css
, nr_pages
);
2453 static void lock_page_lru(struct page
*page
, int *isolated
)
2455 pg_data_t
*pgdat
= page_pgdat(page
);
2457 spin_lock_irq(&pgdat
->lru_lock
);
2458 if (PageLRU(page
)) {
2459 struct lruvec
*lruvec
;
2461 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
2463 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2469 static void unlock_page_lru(struct page
*page
, int isolated
)
2471 pg_data_t
*pgdat
= page_pgdat(page
);
2474 struct lruvec
*lruvec
;
2476 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
2477 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2479 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2481 spin_unlock_irq(&pgdat
->lru_lock
);
2484 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2489 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2492 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2493 * may already be on some other mem_cgroup's LRU. Take care of it.
2496 lock_page_lru(page
, &isolated
);
2499 * Nobody should be changing or seriously looking at
2500 * page->mem_cgroup at this point:
2502 * - the page is uncharged
2504 * - the page is off-LRU
2506 * - an anonymous fault has exclusive page access, except for
2507 * a locked page table
2509 * - a page cache insertion, a swapin fault, or a migration
2510 * have the page locked
2512 page
->mem_cgroup
= memcg
;
2515 unlock_page_lru(page
, isolated
);
2518 #ifdef CONFIG_MEMCG_KMEM
2519 static int memcg_alloc_cache_id(void)
2524 id
= ida_simple_get(&memcg_cache_ida
,
2525 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2529 if (id
< memcg_nr_cache_ids
)
2533 * There's no space for the new id in memcg_caches arrays,
2534 * so we have to grow them.
2536 down_write(&memcg_cache_ids_sem
);
2538 size
= 2 * (id
+ 1);
2539 if (size
< MEMCG_CACHES_MIN_SIZE
)
2540 size
= MEMCG_CACHES_MIN_SIZE
;
2541 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2542 size
= MEMCG_CACHES_MAX_SIZE
;
2544 err
= memcg_update_all_caches(size
);
2546 err
= memcg_update_all_list_lrus(size
);
2548 memcg_nr_cache_ids
= size
;
2550 up_write(&memcg_cache_ids_sem
);
2553 ida_simple_remove(&memcg_cache_ida
, id
);
2559 static void memcg_free_cache_id(int id
)
2561 ida_simple_remove(&memcg_cache_ida
, id
);
2564 struct memcg_kmem_cache_create_work
{
2565 struct mem_cgroup
*memcg
;
2566 struct kmem_cache
*cachep
;
2567 struct work_struct work
;
2570 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2572 struct memcg_kmem_cache_create_work
*cw
=
2573 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2574 struct mem_cgroup
*memcg
= cw
->memcg
;
2575 struct kmem_cache
*cachep
= cw
->cachep
;
2577 memcg_create_kmem_cache(memcg
, cachep
);
2579 css_put(&memcg
->css
);
2584 * Enqueue the creation of a per-memcg kmem_cache.
2586 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2587 struct kmem_cache
*cachep
)
2589 struct memcg_kmem_cache_create_work
*cw
;
2591 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
| __GFP_NOWARN
);
2595 css_get(&memcg
->css
);
2598 cw
->cachep
= cachep
;
2599 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2601 queue_work(memcg_kmem_cache_wq
, &cw
->work
);
2604 static inline bool memcg_kmem_bypass(void)
2606 if (in_interrupt() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
2612 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2613 * @cachep: the original global kmem cache
2615 * Return the kmem_cache we're supposed to use for a slab allocation.
2616 * We try to use the current memcg's version of the cache.
2618 * If the cache does not exist yet, if we are the first user of it, we
2619 * create it asynchronously in a workqueue and let the current allocation
2620 * go through with the original cache.
2622 * This function takes a reference to the cache it returns to assure it
2623 * won't get destroyed while we are working with it. Once the caller is
2624 * done with it, memcg_kmem_put_cache() must be called to release the
2627 struct kmem_cache
*memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2629 struct mem_cgroup
*memcg
;
2630 struct kmem_cache
*memcg_cachep
;
2633 VM_BUG_ON(!is_root_cache(cachep
));
2635 if (memcg_kmem_bypass())
2638 memcg
= get_mem_cgroup_from_current();
2639 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2643 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2644 if (likely(memcg_cachep
))
2645 return memcg_cachep
;
2648 * If we are in a safe context (can wait, and not in interrupt
2649 * context), we could be be predictable and return right away.
2650 * This would guarantee that the allocation being performed
2651 * already belongs in the new cache.
2653 * However, there are some clashes that can arrive from locking.
2654 * For instance, because we acquire the slab_mutex while doing
2655 * memcg_create_kmem_cache, this means no further allocation
2656 * could happen with the slab_mutex held. So it's better to
2659 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2661 css_put(&memcg
->css
);
2666 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2667 * @cachep: the cache returned by memcg_kmem_get_cache
2669 void memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2671 if (!is_root_cache(cachep
))
2672 css_put(&cachep
->memcg_params
.memcg
->css
);
2676 * __memcg_kmem_charge_memcg: charge a kmem page
2677 * @page: page to charge
2678 * @gfp: reclaim mode
2679 * @order: allocation order
2680 * @memcg: memory cgroup to charge
2682 * Returns 0 on success, an error code on failure.
2684 int __memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2685 struct mem_cgroup
*memcg
)
2687 unsigned int nr_pages
= 1 << order
;
2688 struct page_counter
*counter
;
2691 ret
= try_charge(memcg
, gfp
, nr_pages
);
2695 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2696 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2697 cancel_charge(memcg
, nr_pages
);
2701 page
->mem_cgroup
= memcg
;
2707 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2708 * @page: page to charge
2709 * @gfp: reclaim mode
2710 * @order: allocation order
2712 * Returns 0 on success, an error code on failure.
2714 int __memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2716 struct mem_cgroup
*memcg
;
2719 if (memcg_kmem_bypass())
2722 memcg
= get_mem_cgroup_from_current();
2723 if (!mem_cgroup_is_root(memcg
)) {
2724 ret
= __memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2726 __SetPageKmemcg(page
);
2728 css_put(&memcg
->css
);
2732 * __memcg_kmem_uncharge: uncharge a kmem page
2733 * @page: page to uncharge
2734 * @order: allocation order
2736 void __memcg_kmem_uncharge(struct page
*page
, int order
)
2738 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2739 unsigned int nr_pages
= 1 << order
;
2744 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2746 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2747 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2749 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2750 if (do_memsw_account())
2751 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2753 page
->mem_cgroup
= NULL
;
2755 /* slab pages do not have PageKmemcg flag set */
2756 if (PageKmemcg(page
))
2757 __ClearPageKmemcg(page
);
2759 css_put_many(&memcg
->css
, nr_pages
);
2761 #endif /* CONFIG_MEMCG_KMEM */
2763 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2766 * Because tail pages are not marked as "used", set it. We're under
2767 * pgdat->lru_lock and migration entries setup in all page mappings.
2769 void mem_cgroup_split_huge_fixup(struct page
*head
)
2773 if (mem_cgroup_disabled())
2776 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2777 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2779 __mod_memcg_state(head
->mem_cgroup
, MEMCG_RSS_HUGE
, -HPAGE_PMD_NR
);
2781 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2783 #ifdef CONFIG_MEMCG_SWAP
2785 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2786 * @entry: swap entry to be moved
2787 * @from: mem_cgroup which the entry is moved from
2788 * @to: mem_cgroup which the entry is moved to
2790 * It succeeds only when the swap_cgroup's record for this entry is the same
2791 * as the mem_cgroup's id of @from.
2793 * Returns 0 on success, -EINVAL on failure.
2795 * The caller must have charged to @to, IOW, called page_counter_charge() about
2796 * both res and memsw, and called css_get().
2798 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2799 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2801 unsigned short old_id
, new_id
;
2803 old_id
= mem_cgroup_id(from
);
2804 new_id
= mem_cgroup_id(to
);
2806 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2807 mod_memcg_state(from
, MEMCG_SWAP
, -1);
2808 mod_memcg_state(to
, MEMCG_SWAP
, 1);
2814 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2815 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2821 static DEFINE_MUTEX(memcg_max_mutex
);
2823 static int mem_cgroup_resize_max(struct mem_cgroup
*memcg
,
2824 unsigned long max
, bool memsw
)
2826 bool enlarge
= false;
2827 bool drained
= false;
2829 bool limits_invariant
;
2830 struct page_counter
*counter
= memsw
? &memcg
->memsw
: &memcg
->memory
;
2833 if (signal_pending(current
)) {
2838 mutex_lock(&memcg_max_mutex
);
2840 * Make sure that the new limit (memsw or memory limit) doesn't
2841 * break our basic invariant rule memory.max <= memsw.max.
2843 limits_invariant
= memsw
? max
>= memcg
->memory
.max
:
2844 max
<= memcg
->memsw
.max
;
2845 if (!limits_invariant
) {
2846 mutex_unlock(&memcg_max_mutex
);
2850 if (max
> counter
->max
)
2852 ret
= page_counter_set_max(counter
, max
);
2853 mutex_unlock(&memcg_max_mutex
);
2859 drain_all_stock(memcg
);
2864 if (!try_to_free_mem_cgroup_pages(memcg
, 1,
2865 GFP_KERNEL
, !memsw
)) {
2871 if (!ret
&& enlarge
)
2872 memcg_oom_recover(memcg
);
2877 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
2879 unsigned long *total_scanned
)
2881 unsigned long nr_reclaimed
= 0;
2882 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
2883 unsigned long reclaimed
;
2885 struct mem_cgroup_tree_per_node
*mctz
;
2886 unsigned long excess
;
2887 unsigned long nr_scanned
;
2892 mctz
= soft_limit_tree_node(pgdat
->node_id
);
2895 * Do not even bother to check the largest node if the root
2896 * is empty. Do it lockless to prevent lock bouncing. Races
2897 * are acceptable as soft limit is best effort anyway.
2899 if (!mctz
|| RB_EMPTY_ROOT(&mctz
->rb_root
))
2903 * This loop can run a while, specially if mem_cgroup's continuously
2904 * keep exceeding their soft limit and putting the system under
2911 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2916 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
2917 gfp_mask
, &nr_scanned
);
2918 nr_reclaimed
+= reclaimed
;
2919 *total_scanned
+= nr_scanned
;
2920 spin_lock_irq(&mctz
->lock
);
2921 __mem_cgroup_remove_exceeded(mz
, mctz
);
2924 * If we failed to reclaim anything from this memory cgroup
2925 * it is time to move on to the next cgroup
2929 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2931 excess
= soft_limit_excess(mz
->memcg
);
2933 * One school of thought says that we should not add
2934 * back the node to the tree if reclaim returns 0.
2935 * But our reclaim could return 0, simply because due
2936 * to priority we are exposing a smaller subset of
2937 * memory to reclaim from. Consider this as a longer
2940 /* If excess == 0, no tree ops */
2941 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2942 spin_unlock_irq(&mctz
->lock
);
2943 css_put(&mz
->memcg
->css
);
2946 * Could not reclaim anything and there are no more
2947 * mem cgroups to try or we seem to be looping without
2948 * reclaiming anything.
2950 if (!nr_reclaimed
&&
2952 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2954 } while (!nr_reclaimed
);
2956 css_put(&next_mz
->memcg
->css
);
2957 return nr_reclaimed
;
2961 * Test whether @memcg has children, dead or alive. Note that this
2962 * function doesn't care whether @memcg has use_hierarchy enabled and
2963 * returns %true if there are child csses according to the cgroup
2964 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2966 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2971 ret
= css_next_child(NULL
, &memcg
->css
);
2977 * Reclaims as many pages from the given memcg as possible.
2979 * Caller is responsible for holding css reference for memcg.
2981 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2983 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2985 /* we call try-to-free pages for make this cgroup empty */
2986 lru_add_drain_all();
2988 drain_all_stock(memcg
);
2990 /* try to free all pages in this cgroup */
2991 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2994 if (signal_pending(current
))
2997 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
3001 /* maybe some writeback is necessary */
3002 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3010 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3011 char *buf
, size_t nbytes
,
3014 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3016 if (mem_cgroup_is_root(memcg
))
3018 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3021 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3024 return mem_cgroup_from_css(css
)->use_hierarchy
;
3027 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3028 struct cftype
*cft
, u64 val
)
3031 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3032 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
3034 if (memcg
->use_hierarchy
== val
)
3038 * If parent's use_hierarchy is set, we can't make any modifications
3039 * in the child subtrees. If it is unset, then the change can
3040 * occur, provided the current cgroup has no children.
3042 * For the root cgroup, parent_mem is NULL, we allow value to be
3043 * set if there are no children.
3045 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
3046 (val
== 1 || val
== 0)) {
3047 if (!memcg_has_children(memcg
))
3048 memcg
->use_hierarchy
= val
;
3057 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3061 if (mem_cgroup_is_root(memcg
)) {
3062 val
= memcg_page_state(memcg
, MEMCG_CACHE
) +
3063 memcg_page_state(memcg
, MEMCG_RSS
);
3065 val
+= memcg_page_state(memcg
, MEMCG_SWAP
);
3068 val
= page_counter_read(&memcg
->memory
);
3070 val
= page_counter_read(&memcg
->memsw
);
3083 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3086 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3087 struct page_counter
*counter
;
3089 switch (MEMFILE_TYPE(cft
->private)) {
3091 counter
= &memcg
->memory
;
3094 counter
= &memcg
->memsw
;
3097 counter
= &memcg
->kmem
;
3100 counter
= &memcg
->tcpmem
;
3106 switch (MEMFILE_ATTR(cft
->private)) {
3108 if (counter
== &memcg
->memory
)
3109 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
3110 if (counter
== &memcg
->memsw
)
3111 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
3112 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3114 return (u64
)counter
->max
* PAGE_SIZE
;
3116 return (u64
)counter
->watermark
* PAGE_SIZE
;
3118 return counter
->failcnt
;
3119 case RES_SOFT_LIMIT
:
3120 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
3126 #ifdef CONFIG_MEMCG_KMEM
3127 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3131 if (cgroup_memory_nokmem
)
3134 BUG_ON(memcg
->kmemcg_id
>= 0);
3135 BUG_ON(memcg
->kmem_state
);
3137 memcg_id
= memcg_alloc_cache_id();
3141 static_branch_inc(&memcg_kmem_enabled_key
);
3143 * A memory cgroup is considered kmem-online as soon as it gets
3144 * kmemcg_id. Setting the id after enabling static branching will
3145 * guarantee no one starts accounting before all call sites are
3148 memcg
->kmemcg_id
= memcg_id
;
3149 memcg
->kmem_state
= KMEM_ONLINE
;
3150 INIT_LIST_HEAD(&memcg
->kmem_caches
);
3155 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3157 struct cgroup_subsys_state
*css
;
3158 struct mem_cgroup
*parent
, *child
;
3161 if (memcg
->kmem_state
!= KMEM_ONLINE
)
3164 * Clear the online state before clearing memcg_caches array
3165 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3166 * guarantees that no cache will be created for this cgroup
3167 * after we are done (see memcg_create_kmem_cache()).
3169 memcg
->kmem_state
= KMEM_ALLOCATED
;
3171 memcg_deactivate_kmem_caches(memcg
);
3173 kmemcg_id
= memcg
->kmemcg_id
;
3174 BUG_ON(kmemcg_id
< 0);
3176 parent
= parent_mem_cgroup(memcg
);
3178 parent
= root_mem_cgroup
;
3181 * Change kmemcg_id of this cgroup and all its descendants to the
3182 * parent's id, and then move all entries from this cgroup's list_lrus
3183 * to ones of the parent. After we have finished, all list_lrus
3184 * corresponding to this cgroup are guaranteed to remain empty. The
3185 * ordering is imposed by list_lru_node->lock taken by
3186 * memcg_drain_all_list_lrus().
3188 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3189 css_for_each_descendant_pre(css
, &memcg
->css
) {
3190 child
= mem_cgroup_from_css(css
);
3191 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
3192 child
->kmemcg_id
= parent
->kmemcg_id
;
3193 if (!memcg
->use_hierarchy
)
3198 memcg_drain_all_list_lrus(kmemcg_id
, parent
);
3200 memcg_free_cache_id(kmemcg_id
);
3203 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3205 /* css_alloc() failed, offlining didn't happen */
3206 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
3207 memcg_offline_kmem(memcg
);
3209 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
3210 memcg_destroy_kmem_caches(memcg
);
3211 static_branch_dec(&memcg_kmem_enabled_key
);
3212 WARN_ON(page_counter_read(&memcg
->kmem
));
3216 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3220 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3223 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3226 #endif /* CONFIG_MEMCG_KMEM */
3228 static int memcg_update_kmem_max(struct mem_cgroup
*memcg
,
3233 mutex_lock(&memcg_max_mutex
);
3234 ret
= page_counter_set_max(&memcg
->kmem
, max
);
3235 mutex_unlock(&memcg_max_mutex
);
3239 static int memcg_update_tcp_max(struct mem_cgroup
*memcg
, unsigned long max
)
3243 mutex_lock(&memcg_max_mutex
);
3245 ret
= page_counter_set_max(&memcg
->tcpmem
, max
);
3249 if (!memcg
->tcpmem_active
) {
3251 * The active flag needs to be written after the static_key
3252 * update. This is what guarantees that the socket activation
3253 * function is the last one to run. See mem_cgroup_sk_alloc()
3254 * for details, and note that we don't mark any socket as
3255 * belonging to this memcg until that flag is up.
3257 * We need to do this, because static_keys will span multiple
3258 * sites, but we can't control their order. If we mark a socket
3259 * as accounted, but the accounting functions are not patched in
3260 * yet, we'll lose accounting.
3262 * We never race with the readers in mem_cgroup_sk_alloc(),
3263 * because when this value change, the code to process it is not
3266 static_branch_inc(&memcg_sockets_enabled_key
);
3267 memcg
->tcpmem_active
= true;
3270 mutex_unlock(&memcg_max_mutex
);
3275 * The user of this function is...
3278 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3279 char *buf
, size_t nbytes
, loff_t off
)
3281 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3282 unsigned long nr_pages
;
3285 buf
= strstrip(buf
);
3286 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3290 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3292 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3296 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3298 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, false);
3301 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, true);
3304 ret
= memcg_update_kmem_max(memcg
, nr_pages
);
3307 ret
= memcg_update_tcp_max(memcg
, nr_pages
);
3311 case RES_SOFT_LIMIT
:
3312 memcg
->soft_limit
= nr_pages
;
3316 return ret
?: nbytes
;
3319 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3320 size_t nbytes
, loff_t off
)
3322 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3323 struct page_counter
*counter
;
3325 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3327 counter
= &memcg
->memory
;
3330 counter
= &memcg
->memsw
;
3333 counter
= &memcg
->kmem
;
3336 counter
= &memcg
->tcpmem
;
3342 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3344 page_counter_reset_watermark(counter
);
3347 counter
->failcnt
= 0;
3356 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3359 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3363 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3364 struct cftype
*cft
, u64 val
)
3366 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3368 if (val
& ~MOVE_MASK
)
3372 * No kind of locking is needed in here, because ->can_attach() will
3373 * check this value once in the beginning of the process, and then carry
3374 * on with stale data. This means that changes to this value will only
3375 * affect task migrations starting after the change.
3377 memcg
->move_charge_at_immigrate
= val
;
3381 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3382 struct cftype
*cft
, u64 val
)
3390 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3391 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3392 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3394 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
3395 int nid
, unsigned int lru_mask
)
3397 struct lruvec
*lruvec
= mem_cgroup_lruvec(NODE_DATA(nid
), memcg
);
3398 unsigned long nr
= 0;
3401 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
3404 if (!(BIT(lru
) & lru_mask
))
3406 nr
+= lruvec_page_state_local(lruvec
, NR_LRU_BASE
+ lru
);
3411 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
3412 unsigned int lru_mask
)
3414 unsigned long nr
= 0;
3418 if (!(BIT(lru
) & lru_mask
))
3420 nr
+= memcg_page_state_local(memcg
, NR_LRU_BASE
+ lru
);
3425 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3429 unsigned int lru_mask
;
3432 static const struct numa_stat stats
[] = {
3433 { "total", LRU_ALL
},
3434 { "file", LRU_ALL_FILE
},
3435 { "anon", LRU_ALL_ANON
},
3436 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3438 const struct numa_stat
*stat
;
3441 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
3443 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3444 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3445 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3446 for_each_node_state(nid
, N_MEMORY
) {
3447 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3449 seq_printf(m
, " N%d=%lu", nid
, nr
);
3454 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3455 struct mem_cgroup
*iter
;
3458 for_each_mem_cgroup_tree(iter
, memcg
)
3459 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3460 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3461 for_each_node_state(nid
, N_MEMORY
) {
3463 for_each_mem_cgroup_tree(iter
, memcg
)
3464 nr
+= mem_cgroup_node_nr_lru_pages(
3465 iter
, nid
, stat
->lru_mask
);
3466 seq_printf(m
, " N%d=%lu", nid
, nr
);
3473 #endif /* CONFIG_NUMA */
3475 /* Universal VM events cgroup1 shows, original sort order */
3476 static const unsigned int memcg1_events
[] = {
3483 static const char *const memcg1_event_names
[] = {
3490 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3492 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
3493 unsigned long memory
, memsw
;
3494 struct mem_cgroup
*mi
;
3497 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names
) != ARRAY_SIZE(memcg1_stats
));
3498 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3500 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
3501 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
3503 seq_printf(m
, "%s %lu\n", memcg1_stat_names
[i
],
3504 memcg_page_state_local(memcg
, memcg1_stats
[i
]) *
3508 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
3509 seq_printf(m
, "%s %lu\n", memcg1_event_names
[i
],
3510 memcg_events_local(memcg
, memcg1_events
[i
]));
3512 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3513 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3514 memcg_page_state_local(memcg
, NR_LRU_BASE
+ i
) *
3517 /* Hierarchical information */
3518 memory
= memsw
= PAGE_COUNTER_MAX
;
3519 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3520 memory
= min(memory
, mi
->memory
.max
);
3521 memsw
= min(memsw
, mi
->memsw
.max
);
3523 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3524 (u64
)memory
* PAGE_SIZE
);
3525 if (do_memsw_account())
3526 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3527 (u64
)memsw
* PAGE_SIZE
);
3529 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
3530 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
3532 seq_printf(m
, "total_%s %llu\n", memcg1_stat_names
[i
],
3533 (u64
)memcg_page_state(memcg
, i
) * PAGE_SIZE
);
3536 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
3537 seq_printf(m
, "total_%s %llu\n", memcg1_event_names
[i
],
3538 (u64
)memcg_events(memcg
, i
));
3540 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3541 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
],
3542 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
3545 #ifdef CONFIG_DEBUG_VM
3548 struct mem_cgroup_per_node
*mz
;
3549 struct zone_reclaim_stat
*rstat
;
3550 unsigned long recent_rotated
[2] = {0, 0};
3551 unsigned long recent_scanned
[2] = {0, 0};
3553 for_each_online_pgdat(pgdat
) {
3554 mz
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
3555 rstat
= &mz
->lruvec
.reclaim_stat
;
3557 recent_rotated
[0] += rstat
->recent_rotated
[0];
3558 recent_rotated
[1] += rstat
->recent_rotated
[1];
3559 recent_scanned
[0] += rstat
->recent_scanned
[0];
3560 recent_scanned
[1] += rstat
->recent_scanned
[1];
3562 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3563 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3564 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3565 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3572 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3575 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3577 return mem_cgroup_swappiness(memcg
);
3580 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3581 struct cftype
*cft
, u64 val
)
3583 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3589 memcg
->swappiness
= val
;
3591 vm_swappiness
= val
;
3596 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3598 struct mem_cgroup_threshold_ary
*t
;
3599 unsigned long usage
;
3604 t
= rcu_dereference(memcg
->thresholds
.primary
);
3606 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3611 usage
= mem_cgroup_usage(memcg
, swap
);
3614 * current_threshold points to threshold just below or equal to usage.
3615 * If it's not true, a threshold was crossed after last
3616 * call of __mem_cgroup_threshold().
3618 i
= t
->current_threshold
;
3621 * Iterate backward over array of thresholds starting from
3622 * current_threshold and check if a threshold is crossed.
3623 * If none of thresholds below usage is crossed, we read
3624 * only one element of the array here.
3626 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3627 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3629 /* i = current_threshold + 1 */
3633 * Iterate forward over array of thresholds starting from
3634 * current_threshold+1 and check if a threshold is crossed.
3635 * If none of thresholds above usage is crossed, we read
3636 * only one element of the array here.
3638 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3639 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3641 /* Update current_threshold */
3642 t
->current_threshold
= i
- 1;
3647 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3650 __mem_cgroup_threshold(memcg
, false);
3651 if (do_memsw_account())
3652 __mem_cgroup_threshold(memcg
, true);
3654 memcg
= parent_mem_cgroup(memcg
);
3658 static int compare_thresholds(const void *a
, const void *b
)
3660 const struct mem_cgroup_threshold
*_a
= a
;
3661 const struct mem_cgroup_threshold
*_b
= b
;
3663 if (_a
->threshold
> _b
->threshold
)
3666 if (_a
->threshold
< _b
->threshold
)
3672 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3674 struct mem_cgroup_eventfd_list
*ev
;
3676 spin_lock(&memcg_oom_lock
);
3678 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3679 eventfd_signal(ev
->eventfd
, 1);
3681 spin_unlock(&memcg_oom_lock
);
3685 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3687 struct mem_cgroup
*iter
;
3689 for_each_mem_cgroup_tree(iter
, memcg
)
3690 mem_cgroup_oom_notify_cb(iter
);
3693 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3694 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3696 struct mem_cgroup_thresholds
*thresholds
;
3697 struct mem_cgroup_threshold_ary
*new;
3698 unsigned long threshold
;
3699 unsigned long usage
;
3702 ret
= page_counter_memparse(args
, "-1", &threshold
);
3706 mutex_lock(&memcg
->thresholds_lock
);
3709 thresholds
= &memcg
->thresholds
;
3710 usage
= mem_cgroup_usage(memcg
, false);
3711 } else if (type
== _MEMSWAP
) {
3712 thresholds
= &memcg
->memsw_thresholds
;
3713 usage
= mem_cgroup_usage(memcg
, true);
3717 /* Check if a threshold crossed before adding a new one */
3718 if (thresholds
->primary
)
3719 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3721 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3723 /* Allocate memory for new array of thresholds */
3724 new = kmalloc(struct_size(new, entries
, size
), GFP_KERNEL
);
3731 /* Copy thresholds (if any) to new array */
3732 if (thresholds
->primary
) {
3733 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3734 sizeof(struct mem_cgroup_threshold
));
3737 /* Add new threshold */
3738 new->entries
[size
- 1].eventfd
= eventfd
;
3739 new->entries
[size
- 1].threshold
= threshold
;
3741 /* Sort thresholds. Registering of new threshold isn't time-critical */
3742 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3743 compare_thresholds
, NULL
);
3745 /* Find current threshold */
3746 new->current_threshold
= -1;
3747 for (i
= 0; i
< size
; i
++) {
3748 if (new->entries
[i
].threshold
<= usage
) {
3750 * new->current_threshold will not be used until
3751 * rcu_assign_pointer(), so it's safe to increment
3754 ++new->current_threshold
;
3759 /* Free old spare buffer and save old primary buffer as spare */
3760 kfree(thresholds
->spare
);
3761 thresholds
->spare
= thresholds
->primary
;
3763 rcu_assign_pointer(thresholds
->primary
, new);
3765 /* To be sure that nobody uses thresholds */
3769 mutex_unlock(&memcg
->thresholds_lock
);
3774 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3775 struct eventfd_ctx
*eventfd
, const char *args
)
3777 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3780 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3781 struct eventfd_ctx
*eventfd
, const char *args
)
3783 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3786 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3787 struct eventfd_ctx
*eventfd
, enum res_type type
)
3789 struct mem_cgroup_thresholds
*thresholds
;
3790 struct mem_cgroup_threshold_ary
*new;
3791 unsigned long usage
;
3794 mutex_lock(&memcg
->thresholds_lock
);
3797 thresholds
= &memcg
->thresholds
;
3798 usage
= mem_cgroup_usage(memcg
, false);
3799 } else if (type
== _MEMSWAP
) {
3800 thresholds
= &memcg
->memsw_thresholds
;
3801 usage
= mem_cgroup_usage(memcg
, true);
3805 if (!thresholds
->primary
)
3808 /* Check if a threshold crossed before removing */
3809 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3811 /* Calculate new number of threshold */
3813 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3814 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3818 new = thresholds
->spare
;
3820 /* Set thresholds array to NULL if we don't have thresholds */
3829 /* Copy thresholds and find current threshold */
3830 new->current_threshold
= -1;
3831 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3832 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3835 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3836 if (new->entries
[j
].threshold
<= usage
) {
3838 * new->current_threshold will not be used
3839 * until rcu_assign_pointer(), so it's safe to increment
3842 ++new->current_threshold
;
3848 /* Swap primary and spare array */
3849 thresholds
->spare
= thresholds
->primary
;
3851 rcu_assign_pointer(thresholds
->primary
, new);
3853 /* To be sure that nobody uses thresholds */
3856 /* If all events are unregistered, free the spare array */
3858 kfree(thresholds
->spare
);
3859 thresholds
->spare
= NULL
;
3862 mutex_unlock(&memcg
->thresholds_lock
);
3865 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3866 struct eventfd_ctx
*eventfd
)
3868 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3871 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3872 struct eventfd_ctx
*eventfd
)
3874 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3877 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3878 struct eventfd_ctx
*eventfd
, const char *args
)
3880 struct mem_cgroup_eventfd_list
*event
;
3882 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3886 spin_lock(&memcg_oom_lock
);
3888 event
->eventfd
= eventfd
;
3889 list_add(&event
->list
, &memcg
->oom_notify
);
3891 /* already in OOM ? */
3892 if (memcg
->under_oom
)
3893 eventfd_signal(eventfd
, 1);
3894 spin_unlock(&memcg_oom_lock
);
3899 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3900 struct eventfd_ctx
*eventfd
)
3902 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3904 spin_lock(&memcg_oom_lock
);
3906 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3907 if (ev
->eventfd
== eventfd
) {
3908 list_del(&ev
->list
);
3913 spin_unlock(&memcg_oom_lock
);
3916 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3918 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(sf
);
3920 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3921 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3922 seq_printf(sf
, "oom_kill %lu\n",
3923 atomic_long_read(&memcg
->memory_events
[MEMCG_OOM_KILL
]));
3927 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3928 struct cftype
*cft
, u64 val
)
3930 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3932 /* cannot set to root cgroup and only 0 and 1 are allowed */
3933 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3936 memcg
->oom_kill_disable
= val
;
3938 memcg_oom_recover(memcg
);
3943 #ifdef CONFIG_CGROUP_WRITEBACK
3945 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3947 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3950 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3952 wb_domain_exit(&memcg
->cgwb_domain
);
3955 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3957 wb_domain_size_changed(&memcg
->cgwb_domain
);
3960 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3962 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3964 if (!memcg
->css
.parent
)
3967 return &memcg
->cgwb_domain
;
3971 * idx can be of type enum memcg_stat_item or node_stat_item.
3972 * Keep in sync with memcg_exact_page().
3974 static unsigned long memcg_exact_page_state(struct mem_cgroup
*memcg
, int idx
)
3976 long x
= atomic_long_read(&memcg
->vmstats
[idx
]);
3979 for_each_online_cpu(cpu
)
3980 x
+= per_cpu_ptr(memcg
->vmstats_percpu
, cpu
)->stat
[idx
];
3987 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3988 * @wb: bdi_writeback in question
3989 * @pfilepages: out parameter for number of file pages
3990 * @pheadroom: out parameter for number of allocatable pages according to memcg
3991 * @pdirty: out parameter for number of dirty pages
3992 * @pwriteback: out parameter for number of pages under writeback
3994 * Determine the numbers of file, headroom, dirty, and writeback pages in
3995 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3996 * is a bit more involved.
3998 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3999 * headroom is calculated as the lowest headroom of itself and the
4000 * ancestors. Note that this doesn't consider the actual amount of
4001 * available memory in the system. The caller should further cap
4002 * *@pheadroom accordingly.
4004 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
4005 unsigned long *pheadroom
, unsigned long *pdirty
,
4006 unsigned long *pwriteback
)
4008 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4009 struct mem_cgroup
*parent
;
4011 *pdirty
= memcg_exact_page_state(memcg
, NR_FILE_DIRTY
);
4013 /* this should eventually include NR_UNSTABLE_NFS */
4014 *pwriteback
= memcg_exact_page_state(memcg
, NR_WRITEBACK
);
4015 *pfilepages
= memcg_exact_page_state(memcg
, NR_INACTIVE_FILE
) +
4016 memcg_exact_page_state(memcg
, NR_ACTIVE_FILE
);
4017 *pheadroom
= PAGE_COUNTER_MAX
;
4019 while ((parent
= parent_mem_cgroup(memcg
))) {
4020 unsigned long ceiling
= min(memcg
->memory
.max
, memcg
->high
);
4021 unsigned long used
= page_counter_read(&memcg
->memory
);
4023 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
4028 #else /* CONFIG_CGROUP_WRITEBACK */
4030 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4035 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4039 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4043 #endif /* CONFIG_CGROUP_WRITEBACK */
4046 * DO NOT USE IN NEW FILES.
4048 * "cgroup.event_control" implementation.
4050 * This is way over-engineered. It tries to support fully configurable
4051 * events for each user. Such level of flexibility is completely
4052 * unnecessary especially in the light of the planned unified hierarchy.
4054 * Please deprecate this and replace with something simpler if at all
4059 * Unregister event and free resources.
4061 * Gets called from workqueue.
4063 static void memcg_event_remove(struct work_struct
*work
)
4065 struct mem_cgroup_event
*event
=
4066 container_of(work
, struct mem_cgroup_event
, remove
);
4067 struct mem_cgroup
*memcg
= event
->memcg
;
4069 remove_wait_queue(event
->wqh
, &event
->wait
);
4071 event
->unregister_event(memcg
, event
->eventfd
);
4073 /* Notify userspace the event is going away. */
4074 eventfd_signal(event
->eventfd
, 1);
4076 eventfd_ctx_put(event
->eventfd
);
4078 css_put(&memcg
->css
);
4082 * Gets called on EPOLLHUP on eventfd when user closes it.
4084 * Called with wqh->lock held and interrupts disabled.
4086 static int memcg_event_wake(wait_queue_entry_t
*wait
, unsigned mode
,
4087 int sync
, void *key
)
4089 struct mem_cgroup_event
*event
=
4090 container_of(wait
, struct mem_cgroup_event
, wait
);
4091 struct mem_cgroup
*memcg
= event
->memcg
;
4092 __poll_t flags
= key_to_poll(key
);
4094 if (flags
& EPOLLHUP
) {
4096 * If the event has been detached at cgroup removal, we
4097 * can simply return knowing the other side will cleanup
4100 * We can't race against event freeing since the other
4101 * side will require wqh->lock via remove_wait_queue(),
4104 spin_lock(&memcg
->event_list_lock
);
4105 if (!list_empty(&event
->list
)) {
4106 list_del_init(&event
->list
);
4108 * We are in atomic context, but cgroup_event_remove()
4109 * may sleep, so we have to call it in workqueue.
4111 schedule_work(&event
->remove
);
4113 spin_unlock(&memcg
->event_list_lock
);
4119 static void memcg_event_ptable_queue_proc(struct file
*file
,
4120 wait_queue_head_t
*wqh
, poll_table
*pt
)
4122 struct mem_cgroup_event
*event
=
4123 container_of(pt
, struct mem_cgroup_event
, pt
);
4126 add_wait_queue(wqh
, &event
->wait
);
4130 * DO NOT USE IN NEW FILES.
4132 * Parse input and register new cgroup event handler.
4134 * Input must be in format '<event_fd> <control_fd> <args>'.
4135 * Interpretation of args is defined by control file implementation.
4137 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4138 char *buf
, size_t nbytes
, loff_t off
)
4140 struct cgroup_subsys_state
*css
= of_css(of
);
4141 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4142 struct mem_cgroup_event
*event
;
4143 struct cgroup_subsys_state
*cfile_css
;
4144 unsigned int efd
, cfd
;
4151 buf
= strstrip(buf
);
4153 efd
= simple_strtoul(buf
, &endp
, 10);
4158 cfd
= simple_strtoul(buf
, &endp
, 10);
4159 if ((*endp
!= ' ') && (*endp
!= '\0'))
4163 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
4167 event
->memcg
= memcg
;
4168 INIT_LIST_HEAD(&event
->list
);
4169 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
4170 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
4171 INIT_WORK(&event
->remove
, memcg_event_remove
);
4179 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
4180 if (IS_ERR(event
->eventfd
)) {
4181 ret
= PTR_ERR(event
->eventfd
);
4188 goto out_put_eventfd
;
4191 /* the process need read permission on control file */
4192 /* AV: shouldn't we check that it's been opened for read instead? */
4193 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
4198 * Determine the event callbacks and set them in @event. This used
4199 * to be done via struct cftype but cgroup core no longer knows
4200 * about these events. The following is crude but the whole thing
4201 * is for compatibility anyway.
4203 * DO NOT ADD NEW FILES.
4205 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
4207 if (!strcmp(name
, "memory.usage_in_bytes")) {
4208 event
->register_event
= mem_cgroup_usage_register_event
;
4209 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
4210 } else if (!strcmp(name
, "memory.oom_control")) {
4211 event
->register_event
= mem_cgroup_oom_register_event
;
4212 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
4213 } else if (!strcmp(name
, "memory.pressure_level")) {
4214 event
->register_event
= vmpressure_register_event
;
4215 event
->unregister_event
= vmpressure_unregister_event
;
4216 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
4217 event
->register_event
= memsw_cgroup_usage_register_event
;
4218 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
4225 * Verify @cfile should belong to @css. Also, remaining events are
4226 * automatically removed on cgroup destruction but the removal is
4227 * asynchronous, so take an extra ref on @css.
4229 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
4230 &memory_cgrp_subsys
);
4232 if (IS_ERR(cfile_css
))
4234 if (cfile_css
!= css
) {
4239 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
4243 vfs_poll(efile
.file
, &event
->pt
);
4245 spin_lock(&memcg
->event_list_lock
);
4246 list_add(&event
->list
, &memcg
->event_list
);
4247 spin_unlock(&memcg
->event_list_lock
);
4259 eventfd_ctx_put(event
->eventfd
);
4268 static struct cftype mem_cgroup_legacy_files
[] = {
4270 .name
= "usage_in_bytes",
4271 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4272 .read_u64
= mem_cgroup_read_u64
,
4275 .name
= "max_usage_in_bytes",
4276 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4277 .write
= mem_cgroup_reset
,
4278 .read_u64
= mem_cgroup_read_u64
,
4281 .name
= "limit_in_bytes",
4282 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4283 .write
= mem_cgroup_write
,
4284 .read_u64
= mem_cgroup_read_u64
,
4287 .name
= "soft_limit_in_bytes",
4288 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4289 .write
= mem_cgroup_write
,
4290 .read_u64
= mem_cgroup_read_u64
,
4294 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4295 .write
= mem_cgroup_reset
,
4296 .read_u64
= mem_cgroup_read_u64
,
4300 .seq_show
= memcg_stat_show
,
4303 .name
= "force_empty",
4304 .write
= mem_cgroup_force_empty_write
,
4307 .name
= "use_hierarchy",
4308 .write_u64
= mem_cgroup_hierarchy_write
,
4309 .read_u64
= mem_cgroup_hierarchy_read
,
4312 .name
= "cgroup.event_control", /* XXX: for compat */
4313 .write
= memcg_write_event_control
,
4314 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
4317 .name
= "swappiness",
4318 .read_u64
= mem_cgroup_swappiness_read
,
4319 .write_u64
= mem_cgroup_swappiness_write
,
4322 .name
= "move_charge_at_immigrate",
4323 .read_u64
= mem_cgroup_move_charge_read
,
4324 .write_u64
= mem_cgroup_move_charge_write
,
4327 .name
= "oom_control",
4328 .seq_show
= mem_cgroup_oom_control_read
,
4329 .write_u64
= mem_cgroup_oom_control_write
,
4330 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4333 .name
= "pressure_level",
4337 .name
= "numa_stat",
4338 .seq_show
= memcg_numa_stat_show
,
4342 .name
= "kmem.limit_in_bytes",
4343 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4344 .write
= mem_cgroup_write
,
4345 .read_u64
= mem_cgroup_read_u64
,
4348 .name
= "kmem.usage_in_bytes",
4349 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4350 .read_u64
= mem_cgroup_read_u64
,
4353 .name
= "kmem.failcnt",
4354 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4355 .write
= mem_cgroup_reset
,
4356 .read_u64
= mem_cgroup_read_u64
,
4359 .name
= "kmem.max_usage_in_bytes",
4360 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4361 .write
= mem_cgroup_reset
,
4362 .read_u64
= mem_cgroup_read_u64
,
4364 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4366 .name
= "kmem.slabinfo",
4367 .seq_start
= memcg_slab_start
,
4368 .seq_next
= memcg_slab_next
,
4369 .seq_stop
= memcg_slab_stop
,
4370 .seq_show
= memcg_slab_show
,
4374 .name
= "kmem.tcp.limit_in_bytes",
4375 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4376 .write
= mem_cgroup_write
,
4377 .read_u64
= mem_cgroup_read_u64
,
4380 .name
= "kmem.tcp.usage_in_bytes",
4381 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4382 .read_u64
= mem_cgroup_read_u64
,
4385 .name
= "kmem.tcp.failcnt",
4386 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4387 .write
= mem_cgroup_reset
,
4388 .read_u64
= mem_cgroup_read_u64
,
4391 .name
= "kmem.tcp.max_usage_in_bytes",
4392 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4393 .write
= mem_cgroup_reset
,
4394 .read_u64
= mem_cgroup_read_u64
,
4396 { }, /* terminate */
4400 * Private memory cgroup IDR
4402 * Swap-out records and page cache shadow entries need to store memcg
4403 * references in constrained space, so we maintain an ID space that is
4404 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4405 * memory-controlled cgroups to 64k.
4407 * However, there usually are many references to the oflline CSS after
4408 * the cgroup has been destroyed, such as page cache or reclaimable
4409 * slab objects, that don't need to hang on to the ID. We want to keep
4410 * those dead CSS from occupying IDs, or we might quickly exhaust the
4411 * relatively small ID space and prevent the creation of new cgroups
4412 * even when there are much fewer than 64k cgroups - possibly none.
4414 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4415 * be freed and recycled when it's no longer needed, which is usually
4416 * when the CSS is offlined.
4418 * The only exception to that are records of swapped out tmpfs/shmem
4419 * pages that need to be attributed to live ancestors on swapin. But
4420 * those references are manageable from userspace.
4423 static DEFINE_IDR(mem_cgroup_idr
);
4425 static void mem_cgroup_id_remove(struct mem_cgroup
*memcg
)
4427 if (memcg
->id
.id
> 0) {
4428 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4433 static void mem_cgroup_id_get_many(struct mem_cgroup
*memcg
, unsigned int n
)
4435 refcount_add(n
, &memcg
->id
.ref
);
4438 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
4440 if (refcount_sub_and_test(n
, &memcg
->id
.ref
)) {
4441 mem_cgroup_id_remove(memcg
);
4443 /* Memcg ID pins CSS */
4444 css_put(&memcg
->css
);
4448 static inline void mem_cgroup_id_get(struct mem_cgroup
*memcg
)
4450 mem_cgroup_id_get_many(memcg
, 1);
4453 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
4455 mem_cgroup_id_put_many(memcg
, 1);
4459 * mem_cgroup_from_id - look up a memcg from a memcg id
4460 * @id: the memcg id to look up
4462 * Caller must hold rcu_read_lock().
4464 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
4466 WARN_ON_ONCE(!rcu_read_lock_held());
4467 return idr_find(&mem_cgroup_idr
, id
);
4470 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4472 struct mem_cgroup_per_node
*pn
;
4475 * This routine is called against possible nodes.
4476 * But it's BUG to call kmalloc() against offline node.
4478 * TODO: this routine can waste much memory for nodes which will
4479 * never be onlined. It's better to use memory hotplug callback
4482 if (!node_state(node
, N_NORMAL_MEMORY
))
4484 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4488 pn
->lruvec_stat_local
= alloc_percpu(struct lruvec_stat
);
4489 if (!pn
->lruvec_stat_local
) {
4494 pn
->lruvec_stat_cpu
= alloc_percpu(struct lruvec_stat
);
4495 if (!pn
->lruvec_stat_cpu
) {
4496 free_percpu(pn
->lruvec_stat_local
);
4501 lruvec_init(&pn
->lruvec
);
4502 pn
->usage_in_excess
= 0;
4503 pn
->on_tree
= false;
4506 memcg
->nodeinfo
[node
] = pn
;
4510 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4512 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
4517 free_percpu(pn
->lruvec_stat_cpu
);
4518 free_percpu(pn
->lruvec_stat_local
);
4522 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
4527 free_mem_cgroup_per_node_info(memcg
, node
);
4528 free_percpu(memcg
->vmstats_percpu
);
4529 free_percpu(memcg
->vmstats_local
);
4533 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4535 memcg_wb_domain_exit(memcg
);
4536 __mem_cgroup_free(memcg
);
4539 static struct mem_cgroup
*mem_cgroup_alloc(void)
4541 struct mem_cgroup
*memcg
;
4545 size
= sizeof(struct mem_cgroup
);
4546 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4548 memcg
= kzalloc(size
, GFP_KERNEL
);
4552 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
4553 1, MEM_CGROUP_ID_MAX
,
4555 if (memcg
->id
.id
< 0)
4558 memcg
->vmstats_local
= alloc_percpu(struct memcg_vmstats_percpu
);
4559 if (!memcg
->vmstats_local
)
4562 memcg
->vmstats_percpu
= alloc_percpu(struct memcg_vmstats_percpu
);
4563 if (!memcg
->vmstats_percpu
)
4567 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
4570 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4573 INIT_WORK(&memcg
->high_work
, high_work_func
);
4574 memcg
->last_scanned_node
= MAX_NUMNODES
;
4575 INIT_LIST_HEAD(&memcg
->oom_notify
);
4576 mutex_init(&memcg
->thresholds_lock
);
4577 spin_lock_init(&memcg
->move_lock
);
4578 vmpressure_init(&memcg
->vmpressure
);
4579 INIT_LIST_HEAD(&memcg
->event_list
);
4580 spin_lock_init(&memcg
->event_list_lock
);
4581 memcg
->socket_pressure
= jiffies
;
4582 #ifdef CONFIG_MEMCG_KMEM
4583 memcg
->kmemcg_id
= -1;
4585 #ifdef CONFIG_CGROUP_WRITEBACK
4586 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4588 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
4591 mem_cgroup_id_remove(memcg
);
4592 __mem_cgroup_free(memcg
);
4596 static struct cgroup_subsys_state
* __ref
4597 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4599 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
4600 struct mem_cgroup
*memcg
;
4601 long error
= -ENOMEM
;
4603 memcg
= mem_cgroup_alloc();
4605 return ERR_PTR(error
);
4607 memcg
->high
= PAGE_COUNTER_MAX
;
4608 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4610 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4611 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4613 if (parent
&& parent
->use_hierarchy
) {
4614 memcg
->use_hierarchy
= true;
4615 page_counter_init(&memcg
->memory
, &parent
->memory
);
4616 page_counter_init(&memcg
->swap
, &parent
->swap
);
4617 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4618 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4619 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
4621 page_counter_init(&memcg
->memory
, NULL
);
4622 page_counter_init(&memcg
->swap
, NULL
);
4623 page_counter_init(&memcg
->memsw
, NULL
);
4624 page_counter_init(&memcg
->kmem
, NULL
);
4625 page_counter_init(&memcg
->tcpmem
, NULL
);
4627 * Deeper hierachy with use_hierarchy == false doesn't make
4628 * much sense so let cgroup subsystem know about this
4629 * unfortunate state in our controller.
4631 if (parent
!= root_mem_cgroup
)
4632 memory_cgrp_subsys
.broken_hierarchy
= true;
4635 /* The following stuff does not apply to the root */
4637 root_mem_cgroup
= memcg
;
4641 error
= memcg_online_kmem(memcg
);
4645 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4646 static_branch_inc(&memcg_sockets_enabled_key
);
4650 mem_cgroup_id_remove(memcg
);
4651 mem_cgroup_free(memcg
);
4652 return ERR_PTR(-ENOMEM
);
4655 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4657 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4660 * A memcg must be visible for memcg_expand_shrinker_maps()
4661 * by the time the maps are allocated. So, we allocate maps
4662 * here, when for_each_mem_cgroup() can't skip it.
4664 if (memcg_alloc_shrinker_maps(memcg
)) {
4665 mem_cgroup_id_remove(memcg
);
4669 /* Online state pins memcg ID, memcg ID pins CSS */
4670 refcount_set(&memcg
->id
.ref
, 1);
4675 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4677 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4678 struct mem_cgroup_event
*event
, *tmp
;
4681 * Unregister events and notify userspace.
4682 * Notify userspace about cgroup removing only after rmdir of cgroup
4683 * directory to avoid race between userspace and kernelspace.
4685 spin_lock(&memcg
->event_list_lock
);
4686 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4687 list_del_init(&event
->list
);
4688 schedule_work(&event
->remove
);
4690 spin_unlock(&memcg
->event_list_lock
);
4692 page_counter_set_min(&memcg
->memory
, 0);
4693 page_counter_set_low(&memcg
->memory
, 0);
4695 memcg_offline_kmem(memcg
);
4696 wb_memcg_offline(memcg
);
4698 drain_all_stock(memcg
);
4700 mem_cgroup_id_put(memcg
);
4703 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4705 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4707 invalidate_reclaim_iterators(memcg
);
4710 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4712 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4714 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4715 static_branch_dec(&memcg_sockets_enabled_key
);
4717 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
4718 static_branch_dec(&memcg_sockets_enabled_key
);
4720 vmpressure_cleanup(&memcg
->vmpressure
);
4721 cancel_work_sync(&memcg
->high_work
);
4722 mem_cgroup_remove_from_trees(memcg
);
4723 memcg_free_shrinker_maps(memcg
);
4724 memcg_free_kmem(memcg
);
4725 mem_cgroup_free(memcg
);
4729 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4730 * @css: the target css
4732 * Reset the states of the mem_cgroup associated with @css. This is
4733 * invoked when the userland requests disabling on the default hierarchy
4734 * but the memcg is pinned through dependency. The memcg should stop
4735 * applying policies and should revert to the vanilla state as it may be
4736 * made visible again.
4738 * The current implementation only resets the essential configurations.
4739 * This needs to be expanded to cover all the visible parts.
4741 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4743 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4745 page_counter_set_max(&memcg
->memory
, PAGE_COUNTER_MAX
);
4746 page_counter_set_max(&memcg
->swap
, PAGE_COUNTER_MAX
);
4747 page_counter_set_max(&memcg
->memsw
, PAGE_COUNTER_MAX
);
4748 page_counter_set_max(&memcg
->kmem
, PAGE_COUNTER_MAX
);
4749 page_counter_set_max(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
4750 page_counter_set_min(&memcg
->memory
, 0);
4751 page_counter_set_low(&memcg
->memory
, 0);
4752 memcg
->high
= PAGE_COUNTER_MAX
;
4753 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4754 memcg_wb_domain_size_changed(memcg
);
4758 /* Handlers for move charge at task migration. */
4759 static int mem_cgroup_do_precharge(unsigned long count
)
4763 /* Try a single bulk charge without reclaim first, kswapd may wake */
4764 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4766 mc
.precharge
+= count
;
4770 /* Try charges one by one with reclaim, but do not retry */
4772 ret
= try_charge(mc
.to
, GFP_KERNEL
| __GFP_NORETRY
, 1);
4786 enum mc_target_type
{
4793 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4794 unsigned long addr
, pte_t ptent
)
4796 struct page
*page
= _vm_normal_page(vma
, addr
, ptent
, true);
4798 if (!page
|| !page_mapped(page
))
4800 if (PageAnon(page
)) {
4801 if (!(mc
.flags
& MOVE_ANON
))
4804 if (!(mc
.flags
& MOVE_FILE
))
4807 if (!get_page_unless_zero(page
))
4813 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
4814 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4815 pte_t ptent
, swp_entry_t
*entry
)
4817 struct page
*page
= NULL
;
4818 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4820 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4824 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
4825 * a device and because they are not accessible by CPU they are store
4826 * as special swap entry in the CPU page table.
4828 if (is_device_private_entry(ent
)) {
4829 page
= device_private_entry_to_page(ent
);
4831 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
4832 * a refcount of 1 when free (unlike normal page)
4834 if (!page_ref_add_unless(page
, 1, 1))
4840 * Because lookup_swap_cache() updates some statistics counter,
4841 * we call find_get_page() with swapper_space directly.
4843 page
= find_get_page(swap_address_space(ent
), swp_offset(ent
));
4844 if (do_memsw_account())
4845 entry
->val
= ent
.val
;
4850 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4851 pte_t ptent
, swp_entry_t
*entry
)
4857 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4858 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4860 struct page
*page
= NULL
;
4861 struct address_space
*mapping
;
4864 if (!vma
->vm_file
) /* anonymous vma */
4866 if (!(mc
.flags
& MOVE_FILE
))
4869 mapping
= vma
->vm_file
->f_mapping
;
4870 pgoff
= linear_page_index(vma
, addr
);
4872 /* page is moved even if it's not RSS of this task(page-faulted). */
4874 /* shmem/tmpfs may report page out on swap: account for that too. */
4875 if (shmem_mapping(mapping
)) {
4876 page
= find_get_entry(mapping
, pgoff
);
4877 if (xa_is_value(page
)) {
4878 swp_entry_t swp
= radix_to_swp_entry(page
);
4879 if (do_memsw_account())
4881 page
= find_get_page(swap_address_space(swp
),
4885 page
= find_get_page(mapping
, pgoff
);
4887 page
= find_get_page(mapping
, pgoff
);
4893 * mem_cgroup_move_account - move account of the page
4895 * @compound: charge the page as compound or small page
4896 * @from: mem_cgroup which the page is moved from.
4897 * @to: mem_cgroup which the page is moved to. @from != @to.
4899 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4901 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4904 static int mem_cgroup_move_account(struct page
*page
,
4906 struct mem_cgroup
*from
,
4907 struct mem_cgroup
*to
)
4909 unsigned long flags
;
4910 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
4914 VM_BUG_ON(from
== to
);
4915 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4916 VM_BUG_ON(compound
&& !PageTransHuge(page
));
4919 * Prevent mem_cgroup_migrate() from looking at
4920 * page->mem_cgroup of its source page while we change it.
4923 if (!trylock_page(page
))
4927 if (page
->mem_cgroup
!= from
)
4930 anon
= PageAnon(page
);
4932 spin_lock_irqsave(&from
->move_lock
, flags
);
4934 if (!anon
&& page_mapped(page
)) {
4935 __mod_memcg_state(from
, NR_FILE_MAPPED
, -nr_pages
);
4936 __mod_memcg_state(to
, NR_FILE_MAPPED
, nr_pages
);
4940 * move_lock grabbed above and caller set from->moving_account, so
4941 * mod_memcg_page_state will serialize updates to PageDirty.
4942 * So mapping should be stable for dirty pages.
4944 if (!anon
&& PageDirty(page
)) {
4945 struct address_space
*mapping
= page_mapping(page
);
4947 if (mapping_cap_account_dirty(mapping
)) {
4948 __mod_memcg_state(from
, NR_FILE_DIRTY
, -nr_pages
);
4949 __mod_memcg_state(to
, NR_FILE_DIRTY
, nr_pages
);
4953 if (PageWriteback(page
)) {
4954 __mod_memcg_state(from
, NR_WRITEBACK
, -nr_pages
);
4955 __mod_memcg_state(to
, NR_WRITEBACK
, nr_pages
);
4959 * It is safe to change page->mem_cgroup here because the page
4960 * is referenced, charged, and isolated - we can't race with
4961 * uncharging, charging, migration, or LRU putback.
4964 /* caller should have done css_get */
4965 page
->mem_cgroup
= to
;
4966 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4970 local_irq_disable();
4971 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
4972 memcg_check_events(to
, page
);
4973 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
4974 memcg_check_events(from
, page
);
4983 * get_mctgt_type - get target type of moving charge
4984 * @vma: the vma the pte to be checked belongs
4985 * @addr: the address corresponding to the pte to be checked
4986 * @ptent: the pte to be checked
4987 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4990 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4991 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4992 * move charge. if @target is not NULL, the page is stored in target->page
4993 * with extra refcnt got(Callers should handle it).
4994 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4995 * target for charge migration. if @target is not NULL, the entry is stored
4997 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PUBLIC
4998 * or MEMORY_DEVICE_PRIVATE (so ZONE_DEVICE page and thus not on the lru).
4999 * For now we such page is charge like a regular page would be as for all
5000 * intent and purposes it is just special memory taking the place of a
5003 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5005 * Called with pte lock held.
5008 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
5009 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
5011 struct page
*page
= NULL
;
5012 enum mc_target_type ret
= MC_TARGET_NONE
;
5013 swp_entry_t ent
= { .val
= 0 };
5015 if (pte_present(ptent
))
5016 page
= mc_handle_present_pte(vma
, addr
, ptent
);
5017 else if (is_swap_pte(ptent
))
5018 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
5019 else if (pte_none(ptent
))
5020 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
5022 if (!page
&& !ent
.val
)
5026 * Do only loose check w/o serialization.
5027 * mem_cgroup_move_account() checks the page is valid or
5028 * not under LRU exclusion.
5030 if (page
->mem_cgroup
== mc
.from
) {
5031 ret
= MC_TARGET_PAGE
;
5032 if (is_device_private_page(page
) ||
5033 is_device_public_page(page
))
5034 ret
= MC_TARGET_DEVICE
;
5036 target
->page
= page
;
5038 if (!ret
|| !target
)
5042 * There is a swap entry and a page doesn't exist or isn't charged.
5043 * But we cannot move a tail-page in a THP.
5045 if (ent
.val
&& !ret
&& (!page
|| !PageTransCompound(page
)) &&
5046 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
5047 ret
= MC_TARGET_SWAP
;
5054 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5056 * We don't consider PMD mapped swapping or file mapped pages because THP does
5057 * not support them for now.
5058 * Caller should make sure that pmd_trans_huge(pmd) is true.
5060 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5061 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5063 struct page
*page
= NULL
;
5064 enum mc_target_type ret
= MC_TARGET_NONE
;
5066 if (unlikely(is_swap_pmd(pmd
))) {
5067 VM_BUG_ON(thp_migration_supported() &&
5068 !is_pmd_migration_entry(pmd
));
5071 page
= pmd_page(pmd
);
5072 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
5073 if (!(mc
.flags
& MOVE_ANON
))
5075 if (page
->mem_cgroup
== mc
.from
) {
5076 ret
= MC_TARGET_PAGE
;
5079 target
->page
= page
;
5085 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5086 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5088 return MC_TARGET_NONE
;
5092 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5093 unsigned long addr
, unsigned long end
,
5094 struct mm_walk
*walk
)
5096 struct vm_area_struct
*vma
= walk
->vma
;
5100 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5103 * Note their can not be MC_TARGET_DEVICE for now as we do not
5104 * support transparent huge page with MEMORY_DEVICE_PUBLIC or
5105 * MEMORY_DEVICE_PRIVATE but this might change.
5107 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5108 mc
.precharge
+= HPAGE_PMD_NR
;
5113 if (pmd_trans_unstable(pmd
))
5115 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5116 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5117 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5118 mc
.precharge
++; /* increment precharge temporarily */
5119 pte_unmap_unlock(pte
- 1, ptl
);
5125 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5127 unsigned long precharge
;
5129 struct mm_walk mem_cgroup_count_precharge_walk
= {
5130 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
5133 down_read(&mm
->mmap_sem
);
5134 walk_page_range(0, mm
->highest_vm_end
,
5135 &mem_cgroup_count_precharge_walk
);
5136 up_read(&mm
->mmap_sem
);
5138 precharge
= mc
.precharge
;
5144 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
5146 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
5148 VM_BUG_ON(mc
.moving_task
);
5149 mc
.moving_task
= current
;
5150 return mem_cgroup_do_precharge(precharge
);
5153 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5154 static void __mem_cgroup_clear_mc(void)
5156 struct mem_cgroup
*from
= mc
.from
;
5157 struct mem_cgroup
*to
= mc
.to
;
5159 /* we must uncharge all the leftover precharges from mc.to */
5161 cancel_charge(mc
.to
, mc
.precharge
);
5165 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5166 * we must uncharge here.
5168 if (mc
.moved_charge
) {
5169 cancel_charge(mc
.from
, mc
.moved_charge
);
5170 mc
.moved_charge
= 0;
5172 /* we must fixup refcnts and charges */
5173 if (mc
.moved_swap
) {
5174 /* uncharge swap account from the old cgroup */
5175 if (!mem_cgroup_is_root(mc
.from
))
5176 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
5178 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
5181 * we charged both to->memory and to->memsw, so we
5182 * should uncharge to->memory.
5184 if (!mem_cgroup_is_root(mc
.to
))
5185 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
5187 mem_cgroup_id_get_many(mc
.to
, mc
.moved_swap
);
5188 css_put_many(&mc
.to
->css
, mc
.moved_swap
);
5192 memcg_oom_recover(from
);
5193 memcg_oom_recover(to
);
5194 wake_up_all(&mc
.waitq
);
5197 static void mem_cgroup_clear_mc(void)
5199 struct mm_struct
*mm
= mc
.mm
;
5202 * we must clear moving_task before waking up waiters at the end of
5205 mc
.moving_task
= NULL
;
5206 __mem_cgroup_clear_mc();
5207 spin_lock(&mc
.lock
);
5211 spin_unlock(&mc
.lock
);
5216 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5218 struct cgroup_subsys_state
*css
;
5219 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
5220 struct mem_cgroup
*from
;
5221 struct task_struct
*leader
, *p
;
5222 struct mm_struct
*mm
;
5223 unsigned long move_flags
;
5226 /* charge immigration isn't supported on the default hierarchy */
5227 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5231 * Multi-process migrations only happen on the default hierarchy
5232 * where charge immigration is not used. Perform charge
5233 * immigration if @tset contains a leader and whine if there are
5237 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
5240 memcg
= mem_cgroup_from_css(css
);
5246 * We are now commited to this value whatever it is. Changes in this
5247 * tunable will only affect upcoming migrations, not the current one.
5248 * So we need to save it, and keep it going.
5250 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
5254 from
= mem_cgroup_from_task(p
);
5256 VM_BUG_ON(from
== memcg
);
5258 mm
= get_task_mm(p
);
5261 /* We move charges only when we move a owner of the mm */
5262 if (mm
->owner
== p
) {
5265 VM_BUG_ON(mc
.precharge
);
5266 VM_BUG_ON(mc
.moved_charge
);
5267 VM_BUG_ON(mc
.moved_swap
);
5269 spin_lock(&mc
.lock
);
5273 mc
.flags
= move_flags
;
5274 spin_unlock(&mc
.lock
);
5275 /* We set mc.moving_task later */
5277 ret
= mem_cgroup_precharge_mc(mm
);
5279 mem_cgroup_clear_mc();
5286 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5289 mem_cgroup_clear_mc();
5292 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
5293 unsigned long addr
, unsigned long end
,
5294 struct mm_walk
*walk
)
5297 struct vm_area_struct
*vma
= walk
->vma
;
5300 enum mc_target_type target_type
;
5301 union mc_target target
;
5304 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5306 if (mc
.precharge
< HPAGE_PMD_NR
) {
5310 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
5311 if (target_type
== MC_TARGET_PAGE
) {
5313 if (!isolate_lru_page(page
)) {
5314 if (!mem_cgroup_move_account(page
, true,
5316 mc
.precharge
-= HPAGE_PMD_NR
;
5317 mc
.moved_charge
+= HPAGE_PMD_NR
;
5319 putback_lru_page(page
);
5322 } else if (target_type
== MC_TARGET_DEVICE
) {
5324 if (!mem_cgroup_move_account(page
, true,
5326 mc
.precharge
-= HPAGE_PMD_NR
;
5327 mc
.moved_charge
+= HPAGE_PMD_NR
;
5335 if (pmd_trans_unstable(pmd
))
5338 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5339 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
5340 pte_t ptent
= *(pte
++);
5341 bool device
= false;
5347 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
5348 case MC_TARGET_DEVICE
:
5351 case MC_TARGET_PAGE
:
5354 * We can have a part of the split pmd here. Moving it
5355 * can be done but it would be too convoluted so simply
5356 * ignore such a partial THP and keep it in original
5357 * memcg. There should be somebody mapping the head.
5359 if (PageTransCompound(page
))
5361 if (!device
&& isolate_lru_page(page
))
5363 if (!mem_cgroup_move_account(page
, false,
5366 /* we uncharge from mc.from later. */
5370 putback_lru_page(page
);
5371 put
: /* get_mctgt_type() gets the page */
5374 case MC_TARGET_SWAP
:
5376 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
5378 /* we fixup refcnts and charges later. */
5386 pte_unmap_unlock(pte
- 1, ptl
);
5391 * We have consumed all precharges we got in can_attach().
5392 * We try charge one by one, but don't do any additional
5393 * charges to mc.to if we have failed in charge once in attach()
5396 ret
= mem_cgroup_do_precharge(1);
5404 static void mem_cgroup_move_charge(void)
5406 struct mm_walk mem_cgroup_move_charge_walk
= {
5407 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
5411 lru_add_drain_all();
5413 * Signal lock_page_memcg() to take the memcg's move_lock
5414 * while we're moving its pages to another memcg. Then wait
5415 * for already started RCU-only updates to finish.
5417 atomic_inc(&mc
.from
->moving_account
);
5420 if (unlikely(!down_read_trylock(&mc
.mm
->mmap_sem
))) {
5422 * Someone who are holding the mmap_sem might be waiting in
5423 * waitq. So we cancel all extra charges, wake up all waiters,
5424 * and retry. Because we cancel precharges, we might not be able
5425 * to move enough charges, but moving charge is a best-effort
5426 * feature anyway, so it wouldn't be a big problem.
5428 __mem_cgroup_clear_mc();
5433 * When we have consumed all precharges and failed in doing
5434 * additional charge, the page walk just aborts.
5436 walk_page_range(0, mc
.mm
->highest_vm_end
, &mem_cgroup_move_charge_walk
);
5438 up_read(&mc
.mm
->mmap_sem
);
5439 atomic_dec(&mc
.from
->moving_account
);
5442 static void mem_cgroup_move_task(void)
5445 mem_cgroup_move_charge();
5446 mem_cgroup_clear_mc();
5449 #else /* !CONFIG_MMU */
5450 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5454 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5457 static void mem_cgroup_move_task(void)
5463 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5464 * to verify whether we're attached to the default hierarchy on each mount
5467 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5470 * use_hierarchy is forced on the default hierarchy. cgroup core
5471 * guarantees that @root doesn't have any children, so turning it
5472 * on for the root memcg is enough.
5474 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5475 root_mem_cgroup
->use_hierarchy
= true;
5477 root_mem_cgroup
->use_hierarchy
= false;
5480 static int seq_puts_memcg_tunable(struct seq_file
*m
, unsigned long value
)
5482 if (value
== PAGE_COUNTER_MAX
)
5483 seq_puts(m
, "max\n");
5485 seq_printf(m
, "%llu\n", (u64
)value
* PAGE_SIZE
);
5490 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5493 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5495 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
5498 static int memory_min_show(struct seq_file
*m
, void *v
)
5500 return seq_puts_memcg_tunable(m
,
5501 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.min
));
5504 static ssize_t
memory_min_write(struct kernfs_open_file
*of
,
5505 char *buf
, size_t nbytes
, loff_t off
)
5507 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5511 buf
= strstrip(buf
);
5512 err
= page_counter_memparse(buf
, "max", &min
);
5516 page_counter_set_min(&memcg
->memory
, min
);
5521 static int memory_low_show(struct seq_file
*m
, void *v
)
5523 return seq_puts_memcg_tunable(m
,
5524 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.low
));
5527 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5528 char *buf
, size_t nbytes
, loff_t off
)
5530 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5534 buf
= strstrip(buf
);
5535 err
= page_counter_memparse(buf
, "max", &low
);
5539 page_counter_set_low(&memcg
->memory
, low
);
5544 static int memory_high_show(struct seq_file
*m
, void *v
)
5546 return seq_puts_memcg_tunable(m
, READ_ONCE(mem_cgroup_from_seq(m
)->high
));
5549 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5550 char *buf
, size_t nbytes
, loff_t off
)
5552 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5553 unsigned long nr_pages
;
5557 buf
= strstrip(buf
);
5558 err
= page_counter_memparse(buf
, "max", &high
);
5564 nr_pages
= page_counter_read(&memcg
->memory
);
5565 if (nr_pages
> high
)
5566 try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
5569 memcg_wb_domain_size_changed(memcg
);
5573 static int memory_max_show(struct seq_file
*m
, void *v
)
5575 return seq_puts_memcg_tunable(m
,
5576 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.max
));
5579 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5580 char *buf
, size_t nbytes
, loff_t off
)
5582 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5583 unsigned int nr_reclaims
= MEM_CGROUP_RECLAIM_RETRIES
;
5584 bool drained
= false;
5588 buf
= strstrip(buf
);
5589 err
= page_counter_memparse(buf
, "max", &max
);
5593 xchg(&memcg
->memory
.max
, max
);
5596 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
5598 if (nr_pages
<= max
)
5601 if (signal_pending(current
)) {
5607 drain_all_stock(memcg
);
5613 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
5619 memcg_memory_event(memcg
, MEMCG_OOM
);
5620 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
5624 memcg_wb_domain_size_changed(memcg
);
5628 static int memory_events_show(struct seq_file
*m
, void *v
)
5630 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
5632 seq_printf(m
, "low %lu\n",
5633 atomic_long_read(&memcg
->memory_events
[MEMCG_LOW
]));
5634 seq_printf(m
, "high %lu\n",
5635 atomic_long_read(&memcg
->memory_events
[MEMCG_HIGH
]));
5636 seq_printf(m
, "max %lu\n",
5637 atomic_long_read(&memcg
->memory_events
[MEMCG_MAX
]));
5638 seq_printf(m
, "oom %lu\n",
5639 atomic_long_read(&memcg
->memory_events
[MEMCG_OOM
]));
5640 seq_printf(m
, "oom_kill %lu\n",
5641 atomic_long_read(&memcg
->memory_events
[MEMCG_OOM_KILL
]));
5646 static int memory_stat_show(struct seq_file
*m
, void *v
)
5648 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
5652 * Provide statistics on the state of the memory subsystem as
5653 * well as cumulative event counters that show past behavior.
5655 * This list is ordered following a combination of these gradients:
5656 * 1) generic big picture -> specifics and details
5657 * 2) reflecting userspace activity -> reflecting kernel heuristics
5659 * Current memory state:
5662 seq_printf(m
, "anon %llu\n",
5663 (u64
)memcg_page_state(memcg
, MEMCG_RSS
) * PAGE_SIZE
);
5664 seq_printf(m
, "file %llu\n",
5665 (u64
)memcg_page_state(memcg
, MEMCG_CACHE
) * PAGE_SIZE
);
5666 seq_printf(m
, "kernel_stack %llu\n",
5667 (u64
)memcg_page_state(memcg
, MEMCG_KERNEL_STACK_KB
) * 1024);
5668 seq_printf(m
, "slab %llu\n",
5669 (u64
)(memcg_page_state(memcg
, NR_SLAB_RECLAIMABLE
) +
5670 memcg_page_state(memcg
, NR_SLAB_UNRECLAIMABLE
)) *
5672 seq_printf(m
, "sock %llu\n",
5673 (u64
)memcg_page_state(memcg
, MEMCG_SOCK
) * PAGE_SIZE
);
5675 seq_printf(m
, "shmem %llu\n",
5676 (u64
)memcg_page_state(memcg
, NR_SHMEM
) * PAGE_SIZE
);
5677 seq_printf(m
, "file_mapped %llu\n",
5678 (u64
)memcg_page_state(memcg
, NR_FILE_MAPPED
) * PAGE_SIZE
);
5679 seq_printf(m
, "file_dirty %llu\n",
5680 (u64
)memcg_page_state(memcg
, NR_FILE_DIRTY
) * PAGE_SIZE
);
5681 seq_printf(m
, "file_writeback %llu\n",
5682 (u64
)memcg_page_state(memcg
, NR_WRITEBACK
) * PAGE_SIZE
);
5685 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
5686 * with the NR_ANON_THP vm counter, but right now it's a pain in the
5687 * arse because it requires migrating the work out of rmap to a place
5688 * where the page->mem_cgroup is set up and stable.
5690 seq_printf(m
, "anon_thp %llu\n",
5691 (u64
)memcg_page_state(memcg
, MEMCG_RSS_HUGE
) * PAGE_SIZE
);
5693 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5694 seq_printf(m
, "%s %llu\n", mem_cgroup_lru_names
[i
],
5695 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
5698 seq_printf(m
, "slab_reclaimable %llu\n",
5699 (u64
)memcg_page_state(memcg
, NR_SLAB_RECLAIMABLE
) *
5701 seq_printf(m
, "slab_unreclaimable %llu\n",
5702 (u64
)memcg_page_state(memcg
, NR_SLAB_UNRECLAIMABLE
) *
5705 /* Accumulated memory events */
5707 seq_printf(m
, "pgfault %lu\n", memcg_events(memcg
, PGFAULT
));
5708 seq_printf(m
, "pgmajfault %lu\n", memcg_events(memcg
, PGMAJFAULT
));
5710 seq_printf(m
, "workingset_refault %lu\n",
5711 memcg_page_state(memcg
, WORKINGSET_REFAULT
));
5712 seq_printf(m
, "workingset_activate %lu\n",
5713 memcg_page_state(memcg
, WORKINGSET_ACTIVATE
));
5714 seq_printf(m
, "workingset_nodereclaim %lu\n",
5715 memcg_page_state(memcg
, WORKINGSET_NODERECLAIM
));
5717 seq_printf(m
, "pgrefill %lu\n", memcg_events(memcg
, PGREFILL
));
5718 seq_printf(m
, "pgscan %lu\n", memcg_events(memcg
, PGSCAN_KSWAPD
) +
5719 memcg_events(memcg
, PGSCAN_DIRECT
));
5720 seq_printf(m
, "pgsteal %lu\n", memcg_events(memcg
, PGSTEAL_KSWAPD
) +
5721 memcg_events(memcg
, PGSTEAL_DIRECT
));
5722 seq_printf(m
, "pgactivate %lu\n", memcg_events(memcg
, PGACTIVATE
));
5723 seq_printf(m
, "pgdeactivate %lu\n", memcg_events(memcg
, PGDEACTIVATE
));
5724 seq_printf(m
, "pglazyfree %lu\n", memcg_events(memcg
, PGLAZYFREE
));
5725 seq_printf(m
, "pglazyfreed %lu\n", memcg_events(memcg
, PGLAZYFREED
));
5727 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5728 seq_printf(m
, "thp_fault_alloc %lu\n",
5729 memcg_events(memcg
, THP_FAULT_ALLOC
));
5730 seq_printf(m
, "thp_collapse_alloc %lu\n",
5731 memcg_events(memcg
, THP_COLLAPSE_ALLOC
));
5732 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
5737 static int memory_oom_group_show(struct seq_file
*m
, void *v
)
5739 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
5741 seq_printf(m
, "%d\n", memcg
->oom_group
);
5746 static ssize_t
memory_oom_group_write(struct kernfs_open_file
*of
,
5747 char *buf
, size_t nbytes
, loff_t off
)
5749 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5752 buf
= strstrip(buf
);
5756 ret
= kstrtoint(buf
, 0, &oom_group
);
5760 if (oom_group
!= 0 && oom_group
!= 1)
5763 memcg
->oom_group
= oom_group
;
5768 static struct cftype memory_files
[] = {
5771 .flags
= CFTYPE_NOT_ON_ROOT
,
5772 .read_u64
= memory_current_read
,
5776 .flags
= CFTYPE_NOT_ON_ROOT
,
5777 .seq_show
= memory_min_show
,
5778 .write
= memory_min_write
,
5782 .flags
= CFTYPE_NOT_ON_ROOT
,
5783 .seq_show
= memory_low_show
,
5784 .write
= memory_low_write
,
5788 .flags
= CFTYPE_NOT_ON_ROOT
,
5789 .seq_show
= memory_high_show
,
5790 .write
= memory_high_write
,
5794 .flags
= CFTYPE_NOT_ON_ROOT
,
5795 .seq_show
= memory_max_show
,
5796 .write
= memory_max_write
,
5800 .flags
= CFTYPE_NOT_ON_ROOT
,
5801 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
5802 .seq_show
= memory_events_show
,
5806 .flags
= CFTYPE_NOT_ON_ROOT
,
5807 .seq_show
= memory_stat_show
,
5810 .name
= "oom.group",
5811 .flags
= CFTYPE_NOT_ON_ROOT
| CFTYPE_NS_DELEGATABLE
,
5812 .seq_show
= memory_oom_group_show
,
5813 .write
= memory_oom_group_write
,
5818 struct cgroup_subsys memory_cgrp_subsys
= {
5819 .css_alloc
= mem_cgroup_css_alloc
,
5820 .css_online
= mem_cgroup_css_online
,
5821 .css_offline
= mem_cgroup_css_offline
,
5822 .css_released
= mem_cgroup_css_released
,
5823 .css_free
= mem_cgroup_css_free
,
5824 .css_reset
= mem_cgroup_css_reset
,
5825 .can_attach
= mem_cgroup_can_attach
,
5826 .cancel_attach
= mem_cgroup_cancel_attach
,
5827 .post_attach
= mem_cgroup_move_task
,
5828 .bind
= mem_cgroup_bind
,
5829 .dfl_cftypes
= memory_files
,
5830 .legacy_cftypes
= mem_cgroup_legacy_files
,
5835 * mem_cgroup_protected - check if memory consumption is in the normal range
5836 * @root: the top ancestor of the sub-tree being checked
5837 * @memcg: the memory cgroup to check
5839 * WARNING: This function is not stateless! It can only be used as part
5840 * of a top-down tree iteration, not for isolated queries.
5842 * Returns one of the following:
5843 * MEMCG_PROT_NONE: cgroup memory is not protected
5844 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
5845 * an unprotected supply of reclaimable memory from other cgroups.
5846 * MEMCG_PROT_MIN: cgroup memory is protected
5848 * @root is exclusive; it is never protected when looked at directly
5850 * To provide a proper hierarchical behavior, effective memory.min/low values
5851 * are used. Below is the description of how effective memory.low is calculated.
5852 * Effective memory.min values is calculated in the same way.
5854 * Effective memory.low is always equal or less than the original memory.low.
5855 * If there is no memory.low overcommittment (which is always true for
5856 * top-level memory cgroups), these two values are equal.
5857 * Otherwise, it's a part of parent's effective memory.low,
5858 * calculated as a cgroup's memory.low usage divided by sum of sibling's
5859 * memory.low usages, where memory.low usage is the size of actually
5863 * elow = min( memory.low, parent->elow * ------------------ ),
5864 * siblings_low_usage
5866 * | memory.current, if memory.current < memory.low
5871 * Such definition of the effective memory.low provides the expected
5872 * hierarchical behavior: parent's memory.low value is limiting
5873 * children, unprotected memory is reclaimed first and cgroups,
5874 * which are not using their guarantee do not affect actual memory
5877 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
5879 * A A/memory.low = 2G, A/memory.current = 6G
5881 * BC DE B/memory.low = 3G B/memory.current = 2G
5882 * C/memory.low = 1G C/memory.current = 2G
5883 * D/memory.low = 0 D/memory.current = 2G
5884 * E/memory.low = 10G E/memory.current = 0
5886 * and the memory pressure is applied, the following memory distribution
5887 * is expected (approximately):
5889 * A/memory.current = 2G
5891 * B/memory.current = 1.3G
5892 * C/memory.current = 0.6G
5893 * D/memory.current = 0
5894 * E/memory.current = 0
5896 * These calculations require constant tracking of the actual low usages
5897 * (see propagate_protected_usage()), as well as recursive calculation of
5898 * effective memory.low values. But as we do call mem_cgroup_protected()
5899 * path for each memory cgroup top-down from the reclaim,
5900 * it's possible to optimize this part, and save calculated elow
5901 * for next usage. This part is intentionally racy, but it's ok,
5902 * as memory.low is a best-effort mechanism.
5904 enum mem_cgroup_protection
mem_cgroup_protected(struct mem_cgroup
*root
,
5905 struct mem_cgroup
*memcg
)
5907 struct mem_cgroup
*parent
;
5908 unsigned long emin
, parent_emin
;
5909 unsigned long elow
, parent_elow
;
5910 unsigned long usage
;
5912 if (mem_cgroup_disabled())
5913 return MEMCG_PROT_NONE
;
5916 root
= root_mem_cgroup
;
5918 return MEMCG_PROT_NONE
;
5920 usage
= page_counter_read(&memcg
->memory
);
5922 return MEMCG_PROT_NONE
;
5924 emin
= memcg
->memory
.min
;
5925 elow
= memcg
->memory
.low
;
5927 parent
= parent_mem_cgroup(memcg
);
5928 /* No parent means a non-hierarchical mode on v1 memcg */
5930 return MEMCG_PROT_NONE
;
5935 parent_emin
= READ_ONCE(parent
->memory
.emin
);
5936 emin
= min(emin
, parent_emin
);
5937 if (emin
&& parent_emin
) {
5938 unsigned long min_usage
, siblings_min_usage
;
5940 min_usage
= min(usage
, memcg
->memory
.min
);
5941 siblings_min_usage
= atomic_long_read(
5942 &parent
->memory
.children_min_usage
);
5944 if (min_usage
&& siblings_min_usage
)
5945 emin
= min(emin
, parent_emin
* min_usage
/
5946 siblings_min_usage
);
5949 parent_elow
= READ_ONCE(parent
->memory
.elow
);
5950 elow
= min(elow
, parent_elow
);
5951 if (elow
&& parent_elow
) {
5952 unsigned long low_usage
, siblings_low_usage
;
5954 low_usage
= min(usage
, memcg
->memory
.low
);
5955 siblings_low_usage
= atomic_long_read(
5956 &parent
->memory
.children_low_usage
);
5958 if (low_usage
&& siblings_low_usage
)
5959 elow
= min(elow
, parent_elow
* low_usage
/
5960 siblings_low_usage
);
5964 memcg
->memory
.emin
= emin
;
5965 memcg
->memory
.elow
= elow
;
5968 return MEMCG_PROT_MIN
;
5969 else if (usage
<= elow
)
5970 return MEMCG_PROT_LOW
;
5972 return MEMCG_PROT_NONE
;
5976 * mem_cgroup_try_charge - try charging a page
5977 * @page: page to charge
5978 * @mm: mm context of the victim
5979 * @gfp_mask: reclaim mode
5980 * @memcgp: charged memcg return
5981 * @compound: charge the page as compound or small page
5983 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5984 * pages according to @gfp_mask if necessary.
5986 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5987 * Otherwise, an error code is returned.
5989 * After page->mapping has been set up, the caller must finalize the
5990 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5991 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5993 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5994 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
5997 struct mem_cgroup
*memcg
= NULL
;
5998 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
6001 if (mem_cgroup_disabled())
6004 if (PageSwapCache(page
)) {
6006 * Every swap fault against a single page tries to charge the
6007 * page, bail as early as possible. shmem_unuse() encounters
6008 * already charged pages, too. The USED bit is protected by
6009 * the page lock, which serializes swap cache removal, which
6010 * in turn serializes uncharging.
6012 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
6013 if (compound_head(page
)->mem_cgroup
)
6016 if (do_swap_account
) {
6017 swp_entry_t ent
= { .val
= page_private(page
), };
6018 unsigned short id
= lookup_swap_cgroup_id(ent
);
6021 memcg
= mem_cgroup_from_id(id
);
6022 if (memcg
&& !css_tryget_online(&memcg
->css
))
6029 memcg
= get_mem_cgroup_from_mm(mm
);
6031 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
6033 css_put(&memcg
->css
);
6039 int mem_cgroup_try_charge_delay(struct page
*page
, struct mm_struct
*mm
,
6040 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
6043 struct mem_cgroup
*memcg
;
6046 ret
= mem_cgroup_try_charge(page
, mm
, gfp_mask
, memcgp
, compound
);
6048 mem_cgroup_throttle_swaprate(memcg
, page_to_nid(page
), gfp_mask
);
6053 * mem_cgroup_commit_charge - commit a page charge
6054 * @page: page to charge
6055 * @memcg: memcg to charge the page to
6056 * @lrucare: page might be on LRU already
6057 * @compound: charge the page as compound or small page
6059 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6060 * after page->mapping has been set up. This must happen atomically
6061 * as part of the page instantiation, i.e. under the page table lock
6062 * for anonymous pages, under the page lock for page and swap cache.
6064 * In addition, the page must not be on the LRU during the commit, to
6065 * prevent racing with task migration. If it might be, use @lrucare.
6067 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6069 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
6070 bool lrucare
, bool compound
)
6072 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
6074 VM_BUG_ON_PAGE(!page
->mapping
, page
);
6075 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
6077 if (mem_cgroup_disabled())
6080 * Swap faults will attempt to charge the same page multiple
6081 * times. But reuse_swap_page() might have removed the page
6082 * from swapcache already, so we can't check PageSwapCache().
6087 commit_charge(page
, memcg
, lrucare
);
6089 local_irq_disable();
6090 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
6091 memcg_check_events(memcg
, page
);
6094 if (do_memsw_account() && PageSwapCache(page
)) {
6095 swp_entry_t entry
= { .val
= page_private(page
) };
6097 * The swap entry might not get freed for a long time,
6098 * let's not wait for it. The page already received a
6099 * memory+swap charge, drop the swap entry duplicate.
6101 mem_cgroup_uncharge_swap(entry
, nr_pages
);
6106 * mem_cgroup_cancel_charge - cancel a page charge
6107 * @page: page to charge
6108 * @memcg: memcg to charge the page to
6109 * @compound: charge the page as compound or small page
6111 * Cancel a charge transaction started by mem_cgroup_try_charge().
6113 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
6116 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
6118 if (mem_cgroup_disabled())
6121 * Swap faults will attempt to charge the same page multiple
6122 * times. But reuse_swap_page() might have removed the page
6123 * from swapcache already, so we can't check PageSwapCache().
6128 cancel_charge(memcg
, nr_pages
);
6131 struct uncharge_gather
{
6132 struct mem_cgroup
*memcg
;
6133 unsigned long pgpgout
;
6134 unsigned long nr_anon
;
6135 unsigned long nr_file
;
6136 unsigned long nr_kmem
;
6137 unsigned long nr_huge
;
6138 unsigned long nr_shmem
;
6139 struct page
*dummy_page
;
6142 static inline void uncharge_gather_clear(struct uncharge_gather
*ug
)
6144 memset(ug
, 0, sizeof(*ug
));
6147 static void uncharge_batch(const struct uncharge_gather
*ug
)
6149 unsigned long nr_pages
= ug
->nr_anon
+ ug
->nr_file
+ ug
->nr_kmem
;
6150 unsigned long flags
;
6152 if (!mem_cgroup_is_root(ug
->memcg
)) {
6153 page_counter_uncharge(&ug
->memcg
->memory
, nr_pages
);
6154 if (do_memsw_account())
6155 page_counter_uncharge(&ug
->memcg
->memsw
, nr_pages
);
6156 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && ug
->nr_kmem
)
6157 page_counter_uncharge(&ug
->memcg
->kmem
, ug
->nr_kmem
);
6158 memcg_oom_recover(ug
->memcg
);
6161 local_irq_save(flags
);
6162 __mod_memcg_state(ug
->memcg
, MEMCG_RSS
, -ug
->nr_anon
);
6163 __mod_memcg_state(ug
->memcg
, MEMCG_CACHE
, -ug
->nr_file
);
6164 __mod_memcg_state(ug
->memcg
, MEMCG_RSS_HUGE
, -ug
->nr_huge
);
6165 __mod_memcg_state(ug
->memcg
, NR_SHMEM
, -ug
->nr_shmem
);
6166 __count_memcg_events(ug
->memcg
, PGPGOUT
, ug
->pgpgout
);
6167 __this_cpu_add(ug
->memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
6168 memcg_check_events(ug
->memcg
, ug
->dummy_page
);
6169 local_irq_restore(flags
);
6171 if (!mem_cgroup_is_root(ug
->memcg
))
6172 css_put_many(&ug
->memcg
->css
, nr_pages
);
6175 static void uncharge_page(struct page
*page
, struct uncharge_gather
*ug
)
6177 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6178 VM_BUG_ON_PAGE(page_count(page
) && !is_zone_device_page(page
) &&
6179 !PageHWPoison(page
) , page
);
6181 if (!page
->mem_cgroup
)
6185 * Nobody should be changing or seriously looking at
6186 * page->mem_cgroup at this point, we have fully
6187 * exclusive access to the page.
6190 if (ug
->memcg
!= page
->mem_cgroup
) {
6193 uncharge_gather_clear(ug
);
6195 ug
->memcg
= page
->mem_cgroup
;
6198 if (!PageKmemcg(page
)) {
6199 unsigned int nr_pages
= 1;
6201 if (PageTransHuge(page
)) {
6202 nr_pages
<<= compound_order(page
);
6203 ug
->nr_huge
+= nr_pages
;
6206 ug
->nr_anon
+= nr_pages
;
6208 ug
->nr_file
+= nr_pages
;
6209 if (PageSwapBacked(page
))
6210 ug
->nr_shmem
+= nr_pages
;
6214 ug
->nr_kmem
+= 1 << compound_order(page
);
6215 __ClearPageKmemcg(page
);
6218 ug
->dummy_page
= page
;
6219 page
->mem_cgroup
= NULL
;
6222 static void uncharge_list(struct list_head
*page_list
)
6224 struct uncharge_gather ug
;
6225 struct list_head
*next
;
6227 uncharge_gather_clear(&ug
);
6230 * Note that the list can be a single page->lru; hence the
6231 * do-while loop instead of a simple list_for_each_entry().
6233 next
= page_list
->next
;
6237 page
= list_entry(next
, struct page
, lru
);
6238 next
= page
->lru
.next
;
6240 uncharge_page(page
, &ug
);
6241 } while (next
!= page_list
);
6244 uncharge_batch(&ug
);
6248 * mem_cgroup_uncharge - uncharge a page
6249 * @page: page to uncharge
6251 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6252 * mem_cgroup_commit_charge().
6254 void mem_cgroup_uncharge(struct page
*page
)
6256 struct uncharge_gather ug
;
6258 if (mem_cgroup_disabled())
6261 /* Don't touch page->lru of any random page, pre-check: */
6262 if (!page
->mem_cgroup
)
6265 uncharge_gather_clear(&ug
);
6266 uncharge_page(page
, &ug
);
6267 uncharge_batch(&ug
);
6271 * mem_cgroup_uncharge_list - uncharge a list of page
6272 * @page_list: list of pages to uncharge
6274 * Uncharge a list of pages previously charged with
6275 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6277 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
6279 if (mem_cgroup_disabled())
6282 if (!list_empty(page_list
))
6283 uncharge_list(page_list
);
6287 * mem_cgroup_migrate - charge a page's replacement
6288 * @oldpage: currently circulating page
6289 * @newpage: replacement page
6291 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6292 * be uncharged upon free.
6294 * Both pages must be locked, @newpage->mapping must be set up.
6296 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
6298 struct mem_cgroup
*memcg
;
6299 unsigned int nr_pages
;
6301 unsigned long flags
;
6303 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
6304 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
6305 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
6306 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
6309 if (mem_cgroup_disabled())
6312 /* Page cache replacement: new page already charged? */
6313 if (newpage
->mem_cgroup
)
6316 /* Swapcache readahead pages can get replaced before being charged */
6317 memcg
= oldpage
->mem_cgroup
;
6321 /* Force-charge the new page. The old one will be freed soon */
6322 compound
= PageTransHuge(newpage
);
6323 nr_pages
= compound
? hpage_nr_pages(newpage
) : 1;
6325 page_counter_charge(&memcg
->memory
, nr_pages
);
6326 if (do_memsw_account())
6327 page_counter_charge(&memcg
->memsw
, nr_pages
);
6328 css_get_many(&memcg
->css
, nr_pages
);
6330 commit_charge(newpage
, memcg
, false);
6332 local_irq_save(flags
);
6333 mem_cgroup_charge_statistics(memcg
, newpage
, compound
, nr_pages
);
6334 memcg_check_events(memcg
, newpage
);
6335 local_irq_restore(flags
);
6338 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
6339 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
6341 void mem_cgroup_sk_alloc(struct sock
*sk
)
6343 struct mem_cgroup
*memcg
;
6345 if (!mem_cgroup_sockets_enabled
)
6349 * Socket cloning can throw us here with sk_memcg already
6350 * filled. It won't however, necessarily happen from
6351 * process context. So the test for root memcg given
6352 * the current task's memcg won't help us in this case.
6354 * Respecting the original socket's memcg is a better
6355 * decision in this case.
6358 css_get(&sk
->sk_memcg
->css
);
6363 memcg
= mem_cgroup_from_task(current
);
6364 if (memcg
== root_mem_cgroup
)
6366 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
6368 if (css_tryget_online(&memcg
->css
))
6369 sk
->sk_memcg
= memcg
;
6374 void mem_cgroup_sk_free(struct sock
*sk
)
6377 css_put(&sk
->sk_memcg
->css
);
6381 * mem_cgroup_charge_skmem - charge socket memory
6382 * @memcg: memcg to charge
6383 * @nr_pages: number of pages to charge
6385 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6386 * @memcg's configured limit, %false if the charge had to be forced.
6388 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
6390 gfp_t gfp_mask
= GFP_KERNEL
;
6392 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
6393 struct page_counter
*fail
;
6395 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
6396 memcg
->tcpmem_pressure
= 0;
6399 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
6400 memcg
->tcpmem_pressure
= 1;
6404 /* Don't block in the packet receive path */
6406 gfp_mask
= GFP_NOWAIT
;
6408 mod_memcg_state(memcg
, MEMCG_SOCK
, nr_pages
);
6410 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
6413 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
6418 * mem_cgroup_uncharge_skmem - uncharge socket memory
6419 * @memcg: memcg to uncharge
6420 * @nr_pages: number of pages to uncharge
6422 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
6424 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
6425 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
6429 mod_memcg_state(memcg
, MEMCG_SOCK
, -nr_pages
);
6431 refill_stock(memcg
, nr_pages
);
6434 static int __init
cgroup_memory(char *s
)
6438 while ((token
= strsep(&s
, ",")) != NULL
) {
6441 if (!strcmp(token
, "nosocket"))
6442 cgroup_memory_nosocket
= true;
6443 if (!strcmp(token
, "nokmem"))
6444 cgroup_memory_nokmem
= true;
6448 __setup("cgroup.memory=", cgroup_memory
);
6451 * subsys_initcall() for memory controller.
6453 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6454 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6455 * basically everything that doesn't depend on a specific mem_cgroup structure
6456 * should be initialized from here.
6458 static int __init
mem_cgroup_init(void)
6462 #ifdef CONFIG_MEMCG_KMEM
6464 * Kmem cache creation is mostly done with the slab_mutex held,
6465 * so use a workqueue with limited concurrency to avoid stalling
6466 * all worker threads in case lots of cgroups are created and
6467 * destroyed simultaneously.
6469 memcg_kmem_cache_wq
= alloc_workqueue("memcg_kmem_cache", 0, 1);
6470 BUG_ON(!memcg_kmem_cache_wq
);
6473 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD
, "mm/memctrl:dead", NULL
,
6474 memcg_hotplug_cpu_dead
);
6476 for_each_possible_cpu(cpu
)
6477 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
6480 for_each_node(node
) {
6481 struct mem_cgroup_tree_per_node
*rtpn
;
6483 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
6484 node_online(node
) ? node
: NUMA_NO_NODE
);
6486 rtpn
->rb_root
= RB_ROOT
;
6487 rtpn
->rb_rightmost
= NULL
;
6488 spin_lock_init(&rtpn
->lock
);
6489 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6494 subsys_initcall(mem_cgroup_init
);
6496 #ifdef CONFIG_MEMCG_SWAP
6497 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
6499 while (!refcount_inc_not_zero(&memcg
->id
.ref
)) {
6501 * The root cgroup cannot be destroyed, so it's refcount must
6504 if (WARN_ON_ONCE(memcg
== root_mem_cgroup
)) {
6508 memcg
= parent_mem_cgroup(memcg
);
6510 memcg
= root_mem_cgroup
;
6516 * mem_cgroup_swapout - transfer a memsw charge to swap
6517 * @page: page whose memsw charge to transfer
6518 * @entry: swap entry to move the charge to
6520 * Transfer the memsw charge of @page to @entry.
6522 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
6524 struct mem_cgroup
*memcg
, *swap_memcg
;
6525 unsigned int nr_entries
;
6526 unsigned short oldid
;
6528 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6529 VM_BUG_ON_PAGE(page_count(page
), page
);
6531 if (!do_memsw_account())
6534 memcg
= page
->mem_cgroup
;
6536 /* Readahead page, never charged */
6541 * In case the memcg owning these pages has been offlined and doesn't
6542 * have an ID allocated to it anymore, charge the closest online
6543 * ancestor for the swap instead and transfer the memory+swap charge.
6545 swap_memcg
= mem_cgroup_id_get_online(memcg
);
6546 nr_entries
= hpage_nr_pages(page
);
6547 /* Get references for the tail pages, too */
6549 mem_cgroup_id_get_many(swap_memcg
, nr_entries
- 1);
6550 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
),
6552 VM_BUG_ON_PAGE(oldid
, page
);
6553 mod_memcg_state(swap_memcg
, MEMCG_SWAP
, nr_entries
);
6555 page
->mem_cgroup
= NULL
;
6557 if (!mem_cgroup_is_root(memcg
))
6558 page_counter_uncharge(&memcg
->memory
, nr_entries
);
6560 if (memcg
!= swap_memcg
) {
6561 if (!mem_cgroup_is_root(swap_memcg
))
6562 page_counter_charge(&swap_memcg
->memsw
, nr_entries
);
6563 page_counter_uncharge(&memcg
->memsw
, nr_entries
);
6567 * Interrupts should be disabled here because the caller holds the
6568 * i_pages lock which is taken with interrupts-off. It is
6569 * important here to have the interrupts disabled because it is the
6570 * only synchronisation we have for updating the per-CPU variables.
6572 VM_BUG_ON(!irqs_disabled());
6573 mem_cgroup_charge_statistics(memcg
, page
, PageTransHuge(page
),
6575 memcg_check_events(memcg
, page
);
6577 if (!mem_cgroup_is_root(memcg
))
6578 css_put_many(&memcg
->css
, nr_entries
);
6582 * mem_cgroup_try_charge_swap - try charging swap space for a page
6583 * @page: page being added to swap
6584 * @entry: swap entry to charge
6586 * Try to charge @page's memcg for the swap space at @entry.
6588 * Returns 0 on success, -ENOMEM on failure.
6590 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
6592 unsigned int nr_pages
= hpage_nr_pages(page
);
6593 struct page_counter
*counter
;
6594 struct mem_cgroup
*memcg
;
6595 unsigned short oldid
;
6597 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
6600 memcg
= page
->mem_cgroup
;
6602 /* Readahead page, never charged */
6607 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
6611 memcg
= mem_cgroup_id_get_online(memcg
);
6613 if (!mem_cgroup_is_root(memcg
) &&
6614 !page_counter_try_charge(&memcg
->swap
, nr_pages
, &counter
)) {
6615 memcg_memory_event(memcg
, MEMCG_SWAP_MAX
);
6616 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
6617 mem_cgroup_id_put(memcg
);
6621 /* Get references for the tail pages, too */
6623 mem_cgroup_id_get_many(memcg
, nr_pages
- 1);
6624 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
), nr_pages
);
6625 VM_BUG_ON_PAGE(oldid
, page
);
6626 mod_memcg_state(memcg
, MEMCG_SWAP
, nr_pages
);
6632 * mem_cgroup_uncharge_swap - uncharge swap space
6633 * @entry: swap entry to uncharge
6634 * @nr_pages: the amount of swap space to uncharge
6636 void mem_cgroup_uncharge_swap(swp_entry_t entry
, unsigned int nr_pages
)
6638 struct mem_cgroup
*memcg
;
6641 if (!do_swap_account
)
6644 id
= swap_cgroup_record(entry
, 0, nr_pages
);
6646 memcg
= mem_cgroup_from_id(id
);
6648 if (!mem_cgroup_is_root(memcg
)) {
6649 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
6650 page_counter_uncharge(&memcg
->swap
, nr_pages
);
6652 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
6654 mod_memcg_state(memcg
, MEMCG_SWAP
, -nr_pages
);
6655 mem_cgroup_id_put_many(memcg
, nr_pages
);
6660 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
6662 long nr_swap_pages
= get_nr_swap_pages();
6664 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
6665 return nr_swap_pages
;
6666 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
6667 nr_swap_pages
= min_t(long, nr_swap_pages
,
6668 READ_ONCE(memcg
->swap
.max
) -
6669 page_counter_read(&memcg
->swap
));
6670 return nr_swap_pages
;
6673 bool mem_cgroup_swap_full(struct page
*page
)
6675 struct mem_cgroup
*memcg
;
6677 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
6681 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
6684 memcg
= page
->mem_cgroup
;
6688 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
6689 if (page_counter_read(&memcg
->swap
) * 2 >= memcg
->swap
.max
)
6695 /* for remember boot option*/
6696 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6697 static int really_do_swap_account __initdata
= 1;
6699 static int really_do_swap_account __initdata
;
6702 static int __init
enable_swap_account(char *s
)
6704 if (!strcmp(s
, "1"))
6705 really_do_swap_account
= 1;
6706 else if (!strcmp(s
, "0"))
6707 really_do_swap_account
= 0;
6710 __setup("swapaccount=", enable_swap_account
);
6712 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
6715 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6717 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
6720 static int swap_max_show(struct seq_file
*m
, void *v
)
6722 return seq_puts_memcg_tunable(m
,
6723 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.max
));
6726 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
6727 char *buf
, size_t nbytes
, loff_t off
)
6729 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6733 buf
= strstrip(buf
);
6734 err
= page_counter_memparse(buf
, "max", &max
);
6738 xchg(&memcg
->swap
.max
, max
);
6743 static int swap_events_show(struct seq_file
*m
, void *v
)
6745 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6747 seq_printf(m
, "max %lu\n",
6748 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_MAX
]));
6749 seq_printf(m
, "fail %lu\n",
6750 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_FAIL
]));
6755 static struct cftype swap_files
[] = {
6757 .name
= "swap.current",
6758 .flags
= CFTYPE_NOT_ON_ROOT
,
6759 .read_u64
= swap_current_read
,
6763 .flags
= CFTYPE_NOT_ON_ROOT
,
6764 .seq_show
= swap_max_show
,
6765 .write
= swap_max_write
,
6768 .name
= "swap.events",
6769 .flags
= CFTYPE_NOT_ON_ROOT
,
6770 .file_offset
= offsetof(struct mem_cgroup
, swap_events_file
),
6771 .seq_show
= swap_events_show
,
6776 static struct cftype memsw_cgroup_files
[] = {
6778 .name
= "memsw.usage_in_bytes",
6779 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6780 .read_u64
= mem_cgroup_read_u64
,
6783 .name
= "memsw.max_usage_in_bytes",
6784 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6785 .write
= mem_cgroup_reset
,
6786 .read_u64
= mem_cgroup_read_u64
,
6789 .name
= "memsw.limit_in_bytes",
6790 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6791 .write
= mem_cgroup_write
,
6792 .read_u64
= mem_cgroup_read_u64
,
6795 .name
= "memsw.failcnt",
6796 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6797 .write
= mem_cgroup_reset
,
6798 .read_u64
= mem_cgroup_read_u64
,
6800 { }, /* terminate */
6803 static int __init
mem_cgroup_swap_init(void)
6805 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6806 do_swap_account
= 1;
6807 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
6809 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
6810 memsw_cgroup_files
));
6814 subsys_initcall(mem_cgroup_swap_init
);
6816 #endif /* CONFIG_MEMCG_SWAP */