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
24 * Per memcg lru locking
25 * Copyright (C) 2020 Alibaba, Inc, Alex Shi
28 #include <linux/page_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
31 #include <linux/pagewalk.h>
32 #include <linux/sched/mm.h>
33 #include <linux/shmem_fs.h>
34 #include <linux/hugetlb.h>
35 #include <linux/pagemap.h>
36 #include <linux/vm_event_item.h>
37 #include <linux/smp.h>
38 #include <linux/page-flags.h>
39 #include <linux/backing-dev.h>
40 #include <linux/bit_spinlock.h>
41 #include <linux/rcupdate.h>
42 #include <linux/limits.h>
43 #include <linux/export.h>
44 #include <linux/mutex.h>
45 #include <linux/rbtree.h>
46 #include <linux/slab.h>
47 #include <linux/swap.h>
48 #include <linux/swapops.h>
49 #include <linux/spinlock.h>
50 #include <linux/eventfd.h>
51 #include <linux/poll.h>
52 #include <linux/sort.h>
54 #include <linux/seq_file.h>
55 #include <linux/vmpressure.h>
56 #include <linux/mm_inline.h>
57 #include <linux/swap_cgroup.h>
58 #include <linux/cpu.h>
59 #include <linux/oom.h>
60 #include <linux/lockdep.h>
61 #include <linux/file.h>
62 #include <linux/tracehook.h>
63 #include <linux/psi.h>
64 #include <linux/seq_buf.h>
70 #include <linux/uaccess.h>
72 #include <trace/events/vmscan.h>
74 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
75 EXPORT_SYMBOL(memory_cgrp_subsys
);
77 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
79 /* Active memory cgroup to use from an interrupt context */
80 DEFINE_PER_CPU(struct mem_cgroup
*, int_active_memcg
);
82 /* Socket memory accounting disabled? */
83 static bool cgroup_memory_nosocket
;
85 /* Kernel memory accounting disabled? */
86 static bool cgroup_memory_nokmem
;
88 /* Whether the swap controller is active */
89 #ifdef CONFIG_MEMCG_SWAP
90 bool cgroup_memory_noswap __read_mostly
;
92 #define cgroup_memory_noswap 1
95 #ifdef CONFIG_CGROUP_WRITEBACK
96 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq
);
99 /* Whether legacy memory+swap accounting is active */
100 static bool do_memsw_account(void)
102 return !cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_noswap
;
105 #define THRESHOLDS_EVENTS_TARGET 128
106 #define SOFTLIMIT_EVENTS_TARGET 1024
109 * Cgroups above their limits are maintained in a RB-Tree, independent of
110 * their hierarchy representation
113 struct mem_cgroup_tree_per_node
{
114 struct rb_root rb_root
;
115 struct rb_node
*rb_rightmost
;
119 struct mem_cgroup_tree
{
120 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
123 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
126 struct mem_cgroup_eventfd_list
{
127 struct list_head list
;
128 struct eventfd_ctx
*eventfd
;
132 * cgroup_event represents events which userspace want to receive.
134 struct mem_cgroup_event
{
136 * memcg which the event belongs to.
138 struct mem_cgroup
*memcg
;
140 * eventfd to signal userspace about the event.
142 struct eventfd_ctx
*eventfd
;
144 * Each of these stored in a list by the cgroup.
146 struct list_head list
;
148 * register_event() callback will be used to add new userspace
149 * waiter for changes related to this event. Use eventfd_signal()
150 * on eventfd to send notification to userspace.
152 int (*register_event
)(struct mem_cgroup
*memcg
,
153 struct eventfd_ctx
*eventfd
, const char *args
);
155 * unregister_event() callback will be called when userspace closes
156 * the eventfd or on cgroup removing. This callback must be set,
157 * if you want provide notification functionality.
159 void (*unregister_event
)(struct mem_cgroup
*memcg
,
160 struct eventfd_ctx
*eventfd
);
162 * All fields below needed to unregister event when
163 * userspace closes eventfd.
166 wait_queue_head_t
*wqh
;
167 wait_queue_entry_t wait
;
168 struct work_struct remove
;
171 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
172 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
174 /* Stuffs for move charges at task migration. */
176 * Types of charges to be moved.
178 #define MOVE_ANON 0x1U
179 #define MOVE_FILE 0x2U
180 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
182 /* "mc" and its members are protected by cgroup_mutex */
183 static struct move_charge_struct
{
184 spinlock_t lock
; /* for from, to */
185 struct mm_struct
*mm
;
186 struct mem_cgroup
*from
;
187 struct mem_cgroup
*to
;
189 unsigned long precharge
;
190 unsigned long moved_charge
;
191 unsigned long moved_swap
;
192 struct task_struct
*moving_task
; /* a task moving charges */
193 wait_queue_head_t waitq
; /* a waitq for other context */
195 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
196 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
200 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
201 * limit reclaim to prevent infinite loops, if they ever occur.
203 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
204 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
206 /* for encoding cft->private value on file */
215 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
216 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
217 #define MEMFILE_ATTR(val) ((val) & 0xffff)
218 /* Used for OOM nofiier */
219 #define OOM_CONTROL (0)
222 * Iteration constructs for visiting all cgroups (under a tree). If
223 * loops are exited prematurely (break), mem_cgroup_iter_break() must
224 * be used for reference counting.
226 #define for_each_mem_cgroup_tree(iter, root) \
227 for (iter = mem_cgroup_iter(root, NULL, NULL); \
229 iter = mem_cgroup_iter(root, iter, NULL))
231 #define for_each_mem_cgroup(iter) \
232 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
234 iter = mem_cgroup_iter(NULL, iter, NULL))
236 static inline bool should_force_charge(void)
238 return tsk_is_oom_victim(current
) || fatal_signal_pending(current
) ||
239 (current
->flags
& PF_EXITING
);
242 /* Some nice accessors for the vmpressure. */
243 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
246 memcg
= root_mem_cgroup
;
247 return &memcg
->vmpressure
;
250 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
252 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
255 #ifdef CONFIG_MEMCG_KMEM
256 extern spinlock_t css_set_lock
;
258 static void obj_cgroup_release(struct percpu_ref
*ref
)
260 struct obj_cgroup
*objcg
= container_of(ref
, struct obj_cgroup
, refcnt
);
261 struct mem_cgroup
*memcg
;
262 unsigned int nr_bytes
;
263 unsigned int nr_pages
;
267 * At this point all allocated objects are freed, and
268 * objcg->nr_charged_bytes can't have an arbitrary byte value.
269 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
271 * The following sequence can lead to it:
272 * 1) CPU0: objcg == stock->cached_objcg
273 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
274 * PAGE_SIZE bytes are charged
275 * 3) CPU1: a process from another memcg is allocating something,
276 * the stock if flushed,
277 * objcg->nr_charged_bytes = PAGE_SIZE - 92
278 * 5) CPU0: we do release this object,
279 * 92 bytes are added to stock->nr_bytes
280 * 6) CPU0: stock is flushed,
281 * 92 bytes are added to objcg->nr_charged_bytes
283 * In the result, nr_charged_bytes == PAGE_SIZE.
284 * This page will be uncharged in obj_cgroup_release().
286 nr_bytes
= atomic_read(&objcg
->nr_charged_bytes
);
287 WARN_ON_ONCE(nr_bytes
& (PAGE_SIZE
- 1));
288 nr_pages
= nr_bytes
>> PAGE_SHIFT
;
290 spin_lock_irqsave(&css_set_lock
, flags
);
291 memcg
= obj_cgroup_memcg(objcg
);
293 __memcg_kmem_uncharge(memcg
, nr_pages
);
294 list_del(&objcg
->list
);
295 mem_cgroup_put(memcg
);
296 spin_unlock_irqrestore(&css_set_lock
, flags
);
298 percpu_ref_exit(ref
);
299 kfree_rcu(objcg
, rcu
);
302 static struct obj_cgroup
*obj_cgroup_alloc(void)
304 struct obj_cgroup
*objcg
;
307 objcg
= kzalloc(sizeof(struct obj_cgroup
), GFP_KERNEL
);
311 ret
= percpu_ref_init(&objcg
->refcnt
, obj_cgroup_release
, 0,
317 INIT_LIST_HEAD(&objcg
->list
);
321 static void memcg_reparent_objcgs(struct mem_cgroup
*memcg
,
322 struct mem_cgroup
*parent
)
324 struct obj_cgroup
*objcg
, *iter
;
326 objcg
= rcu_replace_pointer(memcg
->objcg
, NULL
, true);
328 spin_lock_irq(&css_set_lock
);
330 /* Move active objcg to the parent's list */
331 xchg(&objcg
->memcg
, parent
);
332 css_get(&parent
->css
);
333 list_add(&objcg
->list
, &parent
->objcg_list
);
335 /* Move already reparented objcgs to the parent's list */
336 list_for_each_entry(iter
, &memcg
->objcg_list
, list
) {
337 css_get(&parent
->css
);
338 xchg(&iter
->memcg
, parent
);
339 css_put(&memcg
->css
);
341 list_splice(&memcg
->objcg_list
, &parent
->objcg_list
);
343 spin_unlock_irq(&css_set_lock
);
345 percpu_ref_kill(&objcg
->refcnt
);
349 * This will be used as a shrinker list's index.
350 * The main reason for not using cgroup id for this:
351 * this works better in sparse environments, where we have a lot of memcgs,
352 * but only a few kmem-limited. Or also, if we have, for instance, 200
353 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
354 * 200 entry array for that.
356 * The current size of the caches array is stored in memcg_nr_cache_ids. It
357 * will double each time we have to increase it.
359 static DEFINE_IDA(memcg_cache_ida
);
360 int memcg_nr_cache_ids
;
362 /* Protects memcg_nr_cache_ids */
363 static DECLARE_RWSEM(memcg_cache_ids_sem
);
365 void memcg_get_cache_ids(void)
367 down_read(&memcg_cache_ids_sem
);
370 void memcg_put_cache_ids(void)
372 up_read(&memcg_cache_ids_sem
);
376 * MIN_SIZE is different than 1, because we would like to avoid going through
377 * the alloc/free process all the time. In a small machine, 4 kmem-limited
378 * cgroups is a reasonable guess. In the future, it could be a parameter or
379 * tunable, but that is strictly not necessary.
381 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
382 * this constant directly from cgroup, but it is understandable that this is
383 * better kept as an internal representation in cgroup.c. In any case, the
384 * cgrp_id space is not getting any smaller, and we don't have to necessarily
385 * increase ours as well if it increases.
387 #define MEMCG_CACHES_MIN_SIZE 4
388 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
391 * A lot of the calls to the cache allocation functions are expected to be
392 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
393 * conditional to this static branch, we'll have to allow modules that does
394 * kmem_cache_alloc and the such to see this symbol as well
396 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
397 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
400 static int memcg_shrinker_map_size
;
401 static DEFINE_MUTEX(memcg_shrinker_map_mutex
);
403 static void memcg_free_shrinker_map_rcu(struct rcu_head
*head
)
405 kvfree(container_of(head
, struct memcg_shrinker_map
, rcu
));
408 static int memcg_expand_one_shrinker_map(struct mem_cgroup
*memcg
,
409 int size
, int old_size
)
411 struct memcg_shrinker_map
*new, *old
;
414 lockdep_assert_held(&memcg_shrinker_map_mutex
);
417 old
= rcu_dereference_protected(
418 mem_cgroup_nodeinfo(memcg
, nid
)->shrinker_map
, true);
419 /* Not yet online memcg */
423 new = kvmalloc_node(sizeof(*new) + size
, GFP_KERNEL
, nid
);
427 /* Set all old bits, clear all new bits */
428 memset(new->map
, (int)0xff, old_size
);
429 memset((void *)new->map
+ old_size
, 0, size
- old_size
);
431 rcu_assign_pointer(memcg
->nodeinfo
[nid
]->shrinker_map
, new);
432 call_rcu(&old
->rcu
, memcg_free_shrinker_map_rcu
);
438 static void memcg_free_shrinker_maps(struct mem_cgroup
*memcg
)
440 struct mem_cgroup_per_node
*pn
;
441 struct memcg_shrinker_map
*map
;
444 if (mem_cgroup_is_root(memcg
))
448 pn
= mem_cgroup_nodeinfo(memcg
, nid
);
449 map
= rcu_dereference_protected(pn
->shrinker_map
, true);
452 rcu_assign_pointer(pn
->shrinker_map
, NULL
);
456 static int memcg_alloc_shrinker_maps(struct mem_cgroup
*memcg
)
458 struct memcg_shrinker_map
*map
;
459 int nid
, size
, ret
= 0;
461 if (mem_cgroup_is_root(memcg
))
464 mutex_lock(&memcg_shrinker_map_mutex
);
465 size
= memcg_shrinker_map_size
;
467 map
= kvzalloc_node(sizeof(*map
) + size
, GFP_KERNEL
, nid
);
469 memcg_free_shrinker_maps(memcg
);
473 rcu_assign_pointer(memcg
->nodeinfo
[nid
]->shrinker_map
, map
);
475 mutex_unlock(&memcg_shrinker_map_mutex
);
480 int memcg_expand_shrinker_maps(int new_id
)
482 int size
, old_size
, ret
= 0;
483 struct mem_cgroup
*memcg
;
485 size
= DIV_ROUND_UP(new_id
+ 1, BITS_PER_LONG
) * sizeof(unsigned long);
486 old_size
= memcg_shrinker_map_size
;
487 if (size
<= old_size
)
490 mutex_lock(&memcg_shrinker_map_mutex
);
491 if (!root_mem_cgroup
)
494 for_each_mem_cgroup(memcg
) {
495 if (mem_cgroup_is_root(memcg
))
497 ret
= memcg_expand_one_shrinker_map(memcg
, size
, old_size
);
499 mem_cgroup_iter_break(NULL
, memcg
);
505 memcg_shrinker_map_size
= size
;
506 mutex_unlock(&memcg_shrinker_map_mutex
);
510 void memcg_set_shrinker_bit(struct mem_cgroup
*memcg
, int nid
, int shrinker_id
)
512 if (shrinker_id
>= 0 && memcg
&& !mem_cgroup_is_root(memcg
)) {
513 struct memcg_shrinker_map
*map
;
516 map
= rcu_dereference(memcg
->nodeinfo
[nid
]->shrinker_map
);
517 /* Pairs with smp mb in shrink_slab() */
518 smp_mb__before_atomic();
519 set_bit(shrinker_id
, map
->map
);
525 * mem_cgroup_css_from_page - css of the memcg associated with a page
526 * @page: page of interest
528 * If memcg is bound to the default hierarchy, css of the memcg associated
529 * with @page is returned. The returned css remains associated with @page
530 * until it is released.
532 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
535 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
537 struct mem_cgroup
*memcg
;
539 memcg
= page_memcg(page
);
541 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
542 memcg
= root_mem_cgroup
;
548 * page_cgroup_ino - return inode number of the memcg a page is charged to
551 * Look up the closest online ancestor of the memory cgroup @page is charged to
552 * and return its inode number or 0 if @page is not charged to any cgroup. It
553 * is safe to call this function without holding a reference to @page.
555 * Note, this function is inherently racy, because there is nothing to prevent
556 * the cgroup inode from getting torn down and potentially reallocated a moment
557 * after page_cgroup_ino() returns, so it only should be used by callers that
558 * do not care (such as procfs interfaces).
560 ino_t
page_cgroup_ino(struct page
*page
)
562 struct mem_cgroup
*memcg
;
563 unsigned long ino
= 0;
566 memcg
= page_memcg_check(page
);
568 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
569 memcg
= parent_mem_cgroup(memcg
);
571 ino
= cgroup_ino(memcg
->css
.cgroup
);
576 static struct mem_cgroup_per_node
*
577 mem_cgroup_page_nodeinfo(struct mem_cgroup
*memcg
, struct page
*page
)
579 int nid
= page_to_nid(page
);
581 return memcg
->nodeinfo
[nid
];
584 static struct mem_cgroup_tree_per_node
*
585 soft_limit_tree_node(int nid
)
587 return soft_limit_tree
.rb_tree_per_node
[nid
];
590 static struct mem_cgroup_tree_per_node
*
591 soft_limit_tree_from_page(struct page
*page
)
593 int nid
= page_to_nid(page
);
595 return soft_limit_tree
.rb_tree_per_node
[nid
];
598 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node
*mz
,
599 struct mem_cgroup_tree_per_node
*mctz
,
600 unsigned long new_usage_in_excess
)
602 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
603 struct rb_node
*parent
= NULL
;
604 struct mem_cgroup_per_node
*mz_node
;
605 bool rightmost
= true;
610 mz
->usage_in_excess
= new_usage_in_excess
;
611 if (!mz
->usage_in_excess
)
615 mz_node
= rb_entry(parent
, struct mem_cgroup_per_node
,
617 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
) {
626 mctz
->rb_rightmost
= &mz
->tree_node
;
628 rb_link_node(&mz
->tree_node
, parent
, p
);
629 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
633 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
634 struct mem_cgroup_tree_per_node
*mctz
)
639 if (&mz
->tree_node
== mctz
->rb_rightmost
)
640 mctz
->rb_rightmost
= rb_prev(&mz
->tree_node
);
642 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
646 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
647 struct mem_cgroup_tree_per_node
*mctz
)
651 spin_lock_irqsave(&mctz
->lock
, flags
);
652 __mem_cgroup_remove_exceeded(mz
, mctz
);
653 spin_unlock_irqrestore(&mctz
->lock
, flags
);
656 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
658 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
659 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
660 unsigned long excess
= 0;
662 if (nr_pages
> soft_limit
)
663 excess
= nr_pages
- soft_limit
;
668 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
670 unsigned long excess
;
671 struct mem_cgroup_per_node
*mz
;
672 struct mem_cgroup_tree_per_node
*mctz
;
674 mctz
= soft_limit_tree_from_page(page
);
678 * Necessary to update all ancestors when hierarchy is used.
679 * because their event counter is not touched.
681 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
682 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
683 excess
= soft_limit_excess(memcg
);
685 * We have to update the tree if mz is on RB-tree or
686 * mem is over its softlimit.
688 if (excess
|| mz
->on_tree
) {
691 spin_lock_irqsave(&mctz
->lock
, flags
);
692 /* if on-tree, remove it */
694 __mem_cgroup_remove_exceeded(mz
, mctz
);
696 * Insert again. mz->usage_in_excess will be updated.
697 * If excess is 0, no tree ops.
699 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
700 spin_unlock_irqrestore(&mctz
->lock
, flags
);
705 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
707 struct mem_cgroup_tree_per_node
*mctz
;
708 struct mem_cgroup_per_node
*mz
;
712 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
713 mctz
= soft_limit_tree_node(nid
);
715 mem_cgroup_remove_exceeded(mz
, mctz
);
719 static struct mem_cgroup_per_node
*
720 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
722 struct mem_cgroup_per_node
*mz
;
726 if (!mctz
->rb_rightmost
)
727 goto done
; /* Nothing to reclaim from */
729 mz
= rb_entry(mctz
->rb_rightmost
,
730 struct mem_cgroup_per_node
, tree_node
);
732 * Remove the node now but someone else can add it back,
733 * we will to add it back at the end of reclaim to its correct
734 * position in the tree.
736 __mem_cgroup_remove_exceeded(mz
, mctz
);
737 if (!soft_limit_excess(mz
->memcg
) ||
738 !css_tryget(&mz
->memcg
->css
))
744 static struct mem_cgroup_per_node
*
745 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
747 struct mem_cgroup_per_node
*mz
;
749 spin_lock_irq(&mctz
->lock
);
750 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
751 spin_unlock_irq(&mctz
->lock
);
756 * __mod_memcg_state - update cgroup memory statistics
757 * @memcg: the memory cgroup
758 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
759 * @val: delta to add to the counter, can be negative
761 void __mod_memcg_state(struct mem_cgroup
*memcg
, int idx
, int val
)
763 long x
, threshold
= MEMCG_CHARGE_BATCH
;
765 if (mem_cgroup_disabled())
768 if (memcg_stat_item_in_bytes(idx
))
769 threshold
<<= PAGE_SHIFT
;
771 x
= val
+ __this_cpu_read(memcg
->vmstats_percpu
->stat
[idx
]);
772 if (unlikely(abs(x
) > threshold
)) {
773 struct mem_cgroup
*mi
;
776 * Batch local counters to keep them in sync with
777 * the hierarchical ones.
779 __this_cpu_add(memcg
->vmstats_local
->stat
[idx
], x
);
780 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
781 atomic_long_add(x
, &mi
->vmstats
[idx
]);
784 __this_cpu_write(memcg
->vmstats_percpu
->stat
[idx
], x
);
787 static struct mem_cgroup_per_node
*
788 parent_nodeinfo(struct mem_cgroup_per_node
*pn
, int nid
)
790 struct mem_cgroup
*parent
;
792 parent
= parent_mem_cgroup(pn
->memcg
);
795 return mem_cgroup_nodeinfo(parent
, nid
);
798 void __mod_memcg_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
801 struct mem_cgroup_per_node
*pn
;
802 struct mem_cgroup
*memcg
;
803 long x
, threshold
= MEMCG_CHARGE_BATCH
;
805 pn
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
809 __mod_memcg_state(memcg
, idx
, val
);
812 __this_cpu_add(pn
->lruvec_stat_local
->count
[idx
], val
);
814 if (vmstat_item_in_bytes(idx
))
815 threshold
<<= PAGE_SHIFT
;
817 x
= val
+ __this_cpu_read(pn
->lruvec_stat_cpu
->count
[idx
]);
818 if (unlikely(abs(x
) > threshold
)) {
819 pg_data_t
*pgdat
= lruvec_pgdat(lruvec
);
820 struct mem_cgroup_per_node
*pi
;
822 for (pi
= pn
; pi
; pi
= parent_nodeinfo(pi
, pgdat
->node_id
))
823 atomic_long_add(x
, &pi
->lruvec_stat
[idx
]);
826 __this_cpu_write(pn
->lruvec_stat_cpu
->count
[idx
], x
);
830 * __mod_lruvec_state - update lruvec memory statistics
831 * @lruvec: the lruvec
832 * @idx: the stat item
833 * @val: delta to add to the counter, can be negative
835 * The lruvec is the intersection of the NUMA node and a cgroup. This
836 * function updates the all three counters that are affected by a
837 * change of state at this level: per-node, per-cgroup, per-lruvec.
839 void __mod_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
843 __mod_node_page_state(lruvec_pgdat(lruvec
), idx
, val
);
845 /* Update memcg and lruvec */
846 if (!mem_cgroup_disabled())
847 __mod_memcg_lruvec_state(lruvec
, idx
, val
);
850 void __mod_lruvec_page_state(struct page
*page
, enum node_stat_item idx
,
853 struct page
*head
= compound_head(page
); /* rmap on tail pages */
854 struct mem_cgroup
*memcg
= page_memcg(head
);
855 pg_data_t
*pgdat
= page_pgdat(page
);
856 struct lruvec
*lruvec
;
858 /* Untracked pages have no memcg, no lruvec. Update only the node */
860 __mod_node_page_state(pgdat
, idx
, val
);
864 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
865 __mod_lruvec_state(lruvec
, idx
, val
);
867 EXPORT_SYMBOL(__mod_lruvec_page_state
);
869 void __mod_lruvec_kmem_state(void *p
, enum node_stat_item idx
, int val
)
871 pg_data_t
*pgdat
= page_pgdat(virt_to_page(p
));
872 struct mem_cgroup
*memcg
;
873 struct lruvec
*lruvec
;
876 memcg
= mem_cgroup_from_obj(p
);
879 * Untracked pages have no memcg, no lruvec. Update only the
880 * node. If we reparent the slab objects to the root memcg,
881 * when we free the slab object, we need to update the per-memcg
882 * vmstats to keep it correct for the root memcg.
885 __mod_node_page_state(pgdat
, idx
, val
);
887 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
888 __mod_lruvec_state(lruvec
, idx
, val
);
894 * __count_memcg_events - account VM events in a cgroup
895 * @memcg: the memory cgroup
896 * @idx: the event item
897 * @count: the number of events that occured
899 void __count_memcg_events(struct mem_cgroup
*memcg
, enum vm_event_item idx
,
904 if (mem_cgroup_disabled())
907 x
= count
+ __this_cpu_read(memcg
->vmstats_percpu
->events
[idx
]);
908 if (unlikely(x
> MEMCG_CHARGE_BATCH
)) {
909 struct mem_cgroup
*mi
;
912 * Batch local counters to keep them in sync with
913 * the hierarchical ones.
915 __this_cpu_add(memcg
->vmstats_local
->events
[idx
], x
);
916 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
917 atomic_long_add(x
, &mi
->vmevents
[idx
]);
920 __this_cpu_write(memcg
->vmstats_percpu
->events
[idx
], x
);
923 static unsigned long memcg_events(struct mem_cgroup
*memcg
, int event
)
925 return atomic_long_read(&memcg
->vmevents
[event
]);
928 static unsigned long memcg_events_local(struct mem_cgroup
*memcg
, int event
)
933 for_each_possible_cpu(cpu
)
934 x
+= per_cpu(memcg
->vmstats_local
->events
[event
], cpu
);
938 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
942 /* pagein of a big page is an event. So, ignore page size */
944 __count_memcg_events(memcg
, PGPGIN
, 1);
946 __count_memcg_events(memcg
, PGPGOUT
, 1);
947 nr_pages
= -nr_pages
; /* for event */
950 __this_cpu_add(memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
953 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
954 enum mem_cgroup_events_target target
)
956 unsigned long val
, next
;
958 val
= __this_cpu_read(memcg
->vmstats_percpu
->nr_page_events
);
959 next
= __this_cpu_read(memcg
->vmstats_percpu
->targets
[target
]);
960 /* from time_after() in jiffies.h */
961 if ((long)(next
- val
) < 0) {
963 case MEM_CGROUP_TARGET_THRESH
:
964 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
966 case MEM_CGROUP_TARGET_SOFTLIMIT
:
967 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
972 __this_cpu_write(memcg
->vmstats_percpu
->targets
[target
], next
);
979 * Check events in order.
982 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
984 /* threshold event is triggered in finer grain than soft limit */
985 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
986 MEM_CGROUP_TARGET_THRESH
))) {
989 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
990 MEM_CGROUP_TARGET_SOFTLIMIT
);
991 mem_cgroup_threshold(memcg
);
992 if (unlikely(do_softlimit
))
993 mem_cgroup_update_tree(memcg
, page
);
997 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1000 * mm_update_next_owner() may clear mm->owner to NULL
1001 * if it races with swapoff, page migration, etc.
1002 * So this can be called with p == NULL.
1007 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
1009 EXPORT_SYMBOL(mem_cgroup_from_task
);
1012 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1013 * @mm: mm from which memcg should be extracted. It can be NULL.
1015 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1016 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1019 struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1021 struct mem_cgroup
*memcg
;
1023 if (mem_cgroup_disabled())
1029 * Page cache insertions can happen withou an
1030 * actual mm context, e.g. during disk probing
1031 * on boot, loopback IO, acct() writes etc.
1034 memcg
= root_mem_cgroup
;
1036 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1037 if (unlikely(!memcg
))
1038 memcg
= root_mem_cgroup
;
1040 } while (!css_tryget(&memcg
->css
));
1044 EXPORT_SYMBOL(get_mem_cgroup_from_mm
);
1047 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1048 * @page: page from which memcg should be extracted.
1050 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1051 * root_mem_cgroup is returned.
1053 struct mem_cgroup
*get_mem_cgroup_from_page(struct page
*page
)
1055 struct mem_cgroup
*memcg
= page_memcg(page
);
1057 if (mem_cgroup_disabled())
1061 /* Page should not get uncharged and freed memcg under us. */
1062 if (!memcg
|| WARN_ON_ONCE(!css_tryget(&memcg
->css
)))
1063 memcg
= root_mem_cgroup
;
1067 EXPORT_SYMBOL(get_mem_cgroup_from_page
);
1069 static __always_inline
struct mem_cgroup
*active_memcg(void)
1072 return this_cpu_read(int_active_memcg
);
1074 return current
->active_memcg
;
1077 static __always_inline
struct mem_cgroup
*get_active_memcg(void)
1079 struct mem_cgroup
*memcg
;
1082 memcg
= active_memcg();
1083 /* remote memcg must hold a ref. */
1084 if (memcg
&& WARN_ON_ONCE(!css_tryget(&memcg
->css
)))
1085 memcg
= root_mem_cgroup
;
1091 static __always_inline
bool memcg_kmem_bypass(void)
1093 /* Allow remote memcg charging from any context. */
1094 if (unlikely(active_memcg()))
1097 /* Memcg to charge can't be determined. */
1098 if (in_interrupt() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
1105 * If active memcg is set, do not fallback to current->mm->memcg.
1107 static __always_inline
struct mem_cgroup
*get_mem_cgroup_from_current(void)
1109 if (memcg_kmem_bypass())
1112 if (unlikely(active_memcg()))
1113 return get_active_memcg();
1115 return get_mem_cgroup_from_mm(current
->mm
);
1119 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1120 * @root: hierarchy root
1121 * @prev: previously returned memcg, NULL on first invocation
1122 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1124 * Returns references to children of the hierarchy below @root, or
1125 * @root itself, or %NULL after a full round-trip.
1127 * Caller must pass the return value in @prev on subsequent
1128 * invocations for reference counting, or use mem_cgroup_iter_break()
1129 * to cancel a hierarchy walk before the round-trip is complete.
1131 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1132 * in the hierarchy among all concurrent reclaimers operating on the
1135 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1136 struct mem_cgroup
*prev
,
1137 struct mem_cgroup_reclaim_cookie
*reclaim
)
1139 struct mem_cgroup_reclaim_iter
*iter
;
1140 struct cgroup_subsys_state
*css
= NULL
;
1141 struct mem_cgroup
*memcg
= NULL
;
1142 struct mem_cgroup
*pos
= NULL
;
1144 if (mem_cgroup_disabled())
1148 root
= root_mem_cgroup
;
1150 if (prev
&& !reclaim
)
1156 struct mem_cgroup_per_node
*mz
;
1158 mz
= mem_cgroup_nodeinfo(root
, reclaim
->pgdat
->node_id
);
1161 if (prev
&& reclaim
->generation
!= iter
->generation
)
1165 pos
= READ_ONCE(iter
->position
);
1166 if (!pos
|| css_tryget(&pos
->css
))
1169 * css reference reached zero, so iter->position will
1170 * be cleared by ->css_released. However, we should not
1171 * rely on this happening soon, because ->css_released
1172 * is called from a work queue, and by busy-waiting we
1173 * might block it. So we clear iter->position right
1176 (void)cmpxchg(&iter
->position
, pos
, NULL
);
1184 css
= css_next_descendant_pre(css
, &root
->css
);
1187 * Reclaimers share the hierarchy walk, and a
1188 * new one might jump in right at the end of
1189 * the hierarchy - make sure they see at least
1190 * one group and restart from the beginning.
1198 * Verify the css and acquire a reference. The root
1199 * is provided by the caller, so we know it's alive
1200 * and kicking, and don't take an extra reference.
1202 memcg
= mem_cgroup_from_css(css
);
1204 if (css
== &root
->css
)
1207 if (css_tryget(css
))
1215 * The position could have already been updated by a competing
1216 * thread, so check that the value hasn't changed since we read
1217 * it to avoid reclaiming from the same cgroup twice.
1219 (void)cmpxchg(&iter
->position
, pos
, memcg
);
1227 reclaim
->generation
= iter
->generation
;
1232 if (prev
&& prev
!= root
)
1233 css_put(&prev
->css
);
1239 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1240 * @root: hierarchy root
1241 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1243 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1244 struct mem_cgroup
*prev
)
1247 root
= root_mem_cgroup
;
1248 if (prev
&& prev
!= root
)
1249 css_put(&prev
->css
);
1252 static void __invalidate_reclaim_iterators(struct mem_cgroup
*from
,
1253 struct mem_cgroup
*dead_memcg
)
1255 struct mem_cgroup_reclaim_iter
*iter
;
1256 struct mem_cgroup_per_node
*mz
;
1259 for_each_node(nid
) {
1260 mz
= mem_cgroup_nodeinfo(from
, nid
);
1262 cmpxchg(&iter
->position
, dead_memcg
, NULL
);
1266 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
1268 struct mem_cgroup
*memcg
= dead_memcg
;
1269 struct mem_cgroup
*last
;
1272 __invalidate_reclaim_iterators(memcg
, dead_memcg
);
1274 } while ((memcg
= parent_mem_cgroup(memcg
)));
1277 * When cgruop1 non-hierarchy mode is used,
1278 * parent_mem_cgroup() does not walk all the way up to the
1279 * cgroup root (root_mem_cgroup). So we have to handle
1280 * dead_memcg from cgroup root separately.
1282 if (last
!= root_mem_cgroup
)
1283 __invalidate_reclaim_iterators(root_mem_cgroup
,
1288 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1289 * @memcg: hierarchy root
1290 * @fn: function to call for each task
1291 * @arg: argument passed to @fn
1293 * This function iterates over tasks attached to @memcg or to any of its
1294 * descendants and calls @fn for each task. If @fn returns a non-zero
1295 * value, the function breaks the iteration loop and returns the value.
1296 * Otherwise, it will iterate over all tasks and return 0.
1298 * This function must not be called for the root memory cgroup.
1300 int mem_cgroup_scan_tasks(struct mem_cgroup
*memcg
,
1301 int (*fn
)(struct task_struct
*, void *), void *arg
)
1303 struct mem_cgroup
*iter
;
1306 BUG_ON(memcg
== root_mem_cgroup
);
1308 for_each_mem_cgroup_tree(iter
, memcg
) {
1309 struct css_task_iter it
;
1310 struct task_struct
*task
;
1312 css_task_iter_start(&iter
->css
, CSS_TASK_ITER_PROCS
, &it
);
1313 while (!ret
&& (task
= css_task_iter_next(&it
)))
1314 ret
= fn(task
, arg
);
1315 css_task_iter_end(&it
);
1317 mem_cgroup_iter_break(memcg
, iter
);
1324 #ifdef CONFIG_DEBUG_VM
1325 void lruvec_memcg_debug(struct lruvec
*lruvec
, struct page
*page
)
1327 struct mem_cgroup
*memcg
;
1329 if (mem_cgroup_disabled())
1332 memcg
= page_memcg(page
);
1335 VM_BUG_ON_PAGE(lruvec_memcg(lruvec
) != root_mem_cgroup
, page
);
1337 VM_BUG_ON_PAGE(lruvec_memcg(lruvec
) != memcg
, page
);
1342 * lock_page_lruvec - lock and return lruvec for a given page.
1345 * This series functions should be used in either conditions:
1346 * PageLRU is cleared or unset
1347 * or page->_refcount is zero
1348 * or page is locked.
1350 struct lruvec
*lock_page_lruvec(struct page
*page
)
1352 struct lruvec
*lruvec
;
1353 struct pglist_data
*pgdat
= page_pgdat(page
);
1356 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1357 spin_lock(&lruvec
->lru_lock
);
1360 lruvec_memcg_debug(lruvec
, page
);
1365 struct lruvec
*lock_page_lruvec_irq(struct page
*page
)
1367 struct lruvec
*lruvec
;
1368 struct pglist_data
*pgdat
= page_pgdat(page
);
1371 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1372 spin_lock_irq(&lruvec
->lru_lock
);
1375 lruvec_memcg_debug(lruvec
, page
);
1380 struct lruvec
*lock_page_lruvec_irqsave(struct page
*page
, unsigned long *flags
)
1382 struct lruvec
*lruvec
;
1383 struct pglist_data
*pgdat
= page_pgdat(page
);
1386 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1387 spin_lock_irqsave(&lruvec
->lru_lock
, *flags
);
1390 lruvec_memcg_debug(lruvec
, page
);
1396 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1397 * @lruvec: mem_cgroup per zone lru vector
1398 * @lru: index of lru list the page is sitting on
1399 * @zid: zone id of the accounted pages
1400 * @nr_pages: positive when adding or negative when removing
1402 * This function must be called under lru_lock, just before a page is added
1403 * to or just after a page is removed from an lru list (that ordering being
1404 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1406 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1407 int zid
, int nr_pages
)
1409 struct mem_cgroup_per_node
*mz
;
1410 unsigned long *lru_size
;
1413 if (mem_cgroup_disabled())
1416 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
1417 lru_size
= &mz
->lru_zone_size
[zid
][lru
];
1420 *lru_size
+= nr_pages
;
1423 if (WARN_ONCE(size
< 0,
1424 "%s(%p, %d, %d): lru_size %ld\n",
1425 __func__
, lruvec
, lru
, nr_pages
, size
)) {
1431 *lru_size
+= nr_pages
;
1435 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1436 * @memcg: the memory cgroup
1438 * Returns the maximum amount of memory @mem can be charged with, in
1441 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1443 unsigned long margin
= 0;
1444 unsigned long count
;
1445 unsigned long limit
;
1447 count
= page_counter_read(&memcg
->memory
);
1448 limit
= READ_ONCE(memcg
->memory
.max
);
1450 margin
= limit
- count
;
1452 if (do_memsw_account()) {
1453 count
= page_counter_read(&memcg
->memsw
);
1454 limit
= READ_ONCE(memcg
->memsw
.max
);
1456 margin
= min(margin
, limit
- count
);
1465 * A routine for checking "mem" is under move_account() or not.
1467 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1468 * moving cgroups. This is for waiting at high-memory pressure
1471 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1473 struct mem_cgroup
*from
;
1474 struct mem_cgroup
*to
;
1477 * Unlike task_move routines, we access mc.to, mc.from not under
1478 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1480 spin_lock(&mc
.lock
);
1486 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1487 mem_cgroup_is_descendant(to
, memcg
);
1489 spin_unlock(&mc
.lock
);
1493 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1495 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1496 if (mem_cgroup_under_move(memcg
)) {
1498 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1499 /* moving charge context might have finished. */
1502 finish_wait(&mc
.waitq
, &wait
);
1509 struct memory_stat
{
1515 static struct memory_stat memory_stats
[] = {
1516 { "anon", PAGE_SIZE
, NR_ANON_MAPPED
},
1517 { "file", PAGE_SIZE
, NR_FILE_PAGES
},
1518 { "kernel_stack", 1024, NR_KERNEL_STACK_KB
},
1519 { "pagetables", PAGE_SIZE
, NR_PAGETABLE
},
1520 { "percpu", 1, MEMCG_PERCPU_B
},
1521 { "sock", PAGE_SIZE
, MEMCG_SOCK
},
1522 { "shmem", PAGE_SIZE
, NR_SHMEM
},
1523 { "file_mapped", PAGE_SIZE
, NR_FILE_MAPPED
},
1524 { "file_dirty", PAGE_SIZE
, NR_FILE_DIRTY
},
1525 { "file_writeback", PAGE_SIZE
, NR_WRITEBACK
},
1526 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1528 * The ratio will be initialized in memory_stats_init(). Because
1529 * on some architectures, the macro of HPAGE_PMD_SIZE is not
1530 * constant(e.g. powerpc).
1532 { "anon_thp", 0, NR_ANON_THPS
},
1533 { "file_thp", 0, NR_FILE_THPS
},
1534 { "shmem_thp", 0, NR_SHMEM_THPS
},
1536 { "inactive_anon", PAGE_SIZE
, NR_INACTIVE_ANON
},
1537 { "active_anon", PAGE_SIZE
, NR_ACTIVE_ANON
},
1538 { "inactive_file", PAGE_SIZE
, NR_INACTIVE_FILE
},
1539 { "active_file", PAGE_SIZE
, NR_ACTIVE_FILE
},
1540 { "unevictable", PAGE_SIZE
, NR_UNEVICTABLE
},
1543 * Note: The slab_reclaimable and slab_unreclaimable must be
1544 * together and slab_reclaimable must be in front.
1546 { "slab_reclaimable", 1, NR_SLAB_RECLAIMABLE_B
},
1547 { "slab_unreclaimable", 1, NR_SLAB_UNRECLAIMABLE_B
},
1549 /* The memory events */
1550 { "workingset_refault_anon", 1, WORKINGSET_REFAULT_ANON
},
1551 { "workingset_refault_file", 1, WORKINGSET_REFAULT_FILE
},
1552 { "workingset_activate_anon", 1, WORKINGSET_ACTIVATE_ANON
},
1553 { "workingset_activate_file", 1, WORKINGSET_ACTIVATE_FILE
},
1554 { "workingset_restore_anon", 1, WORKINGSET_RESTORE_ANON
},
1555 { "workingset_restore_file", 1, WORKINGSET_RESTORE_FILE
},
1556 { "workingset_nodereclaim", 1, WORKINGSET_NODERECLAIM
},
1559 static int __init
memory_stats_init(void)
1563 for (i
= 0; i
< ARRAY_SIZE(memory_stats
); i
++) {
1564 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1565 if (memory_stats
[i
].idx
== NR_ANON_THPS
||
1566 memory_stats
[i
].idx
== NR_FILE_THPS
||
1567 memory_stats
[i
].idx
== NR_SHMEM_THPS
)
1568 memory_stats
[i
].ratio
= HPAGE_PMD_SIZE
;
1570 VM_BUG_ON(!memory_stats
[i
].ratio
);
1571 VM_BUG_ON(memory_stats
[i
].idx
>= MEMCG_NR_STAT
);
1576 pure_initcall(memory_stats_init
);
1578 static char *memory_stat_format(struct mem_cgroup
*memcg
)
1583 seq_buf_init(&s
, kmalloc(PAGE_SIZE
, GFP_KERNEL
), PAGE_SIZE
);
1588 * Provide statistics on the state of the memory subsystem as
1589 * well as cumulative event counters that show past behavior.
1591 * This list is ordered following a combination of these gradients:
1592 * 1) generic big picture -> specifics and details
1593 * 2) reflecting userspace activity -> reflecting kernel heuristics
1595 * Current memory state:
1598 for (i
= 0; i
< ARRAY_SIZE(memory_stats
); i
++) {
1601 size
= memcg_page_state(memcg
, memory_stats
[i
].idx
);
1602 size
*= memory_stats
[i
].ratio
;
1603 seq_buf_printf(&s
, "%s %llu\n", memory_stats
[i
].name
, size
);
1605 if (unlikely(memory_stats
[i
].idx
== NR_SLAB_UNRECLAIMABLE_B
)) {
1606 size
= memcg_page_state(memcg
, NR_SLAB_RECLAIMABLE_B
) +
1607 memcg_page_state(memcg
, NR_SLAB_UNRECLAIMABLE_B
);
1608 seq_buf_printf(&s
, "slab %llu\n", size
);
1612 /* Accumulated memory events */
1614 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGFAULT
),
1615 memcg_events(memcg
, PGFAULT
));
1616 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGMAJFAULT
),
1617 memcg_events(memcg
, PGMAJFAULT
));
1618 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGREFILL
),
1619 memcg_events(memcg
, PGREFILL
));
1620 seq_buf_printf(&s
, "pgscan %lu\n",
1621 memcg_events(memcg
, PGSCAN_KSWAPD
) +
1622 memcg_events(memcg
, PGSCAN_DIRECT
));
1623 seq_buf_printf(&s
, "pgsteal %lu\n",
1624 memcg_events(memcg
, PGSTEAL_KSWAPD
) +
1625 memcg_events(memcg
, PGSTEAL_DIRECT
));
1626 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGACTIVATE
),
1627 memcg_events(memcg
, PGACTIVATE
));
1628 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGDEACTIVATE
),
1629 memcg_events(memcg
, PGDEACTIVATE
));
1630 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREE
),
1631 memcg_events(memcg
, PGLAZYFREE
));
1632 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREED
),
1633 memcg_events(memcg
, PGLAZYFREED
));
1635 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1636 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC
),
1637 memcg_events(memcg
, THP_FAULT_ALLOC
));
1638 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC
),
1639 memcg_events(memcg
, THP_COLLAPSE_ALLOC
));
1640 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1642 /* The above should easily fit into one page */
1643 WARN_ON_ONCE(seq_buf_has_overflowed(&s
));
1648 #define K(x) ((x) << (PAGE_SHIFT-10))
1650 * mem_cgroup_print_oom_context: Print OOM information relevant to
1651 * memory controller.
1652 * @memcg: The memory cgroup that went over limit
1653 * @p: Task that is going to be killed
1655 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1658 void mem_cgroup_print_oom_context(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1663 pr_cont(",oom_memcg=");
1664 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1666 pr_cont(",global_oom");
1668 pr_cont(",task_memcg=");
1669 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1675 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1676 * memory controller.
1677 * @memcg: The memory cgroup that went over limit
1679 void mem_cgroup_print_oom_meminfo(struct mem_cgroup
*memcg
)
1683 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1684 K((u64
)page_counter_read(&memcg
->memory
)),
1685 K((u64
)READ_ONCE(memcg
->memory
.max
)), memcg
->memory
.failcnt
);
1686 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1687 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1688 K((u64
)page_counter_read(&memcg
->swap
)),
1689 K((u64
)READ_ONCE(memcg
->swap
.max
)), memcg
->swap
.failcnt
);
1691 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1692 K((u64
)page_counter_read(&memcg
->memsw
)),
1693 K((u64
)memcg
->memsw
.max
), memcg
->memsw
.failcnt
);
1694 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1695 K((u64
)page_counter_read(&memcg
->kmem
)),
1696 K((u64
)memcg
->kmem
.max
), memcg
->kmem
.failcnt
);
1699 pr_info("Memory cgroup stats for ");
1700 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1702 buf
= memory_stat_format(memcg
);
1710 * Return the memory (and swap, if configured) limit for a memcg.
1712 unsigned long mem_cgroup_get_max(struct mem_cgroup
*memcg
)
1714 unsigned long max
= READ_ONCE(memcg
->memory
.max
);
1716 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
1717 if (mem_cgroup_swappiness(memcg
))
1718 max
+= min(READ_ONCE(memcg
->swap
.max
),
1719 (unsigned long)total_swap_pages
);
1721 if (mem_cgroup_swappiness(memcg
)) {
1722 /* Calculate swap excess capacity from memsw limit */
1723 unsigned long swap
= READ_ONCE(memcg
->memsw
.max
) - max
;
1725 max
+= min(swap
, (unsigned long)total_swap_pages
);
1731 unsigned long mem_cgroup_size(struct mem_cgroup
*memcg
)
1733 return page_counter_read(&memcg
->memory
);
1736 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1739 struct oom_control oc
= {
1743 .gfp_mask
= gfp_mask
,
1748 if (mutex_lock_killable(&oom_lock
))
1751 if (mem_cgroup_margin(memcg
) >= (1 << order
))
1755 * A few threads which were not waiting at mutex_lock_killable() can
1756 * fail to bail out. Therefore, check again after holding oom_lock.
1758 ret
= should_force_charge() || out_of_memory(&oc
);
1761 mutex_unlock(&oom_lock
);
1765 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1768 unsigned long *total_scanned
)
1770 struct mem_cgroup
*victim
= NULL
;
1773 unsigned long excess
;
1774 unsigned long nr_scanned
;
1775 struct mem_cgroup_reclaim_cookie reclaim
= {
1779 excess
= soft_limit_excess(root_memcg
);
1782 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1787 * If we have not been able to reclaim
1788 * anything, it might because there are
1789 * no reclaimable pages under this hierarchy
1794 * We want to do more targeted reclaim.
1795 * excess >> 2 is not to excessive so as to
1796 * reclaim too much, nor too less that we keep
1797 * coming back to reclaim from this cgroup
1799 if (total
>= (excess
>> 2) ||
1800 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1805 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1806 pgdat
, &nr_scanned
);
1807 *total_scanned
+= nr_scanned
;
1808 if (!soft_limit_excess(root_memcg
))
1811 mem_cgroup_iter_break(root_memcg
, victim
);
1815 #ifdef CONFIG_LOCKDEP
1816 static struct lockdep_map memcg_oom_lock_dep_map
= {
1817 .name
= "memcg_oom_lock",
1821 static DEFINE_SPINLOCK(memcg_oom_lock
);
1824 * Check OOM-Killer is already running under our hierarchy.
1825 * If someone is running, return false.
1827 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1829 struct mem_cgroup
*iter
, *failed
= NULL
;
1831 spin_lock(&memcg_oom_lock
);
1833 for_each_mem_cgroup_tree(iter
, memcg
) {
1834 if (iter
->oom_lock
) {
1836 * this subtree of our hierarchy is already locked
1837 * so we cannot give a lock.
1840 mem_cgroup_iter_break(memcg
, iter
);
1843 iter
->oom_lock
= true;
1848 * OK, we failed to lock the whole subtree so we have
1849 * to clean up what we set up to the failing subtree
1851 for_each_mem_cgroup_tree(iter
, memcg
) {
1852 if (iter
== failed
) {
1853 mem_cgroup_iter_break(memcg
, iter
);
1856 iter
->oom_lock
= false;
1859 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1861 spin_unlock(&memcg_oom_lock
);
1866 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1868 struct mem_cgroup
*iter
;
1870 spin_lock(&memcg_oom_lock
);
1871 mutex_release(&memcg_oom_lock_dep_map
, _RET_IP_
);
1872 for_each_mem_cgroup_tree(iter
, memcg
)
1873 iter
->oom_lock
= false;
1874 spin_unlock(&memcg_oom_lock
);
1877 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1879 struct mem_cgroup
*iter
;
1881 spin_lock(&memcg_oom_lock
);
1882 for_each_mem_cgroup_tree(iter
, memcg
)
1884 spin_unlock(&memcg_oom_lock
);
1887 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1889 struct mem_cgroup
*iter
;
1892 * Be careful about under_oom underflows becase a child memcg
1893 * could have been added after mem_cgroup_mark_under_oom.
1895 spin_lock(&memcg_oom_lock
);
1896 for_each_mem_cgroup_tree(iter
, memcg
)
1897 if (iter
->under_oom
> 0)
1899 spin_unlock(&memcg_oom_lock
);
1902 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1904 struct oom_wait_info
{
1905 struct mem_cgroup
*memcg
;
1906 wait_queue_entry_t wait
;
1909 static int memcg_oom_wake_function(wait_queue_entry_t
*wait
,
1910 unsigned mode
, int sync
, void *arg
)
1912 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1913 struct mem_cgroup
*oom_wait_memcg
;
1914 struct oom_wait_info
*oom_wait_info
;
1916 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1917 oom_wait_memcg
= oom_wait_info
->memcg
;
1919 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1920 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1922 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1925 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1928 * For the following lockless ->under_oom test, the only required
1929 * guarantee is that it must see the state asserted by an OOM when
1930 * this function is called as a result of userland actions
1931 * triggered by the notification of the OOM. This is trivially
1932 * achieved by invoking mem_cgroup_mark_under_oom() before
1933 * triggering notification.
1935 if (memcg
&& memcg
->under_oom
)
1936 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1946 static enum oom_status
mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1948 enum oom_status ret
;
1951 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1954 memcg_memory_event(memcg
, MEMCG_OOM
);
1957 * We are in the middle of the charge context here, so we
1958 * don't want to block when potentially sitting on a callstack
1959 * that holds all kinds of filesystem and mm locks.
1961 * cgroup1 allows disabling the OOM killer and waiting for outside
1962 * handling until the charge can succeed; remember the context and put
1963 * the task to sleep at the end of the page fault when all locks are
1966 * On the other hand, in-kernel OOM killer allows for an async victim
1967 * memory reclaim (oom_reaper) and that means that we are not solely
1968 * relying on the oom victim to make a forward progress and we can
1969 * invoke the oom killer here.
1971 * Please note that mem_cgroup_out_of_memory might fail to find a
1972 * victim and then we have to bail out from the charge path.
1974 if (memcg
->oom_kill_disable
) {
1975 if (!current
->in_user_fault
)
1977 css_get(&memcg
->css
);
1978 current
->memcg_in_oom
= memcg
;
1979 current
->memcg_oom_gfp_mask
= mask
;
1980 current
->memcg_oom_order
= order
;
1985 mem_cgroup_mark_under_oom(memcg
);
1987 locked
= mem_cgroup_oom_trylock(memcg
);
1990 mem_cgroup_oom_notify(memcg
);
1992 mem_cgroup_unmark_under_oom(memcg
);
1993 if (mem_cgroup_out_of_memory(memcg
, mask
, order
))
1999 mem_cgroup_oom_unlock(memcg
);
2005 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2006 * @handle: actually kill/wait or just clean up the OOM state
2008 * This has to be called at the end of a page fault if the memcg OOM
2009 * handler was enabled.
2011 * Memcg supports userspace OOM handling where failed allocations must
2012 * sleep on a waitqueue until the userspace task resolves the
2013 * situation. Sleeping directly in the charge context with all kinds
2014 * of locks held is not a good idea, instead we remember an OOM state
2015 * in the task and mem_cgroup_oom_synchronize() has to be called at
2016 * the end of the page fault to complete the OOM handling.
2018 * Returns %true if an ongoing memcg OOM situation was detected and
2019 * completed, %false otherwise.
2021 bool mem_cgroup_oom_synchronize(bool handle
)
2023 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
2024 struct oom_wait_info owait
;
2027 /* OOM is global, do not handle */
2034 owait
.memcg
= memcg
;
2035 owait
.wait
.flags
= 0;
2036 owait
.wait
.func
= memcg_oom_wake_function
;
2037 owait
.wait
.private = current
;
2038 INIT_LIST_HEAD(&owait
.wait
.entry
);
2040 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2041 mem_cgroup_mark_under_oom(memcg
);
2043 locked
= mem_cgroup_oom_trylock(memcg
);
2046 mem_cgroup_oom_notify(memcg
);
2048 if (locked
&& !memcg
->oom_kill_disable
) {
2049 mem_cgroup_unmark_under_oom(memcg
);
2050 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2051 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
2052 current
->memcg_oom_order
);
2055 mem_cgroup_unmark_under_oom(memcg
);
2056 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2060 mem_cgroup_oom_unlock(memcg
);
2062 * There is no guarantee that an OOM-lock contender
2063 * sees the wakeups triggered by the OOM kill
2064 * uncharges. Wake any sleepers explicitely.
2066 memcg_oom_recover(memcg
);
2069 current
->memcg_in_oom
= NULL
;
2070 css_put(&memcg
->css
);
2075 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2076 * @victim: task to be killed by the OOM killer
2077 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2079 * Returns a pointer to a memory cgroup, which has to be cleaned up
2080 * by killing all belonging OOM-killable tasks.
2082 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2084 struct mem_cgroup
*mem_cgroup_get_oom_group(struct task_struct
*victim
,
2085 struct mem_cgroup
*oom_domain
)
2087 struct mem_cgroup
*oom_group
= NULL
;
2088 struct mem_cgroup
*memcg
;
2090 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2094 oom_domain
= root_mem_cgroup
;
2098 memcg
= mem_cgroup_from_task(victim
);
2099 if (memcg
== root_mem_cgroup
)
2103 * If the victim task has been asynchronously moved to a different
2104 * memory cgroup, we might end up killing tasks outside oom_domain.
2105 * In this case it's better to ignore memory.group.oom.
2107 if (unlikely(!mem_cgroup_is_descendant(memcg
, oom_domain
)))
2111 * Traverse the memory cgroup hierarchy from the victim task's
2112 * cgroup up to the OOMing cgroup (or root) to find the
2113 * highest-level memory cgroup with oom.group set.
2115 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
2116 if (memcg
->oom_group
)
2119 if (memcg
== oom_domain
)
2124 css_get(&oom_group
->css
);
2131 void mem_cgroup_print_oom_group(struct mem_cgroup
*memcg
)
2133 pr_info("Tasks in ");
2134 pr_cont_cgroup_path(memcg
->css
.cgroup
);
2135 pr_cont(" are going to be killed due to memory.oom.group set\n");
2139 * lock_page_memcg - lock a page and memcg binding
2142 * This function protects unlocked LRU pages from being moved to
2145 * It ensures lifetime of the returned memcg. Caller is responsible
2146 * for the lifetime of the page; __unlock_page_memcg() is available
2147 * when @page might get freed inside the locked section.
2149 struct mem_cgroup
*lock_page_memcg(struct page
*page
)
2151 struct page
*head
= compound_head(page
); /* rmap on tail pages */
2152 struct mem_cgroup
*memcg
;
2153 unsigned long flags
;
2156 * The RCU lock is held throughout the transaction. The fast
2157 * path can get away without acquiring the memcg->move_lock
2158 * because page moving starts with an RCU grace period.
2160 * The RCU lock also protects the memcg from being freed when
2161 * the page state that is going to change is the only thing
2162 * preventing the page itself from being freed. E.g. writeback
2163 * doesn't hold a page reference and relies on PG_writeback to
2164 * keep off truncation, migration and so forth.
2168 if (mem_cgroup_disabled())
2171 memcg
= page_memcg(head
);
2172 if (unlikely(!memcg
))
2175 #ifdef CONFIG_PROVE_LOCKING
2176 local_irq_save(flags
);
2177 might_lock(&memcg
->move_lock
);
2178 local_irq_restore(flags
);
2181 if (atomic_read(&memcg
->moving_account
) <= 0)
2184 spin_lock_irqsave(&memcg
->move_lock
, flags
);
2185 if (memcg
!= page_memcg(head
)) {
2186 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2191 * When charge migration first begins, we can have locked and
2192 * unlocked page stat updates happening concurrently. Track
2193 * the task who has the lock for unlock_page_memcg().
2195 memcg
->move_lock_task
= current
;
2196 memcg
->move_lock_flags
= flags
;
2200 EXPORT_SYMBOL(lock_page_memcg
);
2203 * __unlock_page_memcg - unlock and unpin a memcg
2206 * Unlock and unpin a memcg returned by lock_page_memcg().
2208 void __unlock_page_memcg(struct mem_cgroup
*memcg
)
2210 if (memcg
&& memcg
->move_lock_task
== current
) {
2211 unsigned long flags
= memcg
->move_lock_flags
;
2213 memcg
->move_lock_task
= NULL
;
2214 memcg
->move_lock_flags
= 0;
2216 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2223 * unlock_page_memcg - unlock a page and memcg binding
2226 void unlock_page_memcg(struct page
*page
)
2228 struct page
*head
= compound_head(page
);
2230 __unlock_page_memcg(page_memcg(head
));
2232 EXPORT_SYMBOL(unlock_page_memcg
);
2234 struct memcg_stock_pcp
{
2235 struct mem_cgroup
*cached
; /* this never be root cgroup */
2236 unsigned int nr_pages
;
2238 #ifdef CONFIG_MEMCG_KMEM
2239 struct obj_cgroup
*cached_objcg
;
2240 unsigned int nr_bytes
;
2243 struct work_struct work
;
2244 unsigned long flags
;
2245 #define FLUSHING_CACHED_CHARGE 0
2247 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2248 static DEFINE_MUTEX(percpu_charge_mutex
);
2250 #ifdef CONFIG_MEMCG_KMEM
2251 static void drain_obj_stock(struct memcg_stock_pcp
*stock
);
2252 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
2253 struct mem_cgroup
*root_memcg
);
2256 static inline void drain_obj_stock(struct memcg_stock_pcp
*stock
)
2259 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
2260 struct mem_cgroup
*root_memcg
)
2267 * consume_stock: Try to consume stocked charge on this cpu.
2268 * @memcg: memcg to consume from.
2269 * @nr_pages: how many pages to charge.
2271 * The charges will only happen if @memcg matches the current cpu's memcg
2272 * stock, and at least @nr_pages are available in that stock. Failure to
2273 * service an allocation will refill the stock.
2275 * returns true if successful, false otherwise.
2277 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2279 struct memcg_stock_pcp
*stock
;
2280 unsigned long flags
;
2283 if (nr_pages
> MEMCG_CHARGE_BATCH
)
2286 local_irq_save(flags
);
2288 stock
= this_cpu_ptr(&memcg_stock
);
2289 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
2290 stock
->nr_pages
-= nr_pages
;
2294 local_irq_restore(flags
);
2300 * Returns stocks cached in percpu and reset cached information.
2302 static void drain_stock(struct memcg_stock_pcp
*stock
)
2304 struct mem_cgroup
*old
= stock
->cached
;
2309 if (stock
->nr_pages
) {
2310 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2311 if (do_memsw_account())
2312 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2313 stock
->nr_pages
= 0;
2317 stock
->cached
= NULL
;
2320 static void drain_local_stock(struct work_struct
*dummy
)
2322 struct memcg_stock_pcp
*stock
;
2323 unsigned long flags
;
2326 * The only protection from memory hotplug vs. drain_stock races is
2327 * that we always operate on local CPU stock here with IRQ disabled
2329 local_irq_save(flags
);
2331 stock
= this_cpu_ptr(&memcg_stock
);
2332 drain_obj_stock(stock
);
2334 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2336 local_irq_restore(flags
);
2340 * Cache charges(val) to local per_cpu area.
2341 * This will be consumed by consume_stock() function, later.
2343 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2345 struct memcg_stock_pcp
*stock
;
2346 unsigned long flags
;
2348 local_irq_save(flags
);
2350 stock
= this_cpu_ptr(&memcg_stock
);
2351 if (stock
->cached
!= memcg
) { /* reset if necessary */
2353 css_get(&memcg
->css
);
2354 stock
->cached
= memcg
;
2356 stock
->nr_pages
+= nr_pages
;
2358 if (stock
->nr_pages
> MEMCG_CHARGE_BATCH
)
2361 local_irq_restore(flags
);
2365 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2366 * of the hierarchy under it.
2368 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2372 /* If someone's already draining, avoid adding running more workers. */
2373 if (!mutex_trylock(&percpu_charge_mutex
))
2376 * Notify other cpus that system-wide "drain" is running
2377 * We do not care about races with the cpu hotplug because cpu down
2378 * as well as workers from this path always operate on the local
2379 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2382 for_each_online_cpu(cpu
) {
2383 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2384 struct mem_cgroup
*memcg
;
2388 memcg
= stock
->cached
;
2389 if (memcg
&& stock
->nr_pages
&&
2390 mem_cgroup_is_descendant(memcg
, root_memcg
))
2392 if (obj_stock_flush_required(stock
, root_memcg
))
2397 !test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2399 drain_local_stock(&stock
->work
);
2401 schedule_work_on(cpu
, &stock
->work
);
2405 mutex_unlock(&percpu_charge_mutex
);
2408 static int memcg_hotplug_cpu_dead(unsigned int cpu
)
2410 struct memcg_stock_pcp
*stock
;
2411 struct mem_cgroup
*memcg
, *mi
;
2413 stock
= &per_cpu(memcg_stock
, cpu
);
2416 for_each_mem_cgroup(memcg
) {
2419 for (i
= 0; i
< MEMCG_NR_STAT
; i
++) {
2423 x
= this_cpu_xchg(memcg
->vmstats_percpu
->stat
[i
], 0);
2425 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2426 atomic_long_add(x
, &memcg
->vmstats
[i
]);
2428 if (i
>= NR_VM_NODE_STAT_ITEMS
)
2431 for_each_node(nid
) {
2432 struct mem_cgroup_per_node
*pn
;
2434 pn
= mem_cgroup_nodeinfo(memcg
, nid
);
2435 x
= this_cpu_xchg(pn
->lruvec_stat_cpu
->count
[i
], 0);
2438 atomic_long_add(x
, &pn
->lruvec_stat
[i
]);
2439 } while ((pn
= parent_nodeinfo(pn
, nid
)));
2443 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++) {
2446 x
= this_cpu_xchg(memcg
->vmstats_percpu
->events
[i
], 0);
2448 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2449 atomic_long_add(x
, &memcg
->vmevents
[i
]);
2456 static unsigned long reclaim_high(struct mem_cgroup
*memcg
,
2457 unsigned int nr_pages
,
2460 unsigned long nr_reclaimed
= 0;
2463 unsigned long pflags
;
2465 if (page_counter_read(&memcg
->memory
) <=
2466 READ_ONCE(memcg
->memory
.high
))
2469 memcg_memory_event(memcg
, MEMCG_HIGH
);
2471 psi_memstall_enter(&pflags
);
2472 nr_reclaimed
+= try_to_free_mem_cgroup_pages(memcg
, nr_pages
,
2474 psi_memstall_leave(&pflags
);
2475 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2476 !mem_cgroup_is_root(memcg
));
2478 return nr_reclaimed
;
2481 static void high_work_func(struct work_struct
*work
)
2483 struct mem_cgroup
*memcg
;
2485 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
2486 reclaim_high(memcg
, MEMCG_CHARGE_BATCH
, GFP_KERNEL
);
2490 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2491 * enough to still cause a significant slowdown in most cases, while still
2492 * allowing diagnostics and tracing to proceed without becoming stuck.
2494 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2497 * When calculating the delay, we use these either side of the exponentiation to
2498 * maintain precision and scale to a reasonable number of jiffies (see the table
2501 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2502 * overage ratio to a delay.
2503 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2504 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2505 * to produce a reasonable delay curve.
2507 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2508 * reasonable delay curve compared to precision-adjusted overage, not
2509 * penalising heavily at first, but still making sure that growth beyond the
2510 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2511 * example, with a high of 100 megabytes:
2513 * +-------+------------------------+
2514 * | usage | time to allocate in ms |
2515 * +-------+------------------------+
2537 * +-------+------------------------+
2539 #define MEMCG_DELAY_PRECISION_SHIFT 20
2540 #define MEMCG_DELAY_SCALING_SHIFT 14
2542 static u64
calculate_overage(unsigned long usage
, unsigned long high
)
2550 * Prevent division by 0 in overage calculation by acting as if
2551 * it was a threshold of 1 page
2553 high
= max(high
, 1UL);
2555 overage
= usage
- high
;
2556 overage
<<= MEMCG_DELAY_PRECISION_SHIFT
;
2557 return div64_u64(overage
, high
);
2560 static u64
mem_find_max_overage(struct mem_cgroup
*memcg
)
2562 u64 overage
, max_overage
= 0;
2565 overage
= calculate_overage(page_counter_read(&memcg
->memory
),
2566 READ_ONCE(memcg
->memory
.high
));
2567 max_overage
= max(overage
, max_overage
);
2568 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2569 !mem_cgroup_is_root(memcg
));
2574 static u64
swap_find_max_overage(struct mem_cgroup
*memcg
)
2576 u64 overage
, max_overage
= 0;
2579 overage
= calculate_overage(page_counter_read(&memcg
->swap
),
2580 READ_ONCE(memcg
->swap
.high
));
2582 memcg_memory_event(memcg
, MEMCG_SWAP_HIGH
);
2583 max_overage
= max(overage
, max_overage
);
2584 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2585 !mem_cgroup_is_root(memcg
));
2591 * Get the number of jiffies that we should penalise a mischievous cgroup which
2592 * is exceeding its memory.high by checking both it and its ancestors.
2594 static unsigned long calculate_high_delay(struct mem_cgroup
*memcg
,
2595 unsigned int nr_pages
,
2598 unsigned long penalty_jiffies
;
2604 * We use overage compared to memory.high to calculate the number of
2605 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2606 * fairly lenient on small overages, and increasingly harsh when the
2607 * memcg in question makes it clear that it has no intention of stopping
2608 * its crazy behaviour, so we exponentially increase the delay based on
2611 penalty_jiffies
= max_overage
* max_overage
* HZ
;
2612 penalty_jiffies
>>= MEMCG_DELAY_PRECISION_SHIFT
;
2613 penalty_jiffies
>>= MEMCG_DELAY_SCALING_SHIFT
;
2616 * Factor in the task's own contribution to the overage, such that four
2617 * N-sized allocations are throttled approximately the same as one
2618 * 4N-sized allocation.
2620 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2621 * larger the current charge patch is than that.
2623 return penalty_jiffies
* nr_pages
/ MEMCG_CHARGE_BATCH
;
2627 * Scheduled by try_charge() to be executed from the userland return path
2628 * and reclaims memory over the high limit.
2630 void mem_cgroup_handle_over_high(void)
2632 unsigned long penalty_jiffies
;
2633 unsigned long pflags
;
2634 unsigned long nr_reclaimed
;
2635 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
2636 int nr_retries
= MAX_RECLAIM_RETRIES
;
2637 struct mem_cgroup
*memcg
;
2638 bool in_retry
= false;
2640 if (likely(!nr_pages
))
2643 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2644 current
->memcg_nr_pages_over_high
= 0;
2648 * The allocating task should reclaim at least the batch size, but for
2649 * subsequent retries we only want to do what's necessary to prevent oom
2650 * or breaching resource isolation.
2652 * This is distinct from memory.max or page allocator behaviour because
2653 * memory.high is currently batched, whereas memory.max and the page
2654 * allocator run every time an allocation is made.
2656 nr_reclaimed
= reclaim_high(memcg
,
2657 in_retry
? SWAP_CLUSTER_MAX
: nr_pages
,
2661 * memory.high is breached and reclaim is unable to keep up. Throttle
2662 * allocators proactively to slow down excessive growth.
2664 penalty_jiffies
= calculate_high_delay(memcg
, nr_pages
,
2665 mem_find_max_overage(memcg
));
2667 penalty_jiffies
+= calculate_high_delay(memcg
, nr_pages
,
2668 swap_find_max_overage(memcg
));
2671 * Clamp the max delay per usermode return so as to still keep the
2672 * application moving forwards and also permit diagnostics, albeit
2675 penalty_jiffies
= min(penalty_jiffies
, MEMCG_MAX_HIGH_DELAY_JIFFIES
);
2678 * Don't sleep if the amount of jiffies this memcg owes us is so low
2679 * that it's not even worth doing, in an attempt to be nice to those who
2680 * go only a small amount over their memory.high value and maybe haven't
2681 * been aggressively reclaimed enough yet.
2683 if (penalty_jiffies
<= HZ
/ 100)
2687 * If reclaim is making forward progress but we're still over
2688 * memory.high, we want to encourage that rather than doing allocator
2691 if (nr_reclaimed
|| nr_retries
--) {
2697 * If we exit early, we're guaranteed to die (since
2698 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2699 * need to account for any ill-begotten jiffies to pay them off later.
2701 psi_memstall_enter(&pflags
);
2702 schedule_timeout_killable(penalty_jiffies
);
2703 psi_memstall_leave(&pflags
);
2706 css_put(&memcg
->css
);
2709 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2710 unsigned int nr_pages
)
2712 unsigned int batch
= max(MEMCG_CHARGE_BATCH
, nr_pages
);
2713 int nr_retries
= MAX_RECLAIM_RETRIES
;
2714 struct mem_cgroup
*mem_over_limit
;
2715 struct page_counter
*counter
;
2716 enum oom_status oom_status
;
2717 unsigned long nr_reclaimed
;
2718 bool may_swap
= true;
2719 bool drained
= false;
2720 unsigned long pflags
;
2722 if (mem_cgroup_is_root(memcg
))
2725 if (consume_stock(memcg
, nr_pages
))
2728 if (!do_memsw_account() ||
2729 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2730 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2732 if (do_memsw_account())
2733 page_counter_uncharge(&memcg
->memsw
, batch
);
2734 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2736 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2740 if (batch
> nr_pages
) {
2746 * Memcg doesn't have a dedicated reserve for atomic
2747 * allocations. But like the global atomic pool, we need to
2748 * put the burden of reclaim on regular allocation requests
2749 * and let these go through as privileged allocations.
2751 if (gfp_mask
& __GFP_ATOMIC
)
2755 * Unlike in global OOM situations, memcg is not in a physical
2756 * memory shortage. Allow dying and OOM-killed tasks to
2757 * bypass the last charges so that they can exit quickly and
2758 * free their memory.
2760 if (unlikely(should_force_charge()))
2764 * Prevent unbounded recursion when reclaim operations need to
2765 * allocate memory. This might exceed the limits temporarily,
2766 * but we prefer facilitating memory reclaim and getting back
2767 * under the limit over triggering OOM kills in these cases.
2769 if (unlikely(current
->flags
& PF_MEMALLOC
))
2772 if (unlikely(task_in_memcg_oom(current
)))
2775 if (!gfpflags_allow_blocking(gfp_mask
))
2778 memcg_memory_event(mem_over_limit
, MEMCG_MAX
);
2780 psi_memstall_enter(&pflags
);
2781 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2782 gfp_mask
, may_swap
);
2783 psi_memstall_leave(&pflags
);
2785 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2789 drain_all_stock(mem_over_limit
);
2794 if (gfp_mask
& __GFP_NORETRY
)
2797 * Even though the limit is exceeded at this point, reclaim
2798 * may have been able to free some pages. Retry the charge
2799 * before killing the task.
2801 * Only for regular pages, though: huge pages are rather
2802 * unlikely to succeed so close to the limit, and we fall back
2803 * to regular pages anyway in case of failure.
2805 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2808 * At task move, charge accounts can be doubly counted. So, it's
2809 * better to wait until the end of task_move if something is going on.
2811 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2817 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
2820 if (gfp_mask
& __GFP_NOFAIL
)
2823 if (fatal_signal_pending(current
))
2827 * keep retrying as long as the memcg oom killer is able to make
2828 * a forward progress or bypass the charge if the oom killer
2829 * couldn't make any progress.
2831 oom_status
= mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2832 get_order(nr_pages
* PAGE_SIZE
));
2833 switch (oom_status
) {
2835 nr_retries
= MAX_RECLAIM_RETRIES
;
2843 if (!(gfp_mask
& __GFP_NOFAIL
))
2847 * The allocation either can't fail or will lead to more memory
2848 * being freed very soon. Allow memory usage go over the limit
2849 * temporarily by force charging it.
2851 page_counter_charge(&memcg
->memory
, nr_pages
);
2852 if (do_memsw_account())
2853 page_counter_charge(&memcg
->memsw
, nr_pages
);
2858 if (batch
> nr_pages
)
2859 refill_stock(memcg
, batch
- nr_pages
);
2862 * If the hierarchy is above the normal consumption range, schedule
2863 * reclaim on returning to userland. We can perform reclaim here
2864 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2865 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2866 * not recorded as it most likely matches current's and won't
2867 * change in the meantime. As high limit is checked again before
2868 * reclaim, the cost of mismatch is negligible.
2871 bool mem_high
, swap_high
;
2873 mem_high
= page_counter_read(&memcg
->memory
) >
2874 READ_ONCE(memcg
->memory
.high
);
2875 swap_high
= page_counter_read(&memcg
->swap
) >
2876 READ_ONCE(memcg
->swap
.high
);
2878 /* Don't bother a random interrupted task */
2879 if (in_interrupt()) {
2881 schedule_work(&memcg
->high_work
);
2887 if (mem_high
|| swap_high
) {
2889 * The allocating tasks in this cgroup will need to do
2890 * reclaim or be throttled to prevent further growth
2891 * of the memory or swap footprints.
2893 * Target some best-effort fairness between the tasks,
2894 * and distribute reclaim work and delay penalties
2895 * based on how much each task is actually allocating.
2897 current
->memcg_nr_pages_over_high
+= batch
;
2898 set_notify_resume(current
);
2901 } while ((memcg
= parent_mem_cgroup(memcg
)));
2906 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2907 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2909 if (mem_cgroup_is_root(memcg
))
2912 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2913 if (do_memsw_account())
2914 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2918 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
)
2920 VM_BUG_ON_PAGE(page_memcg(page
), page
);
2922 * Any of the following ensures page's memcg stability:
2926 * - lock_page_memcg()
2927 * - exclusive reference
2929 page
->memcg_data
= (unsigned long)memcg
;
2932 #ifdef CONFIG_MEMCG_KMEM
2934 * The allocated objcg pointers array is not accounted directly.
2935 * Moreover, it should not come from DMA buffer and is not readily
2936 * reclaimable. So those GFP bits should be masked off.
2938 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | __GFP_ACCOUNT)
2940 int memcg_alloc_page_obj_cgroups(struct page
*page
, struct kmem_cache
*s
,
2943 unsigned int objects
= objs_per_slab_page(s
, page
);
2946 gfp
&= ~OBJCGS_CLEAR_MASK
;
2947 vec
= kcalloc_node(objects
, sizeof(struct obj_cgroup
*), gfp
,
2952 if (!set_page_objcgs(page
, vec
))
2955 kmemleak_not_leak(vec
);
2961 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2963 * A passed kernel object can be a slab object or a generic kernel page, so
2964 * different mechanisms for getting the memory cgroup pointer should be used.
2965 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2966 * can not know for sure how the kernel object is implemented.
2967 * mem_cgroup_from_obj() can be safely used in such cases.
2969 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2970 * cgroup_mutex, etc.
2972 struct mem_cgroup
*mem_cgroup_from_obj(void *p
)
2976 if (mem_cgroup_disabled())
2979 page
= virt_to_head_page(p
);
2982 * Slab objects are accounted individually, not per-page.
2983 * Memcg membership data for each individual object is saved in
2984 * the page->obj_cgroups.
2986 if (page_objcgs_check(page
)) {
2987 struct obj_cgroup
*objcg
;
2990 off
= obj_to_index(page
->slab_cache
, page
, p
);
2991 objcg
= page_objcgs(page
)[off
];
2993 return obj_cgroup_memcg(objcg
);
2999 * page_memcg_check() is used here, because page_has_obj_cgroups()
3000 * check above could fail because the object cgroups vector wasn't set
3001 * at that moment, but it can be set concurrently.
3002 * page_memcg_check(page) will guarantee that a proper memory
3003 * cgroup pointer or NULL will be returned.
3005 return page_memcg_check(page
);
3008 __always_inline
struct obj_cgroup
*get_obj_cgroup_from_current(void)
3010 struct obj_cgroup
*objcg
= NULL
;
3011 struct mem_cgroup
*memcg
;
3013 if (memcg_kmem_bypass())
3017 if (unlikely(active_memcg()))
3018 memcg
= active_memcg();
3020 memcg
= mem_cgroup_from_task(current
);
3022 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
)) {
3023 objcg
= rcu_dereference(memcg
->objcg
);
3024 if (objcg
&& obj_cgroup_tryget(objcg
))
3033 static int memcg_alloc_cache_id(void)
3038 id
= ida_simple_get(&memcg_cache_ida
,
3039 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
3043 if (id
< memcg_nr_cache_ids
)
3047 * There's no space for the new id in memcg_caches arrays,
3048 * so we have to grow them.
3050 down_write(&memcg_cache_ids_sem
);
3052 size
= 2 * (id
+ 1);
3053 if (size
< MEMCG_CACHES_MIN_SIZE
)
3054 size
= MEMCG_CACHES_MIN_SIZE
;
3055 else if (size
> MEMCG_CACHES_MAX_SIZE
)
3056 size
= MEMCG_CACHES_MAX_SIZE
;
3058 err
= memcg_update_all_list_lrus(size
);
3060 memcg_nr_cache_ids
= size
;
3062 up_write(&memcg_cache_ids_sem
);
3065 ida_simple_remove(&memcg_cache_ida
, id
);
3071 static void memcg_free_cache_id(int id
)
3073 ida_simple_remove(&memcg_cache_ida
, id
);
3077 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3078 * @memcg: memory cgroup to charge
3079 * @gfp: reclaim mode
3080 * @nr_pages: number of pages to charge
3082 * Returns 0 on success, an error code on failure.
3084 int __memcg_kmem_charge(struct mem_cgroup
*memcg
, gfp_t gfp
,
3085 unsigned int nr_pages
)
3087 struct page_counter
*counter
;
3090 ret
= try_charge(memcg
, gfp
, nr_pages
);
3094 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
3095 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
3098 * Enforce __GFP_NOFAIL allocation because callers are not
3099 * prepared to see failures and likely do not have any failure
3102 if (gfp
& __GFP_NOFAIL
) {
3103 page_counter_charge(&memcg
->kmem
, nr_pages
);
3106 cancel_charge(memcg
, nr_pages
);
3113 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3114 * @memcg: memcg to uncharge
3115 * @nr_pages: number of pages to uncharge
3117 void __memcg_kmem_uncharge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
3119 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
3120 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
3122 refill_stock(memcg
, nr_pages
);
3126 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3127 * @page: page to charge
3128 * @gfp: reclaim mode
3129 * @order: allocation order
3131 * Returns 0 on success, an error code on failure.
3133 int __memcg_kmem_charge_page(struct page
*page
, gfp_t gfp
, int order
)
3135 struct mem_cgroup
*memcg
;
3138 memcg
= get_mem_cgroup_from_current();
3139 if (memcg
&& !mem_cgroup_is_root(memcg
)) {
3140 ret
= __memcg_kmem_charge(memcg
, gfp
, 1 << order
);
3142 page
->memcg_data
= (unsigned long)memcg
|
3146 css_put(&memcg
->css
);
3152 * __memcg_kmem_uncharge_page: uncharge a kmem page
3153 * @page: page to uncharge
3154 * @order: allocation order
3156 void __memcg_kmem_uncharge_page(struct page
*page
, int order
)
3158 struct mem_cgroup
*memcg
= page_memcg(page
);
3159 unsigned int nr_pages
= 1 << order
;
3164 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
3165 __memcg_kmem_uncharge(memcg
, nr_pages
);
3166 page
->memcg_data
= 0;
3167 css_put(&memcg
->css
);
3170 static bool consume_obj_stock(struct obj_cgroup
*objcg
, unsigned int nr_bytes
)
3172 struct memcg_stock_pcp
*stock
;
3173 unsigned long flags
;
3176 local_irq_save(flags
);
3178 stock
= this_cpu_ptr(&memcg_stock
);
3179 if (objcg
== stock
->cached_objcg
&& stock
->nr_bytes
>= nr_bytes
) {
3180 stock
->nr_bytes
-= nr_bytes
;
3184 local_irq_restore(flags
);
3189 static void drain_obj_stock(struct memcg_stock_pcp
*stock
)
3191 struct obj_cgroup
*old
= stock
->cached_objcg
;
3196 if (stock
->nr_bytes
) {
3197 unsigned int nr_pages
= stock
->nr_bytes
>> PAGE_SHIFT
;
3198 unsigned int nr_bytes
= stock
->nr_bytes
& (PAGE_SIZE
- 1);
3201 struct mem_cgroup
*memcg
;
3205 memcg
= obj_cgroup_memcg(old
);
3206 if (unlikely(!css_tryget(&memcg
->css
)))
3210 __memcg_kmem_uncharge(memcg
, nr_pages
);
3211 css_put(&memcg
->css
);
3215 * The leftover is flushed to the centralized per-memcg value.
3216 * On the next attempt to refill obj stock it will be moved
3217 * to a per-cpu stock (probably, on an other CPU), see
3218 * refill_obj_stock().
3220 * How often it's flushed is a trade-off between the memory
3221 * limit enforcement accuracy and potential CPU contention,
3222 * so it might be changed in the future.
3224 atomic_add(nr_bytes
, &old
->nr_charged_bytes
);
3225 stock
->nr_bytes
= 0;
3228 obj_cgroup_put(old
);
3229 stock
->cached_objcg
= NULL
;
3232 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
3233 struct mem_cgroup
*root_memcg
)
3235 struct mem_cgroup
*memcg
;
3237 if (stock
->cached_objcg
) {
3238 memcg
= obj_cgroup_memcg(stock
->cached_objcg
);
3239 if (memcg
&& mem_cgroup_is_descendant(memcg
, root_memcg
))
3246 static void refill_obj_stock(struct obj_cgroup
*objcg
, unsigned int nr_bytes
)
3248 struct memcg_stock_pcp
*stock
;
3249 unsigned long flags
;
3251 local_irq_save(flags
);
3253 stock
= this_cpu_ptr(&memcg_stock
);
3254 if (stock
->cached_objcg
!= objcg
) { /* reset if necessary */
3255 drain_obj_stock(stock
);
3256 obj_cgroup_get(objcg
);
3257 stock
->cached_objcg
= objcg
;
3258 stock
->nr_bytes
= atomic_xchg(&objcg
->nr_charged_bytes
, 0);
3260 stock
->nr_bytes
+= nr_bytes
;
3262 if (stock
->nr_bytes
> PAGE_SIZE
)
3263 drain_obj_stock(stock
);
3265 local_irq_restore(flags
);
3268 int obj_cgroup_charge(struct obj_cgroup
*objcg
, gfp_t gfp
, size_t size
)
3270 struct mem_cgroup
*memcg
;
3271 unsigned int nr_pages
, nr_bytes
;
3274 if (consume_obj_stock(objcg
, size
))
3278 * In theory, memcg->nr_charged_bytes can have enough
3279 * pre-charged bytes to satisfy the allocation. However,
3280 * flushing memcg->nr_charged_bytes requires two atomic
3281 * operations, and memcg->nr_charged_bytes can't be big,
3282 * so it's better to ignore it and try grab some new pages.
3283 * memcg->nr_charged_bytes will be flushed in
3284 * refill_obj_stock(), called from this function or
3285 * independently later.
3289 memcg
= obj_cgroup_memcg(objcg
);
3290 if (unlikely(!css_tryget(&memcg
->css
)))
3294 nr_pages
= size
>> PAGE_SHIFT
;
3295 nr_bytes
= size
& (PAGE_SIZE
- 1);
3300 ret
= __memcg_kmem_charge(memcg
, gfp
, nr_pages
);
3301 if (!ret
&& nr_bytes
)
3302 refill_obj_stock(objcg
, PAGE_SIZE
- nr_bytes
);
3304 css_put(&memcg
->css
);
3308 void obj_cgroup_uncharge(struct obj_cgroup
*objcg
, size_t size
)
3310 refill_obj_stock(objcg
, size
);
3313 #endif /* CONFIG_MEMCG_KMEM */
3316 * Because page_memcg(head) is not set on tails, set it now.
3318 void split_page_memcg(struct page
*head
, unsigned int nr
)
3320 struct mem_cgroup
*memcg
= page_memcg(head
);
3323 if (mem_cgroup_disabled() || !memcg
)
3326 for (i
= 1; i
< nr
; i
++)
3327 head
[i
].memcg_data
= head
->memcg_data
;
3328 css_get_many(&memcg
->css
, nr
- 1);
3331 #ifdef CONFIG_MEMCG_SWAP
3333 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3334 * @entry: swap entry to be moved
3335 * @from: mem_cgroup which the entry is moved from
3336 * @to: mem_cgroup which the entry is moved to
3338 * It succeeds only when the swap_cgroup's record for this entry is the same
3339 * as the mem_cgroup's id of @from.
3341 * Returns 0 on success, -EINVAL on failure.
3343 * The caller must have charged to @to, IOW, called page_counter_charge() about
3344 * both res and memsw, and called css_get().
3346 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
3347 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3349 unsigned short old_id
, new_id
;
3351 old_id
= mem_cgroup_id(from
);
3352 new_id
= mem_cgroup_id(to
);
3354 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
3355 mod_memcg_state(from
, MEMCG_SWAP
, -1);
3356 mod_memcg_state(to
, MEMCG_SWAP
, 1);
3362 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
3363 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3369 static DEFINE_MUTEX(memcg_max_mutex
);
3371 static int mem_cgroup_resize_max(struct mem_cgroup
*memcg
,
3372 unsigned long max
, bool memsw
)
3374 bool enlarge
= false;
3375 bool drained
= false;
3377 bool limits_invariant
;
3378 struct page_counter
*counter
= memsw
? &memcg
->memsw
: &memcg
->memory
;
3381 if (signal_pending(current
)) {
3386 mutex_lock(&memcg_max_mutex
);
3388 * Make sure that the new limit (memsw or memory limit) doesn't
3389 * break our basic invariant rule memory.max <= memsw.max.
3391 limits_invariant
= memsw
? max
>= READ_ONCE(memcg
->memory
.max
) :
3392 max
<= memcg
->memsw
.max
;
3393 if (!limits_invariant
) {
3394 mutex_unlock(&memcg_max_mutex
);
3398 if (max
> counter
->max
)
3400 ret
= page_counter_set_max(counter
, max
);
3401 mutex_unlock(&memcg_max_mutex
);
3407 drain_all_stock(memcg
);
3412 if (!try_to_free_mem_cgroup_pages(memcg
, 1,
3413 GFP_KERNEL
, !memsw
)) {
3419 if (!ret
&& enlarge
)
3420 memcg_oom_recover(memcg
);
3425 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
3427 unsigned long *total_scanned
)
3429 unsigned long nr_reclaimed
= 0;
3430 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
3431 unsigned long reclaimed
;
3433 struct mem_cgroup_tree_per_node
*mctz
;
3434 unsigned long excess
;
3435 unsigned long nr_scanned
;
3440 mctz
= soft_limit_tree_node(pgdat
->node_id
);
3443 * Do not even bother to check the largest node if the root
3444 * is empty. Do it lockless to prevent lock bouncing. Races
3445 * are acceptable as soft limit is best effort anyway.
3447 if (!mctz
|| RB_EMPTY_ROOT(&mctz
->rb_root
))
3451 * This loop can run a while, specially if mem_cgroup's continuously
3452 * keep exceeding their soft limit and putting the system under
3459 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3464 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
3465 gfp_mask
, &nr_scanned
);
3466 nr_reclaimed
+= reclaimed
;
3467 *total_scanned
+= nr_scanned
;
3468 spin_lock_irq(&mctz
->lock
);
3469 __mem_cgroup_remove_exceeded(mz
, mctz
);
3472 * If we failed to reclaim anything from this memory cgroup
3473 * it is time to move on to the next cgroup
3477 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
3479 excess
= soft_limit_excess(mz
->memcg
);
3481 * One school of thought says that we should not add
3482 * back the node to the tree if reclaim returns 0.
3483 * But our reclaim could return 0, simply because due
3484 * to priority we are exposing a smaller subset of
3485 * memory to reclaim from. Consider this as a longer
3488 /* If excess == 0, no tree ops */
3489 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
3490 spin_unlock_irq(&mctz
->lock
);
3491 css_put(&mz
->memcg
->css
);
3494 * Could not reclaim anything and there are no more
3495 * mem cgroups to try or we seem to be looping without
3496 * reclaiming anything.
3498 if (!nr_reclaimed
&&
3500 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3502 } while (!nr_reclaimed
);
3504 css_put(&next_mz
->memcg
->css
);
3505 return nr_reclaimed
;
3509 * Reclaims as many pages from the given memcg as possible.
3511 * Caller is responsible for holding css reference for memcg.
3513 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
3515 int nr_retries
= MAX_RECLAIM_RETRIES
;
3517 /* we call try-to-free pages for make this cgroup empty */
3518 lru_add_drain_all();
3520 drain_all_stock(memcg
);
3522 /* try to free all pages in this cgroup */
3523 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
3526 if (signal_pending(current
))
3529 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
3533 /* maybe some writeback is necessary */
3534 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3542 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3543 char *buf
, size_t nbytes
,
3546 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3548 if (mem_cgroup_is_root(memcg
))
3550 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3553 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3559 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3560 struct cftype
*cft
, u64 val
)
3565 pr_warn_once("Non-hierarchical mode is deprecated. "
3566 "Please report your usecase to linux-mm@kvack.org if you "
3567 "depend on this functionality.\n");
3572 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3576 if (mem_cgroup_is_root(memcg
)) {
3577 val
= memcg_page_state(memcg
, NR_FILE_PAGES
) +
3578 memcg_page_state(memcg
, NR_ANON_MAPPED
);
3580 val
+= memcg_page_state(memcg
, MEMCG_SWAP
);
3583 val
= page_counter_read(&memcg
->memory
);
3585 val
= page_counter_read(&memcg
->memsw
);
3598 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3601 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3602 struct page_counter
*counter
;
3604 switch (MEMFILE_TYPE(cft
->private)) {
3606 counter
= &memcg
->memory
;
3609 counter
= &memcg
->memsw
;
3612 counter
= &memcg
->kmem
;
3615 counter
= &memcg
->tcpmem
;
3621 switch (MEMFILE_ATTR(cft
->private)) {
3623 if (counter
== &memcg
->memory
)
3624 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
3625 if (counter
== &memcg
->memsw
)
3626 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
3627 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3629 return (u64
)counter
->max
* PAGE_SIZE
;
3631 return (u64
)counter
->watermark
* PAGE_SIZE
;
3633 return counter
->failcnt
;
3634 case RES_SOFT_LIMIT
:
3635 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
3641 static void memcg_flush_percpu_vmstats(struct mem_cgroup
*memcg
)
3643 unsigned long stat
[MEMCG_NR_STAT
] = {0};
3644 struct mem_cgroup
*mi
;
3647 for_each_online_cpu(cpu
)
3648 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
3649 stat
[i
] += per_cpu(memcg
->vmstats_percpu
->stat
[i
], cpu
);
3651 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
3652 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
3653 atomic_long_add(stat
[i
], &mi
->vmstats
[i
]);
3655 for_each_node(node
) {
3656 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
3657 struct mem_cgroup_per_node
*pi
;
3659 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3662 for_each_online_cpu(cpu
)
3663 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3665 pn
->lruvec_stat_cpu
->count
[i
], cpu
);
3667 for (pi
= pn
; pi
; pi
= parent_nodeinfo(pi
, node
))
3668 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3669 atomic_long_add(stat
[i
], &pi
->lruvec_stat
[i
]);
3673 static void memcg_flush_percpu_vmevents(struct mem_cgroup
*memcg
)
3675 unsigned long events
[NR_VM_EVENT_ITEMS
];
3676 struct mem_cgroup
*mi
;
3679 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3682 for_each_online_cpu(cpu
)
3683 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3684 events
[i
] += per_cpu(memcg
->vmstats_percpu
->events
[i
],
3687 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
3688 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3689 atomic_long_add(events
[i
], &mi
->vmevents
[i
]);
3692 #ifdef CONFIG_MEMCG_KMEM
3693 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3695 struct obj_cgroup
*objcg
;
3698 if (cgroup_memory_nokmem
)
3701 BUG_ON(memcg
->kmemcg_id
>= 0);
3702 BUG_ON(memcg
->kmem_state
);
3704 memcg_id
= memcg_alloc_cache_id();
3708 objcg
= obj_cgroup_alloc();
3710 memcg_free_cache_id(memcg_id
);
3713 objcg
->memcg
= memcg
;
3714 rcu_assign_pointer(memcg
->objcg
, objcg
);
3716 static_branch_enable(&memcg_kmem_enabled_key
);
3718 memcg
->kmemcg_id
= memcg_id
;
3719 memcg
->kmem_state
= KMEM_ONLINE
;
3724 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3726 struct cgroup_subsys_state
*css
;
3727 struct mem_cgroup
*parent
, *child
;
3730 if (memcg
->kmem_state
!= KMEM_ONLINE
)
3733 memcg
->kmem_state
= KMEM_ALLOCATED
;
3735 parent
= parent_mem_cgroup(memcg
);
3737 parent
= root_mem_cgroup
;
3739 memcg_reparent_objcgs(memcg
, parent
);
3741 kmemcg_id
= memcg
->kmemcg_id
;
3742 BUG_ON(kmemcg_id
< 0);
3745 * Change kmemcg_id of this cgroup and all its descendants to the
3746 * parent's id, and then move all entries from this cgroup's list_lrus
3747 * to ones of the parent. After we have finished, all list_lrus
3748 * corresponding to this cgroup are guaranteed to remain empty. The
3749 * ordering is imposed by list_lru_node->lock taken by
3750 * memcg_drain_all_list_lrus().
3752 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3753 css_for_each_descendant_pre(css
, &memcg
->css
) {
3754 child
= mem_cgroup_from_css(css
);
3755 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
3756 child
->kmemcg_id
= parent
->kmemcg_id
;
3760 memcg_drain_all_list_lrus(kmemcg_id
, parent
);
3762 memcg_free_cache_id(kmemcg_id
);
3765 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3767 /* css_alloc() failed, offlining didn't happen */
3768 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
3769 memcg_offline_kmem(memcg
);
3772 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3776 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3779 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3782 #endif /* CONFIG_MEMCG_KMEM */
3784 static int memcg_update_kmem_max(struct mem_cgroup
*memcg
,
3789 mutex_lock(&memcg_max_mutex
);
3790 ret
= page_counter_set_max(&memcg
->kmem
, max
);
3791 mutex_unlock(&memcg_max_mutex
);
3795 static int memcg_update_tcp_max(struct mem_cgroup
*memcg
, unsigned long max
)
3799 mutex_lock(&memcg_max_mutex
);
3801 ret
= page_counter_set_max(&memcg
->tcpmem
, max
);
3805 if (!memcg
->tcpmem_active
) {
3807 * The active flag needs to be written after the static_key
3808 * update. This is what guarantees that the socket activation
3809 * function is the last one to run. See mem_cgroup_sk_alloc()
3810 * for details, and note that we don't mark any socket as
3811 * belonging to this memcg until that flag is up.
3813 * We need to do this, because static_keys will span multiple
3814 * sites, but we can't control their order. If we mark a socket
3815 * as accounted, but the accounting functions are not patched in
3816 * yet, we'll lose accounting.
3818 * We never race with the readers in mem_cgroup_sk_alloc(),
3819 * because when this value change, the code to process it is not
3822 static_branch_inc(&memcg_sockets_enabled_key
);
3823 memcg
->tcpmem_active
= true;
3826 mutex_unlock(&memcg_max_mutex
);
3831 * The user of this function is...
3834 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3835 char *buf
, size_t nbytes
, loff_t off
)
3837 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3838 unsigned long nr_pages
;
3841 buf
= strstrip(buf
);
3842 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3846 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3848 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3852 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3854 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, false);
3857 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, true);
3860 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3861 "Please report your usecase to linux-mm@kvack.org if you "
3862 "depend on this functionality.\n");
3863 ret
= memcg_update_kmem_max(memcg
, nr_pages
);
3866 ret
= memcg_update_tcp_max(memcg
, nr_pages
);
3870 case RES_SOFT_LIMIT
:
3871 memcg
->soft_limit
= nr_pages
;
3875 return ret
?: nbytes
;
3878 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3879 size_t nbytes
, loff_t off
)
3881 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3882 struct page_counter
*counter
;
3884 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3886 counter
= &memcg
->memory
;
3889 counter
= &memcg
->memsw
;
3892 counter
= &memcg
->kmem
;
3895 counter
= &memcg
->tcpmem
;
3901 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3903 page_counter_reset_watermark(counter
);
3906 counter
->failcnt
= 0;
3915 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3918 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3922 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3923 struct cftype
*cft
, u64 val
)
3925 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3927 if (val
& ~MOVE_MASK
)
3931 * No kind of locking is needed in here, because ->can_attach() will
3932 * check this value once in the beginning of the process, and then carry
3933 * on with stale data. This means that changes to this value will only
3934 * affect task migrations starting after the change.
3936 memcg
->move_charge_at_immigrate
= val
;
3940 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3941 struct cftype
*cft
, u64 val
)
3949 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3950 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3951 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3953 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
3954 int nid
, unsigned int lru_mask
, bool tree
)
3956 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, NODE_DATA(nid
));
3957 unsigned long nr
= 0;
3960 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
3963 if (!(BIT(lru
) & lru_mask
))
3966 nr
+= lruvec_page_state(lruvec
, NR_LRU_BASE
+ lru
);
3968 nr
+= lruvec_page_state_local(lruvec
, NR_LRU_BASE
+ lru
);
3973 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
3974 unsigned int lru_mask
,
3977 unsigned long nr
= 0;
3981 if (!(BIT(lru
) & lru_mask
))
3984 nr
+= memcg_page_state(memcg
, NR_LRU_BASE
+ lru
);
3986 nr
+= memcg_page_state_local(memcg
, NR_LRU_BASE
+ lru
);
3991 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3995 unsigned int lru_mask
;
3998 static const struct numa_stat stats
[] = {
3999 { "total", LRU_ALL
},
4000 { "file", LRU_ALL_FILE
},
4001 { "anon", LRU_ALL_ANON
},
4002 { "unevictable", BIT(LRU_UNEVICTABLE
) },
4004 const struct numa_stat
*stat
;
4006 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
4008 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
4009 seq_printf(m
, "%s=%lu", stat
->name
,
4010 mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
,
4012 for_each_node_state(nid
, N_MEMORY
)
4013 seq_printf(m
, " N%d=%lu", nid
,
4014 mem_cgroup_node_nr_lru_pages(memcg
, nid
,
4015 stat
->lru_mask
, false));
4019 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
4021 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
,
4022 mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
,
4024 for_each_node_state(nid
, N_MEMORY
)
4025 seq_printf(m
, " N%d=%lu", nid
,
4026 mem_cgroup_node_nr_lru_pages(memcg
, nid
,
4027 stat
->lru_mask
, true));
4033 #endif /* CONFIG_NUMA */
4035 static const unsigned int memcg1_stats
[] = {
4038 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4048 static const char *const memcg1_stat_names
[] = {
4051 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4061 /* Universal VM events cgroup1 shows, original sort order */
4062 static const unsigned int memcg1_events
[] = {
4069 static int memcg_stat_show(struct seq_file
*m
, void *v
)
4071 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
4072 unsigned long memory
, memsw
;
4073 struct mem_cgroup
*mi
;
4076 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names
) != ARRAY_SIZE(memcg1_stats
));
4078 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
4081 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
4083 nr
= memcg_page_state_local(memcg
, memcg1_stats
[i
]);
4084 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4085 if (memcg1_stats
[i
] == NR_ANON_THPS
)
4088 seq_printf(m
, "%s %lu\n", memcg1_stat_names
[i
], nr
* PAGE_SIZE
);
4091 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
4092 seq_printf(m
, "%s %lu\n", vm_event_name(memcg1_events
[i
]),
4093 memcg_events_local(memcg
, memcg1_events
[i
]));
4095 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
4096 seq_printf(m
, "%s %lu\n", lru_list_name(i
),
4097 memcg_page_state_local(memcg
, NR_LRU_BASE
+ i
) *
4100 /* Hierarchical information */
4101 memory
= memsw
= PAGE_COUNTER_MAX
;
4102 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
4103 memory
= min(memory
, READ_ONCE(mi
->memory
.max
));
4104 memsw
= min(memsw
, READ_ONCE(mi
->memsw
.max
));
4106 seq_printf(m
, "hierarchical_memory_limit %llu\n",
4107 (u64
)memory
* PAGE_SIZE
);
4108 if (do_memsw_account())
4109 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
4110 (u64
)memsw
* PAGE_SIZE
);
4112 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
4115 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
4117 nr
= memcg_page_state(memcg
, memcg1_stats
[i
]);
4118 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4119 if (memcg1_stats
[i
] == NR_ANON_THPS
)
4122 seq_printf(m
, "total_%s %llu\n", memcg1_stat_names
[i
],
4123 (u64
)nr
* PAGE_SIZE
);
4126 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
4127 seq_printf(m
, "total_%s %llu\n",
4128 vm_event_name(memcg1_events
[i
]),
4129 (u64
)memcg_events(memcg
, memcg1_events
[i
]));
4131 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
4132 seq_printf(m
, "total_%s %llu\n", lru_list_name(i
),
4133 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
4136 #ifdef CONFIG_DEBUG_VM
4139 struct mem_cgroup_per_node
*mz
;
4140 unsigned long anon_cost
= 0;
4141 unsigned long file_cost
= 0;
4143 for_each_online_pgdat(pgdat
) {
4144 mz
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
4146 anon_cost
+= mz
->lruvec
.anon_cost
;
4147 file_cost
+= mz
->lruvec
.file_cost
;
4149 seq_printf(m
, "anon_cost %lu\n", anon_cost
);
4150 seq_printf(m
, "file_cost %lu\n", file_cost
);
4157 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
4160 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4162 return mem_cgroup_swappiness(memcg
);
4165 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
4166 struct cftype
*cft
, u64 val
)
4168 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4174 memcg
->swappiness
= val
;
4176 vm_swappiness
= val
;
4181 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
4183 struct mem_cgroup_threshold_ary
*t
;
4184 unsigned long usage
;
4189 t
= rcu_dereference(memcg
->thresholds
.primary
);
4191 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
4196 usage
= mem_cgroup_usage(memcg
, swap
);
4199 * current_threshold points to threshold just below or equal to usage.
4200 * If it's not true, a threshold was crossed after last
4201 * call of __mem_cgroup_threshold().
4203 i
= t
->current_threshold
;
4206 * Iterate backward over array of thresholds starting from
4207 * current_threshold and check if a threshold is crossed.
4208 * If none of thresholds below usage is crossed, we read
4209 * only one element of the array here.
4211 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
4212 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4214 /* i = current_threshold + 1 */
4218 * Iterate forward over array of thresholds starting from
4219 * current_threshold+1 and check if a threshold is crossed.
4220 * If none of thresholds above usage is crossed, we read
4221 * only one element of the array here.
4223 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
4224 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4226 /* Update current_threshold */
4227 t
->current_threshold
= i
- 1;
4232 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
4235 __mem_cgroup_threshold(memcg
, false);
4236 if (do_memsw_account())
4237 __mem_cgroup_threshold(memcg
, true);
4239 memcg
= parent_mem_cgroup(memcg
);
4243 static int compare_thresholds(const void *a
, const void *b
)
4245 const struct mem_cgroup_threshold
*_a
= a
;
4246 const struct mem_cgroup_threshold
*_b
= b
;
4248 if (_a
->threshold
> _b
->threshold
)
4251 if (_a
->threshold
< _b
->threshold
)
4257 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
4259 struct mem_cgroup_eventfd_list
*ev
;
4261 spin_lock(&memcg_oom_lock
);
4263 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
4264 eventfd_signal(ev
->eventfd
, 1);
4266 spin_unlock(&memcg_oom_lock
);
4270 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
4272 struct mem_cgroup
*iter
;
4274 for_each_mem_cgroup_tree(iter
, memcg
)
4275 mem_cgroup_oom_notify_cb(iter
);
4278 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4279 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
4281 struct mem_cgroup_thresholds
*thresholds
;
4282 struct mem_cgroup_threshold_ary
*new;
4283 unsigned long threshold
;
4284 unsigned long usage
;
4287 ret
= page_counter_memparse(args
, "-1", &threshold
);
4291 mutex_lock(&memcg
->thresholds_lock
);
4294 thresholds
= &memcg
->thresholds
;
4295 usage
= mem_cgroup_usage(memcg
, false);
4296 } else if (type
== _MEMSWAP
) {
4297 thresholds
= &memcg
->memsw_thresholds
;
4298 usage
= mem_cgroup_usage(memcg
, true);
4302 /* Check if a threshold crossed before adding a new one */
4303 if (thresholds
->primary
)
4304 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4306 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
4308 /* Allocate memory for new array of thresholds */
4309 new = kmalloc(struct_size(new, entries
, size
), GFP_KERNEL
);
4316 /* Copy thresholds (if any) to new array */
4317 if (thresholds
->primary
)
4318 memcpy(new->entries
, thresholds
->primary
->entries
,
4319 flex_array_size(new, entries
, size
- 1));
4321 /* Add new threshold */
4322 new->entries
[size
- 1].eventfd
= eventfd
;
4323 new->entries
[size
- 1].threshold
= threshold
;
4325 /* Sort thresholds. Registering of new threshold isn't time-critical */
4326 sort(new->entries
, size
, sizeof(*new->entries
),
4327 compare_thresholds
, NULL
);
4329 /* Find current threshold */
4330 new->current_threshold
= -1;
4331 for (i
= 0; i
< size
; i
++) {
4332 if (new->entries
[i
].threshold
<= usage
) {
4334 * new->current_threshold will not be used until
4335 * rcu_assign_pointer(), so it's safe to increment
4338 ++new->current_threshold
;
4343 /* Free old spare buffer and save old primary buffer as spare */
4344 kfree(thresholds
->spare
);
4345 thresholds
->spare
= thresholds
->primary
;
4347 rcu_assign_pointer(thresholds
->primary
, new);
4349 /* To be sure that nobody uses thresholds */
4353 mutex_unlock(&memcg
->thresholds_lock
);
4358 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4359 struct eventfd_ctx
*eventfd
, const char *args
)
4361 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
4364 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4365 struct eventfd_ctx
*eventfd
, const char *args
)
4367 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
4370 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4371 struct eventfd_ctx
*eventfd
, enum res_type type
)
4373 struct mem_cgroup_thresholds
*thresholds
;
4374 struct mem_cgroup_threshold_ary
*new;
4375 unsigned long usage
;
4376 int i
, j
, size
, entries
;
4378 mutex_lock(&memcg
->thresholds_lock
);
4381 thresholds
= &memcg
->thresholds
;
4382 usage
= mem_cgroup_usage(memcg
, false);
4383 } else if (type
== _MEMSWAP
) {
4384 thresholds
= &memcg
->memsw_thresholds
;
4385 usage
= mem_cgroup_usage(memcg
, true);
4389 if (!thresholds
->primary
)
4392 /* Check if a threshold crossed before removing */
4393 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4395 /* Calculate new number of threshold */
4397 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4398 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4404 new = thresholds
->spare
;
4406 /* If no items related to eventfd have been cleared, nothing to do */
4410 /* Set thresholds array to NULL if we don't have thresholds */
4419 /* Copy thresholds and find current threshold */
4420 new->current_threshold
= -1;
4421 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4422 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4425 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4426 if (new->entries
[j
].threshold
<= usage
) {
4428 * new->current_threshold will not be used
4429 * until rcu_assign_pointer(), so it's safe to increment
4432 ++new->current_threshold
;
4438 /* Swap primary and spare array */
4439 thresholds
->spare
= thresholds
->primary
;
4441 rcu_assign_pointer(thresholds
->primary
, new);
4443 /* To be sure that nobody uses thresholds */
4446 /* If all events are unregistered, free the spare array */
4448 kfree(thresholds
->spare
);
4449 thresholds
->spare
= NULL
;
4452 mutex_unlock(&memcg
->thresholds_lock
);
4455 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4456 struct eventfd_ctx
*eventfd
)
4458 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
4461 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4462 struct eventfd_ctx
*eventfd
)
4464 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
4467 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
4468 struct eventfd_ctx
*eventfd
, const char *args
)
4470 struct mem_cgroup_eventfd_list
*event
;
4472 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4476 spin_lock(&memcg_oom_lock
);
4478 event
->eventfd
= eventfd
;
4479 list_add(&event
->list
, &memcg
->oom_notify
);
4481 /* already in OOM ? */
4482 if (memcg
->under_oom
)
4483 eventfd_signal(eventfd
, 1);
4484 spin_unlock(&memcg_oom_lock
);
4489 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
4490 struct eventfd_ctx
*eventfd
)
4492 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4494 spin_lock(&memcg_oom_lock
);
4496 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
4497 if (ev
->eventfd
== eventfd
) {
4498 list_del(&ev
->list
);
4503 spin_unlock(&memcg_oom_lock
);
4506 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
4508 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(sf
);
4510 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
4511 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
4512 seq_printf(sf
, "oom_kill %lu\n",
4513 atomic_long_read(&memcg
->memory_events
[MEMCG_OOM_KILL
]));
4517 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
4518 struct cftype
*cft
, u64 val
)
4520 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4522 /* cannot set to root cgroup and only 0 and 1 are allowed */
4523 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
4526 memcg
->oom_kill_disable
= val
;
4528 memcg_oom_recover(memcg
);
4533 #ifdef CONFIG_CGROUP_WRITEBACK
4535 #include <trace/events/writeback.h>
4537 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4539 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
4542 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4544 wb_domain_exit(&memcg
->cgwb_domain
);
4547 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4549 wb_domain_size_changed(&memcg
->cgwb_domain
);
4552 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
4554 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4556 if (!memcg
->css
.parent
)
4559 return &memcg
->cgwb_domain
;
4563 * idx can be of type enum memcg_stat_item or node_stat_item.
4564 * Keep in sync with memcg_exact_page().
4566 static unsigned long memcg_exact_page_state(struct mem_cgroup
*memcg
, int idx
)
4568 long x
= atomic_long_read(&memcg
->vmstats
[idx
]);
4571 for_each_online_cpu(cpu
)
4572 x
+= per_cpu_ptr(memcg
->vmstats_percpu
, cpu
)->stat
[idx
];
4579 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4580 * @wb: bdi_writeback in question
4581 * @pfilepages: out parameter for number of file pages
4582 * @pheadroom: out parameter for number of allocatable pages according to memcg
4583 * @pdirty: out parameter for number of dirty pages
4584 * @pwriteback: out parameter for number of pages under writeback
4586 * Determine the numbers of file, headroom, dirty, and writeback pages in
4587 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4588 * is a bit more involved.
4590 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4591 * headroom is calculated as the lowest headroom of itself and the
4592 * ancestors. Note that this doesn't consider the actual amount of
4593 * available memory in the system. The caller should further cap
4594 * *@pheadroom accordingly.
4596 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
4597 unsigned long *pheadroom
, unsigned long *pdirty
,
4598 unsigned long *pwriteback
)
4600 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4601 struct mem_cgroup
*parent
;
4603 *pdirty
= memcg_exact_page_state(memcg
, NR_FILE_DIRTY
);
4605 *pwriteback
= memcg_exact_page_state(memcg
, NR_WRITEBACK
);
4606 *pfilepages
= memcg_exact_page_state(memcg
, NR_INACTIVE_FILE
) +
4607 memcg_exact_page_state(memcg
, NR_ACTIVE_FILE
);
4608 *pheadroom
= PAGE_COUNTER_MAX
;
4610 while ((parent
= parent_mem_cgroup(memcg
))) {
4611 unsigned long ceiling
= min(READ_ONCE(memcg
->memory
.max
),
4612 READ_ONCE(memcg
->memory
.high
));
4613 unsigned long used
= page_counter_read(&memcg
->memory
);
4615 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
4621 * Foreign dirty flushing
4623 * There's an inherent mismatch between memcg and writeback. The former
4624 * trackes ownership per-page while the latter per-inode. This was a
4625 * deliberate design decision because honoring per-page ownership in the
4626 * writeback path is complicated, may lead to higher CPU and IO overheads
4627 * and deemed unnecessary given that write-sharing an inode across
4628 * different cgroups isn't a common use-case.
4630 * Combined with inode majority-writer ownership switching, this works well
4631 * enough in most cases but there are some pathological cases. For
4632 * example, let's say there are two cgroups A and B which keep writing to
4633 * different but confined parts of the same inode. B owns the inode and
4634 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4635 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4636 * triggering background writeback. A will be slowed down without a way to
4637 * make writeback of the dirty pages happen.
4639 * Conditions like the above can lead to a cgroup getting repatedly and
4640 * severely throttled after making some progress after each
4641 * dirty_expire_interval while the underyling IO device is almost
4644 * Solving this problem completely requires matching the ownership tracking
4645 * granularities between memcg and writeback in either direction. However,
4646 * the more egregious behaviors can be avoided by simply remembering the
4647 * most recent foreign dirtying events and initiating remote flushes on
4648 * them when local writeback isn't enough to keep the memory clean enough.
4650 * The following two functions implement such mechanism. When a foreign
4651 * page - a page whose memcg and writeback ownerships don't match - is
4652 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4653 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4654 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4655 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4656 * foreign bdi_writebacks which haven't expired. Both the numbers of
4657 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4658 * limited to MEMCG_CGWB_FRN_CNT.
4660 * The mechanism only remembers IDs and doesn't hold any object references.
4661 * As being wrong occasionally doesn't matter, updates and accesses to the
4662 * records are lockless and racy.
4664 void mem_cgroup_track_foreign_dirty_slowpath(struct page
*page
,
4665 struct bdi_writeback
*wb
)
4667 struct mem_cgroup
*memcg
= page_memcg(page
);
4668 struct memcg_cgwb_frn
*frn
;
4669 u64 now
= get_jiffies_64();
4670 u64 oldest_at
= now
;
4674 trace_track_foreign_dirty(page
, wb
);
4677 * Pick the slot to use. If there is already a slot for @wb, keep
4678 * using it. If not replace the oldest one which isn't being
4681 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4682 frn
= &memcg
->cgwb_frn
[i
];
4683 if (frn
->bdi_id
== wb
->bdi
->id
&&
4684 frn
->memcg_id
== wb
->memcg_css
->id
)
4686 if (time_before64(frn
->at
, oldest_at
) &&
4687 atomic_read(&frn
->done
.cnt
) == 1) {
4689 oldest_at
= frn
->at
;
4693 if (i
< MEMCG_CGWB_FRN_CNT
) {
4695 * Re-using an existing one. Update timestamp lazily to
4696 * avoid making the cacheline hot. We want them to be
4697 * reasonably up-to-date and significantly shorter than
4698 * dirty_expire_interval as that's what expires the record.
4699 * Use the shorter of 1s and dirty_expire_interval / 8.
4701 unsigned long update_intv
=
4702 min_t(unsigned long, HZ
,
4703 msecs_to_jiffies(dirty_expire_interval
* 10) / 8);
4705 if (time_before64(frn
->at
, now
- update_intv
))
4707 } else if (oldest
>= 0) {
4708 /* replace the oldest free one */
4709 frn
= &memcg
->cgwb_frn
[oldest
];
4710 frn
->bdi_id
= wb
->bdi
->id
;
4711 frn
->memcg_id
= wb
->memcg_css
->id
;
4716 /* issue foreign writeback flushes for recorded foreign dirtying events */
4717 void mem_cgroup_flush_foreign(struct bdi_writeback
*wb
)
4719 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4720 unsigned long intv
= msecs_to_jiffies(dirty_expire_interval
* 10);
4721 u64 now
= jiffies_64
;
4724 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4725 struct memcg_cgwb_frn
*frn
= &memcg
->cgwb_frn
[i
];
4728 * If the record is older than dirty_expire_interval,
4729 * writeback on it has already started. No need to kick it
4730 * off again. Also, don't start a new one if there's
4731 * already one in flight.
4733 if (time_after64(frn
->at
, now
- intv
) &&
4734 atomic_read(&frn
->done
.cnt
) == 1) {
4736 trace_flush_foreign(wb
, frn
->bdi_id
, frn
->memcg_id
);
4737 cgroup_writeback_by_id(frn
->bdi_id
, frn
->memcg_id
, 0,
4738 WB_REASON_FOREIGN_FLUSH
,
4744 #else /* CONFIG_CGROUP_WRITEBACK */
4746 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4751 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4755 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4759 #endif /* CONFIG_CGROUP_WRITEBACK */
4762 * DO NOT USE IN NEW FILES.
4764 * "cgroup.event_control" implementation.
4766 * This is way over-engineered. It tries to support fully configurable
4767 * events for each user. Such level of flexibility is completely
4768 * unnecessary especially in the light of the planned unified hierarchy.
4770 * Please deprecate this and replace with something simpler if at all
4775 * Unregister event and free resources.
4777 * Gets called from workqueue.
4779 static void memcg_event_remove(struct work_struct
*work
)
4781 struct mem_cgroup_event
*event
=
4782 container_of(work
, struct mem_cgroup_event
, remove
);
4783 struct mem_cgroup
*memcg
= event
->memcg
;
4785 remove_wait_queue(event
->wqh
, &event
->wait
);
4787 event
->unregister_event(memcg
, event
->eventfd
);
4789 /* Notify userspace the event is going away. */
4790 eventfd_signal(event
->eventfd
, 1);
4792 eventfd_ctx_put(event
->eventfd
);
4794 css_put(&memcg
->css
);
4798 * Gets called on EPOLLHUP on eventfd when user closes it.
4800 * Called with wqh->lock held and interrupts disabled.
4802 static int memcg_event_wake(wait_queue_entry_t
*wait
, unsigned mode
,
4803 int sync
, void *key
)
4805 struct mem_cgroup_event
*event
=
4806 container_of(wait
, struct mem_cgroup_event
, wait
);
4807 struct mem_cgroup
*memcg
= event
->memcg
;
4808 __poll_t flags
= key_to_poll(key
);
4810 if (flags
& EPOLLHUP
) {
4812 * If the event has been detached at cgroup removal, we
4813 * can simply return knowing the other side will cleanup
4816 * We can't race against event freeing since the other
4817 * side will require wqh->lock via remove_wait_queue(),
4820 spin_lock(&memcg
->event_list_lock
);
4821 if (!list_empty(&event
->list
)) {
4822 list_del_init(&event
->list
);
4824 * We are in atomic context, but cgroup_event_remove()
4825 * may sleep, so we have to call it in workqueue.
4827 schedule_work(&event
->remove
);
4829 spin_unlock(&memcg
->event_list_lock
);
4835 static void memcg_event_ptable_queue_proc(struct file
*file
,
4836 wait_queue_head_t
*wqh
, poll_table
*pt
)
4838 struct mem_cgroup_event
*event
=
4839 container_of(pt
, struct mem_cgroup_event
, pt
);
4842 add_wait_queue(wqh
, &event
->wait
);
4846 * DO NOT USE IN NEW FILES.
4848 * Parse input and register new cgroup event handler.
4850 * Input must be in format '<event_fd> <control_fd> <args>'.
4851 * Interpretation of args is defined by control file implementation.
4853 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4854 char *buf
, size_t nbytes
, loff_t off
)
4856 struct cgroup_subsys_state
*css
= of_css(of
);
4857 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4858 struct mem_cgroup_event
*event
;
4859 struct cgroup_subsys_state
*cfile_css
;
4860 unsigned int efd
, cfd
;
4867 buf
= strstrip(buf
);
4869 efd
= simple_strtoul(buf
, &endp
, 10);
4874 cfd
= simple_strtoul(buf
, &endp
, 10);
4875 if ((*endp
!= ' ') && (*endp
!= '\0'))
4879 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
4883 event
->memcg
= memcg
;
4884 INIT_LIST_HEAD(&event
->list
);
4885 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
4886 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
4887 INIT_WORK(&event
->remove
, memcg_event_remove
);
4895 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
4896 if (IS_ERR(event
->eventfd
)) {
4897 ret
= PTR_ERR(event
->eventfd
);
4904 goto out_put_eventfd
;
4907 /* the process need read permission on control file */
4908 /* AV: shouldn't we check that it's been opened for read instead? */
4909 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
4914 * Determine the event callbacks and set them in @event. This used
4915 * to be done via struct cftype but cgroup core no longer knows
4916 * about these events. The following is crude but the whole thing
4917 * is for compatibility anyway.
4919 * DO NOT ADD NEW FILES.
4921 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
4923 if (!strcmp(name
, "memory.usage_in_bytes")) {
4924 event
->register_event
= mem_cgroup_usage_register_event
;
4925 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
4926 } else if (!strcmp(name
, "memory.oom_control")) {
4927 event
->register_event
= mem_cgroup_oom_register_event
;
4928 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
4929 } else if (!strcmp(name
, "memory.pressure_level")) {
4930 event
->register_event
= vmpressure_register_event
;
4931 event
->unregister_event
= vmpressure_unregister_event
;
4932 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
4933 event
->register_event
= memsw_cgroup_usage_register_event
;
4934 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
4941 * Verify @cfile should belong to @css. Also, remaining events are
4942 * automatically removed on cgroup destruction but the removal is
4943 * asynchronous, so take an extra ref on @css.
4945 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
4946 &memory_cgrp_subsys
);
4948 if (IS_ERR(cfile_css
))
4950 if (cfile_css
!= css
) {
4955 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
4959 vfs_poll(efile
.file
, &event
->pt
);
4961 spin_lock(&memcg
->event_list_lock
);
4962 list_add(&event
->list
, &memcg
->event_list
);
4963 spin_unlock(&memcg
->event_list_lock
);
4975 eventfd_ctx_put(event
->eventfd
);
4984 static struct cftype mem_cgroup_legacy_files
[] = {
4986 .name
= "usage_in_bytes",
4987 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4988 .read_u64
= mem_cgroup_read_u64
,
4991 .name
= "max_usage_in_bytes",
4992 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4993 .write
= mem_cgroup_reset
,
4994 .read_u64
= mem_cgroup_read_u64
,
4997 .name
= "limit_in_bytes",
4998 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4999 .write
= mem_cgroup_write
,
5000 .read_u64
= mem_cgroup_read_u64
,
5003 .name
= "soft_limit_in_bytes",
5004 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5005 .write
= mem_cgroup_write
,
5006 .read_u64
= mem_cgroup_read_u64
,
5010 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5011 .write
= mem_cgroup_reset
,
5012 .read_u64
= mem_cgroup_read_u64
,
5016 .seq_show
= memcg_stat_show
,
5019 .name
= "force_empty",
5020 .write
= mem_cgroup_force_empty_write
,
5023 .name
= "use_hierarchy",
5024 .write_u64
= mem_cgroup_hierarchy_write
,
5025 .read_u64
= mem_cgroup_hierarchy_read
,
5028 .name
= "cgroup.event_control", /* XXX: for compat */
5029 .write
= memcg_write_event_control
,
5030 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
5033 .name
= "swappiness",
5034 .read_u64
= mem_cgroup_swappiness_read
,
5035 .write_u64
= mem_cgroup_swappiness_write
,
5038 .name
= "move_charge_at_immigrate",
5039 .read_u64
= mem_cgroup_move_charge_read
,
5040 .write_u64
= mem_cgroup_move_charge_write
,
5043 .name
= "oom_control",
5044 .seq_show
= mem_cgroup_oom_control_read
,
5045 .write_u64
= mem_cgroup_oom_control_write
,
5046 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
5049 .name
= "pressure_level",
5053 .name
= "numa_stat",
5054 .seq_show
= memcg_numa_stat_show
,
5058 .name
= "kmem.limit_in_bytes",
5059 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
5060 .write
= mem_cgroup_write
,
5061 .read_u64
= mem_cgroup_read_u64
,
5064 .name
= "kmem.usage_in_bytes",
5065 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
5066 .read_u64
= mem_cgroup_read_u64
,
5069 .name
= "kmem.failcnt",
5070 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
5071 .write
= mem_cgroup_reset
,
5072 .read_u64
= mem_cgroup_read_u64
,
5075 .name
= "kmem.max_usage_in_bytes",
5076 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
5077 .write
= mem_cgroup_reset
,
5078 .read_u64
= mem_cgroup_read_u64
,
5080 #if defined(CONFIG_MEMCG_KMEM) && \
5081 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5083 .name
= "kmem.slabinfo",
5084 .seq_show
= memcg_slab_show
,
5088 .name
= "kmem.tcp.limit_in_bytes",
5089 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
5090 .write
= mem_cgroup_write
,
5091 .read_u64
= mem_cgroup_read_u64
,
5094 .name
= "kmem.tcp.usage_in_bytes",
5095 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
5096 .read_u64
= mem_cgroup_read_u64
,
5099 .name
= "kmem.tcp.failcnt",
5100 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
5101 .write
= mem_cgroup_reset
,
5102 .read_u64
= mem_cgroup_read_u64
,
5105 .name
= "kmem.tcp.max_usage_in_bytes",
5106 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
5107 .write
= mem_cgroup_reset
,
5108 .read_u64
= mem_cgroup_read_u64
,
5110 { }, /* terminate */
5114 * Private memory cgroup IDR
5116 * Swap-out records and page cache shadow entries need to store memcg
5117 * references in constrained space, so we maintain an ID space that is
5118 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5119 * memory-controlled cgroups to 64k.
5121 * However, there usually are many references to the offline CSS after
5122 * the cgroup has been destroyed, such as page cache or reclaimable
5123 * slab objects, that don't need to hang on to the ID. We want to keep
5124 * those dead CSS from occupying IDs, or we might quickly exhaust the
5125 * relatively small ID space and prevent the creation of new cgroups
5126 * even when there are much fewer than 64k cgroups - possibly none.
5128 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5129 * be freed and recycled when it's no longer needed, which is usually
5130 * when the CSS is offlined.
5132 * The only exception to that are records of swapped out tmpfs/shmem
5133 * pages that need to be attributed to live ancestors on swapin. But
5134 * those references are manageable from userspace.
5137 static DEFINE_IDR(mem_cgroup_idr
);
5139 static void mem_cgroup_id_remove(struct mem_cgroup
*memcg
)
5141 if (memcg
->id
.id
> 0) {
5142 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
5147 static void __maybe_unused
mem_cgroup_id_get_many(struct mem_cgroup
*memcg
,
5150 refcount_add(n
, &memcg
->id
.ref
);
5153 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
5155 if (refcount_sub_and_test(n
, &memcg
->id
.ref
)) {
5156 mem_cgroup_id_remove(memcg
);
5158 /* Memcg ID pins CSS */
5159 css_put(&memcg
->css
);
5163 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
5165 mem_cgroup_id_put_many(memcg
, 1);
5169 * mem_cgroup_from_id - look up a memcg from a memcg id
5170 * @id: the memcg id to look up
5172 * Caller must hold rcu_read_lock().
5174 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
5176 WARN_ON_ONCE(!rcu_read_lock_held());
5177 return idr_find(&mem_cgroup_idr
, id
);
5180 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
5182 struct mem_cgroup_per_node
*pn
;
5185 * This routine is called against possible nodes.
5186 * But it's BUG to call kmalloc() against offline node.
5188 * TODO: this routine can waste much memory for nodes which will
5189 * never be onlined. It's better to use memory hotplug callback
5192 if (!node_state(node
, N_NORMAL_MEMORY
))
5194 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
5198 pn
->lruvec_stat_local
= alloc_percpu_gfp(struct lruvec_stat
,
5199 GFP_KERNEL_ACCOUNT
);
5200 if (!pn
->lruvec_stat_local
) {
5205 pn
->lruvec_stat_cpu
= alloc_percpu_gfp(struct lruvec_stat
,
5206 GFP_KERNEL_ACCOUNT
);
5207 if (!pn
->lruvec_stat_cpu
) {
5208 free_percpu(pn
->lruvec_stat_local
);
5213 lruvec_init(&pn
->lruvec
);
5214 pn
->usage_in_excess
= 0;
5215 pn
->on_tree
= false;
5218 memcg
->nodeinfo
[node
] = pn
;
5222 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
5224 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
5229 free_percpu(pn
->lruvec_stat_cpu
);
5230 free_percpu(pn
->lruvec_stat_local
);
5234 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5239 free_mem_cgroup_per_node_info(memcg
, node
);
5240 free_percpu(memcg
->vmstats_percpu
);
5241 free_percpu(memcg
->vmstats_local
);
5245 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
5247 memcg_wb_domain_exit(memcg
);
5249 * Flush percpu vmstats and vmevents to guarantee the value correctness
5250 * on parent's and all ancestor levels.
5252 memcg_flush_percpu_vmstats(memcg
);
5253 memcg_flush_percpu_vmevents(memcg
);
5254 __mem_cgroup_free(memcg
);
5257 static struct mem_cgroup
*mem_cgroup_alloc(void)
5259 struct mem_cgroup
*memcg
;
5262 int __maybe_unused i
;
5263 long error
= -ENOMEM
;
5265 size
= sizeof(struct mem_cgroup
);
5266 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
5268 memcg
= kzalloc(size
, GFP_KERNEL
);
5270 return ERR_PTR(error
);
5272 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
5273 1, MEM_CGROUP_ID_MAX
,
5275 if (memcg
->id
.id
< 0) {
5276 error
= memcg
->id
.id
;
5280 memcg
->vmstats_local
= alloc_percpu_gfp(struct memcg_vmstats_percpu
,
5281 GFP_KERNEL_ACCOUNT
);
5282 if (!memcg
->vmstats_local
)
5285 memcg
->vmstats_percpu
= alloc_percpu_gfp(struct memcg_vmstats_percpu
,
5286 GFP_KERNEL_ACCOUNT
);
5287 if (!memcg
->vmstats_percpu
)
5291 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
5294 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
5297 INIT_WORK(&memcg
->high_work
, high_work_func
);
5298 INIT_LIST_HEAD(&memcg
->oom_notify
);
5299 mutex_init(&memcg
->thresholds_lock
);
5300 spin_lock_init(&memcg
->move_lock
);
5301 vmpressure_init(&memcg
->vmpressure
);
5302 INIT_LIST_HEAD(&memcg
->event_list
);
5303 spin_lock_init(&memcg
->event_list_lock
);
5304 memcg
->socket_pressure
= jiffies
;
5305 #ifdef CONFIG_MEMCG_KMEM
5306 memcg
->kmemcg_id
= -1;
5307 INIT_LIST_HEAD(&memcg
->objcg_list
);
5309 #ifdef CONFIG_CGROUP_WRITEBACK
5310 INIT_LIST_HEAD(&memcg
->cgwb_list
);
5311 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5312 memcg
->cgwb_frn
[i
].done
=
5313 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq
);
5315 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5316 spin_lock_init(&memcg
->deferred_split_queue
.split_queue_lock
);
5317 INIT_LIST_HEAD(&memcg
->deferred_split_queue
.split_queue
);
5318 memcg
->deferred_split_queue
.split_queue_len
= 0;
5320 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
5323 mem_cgroup_id_remove(memcg
);
5324 __mem_cgroup_free(memcg
);
5325 return ERR_PTR(error
);
5328 static struct cgroup_subsys_state
* __ref
5329 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
5331 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
5332 struct mem_cgroup
*memcg
, *old_memcg
;
5333 long error
= -ENOMEM
;
5335 old_memcg
= set_active_memcg(parent
);
5336 memcg
= mem_cgroup_alloc();
5337 set_active_memcg(old_memcg
);
5339 return ERR_CAST(memcg
);
5341 page_counter_set_high(&memcg
->memory
, PAGE_COUNTER_MAX
);
5342 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5343 page_counter_set_high(&memcg
->swap
, PAGE_COUNTER_MAX
);
5345 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
5346 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
5348 page_counter_init(&memcg
->memory
, &parent
->memory
);
5349 page_counter_init(&memcg
->swap
, &parent
->swap
);
5350 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
5351 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
5353 page_counter_init(&memcg
->memory
, NULL
);
5354 page_counter_init(&memcg
->swap
, NULL
);
5355 page_counter_init(&memcg
->kmem
, NULL
);
5356 page_counter_init(&memcg
->tcpmem
, NULL
);
5358 root_mem_cgroup
= memcg
;
5362 /* The following stuff does not apply to the root */
5363 error
= memcg_online_kmem(memcg
);
5367 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5368 static_branch_inc(&memcg_sockets_enabled_key
);
5372 mem_cgroup_id_remove(memcg
);
5373 mem_cgroup_free(memcg
);
5374 return ERR_PTR(error
);
5377 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
5379 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5382 * A memcg must be visible for memcg_expand_shrinker_maps()
5383 * by the time the maps are allocated. So, we allocate maps
5384 * here, when for_each_mem_cgroup() can't skip it.
5386 if (memcg_alloc_shrinker_maps(memcg
)) {
5387 mem_cgroup_id_remove(memcg
);
5391 /* Online state pins memcg ID, memcg ID pins CSS */
5392 refcount_set(&memcg
->id
.ref
, 1);
5397 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
5399 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5400 struct mem_cgroup_event
*event
, *tmp
;
5403 * Unregister events and notify userspace.
5404 * Notify userspace about cgroup removing only after rmdir of cgroup
5405 * directory to avoid race between userspace and kernelspace.
5407 spin_lock(&memcg
->event_list_lock
);
5408 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
5409 list_del_init(&event
->list
);
5410 schedule_work(&event
->remove
);
5412 spin_unlock(&memcg
->event_list_lock
);
5414 page_counter_set_min(&memcg
->memory
, 0);
5415 page_counter_set_low(&memcg
->memory
, 0);
5417 memcg_offline_kmem(memcg
);
5418 wb_memcg_offline(memcg
);
5420 drain_all_stock(memcg
);
5422 mem_cgroup_id_put(memcg
);
5425 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
5427 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5429 invalidate_reclaim_iterators(memcg
);
5432 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
5434 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5435 int __maybe_unused i
;
5437 #ifdef CONFIG_CGROUP_WRITEBACK
5438 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5439 wb_wait_for_completion(&memcg
->cgwb_frn
[i
].done
);
5441 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5442 static_branch_dec(&memcg_sockets_enabled_key
);
5444 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
5445 static_branch_dec(&memcg_sockets_enabled_key
);
5447 vmpressure_cleanup(&memcg
->vmpressure
);
5448 cancel_work_sync(&memcg
->high_work
);
5449 mem_cgroup_remove_from_trees(memcg
);
5450 memcg_free_shrinker_maps(memcg
);
5451 memcg_free_kmem(memcg
);
5452 mem_cgroup_free(memcg
);
5456 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5457 * @css: the target css
5459 * Reset the states of the mem_cgroup associated with @css. This is
5460 * invoked when the userland requests disabling on the default hierarchy
5461 * but the memcg is pinned through dependency. The memcg should stop
5462 * applying policies and should revert to the vanilla state as it may be
5463 * made visible again.
5465 * The current implementation only resets the essential configurations.
5466 * This needs to be expanded to cover all the visible parts.
5468 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
5470 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5472 page_counter_set_max(&memcg
->memory
, PAGE_COUNTER_MAX
);
5473 page_counter_set_max(&memcg
->swap
, PAGE_COUNTER_MAX
);
5474 page_counter_set_max(&memcg
->kmem
, PAGE_COUNTER_MAX
);
5475 page_counter_set_max(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
5476 page_counter_set_min(&memcg
->memory
, 0);
5477 page_counter_set_low(&memcg
->memory
, 0);
5478 page_counter_set_high(&memcg
->memory
, PAGE_COUNTER_MAX
);
5479 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5480 page_counter_set_high(&memcg
->swap
, PAGE_COUNTER_MAX
);
5481 memcg_wb_domain_size_changed(memcg
);
5485 /* Handlers for move charge at task migration. */
5486 static int mem_cgroup_do_precharge(unsigned long count
)
5490 /* Try a single bulk charge without reclaim first, kswapd may wake */
5491 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
5493 mc
.precharge
+= count
;
5497 /* Try charges one by one with reclaim, but do not retry */
5499 ret
= try_charge(mc
.to
, GFP_KERNEL
| __GFP_NORETRY
, 1);
5513 enum mc_target_type
{
5520 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
5521 unsigned long addr
, pte_t ptent
)
5523 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
5525 if (!page
|| !page_mapped(page
))
5527 if (PageAnon(page
)) {
5528 if (!(mc
.flags
& MOVE_ANON
))
5531 if (!(mc
.flags
& MOVE_FILE
))
5534 if (!get_page_unless_zero(page
))
5540 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5541 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5542 pte_t ptent
, swp_entry_t
*entry
)
5544 struct page
*page
= NULL
;
5545 swp_entry_t ent
= pte_to_swp_entry(ptent
);
5547 if (!(mc
.flags
& MOVE_ANON
))
5551 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5552 * a device and because they are not accessible by CPU they are store
5553 * as special swap entry in the CPU page table.
5555 if (is_device_private_entry(ent
)) {
5556 page
= device_private_entry_to_page(ent
);
5558 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5559 * a refcount of 1 when free (unlike normal page)
5561 if (!page_ref_add_unless(page
, 1, 1))
5566 if (non_swap_entry(ent
))
5570 * Because lookup_swap_cache() updates some statistics counter,
5571 * we call find_get_page() with swapper_space directly.
5573 page
= find_get_page(swap_address_space(ent
), swp_offset(ent
));
5574 entry
->val
= ent
.val
;
5579 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5580 pte_t ptent
, swp_entry_t
*entry
)
5586 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
5587 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5589 if (!vma
->vm_file
) /* anonymous vma */
5591 if (!(mc
.flags
& MOVE_FILE
))
5594 /* page is moved even if it's not RSS of this task(page-faulted). */
5595 /* shmem/tmpfs may report page out on swap: account for that too. */
5596 return find_get_incore_page(vma
->vm_file
->f_mapping
,
5597 linear_page_index(vma
, addr
));
5601 * mem_cgroup_move_account - move account of the page
5603 * @compound: charge the page as compound or small page
5604 * @from: mem_cgroup which the page is moved from.
5605 * @to: mem_cgroup which the page is moved to. @from != @to.
5607 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5609 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5612 static int mem_cgroup_move_account(struct page
*page
,
5614 struct mem_cgroup
*from
,
5615 struct mem_cgroup
*to
)
5617 struct lruvec
*from_vec
, *to_vec
;
5618 struct pglist_data
*pgdat
;
5619 unsigned int nr_pages
= compound
? thp_nr_pages(page
) : 1;
5622 VM_BUG_ON(from
== to
);
5623 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5624 VM_BUG_ON(compound
&& !PageTransHuge(page
));
5627 * Prevent mem_cgroup_migrate() from looking at
5628 * page's memory cgroup of its source page while we change it.
5631 if (!trylock_page(page
))
5635 if (page_memcg(page
) != from
)
5638 pgdat
= page_pgdat(page
);
5639 from_vec
= mem_cgroup_lruvec(from
, pgdat
);
5640 to_vec
= mem_cgroup_lruvec(to
, pgdat
);
5642 lock_page_memcg(page
);
5644 if (PageAnon(page
)) {
5645 if (page_mapped(page
)) {
5646 __mod_lruvec_state(from_vec
, NR_ANON_MAPPED
, -nr_pages
);
5647 __mod_lruvec_state(to_vec
, NR_ANON_MAPPED
, nr_pages
);
5648 if (PageTransHuge(page
)) {
5649 __dec_lruvec_state(from_vec
, NR_ANON_THPS
);
5650 __inc_lruvec_state(to_vec
, NR_ANON_THPS
);
5655 __mod_lruvec_state(from_vec
, NR_FILE_PAGES
, -nr_pages
);
5656 __mod_lruvec_state(to_vec
, NR_FILE_PAGES
, nr_pages
);
5658 if (PageSwapBacked(page
)) {
5659 __mod_lruvec_state(from_vec
, NR_SHMEM
, -nr_pages
);
5660 __mod_lruvec_state(to_vec
, NR_SHMEM
, nr_pages
);
5663 if (page_mapped(page
)) {
5664 __mod_lruvec_state(from_vec
, NR_FILE_MAPPED
, -nr_pages
);
5665 __mod_lruvec_state(to_vec
, NR_FILE_MAPPED
, nr_pages
);
5668 if (PageDirty(page
)) {
5669 struct address_space
*mapping
= page_mapping(page
);
5671 if (mapping_can_writeback(mapping
)) {
5672 __mod_lruvec_state(from_vec
, NR_FILE_DIRTY
,
5674 __mod_lruvec_state(to_vec
, NR_FILE_DIRTY
,
5680 if (PageWriteback(page
)) {
5681 __mod_lruvec_state(from_vec
, NR_WRITEBACK
, -nr_pages
);
5682 __mod_lruvec_state(to_vec
, NR_WRITEBACK
, nr_pages
);
5686 * All state has been migrated, let's switch to the new memcg.
5688 * It is safe to change page's memcg here because the page
5689 * is referenced, charged, isolated, and locked: we can't race
5690 * with (un)charging, migration, LRU putback, or anything else
5691 * that would rely on a stable page's memory cgroup.
5693 * Note that lock_page_memcg is a memcg lock, not a page lock,
5694 * to save space. As soon as we switch page's memory cgroup to a
5695 * new memcg that isn't locked, the above state can change
5696 * concurrently again. Make sure we're truly done with it.
5701 css_put(&from
->css
);
5703 page
->memcg_data
= (unsigned long)to
;
5705 __unlock_page_memcg(from
);
5709 local_irq_disable();
5710 mem_cgroup_charge_statistics(to
, page
, nr_pages
);
5711 memcg_check_events(to
, page
);
5712 mem_cgroup_charge_statistics(from
, page
, -nr_pages
);
5713 memcg_check_events(from
, page
);
5722 * get_mctgt_type - get target type of moving charge
5723 * @vma: the vma the pte to be checked belongs
5724 * @addr: the address corresponding to the pte to be checked
5725 * @ptent: the pte to be checked
5726 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5729 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5730 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5731 * move charge. if @target is not NULL, the page is stored in target->page
5732 * with extra refcnt got(Callers should handle it).
5733 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5734 * target for charge migration. if @target is not NULL, the entry is stored
5736 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5737 * (so ZONE_DEVICE page and thus not on the lru).
5738 * For now we such page is charge like a regular page would be as for all
5739 * intent and purposes it is just special memory taking the place of a
5742 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5744 * Called with pte lock held.
5747 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
5748 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
5750 struct page
*page
= NULL
;
5751 enum mc_target_type ret
= MC_TARGET_NONE
;
5752 swp_entry_t ent
= { .val
= 0 };
5754 if (pte_present(ptent
))
5755 page
= mc_handle_present_pte(vma
, addr
, ptent
);
5756 else if (is_swap_pte(ptent
))
5757 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
5758 else if (pte_none(ptent
))
5759 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
5761 if (!page
&& !ent
.val
)
5765 * Do only loose check w/o serialization.
5766 * mem_cgroup_move_account() checks the page is valid or
5767 * not under LRU exclusion.
5769 if (page_memcg(page
) == mc
.from
) {
5770 ret
= MC_TARGET_PAGE
;
5771 if (is_device_private_page(page
))
5772 ret
= MC_TARGET_DEVICE
;
5774 target
->page
= page
;
5776 if (!ret
|| !target
)
5780 * There is a swap entry and a page doesn't exist or isn't charged.
5781 * But we cannot move a tail-page in a THP.
5783 if (ent
.val
&& !ret
&& (!page
|| !PageTransCompound(page
)) &&
5784 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
5785 ret
= MC_TARGET_SWAP
;
5792 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5794 * We don't consider PMD mapped swapping or file mapped pages because THP does
5795 * not support them for now.
5796 * Caller should make sure that pmd_trans_huge(pmd) is true.
5798 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5799 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5801 struct page
*page
= NULL
;
5802 enum mc_target_type ret
= MC_TARGET_NONE
;
5804 if (unlikely(is_swap_pmd(pmd
))) {
5805 VM_BUG_ON(thp_migration_supported() &&
5806 !is_pmd_migration_entry(pmd
));
5809 page
= pmd_page(pmd
);
5810 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
5811 if (!(mc
.flags
& MOVE_ANON
))
5813 if (page_memcg(page
) == mc
.from
) {
5814 ret
= MC_TARGET_PAGE
;
5817 target
->page
= page
;
5823 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5824 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5826 return MC_TARGET_NONE
;
5830 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5831 unsigned long addr
, unsigned long end
,
5832 struct mm_walk
*walk
)
5834 struct vm_area_struct
*vma
= walk
->vma
;
5838 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5841 * Note their can not be MC_TARGET_DEVICE for now as we do not
5842 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5843 * this might change.
5845 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5846 mc
.precharge
+= HPAGE_PMD_NR
;
5851 if (pmd_trans_unstable(pmd
))
5853 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5854 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5855 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5856 mc
.precharge
++; /* increment precharge temporarily */
5857 pte_unmap_unlock(pte
- 1, ptl
);
5863 static const struct mm_walk_ops precharge_walk_ops
= {
5864 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
5867 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5869 unsigned long precharge
;
5872 walk_page_range(mm
, 0, mm
->highest_vm_end
, &precharge_walk_ops
, NULL
);
5873 mmap_read_unlock(mm
);
5875 precharge
= mc
.precharge
;
5881 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
5883 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
5885 VM_BUG_ON(mc
.moving_task
);
5886 mc
.moving_task
= current
;
5887 return mem_cgroup_do_precharge(precharge
);
5890 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5891 static void __mem_cgroup_clear_mc(void)
5893 struct mem_cgroup
*from
= mc
.from
;
5894 struct mem_cgroup
*to
= mc
.to
;
5896 /* we must uncharge all the leftover precharges from mc.to */
5898 cancel_charge(mc
.to
, mc
.precharge
);
5902 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5903 * we must uncharge here.
5905 if (mc
.moved_charge
) {
5906 cancel_charge(mc
.from
, mc
.moved_charge
);
5907 mc
.moved_charge
= 0;
5909 /* we must fixup refcnts and charges */
5910 if (mc
.moved_swap
) {
5911 /* uncharge swap account from the old cgroup */
5912 if (!mem_cgroup_is_root(mc
.from
))
5913 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
5915 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
5918 * we charged both to->memory and to->memsw, so we
5919 * should uncharge to->memory.
5921 if (!mem_cgroup_is_root(mc
.to
))
5922 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
5926 memcg_oom_recover(from
);
5927 memcg_oom_recover(to
);
5928 wake_up_all(&mc
.waitq
);
5931 static void mem_cgroup_clear_mc(void)
5933 struct mm_struct
*mm
= mc
.mm
;
5936 * we must clear moving_task before waking up waiters at the end of
5939 mc
.moving_task
= NULL
;
5940 __mem_cgroup_clear_mc();
5941 spin_lock(&mc
.lock
);
5945 spin_unlock(&mc
.lock
);
5950 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5952 struct cgroup_subsys_state
*css
;
5953 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
5954 struct mem_cgroup
*from
;
5955 struct task_struct
*leader
, *p
;
5956 struct mm_struct
*mm
;
5957 unsigned long move_flags
;
5960 /* charge immigration isn't supported on the default hierarchy */
5961 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5965 * Multi-process migrations only happen on the default hierarchy
5966 * where charge immigration is not used. Perform charge
5967 * immigration if @tset contains a leader and whine if there are
5971 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
5974 memcg
= mem_cgroup_from_css(css
);
5980 * We are now commited to this value whatever it is. Changes in this
5981 * tunable will only affect upcoming migrations, not the current one.
5982 * So we need to save it, and keep it going.
5984 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
5988 from
= mem_cgroup_from_task(p
);
5990 VM_BUG_ON(from
== memcg
);
5992 mm
= get_task_mm(p
);
5995 /* We move charges only when we move a owner of the mm */
5996 if (mm
->owner
== p
) {
5999 VM_BUG_ON(mc
.precharge
);
6000 VM_BUG_ON(mc
.moved_charge
);
6001 VM_BUG_ON(mc
.moved_swap
);
6003 spin_lock(&mc
.lock
);
6007 mc
.flags
= move_flags
;
6008 spin_unlock(&mc
.lock
);
6009 /* We set mc.moving_task later */
6011 ret
= mem_cgroup_precharge_mc(mm
);
6013 mem_cgroup_clear_mc();
6020 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
6023 mem_cgroup_clear_mc();
6026 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6027 unsigned long addr
, unsigned long end
,
6028 struct mm_walk
*walk
)
6031 struct vm_area_struct
*vma
= walk
->vma
;
6034 enum mc_target_type target_type
;
6035 union mc_target target
;
6038 ptl
= pmd_trans_huge_lock(pmd
, vma
);
6040 if (mc
.precharge
< HPAGE_PMD_NR
) {
6044 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6045 if (target_type
== MC_TARGET_PAGE
) {
6047 if (!isolate_lru_page(page
)) {
6048 if (!mem_cgroup_move_account(page
, true,
6050 mc
.precharge
-= HPAGE_PMD_NR
;
6051 mc
.moved_charge
+= HPAGE_PMD_NR
;
6053 putback_lru_page(page
);
6056 } else if (target_type
== MC_TARGET_DEVICE
) {
6058 if (!mem_cgroup_move_account(page
, true,
6060 mc
.precharge
-= HPAGE_PMD_NR
;
6061 mc
.moved_charge
+= HPAGE_PMD_NR
;
6069 if (pmd_trans_unstable(pmd
))
6072 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6073 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6074 pte_t ptent
= *(pte
++);
6075 bool device
= false;
6081 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6082 case MC_TARGET_DEVICE
:
6085 case MC_TARGET_PAGE
:
6088 * We can have a part of the split pmd here. Moving it
6089 * can be done but it would be too convoluted so simply
6090 * ignore such a partial THP and keep it in original
6091 * memcg. There should be somebody mapping the head.
6093 if (PageTransCompound(page
))
6095 if (!device
&& isolate_lru_page(page
))
6097 if (!mem_cgroup_move_account(page
, false,
6100 /* we uncharge from mc.from later. */
6104 putback_lru_page(page
);
6105 put
: /* get_mctgt_type() gets the page */
6108 case MC_TARGET_SWAP
:
6110 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6112 mem_cgroup_id_get_many(mc
.to
, 1);
6113 /* we fixup other refcnts and charges later. */
6121 pte_unmap_unlock(pte
- 1, ptl
);
6126 * We have consumed all precharges we got in can_attach().
6127 * We try charge one by one, but don't do any additional
6128 * charges to mc.to if we have failed in charge once in attach()
6131 ret
= mem_cgroup_do_precharge(1);
6139 static const struct mm_walk_ops charge_walk_ops
= {
6140 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6143 static void mem_cgroup_move_charge(void)
6145 lru_add_drain_all();
6147 * Signal lock_page_memcg() to take the memcg's move_lock
6148 * while we're moving its pages to another memcg. Then wait
6149 * for already started RCU-only updates to finish.
6151 atomic_inc(&mc
.from
->moving_account
);
6154 if (unlikely(!mmap_read_trylock(mc
.mm
))) {
6156 * Someone who are holding the mmap_lock might be waiting in
6157 * waitq. So we cancel all extra charges, wake up all waiters,
6158 * and retry. Because we cancel precharges, we might not be able
6159 * to move enough charges, but moving charge is a best-effort
6160 * feature anyway, so it wouldn't be a big problem.
6162 __mem_cgroup_clear_mc();
6167 * When we have consumed all precharges and failed in doing
6168 * additional charge, the page walk just aborts.
6170 walk_page_range(mc
.mm
, 0, mc
.mm
->highest_vm_end
, &charge_walk_ops
,
6173 mmap_read_unlock(mc
.mm
);
6174 atomic_dec(&mc
.from
->moving_account
);
6177 static void mem_cgroup_move_task(void)
6180 mem_cgroup_move_charge();
6181 mem_cgroup_clear_mc();
6184 #else /* !CONFIG_MMU */
6185 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
6189 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
6192 static void mem_cgroup_move_task(void)
6197 static int seq_puts_memcg_tunable(struct seq_file
*m
, unsigned long value
)
6199 if (value
== PAGE_COUNTER_MAX
)
6200 seq_puts(m
, "max\n");
6202 seq_printf(m
, "%llu\n", (u64
)value
* PAGE_SIZE
);
6207 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
6210 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6212 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
6215 static int memory_min_show(struct seq_file
*m
, void *v
)
6217 return seq_puts_memcg_tunable(m
,
6218 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.min
));
6221 static ssize_t
memory_min_write(struct kernfs_open_file
*of
,
6222 char *buf
, size_t nbytes
, loff_t off
)
6224 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6228 buf
= strstrip(buf
);
6229 err
= page_counter_memparse(buf
, "max", &min
);
6233 page_counter_set_min(&memcg
->memory
, min
);
6238 static int memory_low_show(struct seq_file
*m
, void *v
)
6240 return seq_puts_memcg_tunable(m
,
6241 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.low
));
6244 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
6245 char *buf
, size_t nbytes
, loff_t off
)
6247 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6251 buf
= strstrip(buf
);
6252 err
= page_counter_memparse(buf
, "max", &low
);
6256 page_counter_set_low(&memcg
->memory
, low
);
6261 static int memory_high_show(struct seq_file
*m
, void *v
)
6263 return seq_puts_memcg_tunable(m
,
6264 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.high
));
6267 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
6268 char *buf
, size_t nbytes
, loff_t off
)
6270 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6271 unsigned int nr_retries
= MAX_RECLAIM_RETRIES
;
6272 bool drained
= false;
6276 buf
= strstrip(buf
);
6277 err
= page_counter_memparse(buf
, "max", &high
);
6281 page_counter_set_high(&memcg
->memory
, high
);
6284 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6285 unsigned long reclaimed
;
6287 if (nr_pages
<= high
)
6290 if (signal_pending(current
))
6294 drain_all_stock(memcg
);
6299 reclaimed
= try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
6302 if (!reclaimed
&& !nr_retries
--)
6306 memcg_wb_domain_size_changed(memcg
);
6310 static int memory_max_show(struct seq_file
*m
, void *v
)
6312 return seq_puts_memcg_tunable(m
,
6313 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.max
));
6316 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
6317 char *buf
, size_t nbytes
, loff_t off
)
6319 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6320 unsigned int nr_reclaims
= MAX_RECLAIM_RETRIES
;
6321 bool drained
= false;
6325 buf
= strstrip(buf
);
6326 err
= page_counter_memparse(buf
, "max", &max
);
6330 xchg(&memcg
->memory
.max
, max
);
6333 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6335 if (nr_pages
<= max
)
6338 if (signal_pending(current
))
6342 drain_all_stock(memcg
);
6348 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
6354 memcg_memory_event(memcg
, MEMCG_OOM
);
6355 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
6359 memcg_wb_domain_size_changed(memcg
);
6363 static void __memory_events_show(struct seq_file
*m
, atomic_long_t
*events
)
6365 seq_printf(m
, "low %lu\n", atomic_long_read(&events
[MEMCG_LOW
]));
6366 seq_printf(m
, "high %lu\n", atomic_long_read(&events
[MEMCG_HIGH
]));
6367 seq_printf(m
, "max %lu\n", atomic_long_read(&events
[MEMCG_MAX
]));
6368 seq_printf(m
, "oom %lu\n", atomic_long_read(&events
[MEMCG_OOM
]));
6369 seq_printf(m
, "oom_kill %lu\n",
6370 atomic_long_read(&events
[MEMCG_OOM_KILL
]));
6373 static int memory_events_show(struct seq_file
*m
, void *v
)
6375 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6377 __memory_events_show(m
, memcg
->memory_events
);
6381 static int memory_events_local_show(struct seq_file
*m
, void *v
)
6383 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6385 __memory_events_show(m
, memcg
->memory_events_local
);
6389 static int memory_stat_show(struct seq_file
*m
, void *v
)
6391 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6394 buf
= memory_stat_format(memcg
);
6403 static int memory_numa_stat_show(struct seq_file
*m
, void *v
)
6406 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6408 for (i
= 0; i
< ARRAY_SIZE(memory_stats
); i
++) {
6411 if (memory_stats
[i
].idx
>= NR_VM_NODE_STAT_ITEMS
)
6414 seq_printf(m
, "%s", memory_stats
[i
].name
);
6415 for_each_node_state(nid
, N_MEMORY
) {
6417 struct lruvec
*lruvec
;
6419 lruvec
= mem_cgroup_lruvec(memcg
, NODE_DATA(nid
));
6420 size
= lruvec_page_state(lruvec
, memory_stats
[i
].idx
);
6421 size
*= memory_stats
[i
].ratio
;
6422 seq_printf(m
, " N%d=%llu", nid
, size
);
6431 static int memory_oom_group_show(struct seq_file
*m
, void *v
)
6433 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6435 seq_printf(m
, "%d\n", memcg
->oom_group
);
6440 static ssize_t
memory_oom_group_write(struct kernfs_open_file
*of
,
6441 char *buf
, size_t nbytes
, loff_t off
)
6443 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6446 buf
= strstrip(buf
);
6450 ret
= kstrtoint(buf
, 0, &oom_group
);
6454 if (oom_group
!= 0 && oom_group
!= 1)
6457 memcg
->oom_group
= oom_group
;
6462 static struct cftype memory_files
[] = {
6465 .flags
= CFTYPE_NOT_ON_ROOT
,
6466 .read_u64
= memory_current_read
,
6470 .flags
= CFTYPE_NOT_ON_ROOT
,
6471 .seq_show
= memory_min_show
,
6472 .write
= memory_min_write
,
6476 .flags
= CFTYPE_NOT_ON_ROOT
,
6477 .seq_show
= memory_low_show
,
6478 .write
= memory_low_write
,
6482 .flags
= CFTYPE_NOT_ON_ROOT
,
6483 .seq_show
= memory_high_show
,
6484 .write
= memory_high_write
,
6488 .flags
= CFTYPE_NOT_ON_ROOT
,
6489 .seq_show
= memory_max_show
,
6490 .write
= memory_max_write
,
6494 .flags
= CFTYPE_NOT_ON_ROOT
,
6495 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
6496 .seq_show
= memory_events_show
,
6499 .name
= "events.local",
6500 .flags
= CFTYPE_NOT_ON_ROOT
,
6501 .file_offset
= offsetof(struct mem_cgroup
, events_local_file
),
6502 .seq_show
= memory_events_local_show
,
6506 .seq_show
= memory_stat_show
,
6510 .name
= "numa_stat",
6511 .seq_show
= memory_numa_stat_show
,
6515 .name
= "oom.group",
6516 .flags
= CFTYPE_NOT_ON_ROOT
| CFTYPE_NS_DELEGATABLE
,
6517 .seq_show
= memory_oom_group_show
,
6518 .write
= memory_oom_group_write
,
6523 struct cgroup_subsys memory_cgrp_subsys
= {
6524 .css_alloc
= mem_cgroup_css_alloc
,
6525 .css_online
= mem_cgroup_css_online
,
6526 .css_offline
= mem_cgroup_css_offline
,
6527 .css_released
= mem_cgroup_css_released
,
6528 .css_free
= mem_cgroup_css_free
,
6529 .css_reset
= mem_cgroup_css_reset
,
6530 .can_attach
= mem_cgroup_can_attach
,
6531 .cancel_attach
= mem_cgroup_cancel_attach
,
6532 .post_attach
= mem_cgroup_move_task
,
6533 .dfl_cftypes
= memory_files
,
6534 .legacy_cftypes
= mem_cgroup_legacy_files
,
6539 * This function calculates an individual cgroup's effective
6540 * protection which is derived from its own memory.min/low, its
6541 * parent's and siblings' settings, as well as the actual memory
6542 * distribution in the tree.
6544 * The following rules apply to the effective protection values:
6546 * 1. At the first level of reclaim, effective protection is equal to
6547 * the declared protection in memory.min and memory.low.
6549 * 2. To enable safe delegation of the protection configuration, at
6550 * subsequent levels the effective protection is capped to the
6551 * parent's effective protection.
6553 * 3. To make complex and dynamic subtrees easier to configure, the
6554 * user is allowed to overcommit the declared protection at a given
6555 * level. If that is the case, the parent's effective protection is
6556 * distributed to the children in proportion to how much protection
6557 * they have declared and how much of it they are utilizing.
6559 * This makes distribution proportional, but also work-conserving:
6560 * if one cgroup claims much more protection than it uses memory,
6561 * the unused remainder is available to its siblings.
6563 * 4. Conversely, when the declared protection is undercommitted at a
6564 * given level, the distribution of the larger parental protection
6565 * budget is NOT proportional. A cgroup's protection from a sibling
6566 * is capped to its own memory.min/low setting.
6568 * 5. However, to allow protecting recursive subtrees from each other
6569 * without having to declare each individual cgroup's fixed share
6570 * of the ancestor's claim to protection, any unutilized -
6571 * "floating" - protection from up the tree is distributed in
6572 * proportion to each cgroup's *usage*. This makes the protection
6573 * neutral wrt sibling cgroups and lets them compete freely over
6574 * the shared parental protection budget, but it protects the
6575 * subtree as a whole from neighboring subtrees.
6577 * Note that 4. and 5. are not in conflict: 4. is about protecting
6578 * against immediate siblings whereas 5. is about protecting against
6579 * neighboring subtrees.
6581 static unsigned long effective_protection(unsigned long usage
,
6582 unsigned long parent_usage
,
6583 unsigned long setting
,
6584 unsigned long parent_effective
,
6585 unsigned long siblings_protected
)
6587 unsigned long protected;
6590 protected = min(usage
, setting
);
6592 * If all cgroups at this level combined claim and use more
6593 * protection then what the parent affords them, distribute
6594 * shares in proportion to utilization.
6596 * We are using actual utilization rather than the statically
6597 * claimed protection in order to be work-conserving: claimed
6598 * but unused protection is available to siblings that would
6599 * otherwise get a smaller chunk than what they claimed.
6601 if (siblings_protected
> parent_effective
)
6602 return protected * parent_effective
/ siblings_protected
;
6605 * Ok, utilized protection of all children is within what the
6606 * parent affords them, so we know whatever this child claims
6607 * and utilizes is effectively protected.
6609 * If there is unprotected usage beyond this value, reclaim
6610 * will apply pressure in proportion to that amount.
6612 * If there is unutilized protection, the cgroup will be fully
6613 * shielded from reclaim, but we do return a smaller value for
6614 * protection than what the group could enjoy in theory. This
6615 * is okay. With the overcommit distribution above, effective
6616 * protection is always dependent on how memory is actually
6617 * consumed among the siblings anyway.
6622 * If the children aren't claiming (all of) the protection
6623 * afforded to them by the parent, distribute the remainder in
6624 * proportion to the (unprotected) memory of each cgroup. That
6625 * way, cgroups that aren't explicitly prioritized wrt each
6626 * other compete freely over the allowance, but they are
6627 * collectively protected from neighboring trees.
6629 * We're using unprotected memory for the weight so that if
6630 * some cgroups DO claim explicit protection, we don't protect
6631 * the same bytes twice.
6633 * Check both usage and parent_usage against the respective
6634 * protected values. One should imply the other, but they
6635 * aren't read atomically - make sure the division is sane.
6637 if (!(cgrp_dfl_root
.flags
& CGRP_ROOT_MEMORY_RECURSIVE_PROT
))
6639 if (parent_effective
> siblings_protected
&&
6640 parent_usage
> siblings_protected
&&
6641 usage
> protected) {
6642 unsigned long unclaimed
;
6644 unclaimed
= parent_effective
- siblings_protected
;
6645 unclaimed
*= usage
- protected;
6646 unclaimed
/= parent_usage
- siblings_protected
;
6655 * mem_cgroup_protected - check if memory consumption is in the normal range
6656 * @root: the top ancestor of the sub-tree being checked
6657 * @memcg: the memory cgroup to check
6659 * WARNING: This function is not stateless! It can only be used as part
6660 * of a top-down tree iteration, not for isolated queries.
6662 void mem_cgroup_calculate_protection(struct mem_cgroup
*root
,
6663 struct mem_cgroup
*memcg
)
6665 unsigned long usage
, parent_usage
;
6666 struct mem_cgroup
*parent
;
6668 if (mem_cgroup_disabled())
6672 root
= root_mem_cgroup
;
6675 * Effective values of the reclaim targets are ignored so they
6676 * can be stale. Have a look at mem_cgroup_protection for more
6678 * TODO: calculation should be more robust so that we do not need
6679 * that special casing.
6684 usage
= page_counter_read(&memcg
->memory
);
6688 parent
= parent_mem_cgroup(memcg
);
6689 /* No parent means a non-hierarchical mode on v1 memcg */
6693 if (parent
== root
) {
6694 memcg
->memory
.emin
= READ_ONCE(memcg
->memory
.min
);
6695 memcg
->memory
.elow
= READ_ONCE(memcg
->memory
.low
);
6699 parent_usage
= page_counter_read(&parent
->memory
);
6701 WRITE_ONCE(memcg
->memory
.emin
, effective_protection(usage
, parent_usage
,
6702 READ_ONCE(memcg
->memory
.min
),
6703 READ_ONCE(parent
->memory
.emin
),
6704 atomic_long_read(&parent
->memory
.children_min_usage
)));
6706 WRITE_ONCE(memcg
->memory
.elow
, effective_protection(usage
, parent_usage
,
6707 READ_ONCE(memcg
->memory
.low
),
6708 READ_ONCE(parent
->memory
.elow
),
6709 atomic_long_read(&parent
->memory
.children_low_usage
)));
6713 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6714 * @page: page to charge
6715 * @mm: mm context of the victim
6716 * @gfp_mask: reclaim mode
6718 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6719 * pages according to @gfp_mask if necessary.
6721 * Returns 0 on success. Otherwise, an error code is returned.
6723 int mem_cgroup_charge(struct page
*page
, struct mm_struct
*mm
, gfp_t gfp_mask
)
6725 unsigned int nr_pages
= thp_nr_pages(page
);
6726 struct mem_cgroup
*memcg
= NULL
;
6729 if (mem_cgroup_disabled())
6732 if (PageSwapCache(page
)) {
6733 swp_entry_t ent
= { .val
= page_private(page
), };
6737 * Every swap fault against a single page tries to charge the
6738 * page, bail as early as possible. shmem_unuse() encounters
6739 * already charged pages, too. page and memcg binding is
6740 * protected by the page lock, which serializes swap cache
6741 * removal, which in turn serializes uncharging.
6743 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
6744 if (page_memcg(compound_head(page
)))
6747 id
= lookup_swap_cgroup_id(ent
);
6749 memcg
= mem_cgroup_from_id(id
);
6750 if (memcg
&& !css_tryget_online(&memcg
->css
))
6756 memcg
= get_mem_cgroup_from_mm(mm
);
6758 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
6762 css_get(&memcg
->css
);
6763 commit_charge(page
, memcg
);
6765 local_irq_disable();
6766 mem_cgroup_charge_statistics(memcg
, page
, nr_pages
);
6767 memcg_check_events(memcg
, page
);
6771 * Cgroup1's unified memory+swap counter has been charged with the
6772 * new swapcache page, finish the transfer by uncharging the swap
6773 * slot. The swap slot would also get uncharged when it dies, but
6774 * it can stick around indefinitely and we'd count the page twice
6777 * Cgroup2 has separate resource counters for memory and swap,
6778 * so this is a non-issue here. Memory and swap charge lifetimes
6779 * correspond 1:1 to page and swap slot lifetimes: we charge the
6780 * page to memory here, and uncharge swap when the slot is freed.
6782 if (do_memsw_account() && PageSwapCache(page
)) {
6783 swp_entry_t entry
= { .val
= page_private(page
) };
6785 * The swap entry might not get freed for a long time,
6786 * let's not wait for it. The page already received a
6787 * memory+swap charge, drop the swap entry duplicate.
6789 mem_cgroup_uncharge_swap(entry
, nr_pages
);
6793 css_put(&memcg
->css
);
6798 struct uncharge_gather
{
6799 struct mem_cgroup
*memcg
;
6800 unsigned long nr_pages
;
6801 unsigned long pgpgout
;
6802 unsigned long nr_kmem
;
6803 struct page
*dummy_page
;
6806 static inline void uncharge_gather_clear(struct uncharge_gather
*ug
)
6808 memset(ug
, 0, sizeof(*ug
));
6811 static void uncharge_batch(const struct uncharge_gather
*ug
)
6813 unsigned long flags
;
6815 if (!mem_cgroup_is_root(ug
->memcg
)) {
6816 page_counter_uncharge(&ug
->memcg
->memory
, ug
->nr_pages
);
6817 if (do_memsw_account())
6818 page_counter_uncharge(&ug
->memcg
->memsw
, ug
->nr_pages
);
6819 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && ug
->nr_kmem
)
6820 page_counter_uncharge(&ug
->memcg
->kmem
, ug
->nr_kmem
);
6821 memcg_oom_recover(ug
->memcg
);
6824 local_irq_save(flags
);
6825 __count_memcg_events(ug
->memcg
, PGPGOUT
, ug
->pgpgout
);
6826 __this_cpu_add(ug
->memcg
->vmstats_percpu
->nr_page_events
, ug
->nr_pages
);
6827 memcg_check_events(ug
->memcg
, ug
->dummy_page
);
6828 local_irq_restore(flags
);
6830 /* drop reference from uncharge_page */
6831 css_put(&ug
->memcg
->css
);
6834 static void uncharge_page(struct page
*page
, struct uncharge_gather
*ug
)
6836 unsigned long nr_pages
;
6838 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6840 if (!page_memcg(page
))
6844 * Nobody should be changing or seriously looking at
6845 * page_memcg(page) at this point, we have fully
6846 * exclusive access to the page.
6849 if (ug
->memcg
!= page_memcg(page
)) {
6852 uncharge_gather_clear(ug
);
6854 ug
->memcg
= page_memcg(page
);
6856 /* pairs with css_put in uncharge_batch */
6857 css_get(&ug
->memcg
->css
);
6860 nr_pages
= compound_nr(page
);
6861 ug
->nr_pages
+= nr_pages
;
6863 if (PageMemcgKmem(page
))
6864 ug
->nr_kmem
+= nr_pages
;
6868 ug
->dummy_page
= page
;
6869 page
->memcg_data
= 0;
6870 css_put(&ug
->memcg
->css
);
6873 static void uncharge_list(struct list_head
*page_list
)
6875 struct uncharge_gather ug
;
6876 struct list_head
*next
;
6878 uncharge_gather_clear(&ug
);
6881 * Note that the list can be a single page->lru; hence the
6882 * do-while loop instead of a simple list_for_each_entry().
6884 next
= page_list
->next
;
6888 page
= list_entry(next
, struct page
, lru
);
6889 next
= page
->lru
.next
;
6891 uncharge_page(page
, &ug
);
6892 } while (next
!= page_list
);
6895 uncharge_batch(&ug
);
6899 * mem_cgroup_uncharge - uncharge a page
6900 * @page: page to uncharge
6902 * Uncharge a page previously charged with mem_cgroup_charge().
6904 void mem_cgroup_uncharge(struct page
*page
)
6906 struct uncharge_gather ug
;
6908 if (mem_cgroup_disabled())
6911 /* Don't touch page->lru of any random page, pre-check: */
6912 if (!page_memcg(page
))
6915 uncharge_gather_clear(&ug
);
6916 uncharge_page(page
, &ug
);
6917 uncharge_batch(&ug
);
6921 * mem_cgroup_uncharge_list - uncharge a list of page
6922 * @page_list: list of pages to uncharge
6924 * Uncharge a list of pages previously charged with
6925 * mem_cgroup_charge().
6927 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
6929 if (mem_cgroup_disabled())
6932 if (!list_empty(page_list
))
6933 uncharge_list(page_list
);
6937 * mem_cgroup_migrate - charge a page's replacement
6938 * @oldpage: currently circulating page
6939 * @newpage: replacement page
6941 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6942 * be uncharged upon free.
6944 * Both pages must be locked, @newpage->mapping must be set up.
6946 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
6948 struct mem_cgroup
*memcg
;
6949 unsigned int nr_pages
;
6950 unsigned long flags
;
6952 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
6953 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
6954 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
6955 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
6958 if (mem_cgroup_disabled())
6961 /* Page cache replacement: new page already charged? */
6962 if (page_memcg(newpage
))
6965 memcg
= page_memcg(oldpage
);
6966 VM_WARN_ON_ONCE_PAGE(!memcg
, oldpage
);
6970 /* Force-charge the new page. The old one will be freed soon */
6971 nr_pages
= thp_nr_pages(newpage
);
6973 page_counter_charge(&memcg
->memory
, nr_pages
);
6974 if (do_memsw_account())
6975 page_counter_charge(&memcg
->memsw
, nr_pages
);
6977 css_get(&memcg
->css
);
6978 commit_charge(newpage
, memcg
);
6980 local_irq_save(flags
);
6981 mem_cgroup_charge_statistics(memcg
, newpage
, nr_pages
);
6982 memcg_check_events(memcg
, newpage
);
6983 local_irq_restore(flags
);
6986 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
6987 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
6989 void mem_cgroup_sk_alloc(struct sock
*sk
)
6991 struct mem_cgroup
*memcg
;
6993 if (!mem_cgroup_sockets_enabled
)
6996 /* Do not associate the sock with unrelated interrupted task's memcg. */
7001 memcg
= mem_cgroup_from_task(current
);
7002 if (memcg
== root_mem_cgroup
)
7004 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
7006 if (css_tryget(&memcg
->css
))
7007 sk
->sk_memcg
= memcg
;
7012 void mem_cgroup_sk_free(struct sock
*sk
)
7015 css_put(&sk
->sk_memcg
->css
);
7019 * mem_cgroup_charge_skmem - charge socket memory
7020 * @memcg: memcg to charge
7021 * @nr_pages: number of pages to charge
7023 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7024 * @memcg's configured limit, %false if the charge had to be forced.
7026 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
7028 gfp_t gfp_mask
= GFP_KERNEL
;
7030 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
7031 struct page_counter
*fail
;
7033 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
7034 memcg
->tcpmem_pressure
= 0;
7037 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
7038 memcg
->tcpmem_pressure
= 1;
7042 /* Don't block in the packet receive path */
7044 gfp_mask
= GFP_NOWAIT
;
7046 mod_memcg_state(memcg
, MEMCG_SOCK
, nr_pages
);
7048 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
7051 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
7056 * mem_cgroup_uncharge_skmem - uncharge socket memory
7057 * @memcg: memcg to uncharge
7058 * @nr_pages: number of pages to uncharge
7060 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
7062 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
7063 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
7067 mod_memcg_state(memcg
, MEMCG_SOCK
, -nr_pages
);
7069 refill_stock(memcg
, nr_pages
);
7072 static int __init
cgroup_memory(char *s
)
7076 while ((token
= strsep(&s
, ",")) != NULL
) {
7079 if (!strcmp(token
, "nosocket"))
7080 cgroup_memory_nosocket
= true;
7081 if (!strcmp(token
, "nokmem"))
7082 cgroup_memory_nokmem
= true;
7086 __setup("cgroup.memory=", cgroup_memory
);
7089 * subsys_initcall() for memory controller.
7091 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7092 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7093 * basically everything that doesn't depend on a specific mem_cgroup structure
7094 * should be initialized from here.
7096 static int __init
mem_cgroup_init(void)
7100 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD
, "mm/memctrl:dead", NULL
,
7101 memcg_hotplug_cpu_dead
);
7103 for_each_possible_cpu(cpu
)
7104 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
7107 for_each_node(node
) {
7108 struct mem_cgroup_tree_per_node
*rtpn
;
7110 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
7111 node_online(node
) ? node
: NUMA_NO_NODE
);
7113 rtpn
->rb_root
= RB_ROOT
;
7114 rtpn
->rb_rightmost
= NULL
;
7115 spin_lock_init(&rtpn
->lock
);
7116 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
7121 subsys_initcall(mem_cgroup_init
);
7123 #ifdef CONFIG_MEMCG_SWAP
7124 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
7126 while (!refcount_inc_not_zero(&memcg
->id
.ref
)) {
7128 * The root cgroup cannot be destroyed, so it's refcount must
7131 if (WARN_ON_ONCE(memcg
== root_mem_cgroup
)) {
7135 memcg
= parent_mem_cgroup(memcg
);
7137 memcg
= root_mem_cgroup
;
7143 * mem_cgroup_swapout - transfer a memsw charge to swap
7144 * @page: page whose memsw charge to transfer
7145 * @entry: swap entry to move the charge to
7147 * Transfer the memsw charge of @page to @entry.
7149 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
7151 struct mem_cgroup
*memcg
, *swap_memcg
;
7152 unsigned int nr_entries
;
7153 unsigned short oldid
;
7155 VM_BUG_ON_PAGE(PageLRU(page
), page
);
7156 VM_BUG_ON_PAGE(page_count(page
), page
);
7158 if (mem_cgroup_disabled())
7161 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7164 memcg
= page_memcg(page
);
7166 VM_WARN_ON_ONCE_PAGE(!memcg
, page
);
7171 * In case the memcg owning these pages has been offlined and doesn't
7172 * have an ID allocated to it anymore, charge the closest online
7173 * ancestor for the swap instead and transfer the memory+swap charge.
7175 swap_memcg
= mem_cgroup_id_get_online(memcg
);
7176 nr_entries
= thp_nr_pages(page
);
7177 /* Get references for the tail pages, too */
7179 mem_cgroup_id_get_many(swap_memcg
, nr_entries
- 1);
7180 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
),
7182 VM_BUG_ON_PAGE(oldid
, page
);
7183 mod_memcg_state(swap_memcg
, MEMCG_SWAP
, nr_entries
);
7185 page
->memcg_data
= 0;
7187 if (!mem_cgroup_is_root(memcg
))
7188 page_counter_uncharge(&memcg
->memory
, nr_entries
);
7190 if (!cgroup_memory_noswap
&& memcg
!= swap_memcg
) {
7191 if (!mem_cgroup_is_root(swap_memcg
))
7192 page_counter_charge(&swap_memcg
->memsw
, nr_entries
);
7193 page_counter_uncharge(&memcg
->memsw
, nr_entries
);
7197 * Interrupts should be disabled here because the caller holds the
7198 * i_pages lock which is taken with interrupts-off. It is
7199 * important here to have the interrupts disabled because it is the
7200 * only synchronisation we have for updating the per-CPU variables.
7202 VM_BUG_ON(!irqs_disabled());
7203 mem_cgroup_charge_statistics(memcg
, page
, -nr_entries
);
7204 memcg_check_events(memcg
, page
);
7206 css_put(&memcg
->css
);
7210 * mem_cgroup_try_charge_swap - try charging swap space for a page
7211 * @page: page being added to swap
7212 * @entry: swap entry to charge
7214 * Try to charge @page's memcg for the swap space at @entry.
7216 * Returns 0 on success, -ENOMEM on failure.
7218 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
7220 unsigned int nr_pages
= thp_nr_pages(page
);
7221 struct page_counter
*counter
;
7222 struct mem_cgroup
*memcg
;
7223 unsigned short oldid
;
7225 if (mem_cgroup_disabled())
7228 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7231 memcg
= page_memcg(page
);
7233 VM_WARN_ON_ONCE_PAGE(!memcg
, page
);
7238 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7242 memcg
= mem_cgroup_id_get_online(memcg
);
7244 if (!cgroup_memory_noswap
&& !mem_cgroup_is_root(memcg
) &&
7245 !page_counter_try_charge(&memcg
->swap
, nr_pages
, &counter
)) {
7246 memcg_memory_event(memcg
, MEMCG_SWAP_MAX
);
7247 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7248 mem_cgroup_id_put(memcg
);
7252 /* Get references for the tail pages, too */
7254 mem_cgroup_id_get_many(memcg
, nr_pages
- 1);
7255 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
), nr_pages
);
7256 VM_BUG_ON_PAGE(oldid
, page
);
7257 mod_memcg_state(memcg
, MEMCG_SWAP
, nr_pages
);
7263 * mem_cgroup_uncharge_swap - uncharge swap space
7264 * @entry: swap entry to uncharge
7265 * @nr_pages: the amount of swap space to uncharge
7267 void mem_cgroup_uncharge_swap(swp_entry_t entry
, unsigned int nr_pages
)
7269 struct mem_cgroup
*memcg
;
7272 id
= swap_cgroup_record(entry
, 0, nr_pages
);
7274 memcg
= mem_cgroup_from_id(id
);
7276 if (!cgroup_memory_noswap
&& !mem_cgroup_is_root(memcg
)) {
7277 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7278 page_counter_uncharge(&memcg
->swap
, nr_pages
);
7280 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
7282 mod_memcg_state(memcg
, MEMCG_SWAP
, -nr_pages
);
7283 mem_cgroup_id_put_many(memcg
, nr_pages
);
7288 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
7290 long nr_swap_pages
= get_nr_swap_pages();
7292 if (cgroup_memory_noswap
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7293 return nr_swap_pages
;
7294 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
7295 nr_swap_pages
= min_t(long, nr_swap_pages
,
7296 READ_ONCE(memcg
->swap
.max
) -
7297 page_counter_read(&memcg
->swap
));
7298 return nr_swap_pages
;
7301 bool mem_cgroup_swap_full(struct page
*page
)
7303 struct mem_cgroup
*memcg
;
7305 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
7309 if (cgroup_memory_noswap
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7312 memcg
= page_memcg(page
);
7316 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
)) {
7317 unsigned long usage
= page_counter_read(&memcg
->swap
);
7319 if (usage
* 2 >= READ_ONCE(memcg
->swap
.high
) ||
7320 usage
* 2 >= READ_ONCE(memcg
->swap
.max
))
7327 static int __init
setup_swap_account(char *s
)
7329 if (!strcmp(s
, "1"))
7330 cgroup_memory_noswap
= false;
7331 else if (!strcmp(s
, "0"))
7332 cgroup_memory_noswap
= true;
7335 __setup("swapaccount=", setup_swap_account
);
7337 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
7340 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
7342 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
7345 static int swap_high_show(struct seq_file
*m
, void *v
)
7347 return seq_puts_memcg_tunable(m
,
7348 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.high
));
7351 static ssize_t
swap_high_write(struct kernfs_open_file
*of
,
7352 char *buf
, size_t nbytes
, loff_t off
)
7354 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7358 buf
= strstrip(buf
);
7359 err
= page_counter_memparse(buf
, "max", &high
);
7363 page_counter_set_high(&memcg
->swap
, high
);
7368 static int swap_max_show(struct seq_file
*m
, void *v
)
7370 return seq_puts_memcg_tunable(m
,
7371 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.max
));
7374 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
7375 char *buf
, size_t nbytes
, loff_t off
)
7377 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7381 buf
= strstrip(buf
);
7382 err
= page_counter_memparse(buf
, "max", &max
);
7386 xchg(&memcg
->swap
.max
, max
);
7391 static int swap_events_show(struct seq_file
*m
, void *v
)
7393 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
7395 seq_printf(m
, "high %lu\n",
7396 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_HIGH
]));
7397 seq_printf(m
, "max %lu\n",
7398 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_MAX
]));
7399 seq_printf(m
, "fail %lu\n",
7400 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_FAIL
]));
7405 static struct cftype swap_files
[] = {
7407 .name
= "swap.current",
7408 .flags
= CFTYPE_NOT_ON_ROOT
,
7409 .read_u64
= swap_current_read
,
7412 .name
= "swap.high",
7413 .flags
= CFTYPE_NOT_ON_ROOT
,
7414 .seq_show
= swap_high_show
,
7415 .write
= swap_high_write
,
7419 .flags
= CFTYPE_NOT_ON_ROOT
,
7420 .seq_show
= swap_max_show
,
7421 .write
= swap_max_write
,
7424 .name
= "swap.events",
7425 .flags
= CFTYPE_NOT_ON_ROOT
,
7426 .file_offset
= offsetof(struct mem_cgroup
, swap_events_file
),
7427 .seq_show
= swap_events_show
,
7432 static struct cftype memsw_files
[] = {
7434 .name
= "memsw.usage_in_bytes",
7435 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
7436 .read_u64
= mem_cgroup_read_u64
,
7439 .name
= "memsw.max_usage_in_bytes",
7440 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
7441 .write
= mem_cgroup_reset
,
7442 .read_u64
= mem_cgroup_read_u64
,
7445 .name
= "memsw.limit_in_bytes",
7446 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
7447 .write
= mem_cgroup_write
,
7448 .read_u64
= mem_cgroup_read_u64
,
7451 .name
= "memsw.failcnt",
7452 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
7453 .write
= mem_cgroup_reset
,
7454 .read_u64
= mem_cgroup_read_u64
,
7456 { }, /* terminate */
7460 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7461 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7462 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7463 * boot parameter. This may result in premature OOPS inside
7464 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7466 static int __init
mem_cgroup_swap_init(void)
7468 /* No memory control -> no swap control */
7469 if (mem_cgroup_disabled())
7470 cgroup_memory_noswap
= true;
7472 if (cgroup_memory_noswap
)
7475 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
, swap_files
));
7476 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
, memsw_files
));
7480 core_initcall(mem_cgroup_swap_init
);
7482 #endif /* CONFIG_MEMCG_SWAP */