1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
25 #include <linux/page_counter.h>
26 #include <linux/memcontrol.h>
27 #include <linux/cgroup.h>
28 #include <linux/pagewalk.h>
29 #include <linux/sched/mm.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/hugetlb.h>
32 #include <linux/pagemap.h>
33 #include <linux/vm_event_item.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
59 #include <linux/tracehook.h>
60 #include <linux/psi.h>
61 #include <linux/seq_buf.h>
67 #include <linux/uaccess.h>
69 #include <trace/events/vmscan.h>
71 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
72 EXPORT_SYMBOL(memory_cgrp_subsys
);
74 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
76 /* Socket memory accounting disabled? */
77 static bool cgroup_memory_nosocket
;
79 /* Kernel memory accounting disabled? */
80 static bool cgroup_memory_nokmem
;
82 /* Whether the swap controller is active */
83 #ifdef CONFIG_MEMCG_SWAP
84 bool cgroup_memory_noswap __read_mostly
;
86 #define cgroup_memory_noswap 1
89 #ifdef CONFIG_CGROUP_WRITEBACK
90 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq
);
93 /* Whether legacy memory+swap accounting is active */
94 static bool do_memsw_account(void)
96 return !cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_noswap
;
99 #define THRESHOLDS_EVENTS_TARGET 128
100 #define SOFTLIMIT_EVENTS_TARGET 1024
103 * Cgroups above their limits are maintained in a RB-Tree, independent of
104 * their hierarchy representation
107 struct mem_cgroup_tree_per_node
{
108 struct rb_root rb_root
;
109 struct rb_node
*rb_rightmost
;
113 struct mem_cgroup_tree
{
114 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
117 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
120 struct mem_cgroup_eventfd_list
{
121 struct list_head list
;
122 struct eventfd_ctx
*eventfd
;
126 * cgroup_event represents events which userspace want to receive.
128 struct mem_cgroup_event
{
130 * memcg which the event belongs to.
132 struct mem_cgroup
*memcg
;
134 * eventfd to signal userspace about the event.
136 struct eventfd_ctx
*eventfd
;
138 * Each of these stored in a list by the cgroup.
140 struct list_head list
;
142 * register_event() callback will be used to add new userspace
143 * waiter for changes related to this event. Use eventfd_signal()
144 * on eventfd to send notification to userspace.
146 int (*register_event
)(struct mem_cgroup
*memcg
,
147 struct eventfd_ctx
*eventfd
, const char *args
);
149 * unregister_event() callback will be called when userspace closes
150 * the eventfd or on cgroup removing. This callback must be set,
151 * if you want provide notification functionality.
153 void (*unregister_event
)(struct mem_cgroup
*memcg
,
154 struct eventfd_ctx
*eventfd
);
156 * All fields below needed to unregister event when
157 * userspace closes eventfd.
160 wait_queue_head_t
*wqh
;
161 wait_queue_entry_t wait
;
162 struct work_struct remove
;
165 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
166 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
168 /* Stuffs for move charges at task migration. */
170 * Types of charges to be moved.
172 #define MOVE_ANON 0x1U
173 #define MOVE_FILE 0x2U
174 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
176 /* "mc" and its members are protected by cgroup_mutex */
177 static struct move_charge_struct
{
178 spinlock_t lock
; /* for from, to */
179 struct mm_struct
*mm
;
180 struct mem_cgroup
*from
;
181 struct mem_cgroup
*to
;
183 unsigned long precharge
;
184 unsigned long moved_charge
;
185 unsigned long moved_swap
;
186 struct task_struct
*moving_task
; /* a task moving charges */
187 wait_queue_head_t waitq
; /* a waitq for other context */
189 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
190 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
194 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
195 * limit reclaim to prevent infinite loops, if they ever occur.
197 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
198 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
201 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
202 MEM_CGROUP_CHARGE_TYPE_ANON
,
203 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
204 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
208 /* for encoding cft->private value on file */
217 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
218 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
219 #define MEMFILE_ATTR(val) ((val) & 0xffff)
220 /* Used for OOM nofiier */
221 #define OOM_CONTROL (0)
224 * Iteration constructs for visiting all cgroups (under a tree). If
225 * loops are exited prematurely (break), mem_cgroup_iter_break() must
226 * be used for reference counting.
228 #define for_each_mem_cgroup_tree(iter, root) \
229 for (iter = mem_cgroup_iter(root, NULL, NULL); \
231 iter = mem_cgroup_iter(root, iter, NULL))
233 #define for_each_mem_cgroup(iter) \
234 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
236 iter = mem_cgroup_iter(NULL, iter, NULL))
238 static inline bool should_force_charge(void)
240 return tsk_is_oom_victim(current
) || fatal_signal_pending(current
) ||
241 (current
->flags
& PF_EXITING
);
244 /* Some nice accessors for the vmpressure. */
245 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
248 memcg
= root_mem_cgroup
;
249 return &memcg
->vmpressure
;
252 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
254 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
257 #ifdef CONFIG_MEMCG_KMEM
258 extern spinlock_t css_set_lock
;
260 static void obj_cgroup_release(struct percpu_ref
*ref
)
262 struct obj_cgroup
*objcg
= container_of(ref
, struct obj_cgroup
, refcnt
);
263 struct mem_cgroup
*memcg
;
264 unsigned int nr_bytes
;
265 unsigned int nr_pages
;
269 * At this point all allocated objects are freed, and
270 * objcg->nr_charged_bytes can't have an arbitrary byte value.
271 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
273 * The following sequence can lead to it:
274 * 1) CPU0: objcg == stock->cached_objcg
275 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
276 * PAGE_SIZE bytes are charged
277 * 3) CPU1: a process from another memcg is allocating something,
278 * the stock if flushed,
279 * objcg->nr_charged_bytes = PAGE_SIZE - 92
280 * 5) CPU0: we do release this object,
281 * 92 bytes are added to stock->nr_bytes
282 * 6) CPU0: stock is flushed,
283 * 92 bytes are added to objcg->nr_charged_bytes
285 * In the result, nr_charged_bytes == PAGE_SIZE.
286 * This page will be uncharged in obj_cgroup_release().
288 nr_bytes
= atomic_read(&objcg
->nr_charged_bytes
);
289 WARN_ON_ONCE(nr_bytes
& (PAGE_SIZE
- 1));
290 nr_pages
= nr_bytes
>> PAGE_SHIFT
;
292 spin_lock_irqsave(&css_set_lock
, flags
);
293 memcg
= obj_cgroup_memcg(objcg
);
295 __memcg_kmem_uncharge(memcg
, nr_pages
);
296 list_del(&objcg
->list
);
297 mem_cgroup_put(memcg
);
298 spin_unlock_irqrestore(&css_set_lock
, flags
);
300 percpu_ref_exit(ref
);
301 kfree_rcu(objcg
, rcu
);
304 static struct obj_cgroup
*obj_cgroup_alloc(void)
306 struct obj_cgroup
*objcg
;
309 objcg
= kzalloc(sizeof(struct obj_cgroup
), GFP_KERNEL
);
313 ret
= percpu_ref_init(&objcg
->refcnt
, obj_cgroup_release
, 0,
319 INIT_LIST_HEAD(&objcg
->list
);
323 static void memcg_reparent_objcgs(struct mem_cgroup
*memcg
,
324 struct mem_cgroup
*parent
)
326 struct obj_cgroup
*objcg
, *iter
;
328 objcg
= rcu_replace_pointer(memcg
->objcg
, NULL
, true);
330 spin_lock_irq(&css_set_lock
);
332 /* Move active objcg to the parent's list */
333 xchg(&objcg
->memcg
, parent
);
334 css_get(&parent
->css
);
335 list_add(&objcg
->list
, &parent
->objcg_list
);
337 /* Move already reparented objcgs to the parent's list */
338 list_for_each_entry(iter
, &memcg
->objcg_list
, list
) {
339 css_get(&parent
->css
);
340 xchg(&iter
->memcg
, parent
);
341 css_put(&memcg
->css
);
343 list_splice(&memcg
->objcg_list
, &parent
->objcg_list
);
345 spin_unlock_irq(&css_set_lock
);
347 percpu_ref_kill(&objcg
->refcnt
);
351 * This will be used as a shrinker list's index.
352 * The main reason for not using cgroup id for this:
353 * this works better in sparse environments, where we have a lot of memcgs,
354 * but only a few kmem-limited. Or also, if we have, for instance, 200
355 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
356 * 200 entry array for that.
358 * The current size of the caches array is stored in memcg_nr_cache_ids. It
359 * will double each time we have to increase it.
361 static DEFINE_IDA(memcg_cache_ida
);
362 int memcg_nr_cache_ids
;
364 /* Protects memcg_nr_cache_ids */
365 static DECLARE_RWSEM(memcg_cache_ids_sem
);
367 void memcg_get_cache_ids(void)
369 down_read(&memcg_cache_ids_sem
);
372 void memcg_put_cache_ids(void)
374 up_read(&memcg_cache_ids_sem
);
378 * MIN_SIZE is different than 1, because we would like to avoid going through
379 * the alloc/free process all the time. In a small machine, 4 kmem-limited
380 * cgroups is a reasonable guess. In the future, it could be a parameter or
381 * tunable, but that is strictly not necessary.
383 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
384 * this constant directly from cgroup, but it is understandable that this is
385 * better kept as an internal representation in cgroup.c. In any case, the
386 * cgrp_id space is not getting any smaller, and we don't have to necessarily
387 * increase ours as well if it increases.
389 #define MEMCG_CACHES_MIN_SIZE 4
390 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
393 * A lot of the calls to the cache allocation functions are expected to be
394 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
395 * conditional to this static branch, we'll have to allow modules that does
396 * kmem_cache_alloc and the such to see this symbol as well
398 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
399 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
402 static int memcg_shrinker_map_size
;
403 static DEFINE_MUTEX(memcg_shrinker_map_mutex
);
405 static void memcg_free_shrinker_map_rcu(struct rcu_head
*head
)
407 kvfree(container_of(head
, struct memcg_shrinker_map
, rcu
));
410 static int memcg_expand_one_shrinker_map(struct mem_cgroup
*memcg
,
411 int size
, int old_size
)
413 struct memcg_shrinker_map
*new, *old
;
416 lockdep_assert_held(&memcg_shrinker_map_mutex
);
419 old
= rcu_dereference_protected(
420 mem_cgroup_nodeinfo(memcg
, nid
)->shrinker_map
, true);
421 /* Not yet online memcg */
425 new = kvmalloc_node(sizeof(*new) + size
, GFP_KERNEL
, nid
);
429 /* Set all old bits, clear all new bits */
430 memset(new->map
, (int)0xff, old_size
);
431 memset((void *)new->map
+ old_size
, 0, size
- old_size
);
433 rcu_assign_pointer(memcg
->nodeinfo
[nid
]->shrinker_map
, new);
434 call_rcu(&old
->rcu
, memcg_free_shrinker_map_rcu
);
440 static void memcg_free_shrinker_maps(struct mem_cgroup
*memcg
)
442 struct mem_cgroup_per_node
*pn
;
443 struct memcg_shrinker_map
*map
;
446 if (mem_cgroup_is_root(memcg
))
450 pn
= mem_cgroup_nodeinfo(memcg
, nid
);
451 map
= rcu_dereference_protected(pn
->shrinker_map
, true);
454 rcu_assign_pointer(pn
->shrinker_map
, NULL
);
458 static int memcg_alloc_shrinker_maps(struct mem_cgroup
*memcg
)
460 struct memcg_shrinker_map
*map
;
461 int nid
, size
, ret
= 0;
463 if (mem_cgroup_is_root(memcg
))
466 mutex_lock(&memcg_shrinker_map_mutex
);
467 size
= memcg_shrinker_map_size
;
469 map
= kvzalloc_node(sizeof(*map
) + size
, GFP_KERNEL
, nid
);
471 memcg_free_shrinker_maps(memcg
);
475 rcu_assign_pointer(memcg
->nodeinfo
[nid
]->shrinker_map
, map
);
477 mutex_unlock(&memcg_shrinker_map_mutex
);
482 int memcg_expand_shrinker_maps(int new_id
)
484 int size
, old_size
, ret
= 0;
485 struct mem_cgroup
*memcg
;
487 size
= DIV_ROUND_UP(new_id
+ 1, BITS_PER_LONG
) * sizeof(unsigned long);
488 old_size
= memcg_shrinker_map_size
;
489 if (size
<= old_size
)
492 mutex_lock(&memcg_shrinker_map_mutex
);
493 if (!root_mem_cgroup
)
496 for_each_mem_cgroup(memcg
) {
497 if (mem_cgroup_is_root(memcg
))
499 ret
= memcg_expand_one_shrinker_map(memcg
, size
, old_size
);
501 mem_cgroup_iter_break(NULL
, memcg
);
507 memcg_shrinker_map_size
= size
;
508 mutex_unlock(&memcg_shrinker_map_mutex
);
512 void memcg_set_shrinker_bit(struct mem_cgroup
*memcg
, int nid
, int shrinker_id
)
514 if (shrinker_id
>= 0 && memcg
&& !mem_cgroup_is_root(memcg
)) {
515 struct memcg_shrinker_map
*map
;
518 map
= rcu_dereference(memcg
->nodeinfo
[nid
]->shrinker_map
);
519 /* Pairs with smp mb in shrink_slab() */
520 smp_mb__before_atomic();
521 set_bit(shrinker_id
, map
->map
);
527 * mem_cgroup_css_from_page - css of the memcg associated with a page
528 * @page: page of interest
530 * If memcg is bound to the default hierarchy, css of the memcg associated
531 * with @page is returned. The returned css remains associated with @page
532 * until it is released.
534 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
537 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
539 struct mem_cgroup
*memcg
;
541 memcg
= page
->mem_cgroup
;
543 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
544 memcg
= root_mem_cgroup
;
550 * page_cgroup_ino - return inode number of the memcg a page is charged to
553 * Look up the closest online ancestor of the memory cgroup @page is charged to
554 * and return its inode number or 0 if @page is not charged to any cgroup. It
555 * is safe to call this function without holding a reference to @page.
557 * Note, this function is inherently racy, because there is nothing to prevent
558 * the cgroup inode from getting torn down and potentially reallocated a moment
559 * after page_cgroup_ino() returns, so it only should be used by callers that
560 * do not care (such as procfs interfaces).
562 ino_t
page_cgroup_ino(struct page
*page
)
564 struct mem_cgroup
*memcg
;
565 unsigned long ino
= 0;
568 memcg
= page
->mem_cgroup
;
571 * The lowest bit set means that memcg isn't a valid
572 * memcg pointer, but a obj_cgroups pointer.
573 * In this case the page is shared and doesn't belong
574 * to any specific memory cgroup.
576 if ((unsigned long) memcg
& 0x1UL
)
579 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
580 memcg
= parent_mem_cgroup(memcg
);
582 ino
= cgroup_ino(memcg
->css
.cgroup
);
587 static struct mem_cgroup_per_node
*
588 mem_cgroup_page_nodeinfo(struct mem_cgroup
*memcg
, struct page
*page
)
590 int nid
= page_to_nid(page
);
592 return memcg
->nodeinfo
[nid
];
595 static struct mem_cgroup_tree_per_node
*
596 soft_limit_tree_node(int nid
)
598 return soft_limit_tree
.rb_tree_per_node
[nid
];
601 static struct mem_cgroup_tree_per_node
*
602 soft_limit_tree_from_page(struct page
*page
)
604 int nid
= page_to_nid(page
);
606 return soft_limit_tree
.rb_tree_per_node
[nid
];
609 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node
*mz
,
610 struct mem_cgroup_tree_per_node
*mctz
,
611 unsigned long new_usage_in_excess
)
613 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
614 struct rb_node
*parent
= NULL
;
615 struct mem_cgroup_per_node
*mz_node
;
616 bool rightmost
= true;
621 mz
->usage_in_excess
= new_usage_in_excess
;
622 if (!mz
->usage_in_excess
)
626 mz_node
= rb_entry(parent
, struct mem_cgroup_per_node
,
628 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
) {
634 * We can't avoid mem cgroups that are over their soft
635 * limit by the same amount
637 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
642 mctz
->rb_rightmost
= &mz
->tree_node
;
644 rb_link_node(&mz
->tree_node
, parent
, p
);
645 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
649 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
650 struct mem_cgroup_tree_per_node
*mctz
)
655 if (&mz
->tree_node
== mctz
->rb_rightmost
)
656 mctz
->rb_rightmost
= rb_prev(&mz
->tree_node
);
658 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
662 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
663 struct mem_cgroup_tree_per_node
*mctz
)
667 spin_lock_irqsave(&mctz
->lock
, flags
);
668 __mem_cgroup_remove_exceeded(mz
, mctz
);
669 spin_unlock_irqrestore(&mctz
->lock
, flags
);
672 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
674 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
675 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
676 unsigned long excess
= 0;
678 if (nr_pages
> soft_limit
)
679 excess
= nr_pages
- soft_limit
;
684 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
686 unsigned long excess
;
687 struct mem_cgroup_per_node
*mz
;
688 struct mem_cgroup_tree_per_node
*mctz
;
690 mctz
= soft_limit_tree_from_page(page
);
694 * Necessary to update all ancestors when hierarchy is used.
695 * because their event counter is not touched.
697 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
698 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
699 excess
= soft_limit_excess(memcg
);
701 * We have to update the tree if mz is on RB-tree or
702 * mem is over its softlimit.
704 if (excess
|| mz
->on_tree
) {
707 spin_lock_irqsave(&mctz
->lock
, flags
);
708 /* if on-tree, remove it */
710 __mem_cgroup_remove_exceeded(mz
, mctz
);
712 * Insert again. mz->usage_in_excess will be updated.
713 * If excess is 0, no tree ops.
715 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
716 spin_unlock_irqrestore(&mctz
->lock
, flags
);
721 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
723 struct mem_cgroup_tree_per_node
*mctz
;
724 struct mem_cgroup_per_node
*mz
;
728 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
729 mctz
= soft_limit_tree_node(nid
);
731 mem_cgroup_remove_exceeded(mz
, mctz
);
735 static struct mem_cgroup_per_node
*
736 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
738 struct mem_cgroup_per_node
*mz
;
742 if (!mctz
->rb_rightmost
)
743 goto done
; /* Nothing to reclaim from */
745 mz
= rb_entry(mctz
->rb_rightmost
,
746 struct mem_cgroup_per_node
, tree_node
);
748 * Remove the node now but someone else can add it back,
749 * we will to add it back at the end of reclaim to its correct
750 * position in the tree.
752 __mem_cgroup_remove_exceeded(mz
, mctz
);
753 if (!soft_limit_excess(mz
->memcg
) ||
754 !css_tryget(&mz
->memcg
->css
))
760 static struct mem_cgroup_per_node
*
761 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
763 struct mem_cgroup_per_node
*mz
;
765 spin_lock_irq(&mctz
->lock
);
766 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
767 spin_unlock_irq(&mctz
->lock
);
772 * __mod_memcg_state - update cgroup memory statistics
773 * @memcg: the memory cgroup
774 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
775 * @val: delta to add to the counter, can be negative
777 void __mod_memcg_state(struct mem_cgroup
*memcg
, int idx
, int val
)
779 long x
, threshold
= MEMCG_CHARGE_BATCH
;
781 if (mem_cgroup_disabled())
784 if (vmstat_item_in_bytes(idx
))
785 threshold
<<= PAGE_SHIFT
;
787 x
= val
+ __this_cpu_read(memcg
->vmstats_percpu
->stat
[idx
]);
788 if (unlikely(abs(x
) > threshold
)) {
789 struct mem_cgroup
*mi
;
792 * Batch local counters to keep them in sync with
793 * the hierarchical ones.
795 __this_cpu_add(memcg
->vmstats_local
->stat
[idx
], x
);
796 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
797 atomic_long_add(x
, &mi
->vmstats
[idx
]);
800 __this_cpu_write(memcg
->vmstats_percpu
->stat
[idx
], x
);
803 static struct mem_cgroup_per_node
*
804 parent_nodeinfo(struct mem_cgroup_per_node
*pn
, int nid
)
806 struct mem_cgroup
*parent
;
808 parent
= parent_mem_cgroup(pn
->memcg
);
811 return mem_cgroup_nodeinfo(parent
, nid
);
814 void __mod_memcg_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
817 struct mem_cgroup_per_node
*pn
;
818 struct mem_cgroup
*memcg
;
819 long x
, threshold
= MEMCG_CHARGE_BATCH
;
821 pn
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
825 __mod_memcg_state(memcg
, idx
, val
);
828 __this_cpu_add(pn
->lruvec_stat_local
->count
[idx
], val
);
830 if (vmstat_item_in_bytes(idx
))
831 threshold
<<= PAGE_SHIFT
;
833 x
= val
+ __this_cpu_read(pn
->lruvec_stat_cpu
->count
[idx
]);
834 if (unlikely(abs(x
) > threshold
)) {
835 pg_data_t
*pgdat
= lruvec_pgdat(lruvec
);
836 struct mem_cgroup_per_node
*pi
;
838 for (pi
= pn
; pi
; pi
= parent_nodeinfo(pi
, pgdat
->node_id
))
839 atomic_long_add(x
, &pi
->lruvec_stat
[idx
]);
842 __this_cpu_write(pn
->lruvec_stat_cpu
->count
[idx
], x
);
846 * __mod_lruvec_state - update lruvec memory statistics
847 * @lruvec: the lruvec
848 * @idx: the stat item
849 * @val: delta to add to the counter, can be negative
851 * The lruvec is the intersection of the NUMA node and a cgroup. This
852 * function updates the all three counters that are affected by a
853 * change of state at this level: per-node, per-cgroup, per-lruvec.
855 void __mod_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
859 __mod_node_page_state(lruvec_pgdat(lruvec
), idx
, val
);
861 /* Update memcg and lruvec */
862 if (!mem_cgroup_disabled())
863 __mod_memcg_lruvec_state(lruvec
, idx
, val
);
866 void __mod_lruvec_slab_state(void *p
, enum node_stat_item idx
, int val
)
868 pg_data_t
*pgdat
= page_pgdat(virt_to_page(p
));
869 struct mem_cgroup
*memcg
;
870 struct lruvec
*lruvec
;
873 memcg
= mem_cgroup_from_obj(p
);
875 /* Untracked pages have no memcg, no lruvec. Update only the node */
876 if (!memcg
|| memcg
== root_mem_cgroup
) {
877 __mod_node_page_state(pgdat
, idx
, val
);
879 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
880 __mod_lruvec_state(lruvec
, idx
, val
);
885 void mod_memcg_obj_state(void *p
, int idx
, int val
)
887 struct mem_cgroup
*memcg
;
890 memcg
= mem_cgroup_from_obj(p
);
892 mod_memcg_state(memcg
, idx
, val
);
897 * __count_memcg_events - account VM events in a cgroup
898 * @memcg: the memory cgroup
899 * @idx: the event item
900 * @count: the number of events that occured
902 void __count_memcg_events(struct mem_cgroup
*memcg
, enum vm_event_item idx
,
907 if (mem_cgroup_disabled())
910 x
= count
+ __this_cpu_read(memcg
->vmstats_percpu
->events
[idx
]);
911 if (unlikely(x
> MEMCG_CHARGE_BATCH
)) {
912 struct mem_cgroup
*mi
;
915 * Batch local counters to keep them in sync with
916 * the hierarchical ones.
918 __this_cpu_add(memcg
->vmstats_local
->events
[idx
], x
);
919 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
920 atomic_long_add(x
, &mi
->vmevents
[idx
]);
923 __this_cpu_write(memcg
->vmstats_percpu
->events
[idx
], x
);
926 static unsigned long memcg_events(struct mem_cgroup
*memcg
, int event
)
928 return atomic_long_read(&memcg
->vmevents
[event
]);
931 static unsigned long memcg_events_local(struct mem_cgroup
*memcg
, int event
)
936 for_each_possible_cpu(cpu
)
937 x
+= per_cpu(memcg
->vmstats_local
->events
[event
], cpu
);
941 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
945 /* pagein of a big page is an event. So, ignore page size */
947 __count_memcg_events(memcg
, PGPGIN
, 1);
949 __count_memcg_events(memcg
, PGPGOUT
, 1);
950 nr_pages
= -nr_pages
; /* for event */
953 __this_cpu_add(memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
956 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
957 enum mem_cgroup_events_target target
)
959 unsigned long val
, next
;
961 val
= __this_cpu_read(memcg
->vmstats_percpu
->nr_page_events
);
962 next
= __this_cpu_read(memcg
->vmstats_percpu
->targets
[target
]);
963 /* from time_after() in jiffies.h */
964 if ((long)(next
- val
) < 0) {
966 case MEM_CGROUP_TARGET_THRESH
:
967 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
969 case MEM_CGROUP_TARGET_SOFTLIMIT
:
970 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
975 __this_cpu_write(memcg
->vmstats_percpu
->targets
[target
], next
);
982 * Check events in order.
985 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
987 /* threshold event is triggered in finer grain than soft limit */
988 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
989 MEM_CGROUP_TARGET_THRESH
))) {
992 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
993 MEM_CGROUP_TARGET_SOFTLIMIT
);
994 mem_cgroup_threshold(memcg
);
995 if (unlikely(do_softlimit
))
996 mem_cgroup_update_tree(memcg
, page
);
1000 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1003 * mm_update_next_owner() may clear mm->owner to NULL
1004 * if it races with swapoff, page migration, etc.
1005 * So this can be called with p == NULL.
1010 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
1012 EXPORT_SYMBOL(mem_cgroup_from_task
);
1015 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1016 * @mm: mm from which memcg should be extracted. It can be NULL.
1018 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1019 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1022 struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1024 struct mem_cgroup
*memcg
;
1026 if (mem_cgroup_disabled())
1032 * Page cache insertions can happen withou an
1033 * actual mm context, e.g. during disk probing
1034 * on boot, loopback IO, acct() writes etc.
1037 memcg
= root_mem_cgroup
;
1039 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1040 if (unlikely(!memcg
))
1041 memcg
= root_mem_cgroup
;
1043 } while (!css_tryget(&memcg
->css
));
1047 EXPORT_SYMBOL(get_mem_cgroup_from_mm
);
1050 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1051 * @page: page from which memcg should be extracted.
1053 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1054 * root_mem_cgroup is returned.
1056 struct mem_cgroup
*get_mem_cgroup_from_page(struct page
*page
)
1058 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
1060 if (mem_cgroup_disabled())
1064 /* Page should not get uncharged and freed memcg under us. */
1065 if (!memcg
|| WARN_ON_ONCE(!css_tryget(&memcg
->css
)))
1066 memcg
= root_mem_cgroup
;
1070 EXPORT_SYMBOL(get_mem_cgroup_from_page
);
1073 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
1075 static __always_inline
struct mem_cgroup
*get_mem_cgroup_from_current(void)
1077 if (unlikely(current
->active_memcg
)) {
1078 struct mem_cgroup
*memcg
;
1081 /* current->active_memcg must hold a ref. */
1082 if (WARN_ON_ONCE(!css_tryget(¤t
->active_memcg
->css
)))
1083 memcg
= root_mem_cgroup
;
1085 memcg
= current
->active_memcg
;
1089 return get_mem_cgroup_from_mm(current
->mm
);
1093 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1094 * @root: hierarchy root
1095 * @prev: previously returned memcg, NULL on first invocation
1096 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1098 * Returns references to children of the hierarchy below @root, or
1099 * @root itself, or %NULL after a full round-trip.
1101 * Caller must pass the return value in @prev on subsequent
1102 * invocations for reference counting, or use mem_cgroup_iter_break()
1103 * to cancel a hierarchy walk before the round-trip is complete.
1105 * Reclaimers can specify a node and a priority level in @reclaim to
1106 * divide up the memcgs in the hierarchy among all concurrent
1107 * reclaimers operating on the same node and priority.
1109 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1110 struct mem_cgroup
*prev
,
1111 struct mem_cgroup_reclaim_cookie
*reclaim
)
1113 struct mem_cgroup_reclaim_iter
*iter
;
1114 struct cgroup_subsys_state
*css
= NULL
;
1115 struct mem_cgroup
*memcg
= NULL
;
1116 struct mem_cgroup
*pos
= NULL
;
1118 if (mem_cgroup_disabled())
1122 root
= root_mem_cgroup
;
1124 if (prev
&& !reclaim
)
1127 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1136 struct mem_cgroup_per_node
*mz
;
1138 mz
= mem_cgroup_nodeinfo(root
, reclaim
->pgdat
->node_id
);
1141 if (prev
&& reclaim
->generation
!= iter
->generation
)
1145 pos
= READ_ONCE(iter
->position
);
1146 if (!pos
|| css_tryget(&pos
->css
))
1149 * css reference reached zero, so iter->position will
1150 * be cleared by ->css_released. However, we should not
1151 * rely on this happening soon, because ->css_released
1152 * is called from a work queue, and by busy-waiting we
1153 * might block it. So we clear iter->position right
1156 (void)cmpxchg(&iter
->position
, pos
, NULL
);
1164 css
= css_next_descendant_pre(css
, &root
->css
);
1167 * Reclaimers share the hierarchy walk, and a
1168 * new one might jump in right at the end of
1169 * the hierarchy - make sure they see at least
1170 * one group and restart from the beginning.
1178 * Verify the css and acquire a reference. The root
1179 * is provided by the caller, so we know it's alive
1180 * and kicking, and don't take an extra reference.
1182 memcg
= mem_cgroup_from_css(css
);
1184 if (css
== &root
->css
)
1187 if (css_tryget(css
))
1195 * The position could have already been updated by a competing
1196 * thread, so check that the value hasn't changed since we read
1197 * it to avoid reclaiming from the same cgroup twice.
1199 (void)cmpxchg(&iter
->position
, pos
, memcg
);
1207 reclaim
->generation
= iter
->generation
;
1213 if (prev
&& prev
!= root
)
1214 css_put(&prev
->css
);
1220 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1221 * @root: hierarchy root
1222 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1224 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1225 struct mem_cgroup
*prev
)
1228 root
= root_mem_cgroup
;
1229 if (prev
&& prev
!= root
)
1230 css_put(&prev
->css
);
1233 static void __invalidate_reclaim_iterators(struct mem_cgroup
*from
,
1234 struct mem_cgroup
*dead_memcg
)
1236 struct mem_cgroup_reclaim_iter
*iter
;
1237 struct mem_cgroup_per_node
*mz
;
1240 for_each_node(nid
) {
1241 mz
= mem_cgroup_nodeinfo(from
, nid
);
1243 cmpxchg(&iter
->position
, dead_memcg
, NULL
);
1247 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
1249 struct mem_cgroup
*memcg
= dead_memcg
;
1250 struct mem_cgroup
*last
;
1253 __invalidate_reclaim_iterators(memcg
, dead_memcg
);
1255 } while ((memcg
= parent_mem_cgroup(memcg
)));
1258 * When cgruop1 non-hierarchy mode is used,
1259 * parent_mem_cgroup() does not walk all the way up to the
1260 * cgroup root (root_mem_cgroup). So we have to handle
1261 * dead_memcg from cgroup root separately.
1263 if (last
!= root_mem_cgroup
)
1264 __invalidate_reclaim_iterators(root_mem_cgroup
,
1269 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1270 * @memcg: hierarchy root
1271 * @fn: function to call for each task
1272 * @arg: argument passed to @fn
1274 * This function iterates over tasks attached to @memcg or to any of its
1275 * descendants and calls @fn for each task. If @fn returns a non-zero
1276 * value, the function breaks the iteration loop and returns the value.
1277 * Otherwise, it will iterate over all tasks and return 0.
1279 * This function must not be called for the root memory cgroup.
1281 int mem_cgroup_scan_tasks(struct mem_cgroup
*memcg
,
1282 int (*fn
)(struct task_struct
*, void *), void *arg
)
1284 struct mem_cgroup
*iter
;
1287 BUG_ON(memcg
== root_mem_cgroup
);
1289 for_each_mem_cgroup_tree(iter
, memcg
) {
1290 struct css_task_iter it
;
1291 struct task_struct
*task
;
1293 css_task_iter_start(&iter
->css
, CSS_TASK_ITER_PROCS
, &it
);
1294 while (!ret
&& (task
= css_task_iter_next(&it
)))
1295 ret
= fn(task
, arg
);
1296 css_task_iter_end(&it
);
1298 mem_cgroup_iter_break(memcg
, iter
);
1306 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1308 * @pgdat: pgdat of the page
1310 * This function relies on page->mem_cgroup being stable - see the
1311 * access rules in commit_charge().
1313 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct pglist_data
*pgdat
)
1315 struct mem_cgroup_per_node
*mz
;
1316 struct mem_cgroup
*memcg
;
1317 struct lruvec
*lruvec
;
1319 if (mem_cgroup_disabled()) {
1320 lruvec
= &pgdat
->__lruvec
;
1324 memcg
= page
->mem_cgroup
;
1326 * Swapcache readahead pages are added to the LRU - and
1327 * possibly migrated - before they are charged.
1330 memcg
= root_mem_cgroup
;
1332 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
1333 lruvec
= &mz
->lruvec
;
1336 * Since a node can be onlined after the mem_cgroup was created,
1337 * we have to be prepared to initialize lruvec->zone here;
1338 * and if offlined then reonlined, we need to reinitialize it.
1340 if (unlikely(lruvec
->pgdat
!= pgdat
))
1341 lruvec
->pgdat
= pgdat
;
1346 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1347 * @lruvec: mem_cgroup per zone lru vector
1348 * @lru: index of lru list the page is sitting on
1349 * @zid: zone id of the accounted pages
1350 * @nr_pages: positive when adding or negative when removing
1352 * This function must be called under lru_lock, just before a page is added
1353 * to or just after a page is removed from an lru list (that ordering being
1354 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1356 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1357 int zid
, int nr_pages
)
1359 struct mem_cgroup_per_node
*mz
;
1360 unsigned long *lru_size
;
1363 if (mem_cgroup_disabled())
1366 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
1367 lru_size
= &mz
->lru_zone_size
[zid
][lru
];
1370 *lru_size
+= nr_pages
;
1373 if (WARN_ONCE(size
< 0,
1374 "%s(%p, %d, %d): lru_size %ld\n",
1375 __func__
, lruvec
, lru
, nr_pages
, size
)) {
1381 *lru_size
+= nr_pages
;
1385 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1386 * @memcg: the memory cgroup
1388 * Returns the maximum amount of memory @mem can be charged with, in
1391 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1393 unsigned long margin
= 0;
1394 unsigned long count
;
1395 unsigned long limit
;
1397 count
= page_counter_read(&memcg
->memory
);
1398 limit
= READ_ONCE(memcg
->memory
.max
);
1400 margin
= limit
- count
;
1402 if (do_memsw_account()) {
1403 count
= page_counter_read(&memcg
->memsw
);
1404 limit
= READ_ONCE(memcg
->memsw
.max
);
1406 margin
= min(margin
, limit
- count
);
1415 * A routine for checking "mem" is under move_account() or not.
1417 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1418 * moving cgroups. This is for waiting at high-memory pressure
1421 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1423 struct mem_cgroup
*from
;
1424 struct mem_cgroup
*to
;
1427 * Unlike task_move routines, we access mc.to, mc.from not under
1428 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1430 spin_lock(&mc
.lock
);
1436 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1437 mem_cgroup_is_descendant(to
, memcg
);
1439 spin_unlock(&mc
.lock
);
1443 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1445 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1446 if (mem_cgroup_under_move(memcg
)) {
1448 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1449 /* moving charge context might have finished. */
1452 finish_wait(&mc
.waitq
, &wait
);
1459 static char *memory_stat_format(struct mem_cgroup
*memcg
)
1464 seq_buf_init(&s
, kmalloc(PAGE_SIZE
, GFP_KERNEL
), PAGE_SIZE
);
1469 * Provide statistics on the state of the memory subsystem as
1470 * well as cumulative event counters that show past behavior.
1472 * This list is ordered following a combination of these gradients:
1473 * 1) generic big picture -> specifics and details
1474 * 2) reflecting userspace activity -> reflecting kernel heuristics
1476 * Current memory state:
1479 seq_buf_printf(&s
, "anon %llu\n",
1480 (u64
)memcg_page_state(memcg
, NR_ANON_MAPPED
) *
1482 seq_buf_printf(&s
, "file %llu\n",
1483 (u64
)memcg_page_state(memcg
, NR_FILE_PAGES
) *
1485 seq_buf_printf(&s
, "kernel_stack %llu\n",
1486 (u64
)memcg_page_state(memcg
, NR_KERNEL_STACK_KB
) *
1488 seq_buf_printf(&s
, "slab %llu\n",
1489 (u64
)(memcg_page_state(memcg
, NR_SLAB_RECLAIMABLE_B
) +
1490 memcg_page_state(memcg
, NR_SLAB_UNRECLAIMABLE_B
)));
1491 seq_buf_printf(&s
, "sock %llu\n",
1492 (u64
)memcg_page_state(memcg
, MEMCG_SOCK
) *
1495 seq_buf_printf(&s
, "shmem %llu\n",
1496 (u64
)memcg_page_state(memcg
, NR_SHMEM
) *
1498 seq_buf_printf(&s
, "file_mapped %llu\n",
1499 (u64
)memcg_page_state(memcg
, NR_FILE_MAPPED
) *
1501 seq_buf_printf(&s
, "file_dirty %llu\n",
1502 (u64
)memcg_page_state(memcg
, NR_FILE_DIRTY
) *
1504 seq_buf_printf(&s
, "file_writeback %llu\n",
1505 (u64
)memcg_page_state(memcg
, NR_WRITEBACK
) *
1508 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1509 seq_buf_printf(&s
, "anon_thp %llu\n",
1510 (u64
)memcg_page_state(memcg
, NR_ANON_THPS
) *
1514 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1515 seq_buf_printf(&s
, "%s %llu\n", lru_list_name(i
),
1516 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
1519 seq_buf_printf(&s
, "slab_reclaimable %llu\n",
1520 (u64
)memcg_page_state(memcg
, NR_SLAB_RECLAIMABLE_B
));
1521 seq_buf_printf(&s
, "slab_unreclaimable %llu\n",
1522 (u64
)memcg_page_state(memcg
, NR_SLAB_UNRECLAIMABLE_B
));
1524 /* Accumulated memory events */
1526 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGFAULT
),
1527 memcg_events(memcg
, PGFAULT
));
1528 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGMAJFAULT
),
1529 memcg_events(memcg
, PGMAJFAULT
));
1531 seq_buf_printf(&s
, "workingset_refault %lu\n",
1532 memcg_page_state(memcg
, WORKINGSET_REFAULT
));
1533 seq_buf_printf(&s
, "workingset_activate %lu\n",
1534 memcg_page_state(memcg
, WORKINGSET_ACTIVATE
));
1535 seq_buf_printf(&s
, "workingset_restore %lu\n",
1536 memcg_page_state(memcg
, WORKINGSET_RESTORE
));
1537 seq_buf_printf(&s
, "workingset_nodereclaim %lu\n",
1538 memcg_page_state(memcg
, WORKINGSET_NODERECLAIM
));
1540 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGREFILL
),
1541 memcg_events(memcg
, PGREFILL
));
1542 seq_buf_printf(&s
, "pgscan %lu\n",
1543 memcg_events(memcg
, PGSCAN_KSWAPD
) +
1544 memcg_events(memcg
, PGSCAN_DIRECT
));
1545 seq_buf_printf(&s
, "pgsteal %lu\n",
1546 memcg_events(memcg
, PGSTEAL_KSWAPD
) +
1547 memcg_events(memcg
, PGSTEAL_DIRECT
));
1548 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGACTIVATE
),
1549 memcg_events(memcg
, PGACTIVATE
));
1550 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGDEACTIVATE
),
1551 memcg_events(memcg
, PGDEACTIVATE
));
1552 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREE
),
1553 memcg_events(memcg
, PGLAZYFREE
));
1554 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREED
),
1555 memcg_events(memcg
, PGLAZYFREED
));
1557 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1558 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC
),
1559 memcg_events(memcg
, THP_FAULT_ALLOC
));
1560 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC
),
1561 memcg_events(memcg
, THP_COLLAPSE_ALLOC
));
1562 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1564 /* The above should easily fit into one page */
1565 WARN_ON_ONCE(seq_buf_has_overflowed(&s
));
1570 #define K(x) ((x) << (PAGE_SHIFT-10))
1572 * mem_cgroup_print_oom_context: Print OOM information relevant to
1573 * memory controller.
1574 * @memcg: The memory cgroup that went over limit
1575 * @p: Task that is going to be killed
1577 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1580 void mem_cgroup_print_oom_context(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1585 pr_cont(",oom_memcg=");
1586 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1588 pr_cont(",global_oom");
1590 pr_cont(",task_memcg=");
1591 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1597 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1598 * memory controller.
1599 * @memcg: The memory cgroup that went over limit
1601 void mem_cgroup_print_oom_meminfo(struct mem_cgroup
*memcg
)
1605 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1606 K((u64
)page_counter_read(&memcg
->memory
)),
1607 K((u64
)READ_ONCE(memcg
->memory
.max
)), memcg
->memory
.failcnt
);
1608 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1609 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1610 K((u64
)page_counter_read(&memcg
->swap
)),
1611 K((u64
)READ_ONCE(memcg
->swap
.max
)), memcg
->swap
.failcnt
);
1613 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1614 K((u64
)page_counter_read(&memcg
->memsw
)),
1615 K((u64
)memcg
->memsw
.max
), memcg
->memsw
.failcnt
);
1616 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1617 K((u64
)page_counter_read(&memcg
->kmem
)),
1618 K((u64
)memcg
->kmem
.max
), memcg
->kmem
.failcnt
);
1621 pr_info("Memory cgroup stats for ");
1622 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1624 buf
= memory_stat_format(memcg
);
1632 * Return the memory (and swap, if configured) limit for a memcg.
1634 unsigned long mem_cgroup_get_max(struct mem_cgroup
*memcg
)
1638 max
= READ_ONCE(memcg
->memory
.max
);
1639 if (mem_cgroup_swappiness(memcg
)) {
1640 unsigned long memsw_max
;
1641 unsigned long swap_max
;
1643 memsw_max
= memcg
->memsw
.max
;
1644 swap_max
= READ_ONCE(memcg
->swap
.max
);
1645 swap_max
= min(swap_max
, (unsigned long)total_swap_pages
);
1646 max
= min(max
+ swap_max
, memsw_max
);
1651 unsigned long mem_cgroup_size(struct mem_cgroup
*memcg
)
1653 return page_counter_read(&memcg
->memory
);
1656 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1659 struct oom_control oc
= {
1663 .gfp_mask
= gfp_mask
,
1668 if (mutex_lock_killable(&oom_lock
))
1671 if (mem_cgroup_margin(memcg
) >= (1 << order
))
1675 * A few threads which were not waiting at mutex_lock_killable() can
1676 * fail to bail out. Therefore, check again after holding oom_lock.
1678 ret
= should_force_charge() || out_of_memory(&oc
);
1681 mutex_unlock(&oom_lock
);
1685 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1688 unsigned long *total_scanned
)
1690 struct mem_cgroup
*victim
= NULL
;
1693 unsigned long excess
;
1694 unsigned long nr_scanned
;
1695 struct mem_cgroup_reclaim_cookie reclaim
= {
1699 excess
= soft_limit_excess(root_memcg
);
1702 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1707 * If we have not been able to reclaim
1708 * anything, it might because there are
1709 * no reclaimable pages under this hierarchy
1714 * We want to do more targeted reclaim.
1715 * excess >> 2 is not to excessive so as to
1716 * reclaim too much, nor too less that we keep
1717 * coming back to reclaim from this cgroup
1719 if (total
>= (excess
>> 2) ||
1720 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1725 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1726 pgdat
, &nr_scanned
);
1727 *total_scanned
+= nr_scanned
;
1728 if (!soft_limit_excess(root_memcg
))
1731 mem_cgroup_iter_break(root_memcg
, victim
);
1735 #ifdef CONFIG_LOCKDEP
1736 static struct lockdep_map memcg_oom_lock_dep_map
= {
1737 .name
= "memcg_oom_lock",
1741 static DEFINE_SPINLOCK(memcg_oom_lock
);
1744 * Check OOM-Killer is already running under our hierarchy.
1745 * If someone is running, return false.
1747 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1749 struct mem_cgroup
*iter
, *failed
= NULL
;
1751 spin_lock(&memcg_oom_lock
);
1753 for_each_mem_cgroup_tree(iter
, memcg
) {
1754 if (iter
->oom_lock
) {
1756 * this subtree of our hierarchy is already locked
1757 * so we cannot give a lock.
1760 mem_cgroup_iter_break(memcg
, iter
);
1763 iter
->oom_lock
= true;
1768 * OK, we failed to lock the whole subtree so we have
1769 * to clean up what we set up to the failing subtree
1771 for_each_mem_cgroup_tree(iter
, memcg
) {
1772 if (iter
== failed
) {
1773 mem_cgroup_iter_break(memcg
, iter
);
1776 iter
->oom_lock
= false;
1779 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1781 spin_unlock(&memcg_oom_lock
);
1786 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1788 struct mem_cgroup
*iter
;
1790 spin_lock(&memcg_oom_lock
);
1791 mutex_release(&memcg_oom_lock_dep_map
, _RET_IP_
);
1792 for_each_mem_cgroup_tree(iter
, memcg
)
1793 iter
->oom_lock
= false;
1794 spin_unlock(&memcg_oom_lock
);
1797 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1799 struct mem_cgroup
*iter
;
1801 spin_lock(&memcg_oom_lock
);
1802 for_each_mem_cgroup_tree(iter
, memcg
)
1804 spin_unlock(&memcg_oom_lock
);
1807 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1809 struct mem_cgroup
*iter
;
1812 * When a new child is created while the hierarchy is under oom,
1813 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1815 spin_lock(&memcg_oom_lock
);
1816 for_each_mem_cgroup_tree(iter
, memcg
)
1817 if (iter
->under_oom
> 0)
1819 spin_unlock(&memcg_oom_lock
);
1822 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1824 struct oom_wait_info
{
1825 struct mem_cgroup
*memcg
;
1826 wait_queue_entry_t wait
;
1829 static int memcg_oom_wake_function(wait_queue_entry_t
*wait
,
1830 unsigned mode
, int sync
, void *arg
)
1832 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1833 struct mem_cgroup
*oom_wait_memcg
;
1834 struct oom_wait_info
*oom_wait_info
;
1836 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1837 oom_wait_memcg
= oom_wait_info
->memcg
;
1839 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1840 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1842 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1845 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1848 * For the following lockless ->under_oom test, the only required
1849 * guarantee is that it must see the state asserted by an OOM when
1850 * this function is called as a result of userland actions
1851 * triggered by the notification of the OOM. This is trivially
1852 * achieved by invoking mem_cgroup_mark_under_oom() before
1853 * triggering notification.
1855 if (memcg
&& memcg
->under_oom
)
1856 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1866 static enum oom_status
mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1868 enum oom_status ret
;
1871 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1874 memcg_memory_event(memcg
, MEMCG_OOM
);
1877 * We are in the middle of the charge context here, so we
1878 * don't want to block when potentially sitting on a callstack
1879 * that holds all kinds of filesystem and mm locks.
1881 * cgroup1 allows disabling the OOM killer and waiting for outside
1882 * handling until the charge can succeed; remember the context and put
1883 * the task to sleep at the end of the page fault when all locks are
1886 * On the other hand, in-kernel OOM killer allows for an async victim
1887 * memory reclaim (oom_reaper) and that means that we are not solely
1888 * relying on the oom victim to make a forward progress and we can
1889 * invoke the oom killer here.
1891 * Please note that mem_cgroup_out_of_memory might fail to find a
1892 * victim and then we have to bail out from the charge path.
1894 if (memcg
->oom_kill_disable
) {
1895 if (!current
->in_user_fault
)
1897 css_get(&memcg
->css
);
1898 current
->memcg_in_oom
= memcg
;
1899 current
->memcg_oom_gfp_mask
= mask
;
1900 current
->memcg_oom_order
= order
;
1905 mem_cgroup_mark_under_oom(memcg
);
1907 locked
= mem_cgroup_oom_trylock(memcg
);
1910 mem_cgroup_oom_notify(memcg
);
1912 mem_cgroup_unmark_under_oom(memcg
);
1913 if (mem_cgroup_out_of_memory(memcg
, mask
, order
))
1919 mem_cgroup_oom_unlock(memcg
);
1925 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1926 * @handle: actually kill/wait or just clean up the OOM state
1928 * This has to be called at the end of a page fault if the memcg OOM
1929 * handler was enabled.
1931 * Memcg supports userspace OOM handling where failed allocations must
1932 * sleep on a waitqueue until the userspace task resolves the
1933 * situation. Sleeping directly in the charge context with all kinds
1934 * of locks held is not a good idea, instead we remember an OOM state
1935 * in the task and mem_cgroup_oom_synchronize() has to be called at
1936 * the end of the page fault to complete the OOM handling.
1938 * Returns %true if an ongoing memcg OOM situation was detected and
1939 * completed, %false otherwise.
1941 bool mem_cgroup_oom_synchronize(bool handle
)
1943 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1944 struct oom_wait_info owait
;
1947 /* OOM is global, do not handle */
1954 owait
.memcg
= memcg
;
1955 owait
.wait
.flags
= 0;
1956 owait
.wait
.func
= memcg_oom_wake_function
;
1957 owait
.wait
.private = current
;
1958 INIT_LIST_HEAD(&owait
.wait
.entry
);
1960 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1961 mem_cgroup_mark_under_oom(memcg
);
1963 locked
= mem_cgroup_oom_trylock(memcg
);
1966 mem_cgroup_oom_notify(memcg
);
1968 if (locked
&& !memcg
->oom_kill_disable
) {
1969 mem_cgroup_unmark_under_oom(memcg
);
1970 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1971 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1972 current
->memcg_oom_order
);
1975 mem_cgroup_unmark_under_oom(memcg
);
1976 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1980 mem_cgroup_oom_unlock(memcg
);
1982 * There is no guarantee that an OOM-lock contender
1983 * sees the wakeups triggered by the OOM kill
1984 * uncharges. Wake any sleepers explicitely.
1986 memcg_oom_recover(memcg
);
1989 current
->memcg_in_oom
= NULL
;
1990 css_put(&memcg
->css
);
1995 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1996 * @victim: task to be killed by the OOM killer
1997 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1999 * Returns a pointer to a memory cgroup, which has to be cleaned up
2000 * by killing all belonging OOM-killable tasks.
2002 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2004 struct mem_cgroup
*mem_cgroup_get_oom_group(struct task_struct
*victim
,
2005 struct mem_cgroup
*oom_domain
)
2007 struct mem_cgroup
*oom_group
= NULL
;
2008 struct mem_cgroup
*memcg
;
2010 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2014 oom_domain
= root_mem_cgroup
;
2018 memcg
= mem_cgroup_from_task(victim
);
2019 if (memcg
== root_mem_cgroup
)
2023 * If the victim task has been asynchronously moved to a different
2024 * memory cgroup, we might end up killing tasks outside oom_domain.
2025 * In this case it's better to ignore memory.group.oom.
2027 if (unlikely(!mem_cgroup_is_descendant(memcg
, oom_domain
)))
2031 * Traverse the memory cgroup hierarchy from the victim task's
2032 * cgroup up to the OOMing cgroup (or root) to find the
2033 * highest-level memory cgroup with oom.group set.
2035 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
2036 if (memcg
->oom_group
)
2039 if (memcg
== oom_domain
)
2044 css_get(&oom_group
->css
);
2051 void mem_cgroup_print_oom_group(struct mem_cgroup
*memcg
)
2053 pr_info("Tasks in ");
2054 pr_cont_cgroup_path(memcg
->css
.cgroup
);
2055 pr_cont(" are going to be killed due to memory.oom.group set\n");
2059 * lock_page_memcg - lock a page->mem_cgroup binding
2062 * This function protects unlocked LRU pages from being moved to
2065 * It ensures lifetime of the returned memcg. Caller is responsible
2066 * for the lifetime of the page; __unlock_page_memcg() is available
2067 * when @page might get freed inside the locked section.
2069 struct mem_cgroup
*lock_page_memcg(struct page
*page
)
2071 struct page
*head
= compound_head(page
); /* rmap on tail pages */
2072 struct mem_cgroup
*memcg
;
2073 unsigned long flags
;
2076 * The RCU lock is held throughout the transaction. The fast
2077 * path can get away without acquiring the memcg->move_lock
2078 * because page moving starts with an RCU grace period.
2080 * The RCU lock also protects the memcg from being freed when
2081 * the page state that is going to change is the only thing
2082 * preventing the page itself from being freed. E.g. writeback
2083 * doesn't hold a page reference and relies on PG_writeback to
2084 * keep off truncation, migration and so forth.
2088 if (mem_cgroup_disabled())
2091 memcg
= head
->mem_cgroup
;
2092 if (unlikely(!memcg
))
2095 if (atomic_read(&memcg
->moving_account
) <= 0)
2098 spin_lock_irqsave(&memcg
->move_lock
, flags
);
2099 if (memcg
!= head
->mem_cgroup
) {
2100 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2105 * When charge migration first begins, we can have locked and
2106 * unlocked page stat updates happening concurrently. Track
2107 * the task who has the lock for unlock_page_memcg().
2109 memcg
->move_lock_task
= current
;
2110 memcg
->move_lock_flags
= flags
;
2114 EXPORT_SYMBOL(lock_page_memcg
);
2117 * __unlock_page_memcg - unlock and unpin a memcg
2120 * Unlock and unpin a memcg returned by lock_page_memcg().
2122 void __unlock_page_memcg(struct mem_cgroup
*memcg
)
2124 if (memcg
&& memcg
->move_lock_task
== current
) {
2125 unsigned long flags
= memcg
->move_lock_flags
;
2127 memcg
->move_lock_task
= NULL
;
2128 memcg
->move_lock_flags
= 0;
2130 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2137 * unlock_page_memcg - unlock a page->mem_cgroup binding
2140 void unlock_page_memcg(struct page
*page
)
2142 struct page
*head
= compound_head(page
);
2144 __unlock_page_memcg(head
->mem_cgroup
);
2146 EXPORT_SYMBOL(unlock_page_memcg
);
2148 struct memcg_stock_pcp
{
2149 struct mem_cgroup
*cached
; /* this never be root cgroup */
2150 unsigned int nr_pages
;
2152 #ifdef CONFIG_MEMCG_KMEM
2153 struct obj_cgroup
*cached_objcg
;
2154 unsigned int nr_bytes
;
2157 struct work_struct work
;
2158 unsigned long flags
;
2159 #define FLUSHING_CACHED_CHARGE 0
2161 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2162 static DEFINE_MUTEX(percpu_charge_mutex
);
2164 #ifdef CONFIG_MEMCG_KMEM
2165 static void drain_obj_stock(struct memcg_stock_pcp
*stock
);
2166 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
2167 struct mem_cgroup
*root_memcg
);
2170 static inline void drain_obj_stock(struct memcg_stock_pcp
*stock
)
2173 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
2174 struct mem_cgroup
*root_memcg
)
2181 * consume_stock: Try to consume stocked charge on this cpu.
2182 * @memcg: memcg to consume from.
2183 * @nr_pages: how many pages to charge.
2185 * The charges will only happen if @memcg matches the current cpu's memcg
2186 * stock, and at least @nr_pages are available in that stock. Failure to
2187 * service an allocation will refill the stock.
2189 * returns true if successful, false otherwise.
2191 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2193 struct memcg_stock_pcp
*stock
;
2194 unsigned long flags
;
2197 if (nr_pages
> MEMCG_CHARGE_BATCH
)
2200 local_irq_save(flags
);
2202 stock
= this_cpu_ptr(&memcg_stock
);
2203 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
2204 stock
->nr_pages
-= nr_pages
;
2208 local_irq_restore(flags
);
2214 * Returns stocks cached in percpu and reset cached information.
2216 static void drain_stock(struct memcg_stock_pcp
*stock
)
2218 struct mem_cgroup
*old
= stock
->cached
;
2223 if (stock
->nr_pages
) {
2224 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2225 if (do_memsw_account())
2226 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2227 stock
->nr_pages
= 0;
2231 stock
->cached
= NULL
;
2234 static void drain_local_stock(struct work_struct
*dummy
)
2236 struct memcg_stock_pcp
*stock
;
2237 unsigned long flags
;
2240 * The only protection from memory hotplug vs. drain_stock races is
2241 * that we always operate on local CPU stock here with IRQ disabled
2243 local_irq_save(flags
);
2245 stock
= this_cpu_ptr(&memcg_stock
);
2246 drain_obj_stock(stock
);
2248 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2250 local_irq_restore(flags
);
2254 * Cache charges(val) to local per_cpu area.
2255 * This will be consumed by consume_stock() function, later.
2257 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2259 struct memcg_stock_pcp
*stock
;
2260 unsigned long flags
;
2262 local_irq_save(flags
);
2264 stock
= this_cpu_ptr(&memcg_stock
);
2265 if (stock
->cached
!= memcg
) { /* reset if necessary */
2267 css_get(&memcg
->css
);
2268 stock
->cached
= memcg
;
2270 stock
->nr_pages
+= nr_pages
;
2272 if (stock
->nr_pages
> MEMCG_CHARGE_BATCH
)
2275 local_irq_restore(flags
);
2279 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2280 * of the hierarchy under it.
2282 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2286 /* If someone's already draining, avoid adding running more workers. */
2287 if (!mutex_trylock(&percpu_charge_mutex
))
2290 * Notify other cpus that system-wide "drain" is running
2291 * We do not care about races with the cpu hotplug because cpu down
2292 * as well as workers from this path always operate on the local
2293 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2296 for_each_online_cpu(cpu
) {
2297 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2298 struct mem_cgroup
*memcg
;
2302 memcg
= stock
->cached
;
2303 if (memcg
&& stock
->nr_pages
&&
2304 mem_cgroup_is_descendant(memcg
, root_memcg
))
2306 if (obj_stock_flush_required(stock
, root_memcg
))
2311 !test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2313 drain_local_stock(&stock
->work
);
2315 schedule_work_on(cpu
, &stock
->work
);
2319 mutex_unlock(&percpu_charge_mutex
);
2322 static int memcg_hotplug_cpu_dead(unsigned int cpu
)
2324 struct memcg_stock_pcp
*stock
;
2325 struct mem_cgroup
*memcg
, *mi
;
2327 stock
= &per_cpu(memcg_stock
, cpu
);
2330 for_each_mem_cgroup(memcg
) {
2333 for (i
= 0; i
< MEMCG_NR_STAT
; i
++) {
2337 x
= this_cpu_xchg(memcg
->vmstats_percpu
->stat
[i
], 0);
2339 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2340 atomic_long_add(x
, &memcg
->vmstats
[i
]);
2342 if (i
>= NR_VM_NODE_STAT_ITEMS
)
2345 for_each_node(nid
) {
2346 struct mem_cgroup_per_node
*pn
;
2348 pn
= mem_cgroup_nodeinfo(memcg
, nid
);
2349 x
= this_cpu_xchg(pn
->lruvec_stat_cpu
->count
[i
], 0);
2352 atomic_long_add(x
, &pn
->lruvec_stat
[i
]);
2353 } while ((pn
= parent_nodeinfo(pn
, nid
)));
2357 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++) {
2360 x
= this_cpu_xchg(memcg
->vmstats_percpu
->events
[i
], 0);
2362 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2363 atomic_long_add(x
, &memcg
->vmevents
[i
]);
2370 static unsigned long reclaim_high(struct mem_cgroup
*memcg
,
2371 unsigned int nr_pages
,
2374 unsigned long nr_reclaimed
= 0;
2377 if (page_counter_read(&memcg
->memory
) <=
2378 READ_ONCE(memcg
->memory
.high
))
2380 memcg_memory_event(memcg
, MEMCG_HIGH
);
2381 nr_reclaimed
+= try_to_free_mem_cgroup_pages(memcg
, nr_pages
,
2383 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2384 !mem_cgroup_is_root(memcg
));
2386 return nr_reclaimed
;
2389 static void high_work_func(struct work_struct
*work
)
2391 struct mem_cgroup
*memcg
;
2393 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
2394 reclaim_high(memcg
, MEMCG_CHARGE_BATCH
, GFP_KERNEL
);
2398 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2399 * enough to still cause a significant slowdown in most cases, while still
2400 * allowing diagnostics and tracing to proceed without becoming stuck.
2402 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2405 * When calculating the delay, we use these either side of the exponentiation to
2406 * maintain precision and scale to a reasonable number of jiffies (see the table
2409 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2410 * overage ratio to a delay.
2411 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2412 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2413 * to produce a reasonable delay curve.
2415 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2416 * reasonable delay curve compared to precision-adjusted overage, not
2417 * penalising heavily at first, but still making sure that growth beyond the
2418 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2419 * example, with a high of 100 megabytes:
2421 * +-------+------------------------+
2422 * | usage | time to allocate in ms |
2423 * +-------+------------------------+
2445 * +-------+------------------------+
2447 #define MEMCG_DELAY_PRECISION_SHIFT 20
2448 #define MEMCG_DELAY_SCALING_SHIFT 14
2450 static u64
calculate_overage(unsigned long usage
, unsigned long high
)
2458 * Prevent division by 0 in overage calculation by acting as if
2459 * it was a threshold of 1 page
2461 high
= max(high
, 1UL);
2463 overage
= usage
- high
;
2464 overage
<<= MEMCG_DELAY_PRECISION_SHIFT
;
2465 return div64_u64(overage
, high
);
2468 static u64
mem_find_max_overage(struct mem_cgroup
*memcg
)
2470 u64 overage
, max_overage
= 0;
2473 overage
= calculate_overage(page_counter_read(&memcg
->memory
),
2474 READ_ONCE(memcg
->memory
.high
));
2475 max_overage
= max(overage
, max_overage
);
2476 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2477 !mem_cgroup_is_root(memcg
));
2482 static u64
swap_find_max_overage(struct mem_cgroup
*memcg
)
2484 u64 overage
, max_overage
= 0;
2487 overage
= calculate_overage(page_counter_read(&memcg
->swap
),
2488 READ_ONCE(memcg
->swap
.high
));
2490 memcg_memory_event(memcg
, MEMCG_SWAP_HIGH
);
2491 max_overage
= max(overage
, max_overage
);
2492 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2493 !mem_cgroup_is_root(memcg
));
2499 * Get the number of jiffies that we should penalise a mischievous cgroup which
2500 * is exceeding its memory.high by checking both it and its ancestors.
2502 static unsigned long calculate_high_delay(struct mem_cgroup
*memcg
,
2503 unsigned int nr_pages
,
2506 unsigned long penalty_jiffies
;
2512 * We use overage compared to memory.high to calculate the number of
2513 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2514 * fairly lenient on small overages, and increasingly harsh when the
2515 * memcg in question makes it clear that it has no intention of stopping
2516 * its crazy behaviour, so we exponentially increase the delay based on
2519 penalty_jiffies
= max_overage
* max_overage
* HZ
;
2520 penalty_jiffies
>>= MEMCG_DELAY_PRECISION_SHIFT
;
2521 penalty_jiffies
>>= MEMCG_DELAY_SCALING_SHIFT
;
2524 * Factor in the task's own contribution to the overage, such that four
2525 * N-sized allocations are throttled approximately the same as one
2526 * 4N-sized allocation.
2528 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2529 * larger the current charge patch is than that.
2531 return penalty_jiffies
* nr_pages
/ MEMCG_CHARGE_BATCH
;
2535 * Scheduled by try_charge() to be executed from the userland return path
2536 * and reclaims memory over the high limit.
2538 void mem_cgroup_handle_over_high(void)
2540 unsigned long penalty_jiffies
;
2541 unsigned long pflags
;
2542 unsigned long nr_reclaimed
;
2543 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
2544 int nr_retries
= MAX_RECLAIM_RETRIES
;
2545 struct mem_cgroup
*memcg
;
2546 bool in_retry
= false;
2548 if (likely(!nr_pages
))
2551 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2552 current
->memcg_nr_pages_over_high
= 0;
2556 * The allocating task should reclaim at least the batch size, but for
2557 * subsequent retries we only want to do what's necessary to prevent oom
2558 * or breaching resource isolation.
2560 * This is distinct from memory.max or page allocator behaviour because
2561 * memory.high is currently batched, whereas memory.max and the page
2562 * allocator run every time an allocation is made.
2564 nr_reclaimed
= reclaim_high(memcg
,
2565 in_retry
? SWAP_CLUSTER_MAX
: nr_pages
,
2569 * memory.high is breached and reclaim is unable to keep up. Throttle
2570 * allocators proactively to slow down excessive growth.
2572 penalty_jiffies
= calculate_high_delay(memcg
, nr_pages
,
2573 mem_find_max_overage(memcg
));
2575 penalty_jiffies
+= calculate_high_delay(memcg
, nr_pages
,
2576 swap_find_max_overage(memcg
));
2579 * Clamp the max delay per usermode return so as to still keep the
2580 * application moving forwards and also permit diagnostics, albeit
2583 penalty_jiffies
= min(penalty_jiffies
, MEMCG_MAX_HIGH_DELAY_JIFFIES
);
2586 * Don't sleep if the amount of jiffies this memcg owes us is so low
2587 * that it's not even worth doing, in an attempt to be nice to those who
2588 * go only a small amount over their memory.high value and maybe haven't
2589 * been aggressively reclaimed enough yet.
2591 if (penalty_jiffies
<= HZ
/ 100)
2595 * If reclaim is making forward progress but we're still over
2596 * memory.high, we want to encourage that rather than doing allocator
2599 if (nr_reclaimed
|| nr_retries
--) {
2605 * If we exit early, we're guaranteed to die (since
2606 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2607 * need to account for any ill-begotten jiffies to pay them off later.
2609 psi_memstall_enter(&pflags
);
2610 schedule_timeout_killable(penalty_jiffies
);
2611 psi_memstall_leave(&pflags
);
2614 css_put(&memcg
->css
);
2617 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2618 unsigned int nr_pages
)
2620 unsigned int batch
= max(MEMCG_CHARGE_BATCH
, nr_pages
);
2621 int nr_retries
= MAX_RECLAIM_RETRIES
;
2622 struct mem_cgroup
*mem_over_limit
;
2623 struct page_counter
*counter
;
2624 unsigned long nr_reclaimed
;
2625 bool may_swap
= true;
2626 bool drained
= false;
2627 enum oom_status oom_status
;
2629 if (mem_cgroup_is_root(memcg
))
2632 if (consume_stock(memcg
, nr_pages
))
2635 if (!do_memsw_account() ||
2636 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2637 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2639 if (do_memsw_account())
2640 page_counter_uncharge(&memcg
->memsw
, batch
);
2641 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2643 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2647 if (batch
> nr_pages
) {
2653 * Memcg doesn't have a dedicated reserve for atomic
2654 * allocations. But like the global atomic pool, we need to
2655 * put the burden of reclaim on regular allocation requests
2656 * and let these go through as privileged allocations.
2658 if (gfp_mask
& __GFP_ATOMIC
)
2662 * Unlike in global OOM situations, memcg is not in a physical
2663 * memory shortage. Allow dying and OOM-killed tasks to
2664 * bypass the last charges so that they can exit quickly and
2665 * free their memory.
2667 if (unlikely(should_force_charge()))
2671 * Prevent unbounded recursion when reclaim operations need to
2672 * allocate memory. This might exceed the limits temporarily,
2673 * but we prefer facilitating memory reclaim and getting back
2674 * under the limit over triggering OOM kills in these cases.
2676 if (unlikely(current
->flags
& PF_MEMALLOC
))
2679 if (unlikely(task_in_memcg_oom(current
)))
2682 if (!gfpflags_allow_blocking(gfp_mask
))
2685 memcg_memory_event(mem_over_limit
, MEMCG_MAX
);
2687 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2688 gfp_mask
, may_swap
);
2690 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2694 drain_all_stock(mem_over_limit
);
2699 if (gfp_mask
& __GFP_NORETRY
)
2702 * Even though the limit is exceeded at this point, reclaim
2703 * may have been able to free some pages. Retry the charge
2704 * before killing the task.
2706 * Only for regular pages, though: huge pages are rather
2707 * unlikely to succeed so close to the limit, and we fall back
2708 * to regular pages anyway in case of failure.
2710 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2713 * At task move, charge accounts can be doubly counted. So, it's
2714 * better to wait until the end of task_move if something is going on.
2716 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2722 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
2725 if (gfp_mask
& __GFP_NOFAIL
)
2728 if (fatal_signal_pending(current
))
2732 * keep retrying as long as the memcg oom killer is able to make
2733 * a forward progress or bypass the charge if the oom killer
2734 * couldn't make any progress.
2736 oom_status
= mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2737 get_order(nr_pages
* PAGE_SIZE
));
2738 switch (oom_status
) {
2740 nr_retries
= MAX_RECLAIM_RETRIES
;
2748 if (!(gfp_mask
& __GFP_NOFAIL
))
2752 * The allocation either can't fail or will lead to more memory
2753 * being freed very soon. Allow memory usage go over the limit
2754 * temporarily by force charging it.
2756 page_counter_charge(&memcg
->memory
, nr_pages
);
2757 if (do_memsw_account())
2758 page_counter_charge(&memcg
->memsw
, nr_pages
);
2763 if (batch
> nr_pages
)
2764 refill_stock(memcg
, batch
- nr_pages
);
2767 * If the hierarchy is above the normal consumption range, schedule
2768 * reclaim on returning to userland. We can perform reclaim here
2769 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2770 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2771 * not recorded as it most likely matches current's and won't
2772 * change in the meantime. As high limit is checked again before
2773 * reclaim, the cost of mismatch is negligible.
2776 bool mem_high
, swap_high
;
2778 mem_high
= page_counter_read(&memcg
->memory
) >
2779 READ_ONCE(memcg
->memory
.high
);
2780 swap_high
= page_counter_read(&memcg
->swap
) >
2781 READ_ONCE(memcg
->swap
.high
);
2783 /* Don't bother a random interrupted task */
2784 if (in_interrupt()) {
2786 schedule_work(&memcg
->high_work
);
2792 if (mem_high
|| swap_high
) {
2794 * The allocating tasks in this cgroup will need to do
2795 * reclaim or be throttled to prevent further growth
2796 * of the memory or swap footprints.
2798 * Target some best-effort fairness between the tasks,
2799 * and distribute reclaim work and delay penalties
2800 * based on how much each task is actually allocating.
2802 current
->memcg_nr_pages_over_high
+= batch
;
2803 set_notify_resume(current
);
2806 } while ((memcg
= parent_mem_cgroup(memcg
)));
2811 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2812 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2814 if (mem_cgroup_is_root(memcg
))
2817 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2818 if (do_memsw_account())
2819 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2823 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
)
2825 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2827 * Any of the following ensures page->mem_cgroup stability:
2831 * - lock_page_memcg()
2832 * - exclusive reference
2834 page
->mem_cgroup
= memcg
;
2837 #ifdef CONFIG_MEMCG_KMEM
2838 int memcg_alloc_page_obj_cgroups(struct page
*page
, struct kmem_cache
*s
,
2841 unsigned int objects
= objs_per_slab_page(s
, page
);
2844 vec
= kcalloc_node(objects
, sizeof(struct obj_cgroup
*), gfp
,
2849 if (cmpxchg(&page
->obj_cgroups
, NULL
,
2850 (struct obj_cgroup
**) ((unsigned long)vec
| 0x1UL
)))
2853 kmemleak_not_leak(vec
);
2859 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2861 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2862 * cgroup_mutex, etc.
2864 struct mem_cgroup
*mem_cgroup_from_obj(void *p
)
2868 if (mem_cgroup_disabled())
2871 page
= virt_to_head_page(p
);
2874 * Slab objects are accounted individually, not per-page.
2875 * Memcg membership data for each individual object is saved in
2876 * the page->obj_cgroups.
2878 if (page_has_obj_cgroups(page
)) {
2879 struct obj_cgroup
*objcg
;
2882 off
= obj_to_index(page
->slab_cache
, page
, p
);
2883 objcg
= page_obj_cgroups(page
)[off
];
2885 return obj_cgroup_memcg(objcg
);
2890 /* All other pages use page->mem_cgroup */
2891 return page
->mem_cgroup
;
2894 __always_inline
struct obj_cgroup
*get_obj_cgroup_from_current(void)
2896 struct obj_cgroup
*objcg
= NULL
;
2897 struct mem_cgroup
*memcg
;
2899 if (unlikely(!current
->mm
&& !current
->active_memcg
))
2903 if (unlikely(current
->active_memcg
))
2904 memcg
= rcu_dereference(current
->active_memcg
);
2906 memcg
= mem_cgroup_from_task(current
);
2908 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
)) {
2909 objcg
= rcu_dereference(memcg
->objcg
);
2910 if (objcg
&& obj_cgroup_tryget(objcg
))
2918 static int memcg_alloc_cache_id(void)
2923 id
= ida_simple_get(&memcg_cache_ida
,
2924 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2928 if (id
< memcg_nr_cache_ids
)
2932 * There's no space for the new id in memcg_caches arrays,
2933 * so we have to grow them.
2935 down_write(&memcg_cache_ids_sem
);
2937 size
= 2 * (id
+ 1);
2938 if (size
< MEMCG_CACHES_MIN_SIZE
)
2939 size
= MEMCG_CACHES_MIN_SIZE
;
2940 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2941 size
= MEMCG_CACHES_MAX_SIZE
;
2943 err
= memcg_update_all_list_lrus(size
);
2945 memcg_nr_cache_ids
= size
;
2947 up_write(&memcg_cache_ids_sem
);
2950 ida_simple_remove(&memcg_cache_ida
, id
);
2956 static void memcg_free_cache_id(int id
)
2958 ida_simple_remove(&memcg_cache_ida
, id
);
2962 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
2963 * @memcg: memory cgroup to charge
2964 * @gfp: reclaim mode
2965 * @nr_pages: number of pages to charge
2967 * Returns 0 on success, an error code on failure.
2969 int __memcg_kmem_charge(struct mem_cgroup
*memcg
, gfp_t gfp
,
2970 unsigned int nr_pages
)
2972 struct page_counter
*counter
;
2975 ret
= try_charge(memcg
, gfp
, nr_pages
);
2979 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2980 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2983 * Enforce __GFP_NOFAIL allocation because callers are not
2984 * prepared to see failures and likely do not have any failure
2987 if (gfp
& __GFP_NOFAIL
) {
2988 page_counter_charge(&memcg
->kmem
, nr_pages
);
2991 cancel_charge(memcg
, nr_pages
);
2998 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
2999 * @memcg: memcg to uncharge
3000 * @nr_pages: number of pages to uncharge
3002 void __memcg_kmem_uncharge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
3004 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
3005 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
3007 page_counter_uncharge(&memcg
->memory
, nr_pages
);
3008 if (do_memsw_account())
3009 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
3013 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3014 * @page: page to charge
3015 * @gfp: reclaim mode
3016 * @order: allocation order
3018 * Returns 0 on success, an error code on failure.
3020 int __memcg_kmem_charge_page(struct page
*page
, gfp_t gfp
, int order
)
3022 struct mem_cgroup
*memcg
;
3025 if (memcg_kmem_bypass())
3028 memcg
= get_mem_cgroup_from_current();
3029 if (!mem_cgroup_is_root(memcg
)) {
3030 ret
= __memcg_kmem_charge(memcg
, gfp
, 1 << order
);
3032 page
->mem_cgroup
= memcg
;
3033 __SetPageKmemcg(page
);
3037 css_put(&memcg
->css
);
3042 * __memcg_kmem_uncharge_page: uncharge a kmem page
3043 * @page: page to uncharge
3044 * @order: allocation order
3046 void __memcg_kmem_uncharge_page(struct page
*page
, int order
)
3048 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
3049 unsigned int nr_pages
= 1 << order
;
3054 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
3055 __memcg_kmem_uncharge(memcg
, nr_pages
);
3056 page
->mem_cgroup
= NULL
;
3057 css_put(&memcg
->css
);
3059 /* slab pages do not have PageKmemcg flag set */
3060 if (PageKmemcg(page
))
3061 __ClearPageKmemcg(page
);
3064 static bool consume_obj_stock(struct obj_cgroup
*objcg
, unsigned int nr_bytes
)
3066 struct memcg_stock_pcp
*stock
;
3067 unsigned long flags
;
3070 local_irq_save(flags
);
3072 stock
= this_cpu_ptr(&memcg_stock
);
3073 if (objcg
== stock
->cached_objcg
&& stock
->nr_bytes
>= nr_bytes
) {
3074 stock
->nr_bytes
-= nr_bytes
;
3078 local_irq_restore(flags
);
3083 static void drain_obj_stock(struct memcg_stock_pcp
*stock
)
3085 struct obj_cgroup
*old
= stock
->cached_objcg
;
3090 if (stock
->nr_bytes
) {
3091 unsigned int nr_pages
= stock
->nr_bytes
>> PAGE_SHIFT
;
3092 unsigned int nr_bytes
= stock
->nr_bytes
& (PAGE_SIZE
- 1);
3096 __memcg_kmem_uncharge(obj_cgroup_memcg(old
), nr_pages
);
3101 * The leftover is flushed to the centralized per-memcg value.
3102 * On the next attempt to refill obj stock it will be moved
3103 * to a per-cpu stock (probably, on an other CPU), see
3104 * refill_obj_stock().
3106 * How often it's flushed is a trade-off between the memory
3107 * limit enforcement accuracy and potential CPU contention,
3108 * so it might be changed in the future.
3110 atomic_add(nr_bytes
, &old
->nr_charged_bytes
);
3111 stock
->nr_bytes
= 0;
3114 obj_cgroup_put(old
);
3115 stock
->cached_objcg
= NULL
;
3118 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
3119 struct mem_cgroup
*root_memcg
)
3121 struct mem_cgroup
*memcg
;
3123 if (stock
->cached_objcg
) {
3124 memcg
= obj_cgroup_memcg(stock
->cached_objcg
);
3125 if (memcg
&& mem_cgroup_is_descendant(memcg
, root_memcg
))
3132 static void refill_obj_stock(struct obj_cgroup
*objcg
, unsigned int nr_bytes
)
3134 struct memcg_stock_pcp
*stock
;
3135 unsigned long flags
;
3137 local_irq_save(flags
);
3139 stock
= this_cpu_ptr(&memcg_stock
);
3140 if (stock
->cached_objcg
!= objcg
) { /* reset if necessary */
3141 drain_obj_stock(stock
);
3142 obj_cgroup_get(objcg
);
3143 stock
->cached_objcg
= objcg
;
3144 stock
->nr_bytes
= atomic_xchg(&objcg
->nr_charged_bytes
, 0);
3146 stock
->nr_bytes
+= nr_bytes
;
3148 if (stock
->nr_bytes
> PAGE_SIZE
)
3149 drain_obj_stock(stock
);
3151 local_irq_restore(flags
);
3154 int obj_cgroup_charge(struct obj_cgroup
*objcg
, gfp_t gfp
, size_t size
)
3156 struct mem_cgroup
*memcg
;
3157 unsigned int nr_pages
, nr_bytes
;
3160 if (consume_obj_stock(objcg
, size
))
3164 * In theory, memcg->nr_charged_bytes can have enough
3165 * pre-charged bytes to satisfy the allocation. However,
3166 * flushing memcg->nr_charged_bytes requires two atomic
3167 * operations, and memcg->nr_charged_bytes can't be big,
3168 * so it's better to ignore it and try grab some new pages.
3169 * memcg->nr_charged_bytes will be flushed in
3170 * refill_obj_stock(), called from this function or
3171 * independently later.
3174 memcg
= obj_cgroup_memcg(objcg
);
3175 css_get(&memcg
->css
);
3178 nr_pages
= size
>> PAGE_SHIFT
;
3179 nr_bytes
= size
& (PAGE_SIZE
- 1);
3184 ret
= __memcg_kmem_charge(memcg
, gfp
, nr_pages
);
3185 if (!ret
&& nr_bytes
)
3186 refill_obj_stock(objcg
, PAGE_SIZE
- nr_bytes
);
3188 css_put(&memcg
->css
);
3192 void obj_cgroup_uncharge(struct obj_cgroup
*objcg
, size_t size
)
3194 refill_obj_stock(objcg
, size
);
3197 #endif /* CONFIG_MEMCG_KMEM */
3199 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3202 * Because tail pages are not marked as "used", set it. We're under
3203 * pgdat->lru_lock and migration entries setup in all page mappings.
3205 void mem_cgroup_split_huge_fixup(struct page
*head
)
3207 struct mem_cgroup
*memcg
= head
->mem_cgroup
;
3210 if (mem_cgroup_disabled())
3213 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3214 css_get(&memcg
->css
);
3215 head
[i
].mem_cgroup
= memcg
;
3218 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3220 #ifdef CONFIG_MEMCG_SWAP
3222 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3223 * @entry: swap entry to be moved
3224 * @from: mem_cgroup which the entry is moved from
3225 * @to: mem_cgroup which the entry is moved to
3227 * It succeeds only when the swap_cgroup's record for this entry is the same
3228 * as the mem_cgroup's id of @from.
3230 * Returns 0 on success, -EINVAL on failure.
3232 * The caller must have charged to @to, IOW, called page_counter_charge() about
3233 * both res and memsw, and called css_get().
3235 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
3236 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3238 unsigned short old_id
, new_id
;
3240 old_id
= mem_cgroup_id(from
);
3241 new_id
= mem_cgroup_id(to
);
3243 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
3244 mod_memcg_state(from
, MEMCG_SWAP
, -1);
3245 mod_memcg_state(to
, MEMCG_SWAP
, 1);
3251 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
3252 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3258 static DEFINE_MUTEX(memcg_max_mutex
);
3260 static int mem_cgroup_resize_max(struct mem_cgroup
*memcg
,
3261 unsigned long max
, bool memsw
)
3263 bool enlarge
= false;
3264 bool drained
= false;
3266 bool limits_invariant
;
3267 struct page_counter
*counter
= memsw
? &memcg
->memsw
: &memcg
->memory
;
3270 if (signal_pending(current
)) {
3275 mutex_lock(&memcg_max_mutex
);
3277 * Make sure that the new limit (memsw or memory limit) doesn't
3278 * break our basic invariant rule memory.max <= memsw.max.
3280 limits_invariant
= memsw
? max
>= READ_ONCE(memcg
->memory
.max
) :
3281 max
<= memcg
->memsw
.max
;
3282 if (!limits_invariant
) {
3283 mutex_unlock(&memcg_max_mutex
);
3287 if (max
> counter
->max
)
3289 ret
= page_counter_set_max(counter
, max
);
3290 mutex_unlock(&memcg_max_mutex
);
3296 drain_all_stock(memcg
);
3301 if (!try_to_free_mem_cgroup_pages(memcg
, 1,
3302 GFP_KERNEL
, !memsw
)) {
3308 if (!ret
&& enlarge
)
3309 memcg_oom_recover(memcg
);
3314 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
3316 unsigned long *total_scanned
)
3318 unsigned long nr_reclaimed
= 0;
3319 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
3320 unsigned long reclaimed
;
3322 struct mem_cgroup_tree_per_node
*mctz
;
3323 unsigned long excess
;
3324 unsigned long nr_scanned
;
3329 mctz
= soft_limit_tree_node(pgdat
->node_id
);
3332 * Do not even bother to check the largest node if the root
3333 * is empty. Do it lockless to prevent lock bouncing. Races
3334 * are acceptable as soft limit is best effort anyway.
3336 if (!mctz
|| RB_EMPTY_ROOT(&mctz
->rb_root
))
3340 * This loop can run a while, specially if mem_cgroup's continuously
3341 * keep exceeding their soft limit and putting the system under
3348 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3353 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
3354 gfp_mask
, &nr_scanned
);
3355 nr_reclaimed
+= reclaimed
;
3356 *total_scanned
+= nr_scanned
;
3357 spin_lock_irq(&mctz
->lock
);
3358 __mem_cgroup_remove_exceeded(mz
, mctz
);
3361 * If we failed to reclaim anything from this memory cgroup
3362 * it is time to move on to the next cgroup
3366 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
3368 excess
= soft_limit_excess(mz
->memcg
);
3370 * One school of thought says that we should not add
3371 * back the node to the tree if reclaim returns 0.
3372 * But our reclaim could return 0, simply because due
3373 * to priority we are exposing a smaller subset of
3374 * memory to reclaim from. Consider this as a longer
3377 /* If excess == 0, no tree ops */
3378 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
3379 spin_unlock_irq(&mctz
->lock
);
3380 css_put(&mz
->memcg
->css
);
3383 * Could not reclaim anything and there are no more
3384 * mem cgroups to try or we seem to be looping without
3385 * reclaiming anything.
3387 if (!nr_reclaimed
&&
3389 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3391 } while (!nr_reclaimed
);
3393 css_put(&next_mz
->memcg
->css
);
3394 return nr_reclaimed
;
3398 * Test whether @memcg has children, dead or alive. Note that this
3399 * function doesn't care whether @memcg has use_hierarchy enabled and
3400 * returns %true if there are child csses according to the cgroup
3401 * hierarchy. Testing use_hierarchy is the caller's responsibility.
3403 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
3408 ret
= css_next_child(NULL
, &memcg
->css
);
3414 * Reclaims as many pages from the given memcg as possible.
3416 * Caller is responsible for holding css reference for memcg.
3418 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
3420 int nr_retries
= MAX_RECLAIM_RETRIES
;
3422 /* we call try-to-free pages for make this cgroup empty */
3423 lru_add_drain_all();
3425 drain_all_stock(memcg
);
3427 /* try to free all pages in this cgroup */
3428 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
3431 if (signal_pending(current
))
3434 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
3438 /* maybe some writeback is necessary */
3439 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3447 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3448 char *buf
, size_t nbytes
,
3451 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3453 if (mem_cgroup_is_root(memcg
))
3455 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3458 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3461 return mem_cgroup_from_css(css
)->use_hierarchy
;
3464 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3465 struct cftype
*cft
, u64 val
)
3468 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3469 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
3471 if (memcg
->use_hierarchy
== val
)
3475 * If parent's use_hierarchy is set, we can't make any modifications
3476 * in the child subtrees. If it is unset, then the change can
3477 * occur, provided the current cgroup has no children.
3479 * For the root cgroup, parent_mem is NULL, we allow value to be
3480 * set if there are no children.
3482 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
3483 (val
== 1 || val
== 0)) {
3484 if (!memcg_has_children(memcg
))
3485 memcg
->use_hierarchy
= val
;
3494 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3498 if (mem_cgroup_is_root(memcg
)) {
3499 val
= memcg_page_state(memcg
, NR_FILE_PAGES
) +
3500 memcg_page_state(memcg
, NR_ANON_MAPPED
);
3502 val
+= memcg_page_state(memcg
, MEMCG_SWAP
);
3505 val
= page_counter_read(&memcg
->memory
);
3507 val
= page_counter_read(&memcg
->memsw
);
3520 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3523 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3524 struct page_counter
*counter
;
3526 switch (MEMFILE_TYPE(cft
->private)) {
3528 counter
= &memcg
->memory
;
3531 counter
= &memcg
->memsw
;
3534 counter
= &memcg
->kmem
;
3537 counter
= &memcg
->tcpmem
;
3543 switch (MEMFILE_ATTR(cft
->private)) {
3545 if (counter
== &memcg
->memory
)
3546 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
3547 if (counter
== &memcg
->memsw
)
3548 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
3549 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3551 return (u64
)counter
->max
* PAGE_SIZE
;
3553 return (u64
)counter
->watermark
* PAGE_SIZE
;
3555 return counter
->failcnt
;
3556 case RES_SOFT_LIMIT
:
3557 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
3563 static void memcg_flush_percpu_vmstats(struct mem_cgroup
*memcg
)
3565 unsigned long stat
[MEMCG_NR_STAT
] = {0};
3566 struct mem_cgroup
*mi
;
3569 for_each_online_cpu(cpu
)
3570 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
3571 stat
[i
] += per_cpu(memcg
->vmstats_percpu
->stat
[i
], cpu
);
3573 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
3574 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
3575 atomic_long_add(stat
[i
], &mi
->vmstats
[i
]);
3577 for_each_node(node
) {
3578 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
3579 struct mem_cgroup_per_node
*pi
;
3581 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3584 for_each_online_cpu(cpu
)
3585 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3587 pn
->lruvec_stat_cpu
->count
[i
], cpu
);
3589 for (pi
= pn
; pi
; pi
= parent_nodeinfo(pi
, node
))
3590 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3591 atomic_long_add(stat
[i
], &pi
->lruvec_stat
[i
]);
3595 static void memcg_flush_percpu_vmevents(struct mem_cgroup
*memcg
)
3597 unsigned long events
[NR_VM_EVENT_ITEMS
];
3598 struct mem_cgroup
*mi
;
3601 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3604 for_each_online_cpu(cpu
)
3605 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3606 events
[i
] += per_cpu(memcg
->vmstats_percpu
->events
[i
],
3609 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
3610 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3611 atomic_long_add(events
[i
], &mi
->vmevents
[i
]);
3614 #ifdef CONFIG_MEMCG_KMEM
3615 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3617 struct obj_cgroup
*objcg
;
3620 if (cgroup_memory_nokmem
)
3623 BUG_ON(memcg
->kmemcg_id
>= 0);
3624 BUG_ON(memcg
->kmem_state
);
3626 memcg_id
= memcg_alloc_cache_id();
3630 objcg
= obj_cgroup_alloc();
3632 memcg_free_cache_id(memcg_id
);
3635 objcg
->memcg
= memcg
;
3636 rcu_assign_pointer(memcg
->objcg
, objcg
);
3638 static_branch_enable(&memcg_kmem_enabled_key
);
3641 * A memory cgroup is considered kmem-online as soon as it gets
3642 * kmemcg_id. Setting the id after enabling static branching will
3643 * guarantee no one starts accounting before all call sites are
3646 memcg
->kmemcg_id
= memcg_id
;
3647 memcg
->kmem_state
= KMEM_ONLINE
;
3652 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3654 struct cgroup_subsys_state
*css
;
3655 struct mem_cgroup
*parent
, *child
;
3658 if (memcg
->kmem_state
!= KMEM_ONLINE
)
3661 memcg
->kmem_state
= KMEM_ALLOCATED
;
3663 parent
= parent_mem_cgroup(memcg
);
3665 parent
= root_mem_cgroup
;
3667 memcg_reparent_objcgs(memcg
, parent
);
3669 kmemcg_id
= memcg
->kmemcg_id
;
3670 BUG_ON(kmemcg_id
< 0);
3673 * Change kmemcg_id of this cgroup and all its descendants to the
3674 * parent's id, and then move all entries from this cgroup's list_lrus
3675 * to ones of the parent. After we have finished, all list_lrus
3676 * corresponding to this cgroup are guaranteed to remain empty. The
3677 * ordering is imposed by list_lru_node->lock taken by
3678 * memcg_drain_all_list_lrus().
3680 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3681 css_for_each_descendant_pre(css
, &memcg
->css
) {
3682 child
= mem_cgroup_from_css(css
);
3683 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
3684 child
->kmemcg_id
= parent
->kmemcg_id
;
3685 if (!memcg
->use_hierarchy
)
3690 memcg_drain_all_list_lrus(kmemcg_id
, parent
);
3692 memcg_free_cache_id(kmemcg_id
);
3695 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3697 /* css_alloc() failed, offlining didn't happen */
3698 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
3699 memcg_offline_kmem(memcg
);
3702 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3706 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3709 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3712 #endif /* CONFIG_MEMCG_KMEM */
3714 static int memcg_update_kmem_max(struct mem_cgroup
*memcg
,
3719 mutex_lock(&memcg_max_mutex
);
3720 ret
= page_counter_set_max(&memcg
->kmem
, max
);
3721 mutex_unlock(&memcg_max_mutex
);
3725 static int memcg_update_tcp_max(struct mem_cgroup
*memcg
, unsigned long max
)
3729 mutex_lock(&memcg_max_mutex
);
3731 ret
= page_counter_set_max(&memcg
->tcpmem
, max
);
3735 if (!memcg
->tcpmem_active
) {
3737 * The active flag needs to be written after the static_key
3738 * update. This is what guarantees that the socket activation
3739 * function is the last one to run. See mem_cgroup_sk_alloc()
3740 * for details, and note that we don't mark any socket as
3741 * belonging to this memcg until that flag is up.
3743 * We need to do this, because static_keys will span multiple
3744 * sites, but we can't control their order. If we mark a socket
3745 * as accounted, but the accounting functions are not patched in
3746 * yet, we'll lose accounting.
3748 * We never race with the readers in mem_cgroup_sk_alloc(),
3749 * because when this value change, the code to process it is not
3752 static_branch_inc(&memcg_sockets_enabled_key
);
3753 memcg
->tcpmem_active
= true;
3756 mutex_unlock(&memcg_max_mutex
);
3761 * The user of this function is...
3764 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3765 char *buf
, size_t nbytes
, loff_t off
)
3767 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3768 unsigned long nr_pages
;
3771 buf
= strstrip(buf
);
3772 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3776 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3778 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3782 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3784 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, false);
3787 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, true);
3790 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3791 "Please report your usecase to linux-mm@kvack.org if you "
3792 "depend on this functionality.\n");
3793 ret
= memcg_update_kmem_max(memcg
, nr_pages
);
3796 ret
= memcg_update_tcp_max(memcg
, nr_pages
);
3800 case RES_SOFT_LIMIT
:
3801 memcg
->soft_limit
= nr_pages
;
3805 return ret
?: nbytes
;
3808 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3809 size_t nbytes
, loff_t off
)
3811 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3812 struct page_counter
*counter
;
3814 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3816 counter
= &memcg
->memory
;
3819 counter
= &memcg
->memsw
;
3822 counter
= &memcg
->kmem
;
3825 counter
= &memcg
->tcpmem
;
3831 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3833 page_counter_reset_watermark(counter
);
3836 counter
->failcnt
= 0;
3845 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3848 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3852 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3853 struct cftype
*cft
, u64 val
)
3855 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3857 if (val
& ~MOVE_MASK
)
3861 * No kind of locking is needed in here, because ->can_attach() will
3862 * check this value once in the beginning of the process, and then carry
3863 * on with stale data. This means that changes to this value will only
3864 * affect task migrations starting after the change.
3866 memcg
->move_charge_at_immigrate
= val
;
3870 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3871 struct cftype
*cft
, u64 val
)
3879 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3880 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3881 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3883 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
3884 int nid
, unsigned int lru_mask
, bool tree
)
3886 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, NODE_DATA(nid
));
3887 unsigned long nr
= 0;
3890 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
3893 if (!(BIT(lru
) & lru_mask
))
3896 nr
+= lruvec_page_state(lruvec
, NR_LRU_BASE
+ lru
);
3898 nr
+= lruvec_page_state_local(lruvec
, NR_LRU_BASE
+ lru
);
3903 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
3904 unsigned int lru_mask
,
3907 unsigned long nr
= 0;
3911 if (!(BIT(lru
) & lru_mask
))
3914 nr
+= memcg_page_state(memcg
, NR_LRU_BASE
+ lru
);
3916 nr
+= memcg_page_state_local(memcg
, NR_LRU_BASE
+ lru
);
3921 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3925 unsigned int lru_mask
;
3928 static const struct numa_stat stats
[] = {
3929 { "total", LRU_ALL
},
3930 { "file", LRU_ALL_FILE
},
3931 { "anon", LRU_ALL_ANON
},
3932 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3934 const struct numa_stat
*stat
;
3936 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
3938 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3939 seq_printf(m
, "%s=%lu", stat
->name
,
3940 mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
,
3942 for_each_node_state(nid
, N_MEMORY
)
3943 seq_printf(m
, " N%d=%lu", nid
,
3944 mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3945 stat
->lru_mask
, false));
3949 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3951 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
,
3952 mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
,
3954 for_each_node_state(nid
, N_MEMORY
)
3955 seq_printf(m
, " N%d=%lu", nid
,
3956 mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3957 stat
->lru_mask
, true));
3963 #endif /* CONFIG_NUMA */
3965 static const unsigned int memcg1_stats
[] = {
3968 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3978 static const char *const memcg1_stat_names
[] = {
3981 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3991 /* Universal VM events cgroup1 shows, original sort order */
3992 static const unsigned int memcg1_events
[] = {
3999 static int memcg_stat_show(struct seq_file
*m
, void *v
)
4001 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
4002 unsigned long memory
, memsw
;
4003 struct mem_cgroup
*mi
;
4006 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names
) != ARRAY_SIZE(memcg1_stats
));
4008 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
4011 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
4013 nr
= memcg_page_state_local(memcg
, memcg1_stats
[i
]);
4014 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4015 if (memcg1_stats
[i
] == NR_ANON_THPS
)
4018 seq_printf(m
, "%s %lu\n", memcg1_stat_names
[i
], nr
* PAGE_SIZE
);
4021 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
4022 seq_printf(m
, "%s %lu\n", vm_event_name(memcg1_events
[i
]),
4023 memcg_events_local(memcg
, memcg1_events
[i
]));
4025 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
4026 seq_printf(m
, "%s %lu\n", lru_list_name(i
),
4027 memcg_page_state_local(memcg
, NR_LRU_BASE
+ i
) *
4030 /* Hierarchical information */
4031 memory
= memsw
= PAGE_COUNTER_MAX
;
4032 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
4033 memory
= min(memory
, READ_ONCE(mi
->memory
.max
));
4034 memsw
= min(memsw
, READ_ONCE(mi
->memsw
.max
));
4036 seq_printf(m
, "hierarchical_memory_limit %llu\n",
4037 (u64
)memory
* PAGE_SIZE
);
4038 if (do_memsw_account())
4039 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
4040 (u64
)memsw
* PAGE_SIZE
);
4042 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
4043 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
4045 seq_printf(m
, "total_%s %llu\n", memcg1_stat_names
[i
],
4046 (u64
)memcg_page_state(memcg
, memcg1_stats
[i
]) *
4050 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
4051 seq_printf(m
, "total_%s %llu\n",
4052 vm_event_name(memcg1_events
[i
]),
4053 (u64
)memcg_events(memcg
, memcg1_events
[i
]));
4055 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
4056 seq_printf(m
, "total_%s %llu\n", lru_list_name(i
),
4057 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
4060 #ifdef CONFIG_DEBUG_VM
4063 struct mem_cgroup_per_node
*mz
;
4064 unsigned long anon_cost
= 0;
4065 unsigned long file_cost
= 0;
4067 for_each_online_pgdat(pgdat
) {
4068 mz
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
4070 anon_cost
+= mz
->lruvec
.anon_cost
;
4071 file_cost
+= mz
->lruvec
.file_cost
;
4073 seq_printf(m
, "anon_cost %lu\n", anon_cost
);
4074 seq_printf(m
, "file_cost %lu\n", file_cost
);
4081 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
4084 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4086 return mem_cgroup_swappiness(memcg
);
4089 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
4090 struct cftype
*cft
, u64 val
)
4092 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4098 memcg
->swappiness
= val
;
4100 vm_swappiness
= val
;
4105 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
4107 struct mem_cgroup_threshold_ary
*t
;
4108 unsigned long usage
;
4113 t
= rcu_dereference(memcg
->thresholds
.primary
);
4115 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
4120 usage
= mem_cgroup_usage(memcg
, swap
);
4123 * current_threshold points to threshold just below or equal to usage.
4124 * If it's not true, a threshold was crossed after last
4125 * call of __mem_cgroup_threshold().
4127 i
= t
->current_threshold
;
4130 * Iterate backward over array of thresholds starting from
4131 * current_threshold and check if a threshold is crossed.
4132 * If none of thresholds below usage is crossed, we read
4133 * only one element of the array here.
4135 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
4136 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4138 /* i = current_threshold + 1 */
4142 * Iterate forward over array of thresholds starting from
4143 * current_threshold+1 and check if a threshold is crossed.
4144 * If none of thresholds above usage is crossed, we read
4145 * only one element of the array here.
4147 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
4148 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4150 /* Update current_threshold */
4151 t
->current_threshold
= i
- 1;
4156 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
4159 __mem_cgroup_threshold(memcg
, false);
4160 if (do_memsw_account())
4161 __mem_cgroup_threshold(memcg
, true);
4163 memcg
= parent_mem_cgroup(memcg
);
4167 static int compare_thresholds(const void *a
, const void *b
)
4169 const struct mem_cgroup_threshold
*_a
= a
;
4170 const struct mem_cgroup_threshold
*_b
= b
;
4172 if (_a
->threshold
> _b
->threshold
)
4175 if (_a
->threshold
< _b
->threshold
)
4181 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
4183 struct mem_cgroup_eventfd_list
*ev
;
4185 spin_lock(&memcg_oom_lock
);
4187 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
4188 eventfd_signal(ev
->eventfd
, 1);
4190 spin_unlock(&memcg_oom_lock
);
4194 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
4196 struct mem_cgroup
*iter
;
4198 for_each_mem_cgroup_tree(iter
, memcg
)
4199 mem_cgroup_oom_notify_cb(iter
);
4202 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4203 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
4205 struct mem_cgroup_thresholds
*thresholds
;
4206 struct mem_cgroup_threshold_ary
*new;
4207 unsigned long threshold
;
4208 unsigned long usage
;
4211 ret
= page_counter_memparse(args
, "-1", &threshold
);
4215 mutex_lock(&memcg
->thresholds_lock
);
4218 thresholds
= &memcg
->thresholds
;
4219 usage
= mem_cgroup_usage(memcg
, false);
4220 } else if (type
== _MEMSWAP
) {
4221 thresholds
= &memcg
->memsw_thresholds
;
4222 usage
= mem_cgroup_usage(memcg
, true);
4226 /* Check if a threshold crossed before adding a new one */
4227 if (thresholds
->primary
)
4228 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4230 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
4232 /* Allocate memory for new array of thresholds */
4233 new = kmalloc(struct_size(new, entries
, size
), GFP_KERNEL
);
4240 /* Copy thresholds (if any) to new array */
4241 if (thresholds
->primary
) {
4242 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
4243 sizeof(struct mem_cgroup_threshold
));
4246 /* Add new threshold */
4247 new->entries
[size
- 1].eventfd
= eventfd
;
4248 new->entries
[size
- 1].threshold
= threshold
;
4250 /* Sort thresholds. Registering of new threshold isn't time-critical */
4251 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
4252 compare_thresholds
, NULL
);
4254 /* Find current threshold */
4255 new->current_threshold
= -1;
4256 for (i
= 0; i
< size
; i
++) {
4257 if (new->entries
[i
].threshold
<= usage
) {
4259 * new->current_threshold will not be used until
4260 * rcu_assign_pointer(), so it's safe to increment
4263 ++new->current_threshold
;
4268 /* Free old spare buffer and save old primary buffer as spare */
4269 kfree(thresholds
->spare
);
4270 thresholds
->spare
= thresholds
->primary
;
4272 rcu_assign_pointer(thresholds
->primary
, new);
4274 /* To be sure that nobody uses thresholds */
4278 mutex_unlock(&memcg
->thresholds_lock
);
4283 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4284 struct eventfd_ctx
*eventfd
, const char *args
)
4286 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
4289 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4290 struct eventfd_ctx
*eventfd
, const char *args
)
4292 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
4295 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4296 struct eventfd_ctx
*eventfd
, enum res_type type
)
4298 struct mem_cgroup_thresholds
*thresholds
;
4299 struct mem_cgroup_threshold_ary
*new;
4300 unsigned long usage
;
4301 int i
, j
, size
, entries
;
4303 mutex_lock(&memcg
->thresholds_lock
);
4306 thresholds
= &memcg
->thresholds
;
4307 usage
= mem_cgroup_usage(memcg
, false);
4308 } else if (type
== _MEMSWAP
) {
4309 thresholds
= &memcg
->memsw_thresholds
;
4310 usage
= mem_cgroup_usage(memcg
, true);
4314 if (!thresholds
->primary
)
4317 /* Check if a threshold crossed before removing */
4318 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4320 /* Calculate new number of threshold */
4322 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4323 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4329 new = thresholds
->spare
;
4331 /* If no items related to eventfd have been cleared, nothing to do */
4335 /* Set thresholds array to NULL if we don't have thresholds */
4344 /* Copy thresholds and find current threshold */
4345 new->current_threshold
= -1;
4346 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4347 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4350 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4351 if (new->entries
[j
].threshold
<= usage
) {
4353 * new->current_threshold will not be used
4354 * until rcu_assign_pointer(), so it's safe to increment
4357 ++new->current_threshold
;
4363 /* Swap primary and spare array */
4364 thresholds
->spare
= thresholds
->primary
;
4366 rcu_assign_pointer(thresholds
->primary
, new);
4368 /* To be sure that nobody uses thresholds */
4371 /* If all events are unregistered, free the spare array */
4373 kfree(thresholds
->spare
);
4374 thresholds
->spare
= NULL
;
4377 mutex_unlock(&memcg
->thresholds_lock
);
4380 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4381 struct eventfd_ctx
*eventfd
)
4383 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
4386 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4387 struct eventfd_ctx
*eventfd
)
4389 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
4392 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
4393 struct eventfd_ctx
*eventfd
, const char *args
)
4395 struct mem_cgroup_eventfd_list
*event
;
4397 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4401 spin_lock(&memcg_oom_lock
);
4403 event
->eventfd
= eventfd
;
4404 list_add(&event
->list
, &memcg
->oom_notify
);
4406 /* already in OOM ? */
4407 if (memcg
->under_oom
)
4408 eventfd_signal(eventfd
, 1);
4409 spin_unlock(&memcg_oom_lock
);
4414 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
4415 struct eventfd_ctx
*eventfd
)
4417 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4419 spin_lock(&memcg_oom_lock
);
4421 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
4422 if (ev
->eventfd
== eventfd
) {
4423 list_del(&ev
->list
);
4428 spin_unlock(&memcg_oom_lock
);
4431 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
4433 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(sf
);
4435 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
4436 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
4437 seq_printf(sf
, "oom_kill %lu\n",
4438 atomic_long_read(&memcg
->memory_events
[MEMCG_OOM_KILL
]));
4442 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
4443 struct cftype
*cft
, u64 val
)
4445 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4447 /* cannot set to root cgroup and only 0 and 1 are allowed */
4448 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
4451 memcg
->oom_kill_disable
= val
;
4453 memcg_oom_recover(memcg
);
4458 #ifdef CONFIG_CGROUP_WRITEBACK
4460 #include <trace/events/writeback.h>
4462 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4464 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
4467 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4469 wb_domain_exit(&memcg
->cgwb_domain
);
4472 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4474 wb_domain_size_changed(&memcg
->cgwb_domain
);
4477 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
4479 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4481 if (!memcg
->css
.parent
)
4484 return &memcg
->cgwb_domain
;
4488 * idx can be of type enum memcg_stat_item or node_stat_item.
4489 * Keep in sync with memcg_exact_page().
4491 static unsigned long memcg_exact_page_state(struct mem_cgroup
*memcg
, int idx
)
4493 long x
= atomic_long_read(&memcg
->vmstats
[idx
]);
4496 for_each_online_cpu(cpu
)
4497 x
+= per_cpu_ptr(memcg
->vmstats_percpu
, cpu
)->stat
[idx
];
4504 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4505 * @wb: bdi_writeback in question
4506 * @pfilepages: out parameter for number of file pages
4507 * @pheadroom: out parameter for number of allocatable pages according to memcg
4508 * @pdirty: out parameter for number of dirty pages
4509 * @pwriteback: out parameter for number of pages under writeback
4511 * Determine the numbers of file, headroom, dirty, and writeback pages in
4512 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4513 * is a bit more involved.
4515 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4516 * headroom is calculated as the lowest headroom of itself and the
4517 * ancestors. Note that this doesn't consider the actual amount of
4518 * available memory in the system. The caller should further cap
4519 * *@pheadroom accordingly.
4521 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
4522 unsigned long *pheadroom
, unsigned long *pdirty
,
4523 unsigned long *pwriteback
)
4525 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4526 struct mem_cgroup
*parent
;
4528 *pdirty
= memcg_exact_page_state(memcg
, NR_FILE_DIRTY
);
4530 *pwriteback
= memcg_exact_page_state(memcg
, NR_WRITEBACK
);
4531 *pfilepages
= memcg_exact_page_state(memcg
, NR_INACTIVE_FILE
) +
4532 memcg_exact_page_state(memcg
, NR_ACTIVE_FILE
);
4533 *pheadroom
= PAGE_COUNTER_MAX
;
4535 while ((parent
= parent_mem_cgroup(memcg
))) {
4536 unsigned long ceiling
= min(READ_ONCE(memcg
->memory
.max
),
4537 READ_ONCE(memcg
->memory
.high
));
4538 unsigned long used
= page_counter_read(&memcg
->memory
);
4540 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
4546 * Foreign dirty flushing
4548 * There's an inherent mismatch between memcg and writeback. The former
4549 * trackes ownership per-page while the latter per-inode. This was a
4550 * deliberate design decision because honoring per-page ownership in the
4551 * writeback path is complicated, may lead to higher CPU and IO overheads
4552 * and deemed unnecessary given that write-sharing an inode across
4553 * different cgroups isn't a common use-case.
4555 * Combined with inode majority-writer ownership switching, this works well
4556 * enough in most cases but there are some pathological cases. For
4557 * example, let's say there are two cgroups A and B which keep writing to
4558 * different but confined parts of the same inode. B owns the inode and
4559 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4560 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4561 * triggering background writeback. A will be slowed down without a way to
4562 * make writeback of the dirty pages happen.
4564 * Conditions like the above can lead to a cgroup getting repatedly and
4565 * severely throttled after making some progress after each
4566 * dirty_expire_interval while the underyling IO device is almost
4569 * Solving this problem completely requires matching the ownership tracking
4570 * granularities between memcg and writeback in either direction. However,
4571 * the more egregious behaviors can be avoided by simply remembering the
4572 * most recent foreign dirtying events and initiating remote flushes on
4573 * them when local writeback isn't enough to keep the memory clean enough.
4575 * The following two functions implement such mechanism. When a foreign
4576 * page - a page whose memcg and writeback ownerships don't match - is
4577 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4578 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4579 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4580 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4581 * foreign bdi_writebacks which haven't expired. Both the numbers of
4582 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4583 * limited to MEMCG_CGWB_FRN_CNT.
4585 * The mechanism only remembers IDs and doesn't hold any object references.
4586 * As being wrong occasionally doesn't matter, updates and accesses to the
4587 * records are lockless and racy.
4589 void mem_cgroup_track_foreign_dirty_slowpath(struct page
*page
,
4590 struct bdi_writeback
*wb
)
4592 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
4593 struct memcg_cgwb_frn
*frn
;
4594 u64 now
= get_jiffies_64();
4595 u64 oldest_at
= now
;
4599 trace_track_foreign_dirty(page
, wb
);
4602 * Pick the slot to use. If there is already a slot for @wb, keep
4603 * using it. If not replace the oldest one which isn't being
4606 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4607 frn
= &memcg
->cgwb_frn
[i
];
4608 if (frn
->bdi_id
== wb
->bdi
->id
&&
4609 frn
->memcg_id
== wb
->memcg_css
->id
)
4611 if (time_before64(frn
->at
, oldest_at
) &&
4612 atomic_read(&frn
->done
.cnt
) == 1) {
4614 oldest_at
= frn
->at
;
4618 if (i
< MEMCG_CGWB_FRN_CNT
) {
4620 * Re-using an existing one. Update timestamp lazily to
4621 * avoid making the cacheline hot. We want them to be
4622 * reasonably up-to-date and significantly shorter than
4623 * dirty_expire_interval as that's what expires the record.
4624 * Use the shorter of 1s and dirty_expire_interval / 8.
4626 unsigned long update_intv
=
4627 min_t(unsigned long, HZ
,
4628 msecs_to_jiffies(dirty_expire_interval
* 10) / 8);
4630 if (time_before64(frn
->at
, now
- update_intv
))
4632 } else if (oldest
>= 0) {
4633 /* replace the oldest free one */
4634 frn
= &memcg
->cgwb_frn
[oldest
];
4635 frn
->bdi_id
= wb
->bdi
->id
;
4636 frn
->memcg_id
= wb
->memcg_css
->id
;
4641 /* issue foreign writeback flushes for recorded foreign dirtying events */
4642 void mem_cgroup_flush_foreign(struct bdi_writeback
*wb
)
4644 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4645 unsigned long intv
= msecs_to_jiffies(dirty_expire_interval
* 10);
4646 u64 now
= jiffies_64
;
4649 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4650 struct memcg_cgwb_frn
*frn
= &memcg
->cgwb_frn
[i
];
4653 * If the record is older than dirty_expire_interval,
4654 * writeback on it has already started. No need to kick it
4655 * off again. Also, don't start a new one if there's
4656 * already one in flight.
4658 if (time_after64(frn
->at
, now
- intv
) &&
4659 atomic_read(&frn
->done
.cnt
) == 1) {
4661 trace_flush_foreign(wb
, frn
->bdi_id
, frn
->memcg_id
);
4662 cgroup_writeback_by_id(frn
->bdi_id
, frn
->memcg_id
, 0,
4663 WB_REASON_FOREIGN_FLUSH
,
4669 #else /* CONFIG_CGROUP_WRITEBACK */
4671 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4676 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4680 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4684 #endif /* CONFIG_CGROUP_WRITEBACK */
4687 * DO NOT USE IN NEW FILES.
4689 * "cgroup.event_control" implementation.
4691 * This is way over-engineered. It tries to support fully configurable
4692 * events for each user. Such level of flexibility is completely
4693 * unnecessary especially in the light of the planned unified hierarchy.
4695 * Please deprecate this and replace with something simpler if at all
4700 * Unregister event and free resources.
4702 * Gets called from workqueue.
4704 static void memcg_event_remove(struct work_struct
*work
)
4706 struct mem_cgroup_event
*event
=
4707 container_of(work
, struct mem_cgroup_event
, remove
);
4708 struct mem_cgroup
*memcg
= event
->memcg
;
4710 remove_wait_queue(event
->wqh
, &event
->wait
);
4712 event
->unregister_event(memcg
, event
->eventfd
);
4714 /* Notify userspace the event is going away. */
4715 eventfd_signal(event
->eventfd
, 1);
4717 eventfd_ctx_put(event
->eventfd
);
4719 css_put(&memcg
->css
);
4723 * Gets called on EPOLLHUP on eventfd when user closes it.
4725 * Called with wqh->lock held and interrupts disabled.
4727 static int memcg_event_wake(wait_queue_entry_t
*wait
, unsigned mode
,
4728 int sync
, void *key
)
4730 struct mem_cgroup_event
*event
=
4731 container_of(wait
, struct mem_cgroup_event
, wait
);
4732 struct mem_cgroup
*memcg
= event
->memcg
;
4733 __poll_t flags
= key_to_poll(key
);
4735 if (flags
& EPOLLHUP
) {
4737 * If the event has been detached at cgroup removal, we
4738 * can simply return knowing the other side will cleanup
4741 * We can't race against event freeing since the other
4742 * side will require wqh->lock via remove_wait_queue(),
4745 spin_lock(&memcg
->event_list_lock
);
4746 if (!list_empty(&event
->list
)) {
4747 list_del_init(&event
->list
);
4749 * We are in atomic context, but cgroup_event_remove()
4750 * may sleep, so we have to call it in workqueue.
4752 schedule_work(&event
->remove
);
4754 spin_unlock(&memcg
->event_list_lock
);
4760 static void memcg_event_ptable_queue_proc(struct file
*file
,
4761 wait_queue_head_t
*wqh
, poll_table
*pt
)
4763 struct mem_cgroup_event
*event
=
4764 container_of(pt
, struct mem_cgroup_event
, pt
);
4767 add_wait_queue(wqh
, &event
->wait
);
4771 * DO NOT USE IN NEW FILES.
4773 * Parse input and register new cgroup event handler.
4775 * Input must be in format '<event_fd> <control_fd> <args>'.
4776 * Interpretation of args is defined by control file implementation.
4778 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4779 char *buf
, size_t nbytes
, loff_t off
)
4781 struct cgroup_subsys_state
*css
= of_css(of
);
4782 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4783 struct mem_cgroup_event
*event
;
4784 struct cgroup_subsys_state
*cfile_css
;
4785 unsigned int efd
, cfd
;
4792 buf
= strstrip(buf
);
4794 efd
= simple_strtoul(buf
, &endp
, 10);
4799 cfd
= simple_strtoul(buf
, &endp
, 10);
4800 if ((*endp
!= ' ') && (*endp
!= '\0'))
4804 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
4808 event
->memcg
= memcg
;
4809 INIT_LIST_HEAD(&event
->list
);
4810 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
4811 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
4812 INIT_WORK(&event
->remove
, memcg_event_remove
);
4820 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
4821 if (IS_ERR(event
->eventfd
)) {
4822 ret
= PTR_ERR(event
->eventfd
);
4829 goto out_put_eventfd
;
4832 /* the process need read permission on control file */
4833 /* AV: shouldn't we check that it's been opened for read instead? */
4834 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
4839 * Determine the event callbacks and set them in @event. This used
4840 * to be done via struct cftype but cgroup core no longer knows
4841 * about these events. The following is crude but the whole thing
4842 * is for compatibility anyway.
4844 * DO NOT ADD NEW FILES.
4846 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
4848 if (!strcmp(name
, "memory.usage_in_bytes")) {
4849 event
->register_event
= mem_cgroup_usage_register_event
;
4850 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
4851 } else if (!strcmp(name
, "memory.oom_control")) {
4852 event
->register_event
= mem_cgroup_oom_register_event
;
4853 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
4854 } else if (!strcmp(name
, "memory.pressure_level")) {
4855 event
->register_event
= vmpressure_register_event
;
4856 event
->unregister_event
= vmpressure_unregister_event
;
4857 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
4858 event
->register_event
= memsw_cgroup_usage_register_event
;
4859 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
4866 * Verify @cfile should belong to @css. Also, remaining events are
4867 * automatically removed on cgroup destruction but the removal is
4868 * asynchronous, so take an extra ref on @css.
4870 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
4871 &memory_cgrp_subsys
);
4873 if (IS_ERR(cfile_css
))
4875 if (cfile_css
!= css
) {
4880 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
4884 vfs_poll(efile
.file
, &event
->pt
);
4886 spin_lock(&memcg
->event_list_lock
);
4887 list_add(&event
->list
, &memcg
->event_list
);
4888 spin_unlock(&memcg
->event_list_lock
);
4900 eventfd_ctx_put(event
->eventfd
);
4909 static struct cftype mem_cgroup_legacy_files
[] = {
4911 .name
= "usage_in_bytes",
4912 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4913 .read_u64
= mem_cgroup_read_u64
,
4916 .name
= "max_usage_in_bytes",
4917 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4918 .write
= mem_cgroup_reset
,
4919 .read_u64
= mem_cgroup_read_u64
,
4922 .name
= "limit_in_bytes",
4923 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4924 .write
= mem_cgroup_write
,
4925 .read_u64
= mem_cgroup_read_u64
,
4928 .name
= "soft_limit_in_bytes",
4929 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4930 .write
= mem_cgroup_write
,
4931 .read_u64
= mem_cgroup_read_u64
,
4935 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4936 .write
= mem_cgroup_reset
,
4937 .read_u64
= mem_cgroup_read_u64
,
4941 .seq_show
= memcg_stat_show
,
4944 .name
= "force_empty",
4945 .write
= mem_cgroup_force_empty_write
,
4948 .name
= "use_hierarchy",
4949 .write_u64
= mem_cgroup_hierarchy_write
,
4950 .read_u64
= mem_cgroup_hierarchy_read
,
4953 .name
= "cgroup.event_control", /* XXX: for compat */
4954 .write
= memcg_write_event_control
,
4955 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
4958 .name
= "swappiness",
4959 .read_u64
= mem_cgroup_swappiness_read
,
4960 .write_u64
= mem_cgroup_swappiness_write
,
4963 .name
= "move_charge_at_immigrate",
4964 .read_u64
= mem_cgroup_move_charge_read
,
4965 .write_u64
= mem_cgroup_move_charge_write
,
4968 .name
= "oom_control",
4969 .seq_show
= mem_cgroup_oom_control_read
,
4970 .write_u64
= mem_cgroup_oom_control_write
,
4971 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4974 .name
= "pressure_level",
4978 .name
= "numa_stat",
4979 .seq_show
= memcg_numa_stat_show
,
4983 .name
= "kmem.limit_in_bytes",
4984 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4985 .write
= mem_cgroup_write
,
4986 .read_u64
= mem_cgroup_read_u64
,
4989 .name
= "kmem.usage_in_bytes",
4990 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4991 .read_u64
= mem_cgroup_read_u64
,
4994 .name
= "kmem.failcnt",
4995 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4996 .write
= mem_cgroup_reset
,
4997 .read_u64
= mem_cgroup_read_u64
,
5000 .name
= "kmem.max_usage_in_bytes",
5001 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
5002 .write
= mem_cgroup_reset
,
5003 .read_u64
= mem_cgroup_read_u64
,
5005 #if defined(CONFIG_MEMCG_KMEM) && \
5006 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5008 .name
= "kmem.slabinfo",
5009 .seq_show
= memcg_slab_show
,
5013 .name
= "kmem.tcp.limit_in_bytes",
5014 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
5015 .write
= mem_cgroup_write
,
5016 .read_u64
= mem_cgroup_read_u64
,
5019 .name
= "kmem.tcp.usage_in_bytes",
5020 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
5021 .read_u64
= mem_cgroup_read_u64
,
5024 .name
= "kmem.tcp.failcnt",
5025 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
5026 .write
= mem_cgroup_reset
,
5027 .read_u64
= mem_cgroup_read_u64
,
5030 .name
= "kmem.tcp.max_usage_in_bytes",
5031 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
5032 .write
= mem_cgroup_reset
,
5033 .read_u64
= mem_cgroup_read_u64
,
5035 { }, /* terminate */
5039 * Private memory cgroup IDR
5041 * Swap-out records and page cache shadow entries need to store memcg
5042 * references in constrained space, so we maintain an ID space that is
5043 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5044 * memory-controlled cgroups to 64k.
5046 * However, there usually are many references to the offline CSS after
5047 * the cgroup has been destroyed, such as page cache or reclaimable
5048 * slab objects, that don't need to hang on to the ID. We want to keep
5049 * those dead CSS from occupying IDs, or we might quickly exhaust the
5050 * relatively small ID space and prevent the creation of new cgroups
5051 * even when there are much fewer than 64k cgroups - possibly none.
5053 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5054 * be freed and recycled when it's no longer needed, which is usually
5055 * when the CSS is offlined.
5057 * The only exception to that are records of swapped out tmpfs/shmem
5058 * pages that need to be attributed to live ancestors on swapin. But
5059 * those references are manageable from userspace.
5062 static DEFINE_IDR(mem_cgroup_idr
);
5064 static void mem_cgroup_id_remove(struct mem_cgroup
*memcg
)
5066 if (memcg
->id
.id
> 0) {
5067 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
5072 static void __maybe_unused
mem_cgroup_id_get_many(struct mem_cgroup
*memcg
,
5075 refcount_add(n
, &memcg
->id
.ref
);
5078 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
5080 if (refcount_sub_and_test(n
, &memcg
->id
.ref
)) {
5081 mem_cgroup_id_remove(memcg
);
5083 /* Memcg ID pins CSS */
5084 css_put(&memcg
->css
);
5088 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
5090 mem_cgroup_id_put_many(memcg
, 1);
5094 * mem_cgroup_from_id - look up a memcg from a memcg id
5095 * @id: the memcg id to look up
5097 * Caller must hold rcu_read_lock().
5099 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
5101 WARN_ON_ONCE(!rcu_read_lock_held());
5102 return idr_find(&mem_cgroup_idr
, id
);
5105 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
5107 struct mem_cgroup_per_node
*pn
;
5110 * This routine is called against possible nodes.
5111 * But it's BUG to call kmalloc() against offline node.
5113 * TODO: this routine can waste much memory for nodes which will
5114 * never be onlined. It's better to use memory hotplug callback
5117 if (!node_state(node
, N_NORMAL_MEMORY
))
5119 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
5123 pn
->lruvec_stat_local
= alloc_percpu(struct lruvec_stat
);
5124 if (!pn
->lruvec_stat_local
) {
5129 pn
->lruvec_stat_cpu
= alloc_percpu(struct lruvec_stat
);
5130 if (!pn
->lruvec_stat_cpu
) {
5131 free_percpu(pn
->lruvec_stat_local
);
5136 lruvec_init(&pn
->lruvec
);
5137 pn
->usage_in_excess
= 0;
5138 pn
->on_tree
= false;
5141 memcg
->nodeinfo
[node
] = pn
;
5145 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
5147 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
5152 free_percpu(pn
->lruvec_stat_cpu
);
5153 free_percpu(pn
->lruvec_stat_local
);
5157 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5162 free_mem_cgroup_per_node_info(memcg
, node
);
5163 free_percpu(memcg
->vmstats_percpu
);
5164 free_percpu(memcg
->vmstats_local
);
5168 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
5170 memcg_wb_domain_exit(memcg
);
5172 * Flush percpu vmstats and vmevents to guarantee the value correctness
5173 * on parent's and all ancestor levels.
5175 memcg_flush_percpu_vmstats(memcg
);
5176 memcg_flush_percpu_vmevents(memcg
);
5177 __mem_cgroup_free(memcg
);
5180 static struct mem_cgroup
*mem_cgroup_alloc(void)
5182 struct mem_cgroup
*memcg
;
5185 int __maybe_unused i
;
5186 long error
= -ENOMEM
;
5188 size
= sizeof(struct mem_cgroup
);
5189 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
5191 memcg
= kzalloc(size
, GFP_KERNEL
);
5193 return ERR_PTR(error
);
5195 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
5196 1, MEM_CGROUP_ID_MAX
,
5198 if (memcg
->id
.id
< 0) {
5199 error
= memcg
->id
.id
;
5203 memcg
->vmstats_local
= alloc_percpu(struct memcg_vmstats_percpu
);
5204 if (!memcg
->vmstats_local
)
5207 memcg
->vmstats_percpu
= alloc_percpu(struct memcg_vmstats_percpu
);
5208 if (!memcg
->vmstats_percpu
)
5212 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
5215 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
5218 INIT_WORK(&memcg
->high_work
, high_work_func
);
5219 INIT_LIST_HEAD(&memcg
->oom_notify
);
5220 mutex_init(&memcg
->thresholds_lock
);
5221 spin_lock_init(&memcg
->move_lock
);
5222 vmpressure_init(&memcg
->vmpressure
);
5223 INIT_LIST_HEAD(&memcg
->event_list
);
5224 spin_lock_init(&memcg
->event_list_lock
);
5225 memcg
->socket_pressure
= jiffies
;
5226 #ifdef CONFIG_MEMCG_KMEM
5227 memcg
->kmemcg_id
= -1;
5228 INIT_LIST_HEAD(&memcg
->objcg_list
);
5230 #ifdef CONFIG_CGROUP_WRITEBACK
5231 INIT_LIST_HEAD(&memcg
->cgwb_list
);
5232 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5233 memcg
->cgwb_frn
[i
].done
=
5234 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq
);
5236 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5237 spin_lock_init(&memcg
->deferred_split_queue
.split_queue_lock
);
5238 INIT_LIST_HEAD(&memcg
->deferred_split_queue
.split_queue
);
5239 memcg
->deferred_split_queue
.split_queue_len
= 0;
5241 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
5244 mem_cgroup_id_remove(memcg
);
5245 __mem_cgroup_free(memcg
);
5246 return ERR_PTR(error
);
5249 static struct cgroup_subsys_state
* __ref
5250 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
5252 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
5253 struct mem_cgroup
*memcg
;
5254 long error
= -ENOMEM
;
5256 memcg
= mem_cgroup_alloc();
5258 return ERR_CAST(memcg
);
5260 page_counter_set_high(&memcg
->memory
, PAGE_COUNTER_MAX
);
5261 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5262 page_counter_set_high(&memcg
->swap
, PAGE_COUNTER_MAX
);
5264 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
5265 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
5267 if (parent
&& parent
->use_hierarchy
) {
5268 memcg
->use_hierarchy
= true;
5269 page_counter_init(&memcg
->memory
, &parent
->memory
);
5270 page_counter_init(&memcg
->swap
, &parent
->swap
);
5271 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
5272 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
5273 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
5275 page_counter_init(&memcg
->memory
, NULL
);
5276 page_counter_init(&memcg
->swap
, NULL
);
5277 page_counter_init(&memcg
->memsw
, NULL
);
5278 page_counter_init(&memcg
->kmem
, NULL
);
5279 page_counter_init(&memcg
->tcpmem
, NULL
);
5281 * Deeper hierachy with use_hierarchy == false doesn't make
5282 * much sense so let cgroup subsystem know about this
5283 * unfortunate state in our controller.
5285 if (parent
!= root_mem_cgroup
)
5286 memory_cgrp_subsys
.broken_hierarchy
= true;
5289 /* The following stuff does not apply to the root */
5291 root_mem_cgroup
= memcg
;
5295 error
= memcg_online_kmem(memcg
);
5299 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5300 static_branch_inc(&memcg_sockets_enabled_key
);
5304 mem_cgroup_id_remove(memcg
);
5305 mem_cgroup_free(memcg
);
5306 return ERR_PTR(error
);
5309 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
5311 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5314 * A memcg must be visible for memcg_expand_shrinker_maps()
5315 * by the time the maps are allocated. So, we allocate maps
5316 * here, when for_each_mem_cgroup() can't skip it.
5318 if (memcg_alloc_shrinker_maps(memcg
)) {
5319 mem_cgroup_id_remove(memcg
);
5323 /* Online state pins memcg ID, memcg ID pins CSS */
5324 refcount_set(&memcg
->id
.ref
, 1);
5329 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
5331 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5332 struct mem_cgroup_event
*event
, *tmp
;
5335 * Unregister events and notify userspace.
5336 * Notify userspace about cgroup removing only after rmdir of cgroup
5337 * directory to avoid race between userspace and kernelspace.
5339 spin_lock(&memcg
->event_list_lock
);
5340 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
5341 list_del_init(&event
->list
);
5342 schedule_work(&event
->remove
);
5344 spin_unlock(&memcg
->event_list_lock
);
5346 page_counter_set_min(&memcg
->memory
, 0);
5347 page_counter_set_low(&memcg
->memory
, 0);
5349 memcg_offline_kmem(memcg
);
5350 wb_memcg_offline(memcg
);
5352 drain_all_stock(memcg
);
5354 mem_cgroup_id_put(memcg
);
5357 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
5359 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5361 invalidate_reclaim_iterators(memcg
);
5364 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
5366 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5367 int __maybe_unused i
;
5369 #ifdef CONFIG_CGROUP_WRITEBACK
5370 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5371 wb_wait_for_completion(&memcg
->cgwb_frn
[i
].done
);
5373 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5374 static_branch_dec(&memcg_sockets_enabled_key
);
5376 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
5377 static_branch_dec(&memcg_sockets_enabled_key
);
5379 vmpressure_cleanup(&memcg
->vmpressure
);
5380 cancel_work_sync(&memcg
->high_work
);
5381 mem_cgroup_remove_from_trees(memcg
);
5382 memcg_free_shrinker_maps(memcg
);
5383 memcg_free_kmem(memcg
);
5384 mem_cgroup_free(memcg
);
5388 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5389 * @css: the target css
5391 * Reset the states of the mem_cgroup associated with @css. This is
5392 * invoked when the userland requests disabling on the default hierarchy
5393 * but the memcg is pinned through dependency. The memcg should stop
5394 * applying policies and should revert to the vanilla state as it may be
5395 * made visible again.
5397 * The current implementation only resets the essential configurations.
5398 * This needs to be expanded to cover all the visible parts.
5400 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
5402 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5404 page_counter_set_max(&memcg
->memory
, PAGE_COUNTER_MAX
);
5405 page_counter_set_max(&memcg
->swap
, PAGE_COUNTER_MAX
);
5406 page_counter_set_max(&memcg
->memsw
, PAGE_COUNTER_MAX
);
5407 page_counter_set_max(&memcg
->kmem
, PAGE_COUNTER_MAX
);
5408 page_counter_set_max(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
5409 page_counter_set_min(&memcg
->memory
, 0);
5410 page_counter_set_low(&memcg
->memory
, 0);
5411 page_counter_set_high(&memcg
->memory
, PAGE_COUNTER_MAX
);
5412 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5413 page_counter_set_high(&memcg
->swap
, PAGE_COUNTER_MAX
);
5414 memcg_wb_domain_size_changed(memcg
);
5418 /* Handlers for move charge at task migration. */
5419 static int mem_cgroup_do_precharge(unsigned long count
)
5423 /* Try a single bulk charge without reclaim first, kswapd may wake */
5424 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
5426 mc
.precharge
+= count
;
5430 /* Try charges one by one with reclaim, but do not retry */
5432 ret
= try_charge(mc
.to
, GFP_KERNEL
| __GFP_NORETRY
, 1);
5446 enum mc_target_type
{
5453 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
5454 unsigned long addr
, pte_t ptent
)
5456 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
5458 if (!page
|| !page_mapped(page
))
5460 if (PageAnon(page
)) {
5461 if (!(mc
.flags
& MOVE_ANON
))
5464 if (!(mc
.flags
& MOVE_FILE
))
5467 if (!get_page_unless_zero(page
))
5473 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5474 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5475 pte_t ptent
, swp_entry_t
*entry
)
5477 struct page
*page
= NULL
;
5478 swp_entry_t ent
= pte_to_swp_entry(ptent
);
5480 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
5484 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5485 * a device and because they are not accessible by CPU they are store
5486 * as special swap entry in the CPU page table.
5488 if (is_device_private_entry(ent
)) {
5489 page
= device_private_entry_to_page(ent
);
5491 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5492 * a refcount of 1 when free (unlike normal page)
5494 if (!page_ref_add_unless(page
, 1, 1))
5500 * Because lookup_swap_cache() updates some statistics counter,
5501 * we call find_get_page() with swapper_space directly.
5503 page
= find_get_page(swap_address_space(ent
), swp_offset(ent
));
5504 entry
->val
= ent
.val
;
5509 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5510 pte_t ptent
, swp_entry_t
*entry
)
5516 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
5517 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5519 struct page
*page
= NULL
;
5520 struct address_space
*mapping
;
5523 if (!vma
->vm_file
) /* anonymous vma */
5525 if (!(mc
.flags
& MOVE_FILE
))
5528 mapping
= vma
->vm_file
->f_mapping
;
5529 pgoff
= linear_page_index(vma
, addr
);
5531 /* page is moved even if it's not RSS of this task(page-faulted). */
5533 /* shmem/tmpfs may report page out on swap: account for that too. */
5534 if (shmem_mapping(mapping
)) {
5535 page
= find_get_entry(mapping
, pgoff
);
5536 if (xa_is_value(page
)) {
5537 swp_entry_t swp
= radix_to_swp_entry(page
);
5539 page
= find_get_page(swap_address_space(swp
),
5543 page
= find_get_page(mapping
, pgoff
);
5545 page
= find_get_page(mapping
, pgoff
);
5551 * mem_cgroup_move_account - move account of the page
5553 * @compound: charge the page as compound or small page
5554 * @from: mem_cgroup which the page is moved from.
5555 * @to: mem_cgroup which the page is moved to. @from != @to.
5557 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5559 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5562 static int mem_cgroup_move_account(struct page
*page
,
5564 struct mem_cgroup
*from
,
5565 struct mem_cgroup
*to
)
5567 struct lruvec
*from_vec
, *to_vec
;
5568 struct pglist_data
*pgdat
;
5569 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5572 VM_BUG_ON(from
== to
);
5573 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5574 VM_BUG_ON(compound
&& !PageTransHuge(page
));
5577 * Prevent mem_cgroup_migrate() from looking at
5578 * page->mem_cgroup of its source page while we change it.
5581 if (!trylock_page(page
))
5585 if (page
->mem_cgroup
!= from
)
5588 pgdat
= page_pgdat(page
);
5589 from_vec
= mem_cgroup_lruvec(from
, pgdat
);
5590 to_vec
= mem_cgroup_lruvec(to
, pgdat
);
5592 lock_page_memcg(page
);
5594 if (PageAnon(page
)) {
5595 if (page_mapped(page
)) {
5596 __mod_lruvec_state(from_vec
, NR_ANON_MAPPED
, -nr_pages
);
5597 __mod_lruvec_state(to_vec
, NR_ANON_MAPPED
, nr_pages
);
5598 if (PageTransHuge(page
)) {
5599 __mod_lruvec_state(from_vec
, NR_ANON_THPS
,
5601 __mod_lruvec_state(to_vec
, NR_ANON_THPS
,
5607 __mod_lruvec_state(from_vec
, NR_FILE_PAGES
, -nr_pages
);
5608 __mod_lruvec_state(to_vec
, NR_FILE_PAGES
, nr_pages
);
5610 if (PageSwapBacked(page
)) {
5611 __mod_lruvec_state(from_vec
, NR_SHMEM
, -nr_pages
);
5612 __mod_lruvec_state(to_vec
, NR_SHMEM
, nr_pages
);
5615 if (page_mapped(page
)) {
5616 __mod_lruvec_state(from_vec
, NR_FILE_MAPPED
, -nr_pages
);
5617 __mod_lruvec_state(to_vec
, NR_FILE_MAPPED
, nr_pages
);
5620 if (PageDirty(page
)) {
5621 struct address_space
*mapping
= page_mapping(page
);
5623 if (mapping_cap_account_dirty(mapping
)) {
5624 __mod_lruvec_state(from_vec
, NR_FILE_DIRTY
,
5626 __mod_lruvec_state(to_vec
, NR_FILE_DIRTY
,
5632 if (PageWriteback(page
)) {
5633 __mod_lruvec_state(from_vec
, NR_WRITEBACK
, -nr_pages
);
5634 __mod_lruvec_state(to_vec
, NR_WRITEBACK
, nr_pages
);
5638 * All state has been migrated, let's switch to the new memcg.
5640 * It is safe to change page->mem_cgroup here because the page
5641 * is referenced, charged, isolated, and locked: we can't race
5642 * with (un)charging, migration, LRU putback, or anything else
5643 * that would rely on a stable page->mem_cgroup.
5645 * Note that lock_page_memcg is a memcg lock, not a page lock,
5646 * to save space. As soon as we switch page->mem_cgroup to a
5647 * new memcg that isn't locked, the above state can change
5648 * concurrently again. Make sure we're truly done with it.
5653 css_put(&from
->css
);
5655 page
->mem_cgroup
= to
;
5657 __unlock_page_memcg(from
);
5661 local_irq_disable();
5662 mem_cgroup_charge_statistics(to
, page
, nr_pages
);
5663 memcg_check_events(to
, page
);
5664 mem_cgroup_charge_statistics(from
, page
, -nr_pages
);
5665 memcg_check_events(from
, page
);
5674 * get_mctgt_type - get target type of moving charge
5675 * @vma: the vma the pte to be checked belongs
5676 * @addr: the address corresponding to the pte to be checked
5677 * @ptent: the pte to be checked
5678 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5681 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5682 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5683 * move charge. if @target is not NULL, the page is stored in target->page
5684 * with extra refcnt got(Callers should handle it).
5685 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5686 * target for charge migration. if @target is not NULL, the entry is stored
5688 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5689 * (so ZONE_DEVICE page and thus not on the lru).
5690 * For now we such page is charge like a regular page would be as for all
5691 * intent and purposes it is just special memory taking the place of a
5694 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5696 * Called with pte lock held.
5699 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
5700 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
5702 struct page
*page
= NULL
;
5703 enum mc_target_type ret
= MC_TARGET_NONE
;
5704 swp_entry_t ent
= { .val
= 0 };
5706 if (pte_present(ptent
))
5707 page
= mc_handle_present_pte(vma
, addr
, ptent
);
5708 else if (is_swap_pte(ptent
))
5709 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
5710 else if (pte_none(ptent
))
5711 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
5713 if (!page
&& !ent
.val
)
5717 * Do only loose check w/o serialization.
5718 * mem_cgroup_move_account() checks the page is valid or
5719 * not under LRU exclusion.
5721 if (page
->mem_cgroup
== mc
.from
) {
5722 ret
= MC_TARGET_PAGE
;
5723 if (is_device_private_page(page
))
5724 ret
= MC_TARGET_DEVICE
;
5726 target
->page
= page
;
5728 if (!ret
|| !target
)
5732 * There is a swap entry and a page doesn't exist or isn't charged.
5733 * But we cannot move a tail-page in a THP.
5735 if (ent
.val
&& !ret
&& (!page
|| !PageTransCompound(page
)) &&
5736 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
5737 ret
= MC_TARGET_SWAP
;
5744 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5746 * We don't consider PMD mapped swapping or file mapped pages because THP does
5747 * not support them for now.
5748 * Caller should make sure that pmd_trans_huge(pmd) is true.
5750 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5751 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5753 struct page
*page
= NULL
;
5754 enum mc_target_type ret
= MC_TARGET_NONE
;
5756 if (unlikely(is_swap_pmd(pmd
))) {
5757 VM_BUG_ON(thp_migration_supported() &&
5758 !is_pmd_migration_entry(pmd
));
5761 page
= pmd_page(pmd
);
5762 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
5763 if (!(mc
.flags
& MOVE_ANON
))
5765 if (page
->mem_cgroup
== mc
.from
) {
5766 ret
= MC_TARGET_PAGE
;
5769 target
->page
= page
;
5775 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5776 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5778 return MC_TARGET_NONE
;
5782 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5783 unsigned long addr
, unsigned long end
,
5784 struct mm_walk
*walk
)
5786 struct vm_area_struct
*vma
= walk
->vma
;
5790 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5793 * Note their can not be MC_TARGET_DEVICE for now as we do not
5794 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5795 * this might change.
5797 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5798 mc
.precharge
+= HPAGE_PMD_NR
;
5803 if (pmd_trans_unstable(pmd
))
5805 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5806 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5807 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5808 mc
.precharge
++; /* increment precharge temporarily */
5809 pte_unmap_unlock(pte
- 1, ptl
);
5815 static const struct mm_walk_ops precharge_walk_ops
= {
5816 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
5819 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5821 unsigned long precharge
;
5824 walk_page_range(mm
, 0, mm
->highest_vm_end
, &precharge_walk_ops
, NULL
);
5825 mmap_read_unlock(mm
);
5827 precharge
= mc
.precharge
;
5833 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
5835 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
5837 VM_BUG_ON(mc
.moving_task
);
5838 mc
.moving_task
= current
;
5839 return mem_cgroup_do_precharge(precharge
);
5842 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5843 static void __mem_cgroup_clear_mc(void)
5845 struct mem_cgroup
*from
= mc
.from
;
5846 struct mem_cgroup
*to
= mc
.to
;
5848 /* we must uncharge all the leftover precharges from mc.to */
5850 cancel_charge(mc
.to
, mc
.precharge
);
5854 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5855 * we must uncharge here.
5857 if (mc
.moved_charge
) {
5858 cancel_charge(mc
.from
, mc
.moved_charge
);
5859 mc
.moved_charge
= 0;
5861 /* we must fixup refcnts and charges */
5862 if (mc
.moved_swap
) {
5863 /* uncharge swap account from the old cgroup */
5864 if (!mem_cgroup_is_root(mc
.from
))
5865 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
5867 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
5870 * we charged both to->memory and to->memsw, so we
5871 * should uncharge to->memory.
5873 if (!mem_cgroup_is_root(mc
.to
))
5874 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
5878 memcg_oom_recover(from
);
5879 memcg_oom_recover(to
);
5880 wake_up_all(&mc
.waitq
);
5883 static void mem_cgroup_clear_mc(void)
5885 struct mm_struct
*mm
= mc
.mm
;
5888 * we must clear moving_task before waking up waiters at the end of
5891 mc
.moving_task
= NULL
;
5892 __mem_cgroup_clear_mc();
5893 spin_lock(&mc
.lock
);
5897 spin_unlock(&mc
.lock
);
5902 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5904 struct cgroup_subsys_state
*css
;
5905 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
5906 struct mem_cgroup
*from
;
5907 struct task_struct
*leader
, *p
;
5908 struct mm_struct
*mm
;
5909 unsigned long move_flags
;
5912 /* charge immigration isn't supported on the default hierarchy */
5913 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5917 * Multi-process migrations only happen on the default hierarchy
5918 * where charge immigration is not used. Perform charge
5919 * immigration if @tset contains a leader and whine if there are
5923 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
5926 memcg
= mem_cgroup_from_css(css
);
5932 * We are now commited to this value whatever it is. Changes in this
5933 * tunable will only affect upcoming migrations, not the current one.
5934 * So we need to save it, and keep it going.
5936 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
5940 from
= mem_cgroup_from_task(p
);
5942 VM_BUG_ON(from
== memcg
);
5944 mm
= get_task_mm(p
);
5947 /* We move charges only when we move a owner of the mm */
5948 if (mm
->owner
== p
) {
5951 VM_BUG_ON(mc
.precharge
);
5952 VM_BUG_ON(mc
.moved_charge
);
5953 VM_BUG_ON(mc
.moved_swap
);
5955 spin_lock(&mc
.lock
);
5959 mc
.flags
= move_flags
;
5960 spin_unlock(&mc
.lock
);
5961 /* We set mc.moving_task later */
5963 ret
= mem_cgroup_precharge_mc(mm
);
5965 mem_cgroup_clear_mc();
5972 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5975 mem_cgroup_clear_mc();
5978 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
5979 unsigned long addr
, unsigned long end
,
5980 struct mm_walk
*walk
)
5983 struct vm_area_struct
*vma
= walk
->vma
;
5986 enum mc_target_type target_type
;
5987 union mc_target target
;
5990 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5992 if (mc
.precharge
< HPAGE_PMD_NR
) {
5996 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
5997 if (target_type
== MC_TARGET_PAGE
) {
5999 if (!isolate_lru_page(page
)) {
6000 if (!mem_cgroup_move_account(page
, true,
6002 mc
.precharge
-= HPAGE_PMD_NR
;
6003 mc
.moved_charge
+= HPAGE_PMD_NR
;
6005 putback_lru_page(page
);
6008 } else if (target_type
== MC_TARGET_DEVICE
) {
6010 if (!mem_cgroup_move_account(page
, true,
6012 mc
.precharge
-= HPAGE_PMD_NR
;
6013 mc
.moved_charge
+= HPAGE_PMD_NR
;
6021 if (pmd_trans_unstable(pmd
))
6024 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6025 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6026 pte_t ptent
= *(pte
++);
6027 bool device
= false;
6033 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6034 case MC_TARGET_DEVICE
:
6037 case MC_TARGET_PAGE
:
6040 * We can have a part of the split pmd here. Moving it
6041 * can be done but it would be too convoluted so simply
6042 * ignore such a partial THP and keep it in original
6043 * memcg. There should be somebody mapping the head.
6045 if (PageTransCompound(page
))
6047 if (!device
&& isolate_lru_page(page
))
6049 if (!mem_cgroup_move_account(page
, false,
6052 /* we uncharge from mc.from later. */
6056 putback_lru_page(page
);
6057 put
: /* get_mctgt_type() gets the page */
6060 case MC_TARGET_SWAP
:
6062 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6064 mem_cgroup_id_get_many(mc
.to
, 1);
6065 /* we fixup other refcnts and charges later. */
6073 pte_unmap_unlock(pte
- 1, ptl
);
6078 * We have consumed all precharges we got in can_attach().
6079 * We try charge one by one, but don't do any additional
6080 * charges to mc.to if we have failed in charge once in attach()
6083 ret
= mem_cgroup_do_precharge(1);
6091 static const struct mm_walk_ops charge_walk_ops
= {
6092 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6095 static void mem_cgroup_move_charge(void)
6097 lru_add_drain_all();
6099 * Signal lock_page_memcg() to take the memcg's move_lock
6100 * while we're moving its pages to another memcg. Then wait
6101 * for already started RCU-only updates to finish.
6103 atomic_inc(&mc
.from
->moving_account
);
6106 if (unlikely(!mmap_read_trylock(mc
.mm
))) {
6108 * Someone who are holding the mmap_lock might be waiting in
6109 * waitq. So we cancel all extra charges, wake up all waiters,
6110 * and retry. Because we cancel precharges, we might not be able
6111 * to move enough charges, but moving charge is a best-effort
6112 * feature anyway, so it wouldn't be a big problem.
6114 __mem_cgroup_clear_mc();
6119 * When we have consumed all precharges and failed in doing
6120 * additional charge, the page walk just aborts.
6122 walk_page_range(mc
.mm
, 0, mc
.mm
->highest_vm_end
, &charge_walk_ops
,
6125 mmap_read_unlock(mc
.mm
);
6126 atomic_dec(&mc
.from
->moving_account
);
6129 static void mem_cgroup_move_task(void)
6132 mem_cgroup_move_charge();
6133 mem_cgroup_clear_mc();
6136 #else /* !CONFIG_MMU */
6137 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
6141 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
6144 static void mem_cgroup_move_task(void)
6150 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6151 * to verify whether we're attached to the default hierarchy on each mount
6154 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
6157 * use_hierarchy is forced on the default hierarchy. cgroup core
6158 * guarantees that @root doesn't have any children, so turning it
6159 * on for the root memcg is enough.
6161 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
6162 root_mem_cgroup
->use_hierarchy
= true;
6164 root_mem_cgroup
->use_hierarchy
= false;
6167 static int seq_puts_memcg_tunable(struct seq_file
*m
, unsigned long value
)
6169 if (value
== PAGE_COUNTER_MAX
)
6170 seq_puts(m
, "max\n");
6172 seq_printf(m
, "%llu\n", (u64
)value
* PAGE_SIZE
);
6177 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
6180 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6182 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
6185 static int memory_min_show(struct seq_file
*m
, void *v
)
6187 return seq_puts_memcg_tunable(m
,
6188 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.min
));
6191 static ssize_t
memory_min_write(struct kernfs_open_file
*of
,
6192 char *buf
, size_t nbytes
, loff_t off
)
6194 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6198 buf
= strstrip(buf
);
6199 err
= page_counter_memparse(buf
, "max", &min
);
6203 page_counter_set_min(&memcg
->memory
, min
);
6208 static int memory_low_show(struct seq_file
*m
, void *v
)
6210 return seq_puts_memcg_tunable(m
,
6211 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.low
));
6214 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
6215 char *buf
, size_t nbytes
, loff_t off
)
6217 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6221 buf
= strstrip(buf
);
6222 err
= page_counter_memparse(buf
, "max", &low
);
6226 page_counter_set_low(&memcg
->memory
, low
);
6231 static int memory_high_show(struct seq_file
*m
, void *v
)
6233 return seq_puts_memcg_tunable(m
,
6234 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.high
));
6237 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
6238 char *buf
, size_t nbytes
, loff_t off
)
6240 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6241 unsigned int nr_retries
= MAX_RECLAIM_RETRIES
;
6242 bool drained
= false;
6246 buf
= strstrip(buf
);
6247 err
= page_counter_memparse(buf
, "max", &high
);
6252 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6253 unsigned long reclaimed
;
6255 if (nr_pages
<= high
)
6258 if (signal_pending(current
))
6262 drain_all_stock(memcg
);
6267 reclaimed
= try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
6270 if (!reclaimed
&& !nr_retries
--)
6274 page_counter_set_high(&memcg
->memory
, high
);
6276 memcg_wb_domain_size_changed(memcg
);
6281 static int memory_max_show(struct seq_file
*m
, void *v
)
6283 return seq_puts_memcg_tunable(m
,
6284 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.max
));
6287 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
6288 char *buf
, size_t nbytes
, loff_t off
)
6290 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6291 unsigned int nr_reclaims
= MAX_RECLAIM_RETRIES
;
6292 bool drained
= false;
6296 buf
= strstrip(buf
);
6297 err
= page_counter_memparse(buf
, "max", &max
);
6301 xchg(&memcg
->memory
.max
, max
);
6304 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6306 if (nr_pages
<= max
)
6309 if (signal_pending(current
))
6313 drain_all_stock(memcg
);
6319 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
6325 memcg_memory_event(memcg
, MEMCG_OOM
);
6326 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
6330 memcg_wb_domain_size_changed(memcg
);
6334 static void __memory_events_show(struct seq_file
*m
, atomic_long_t
*events
)
6336 seq_printf(m
, "low %lu\n", atomic_long_read(&events
[MEMCG_LOW
]));
6337 seq_printf(m
, "high %lu\n", atomic_long_read(&events
[MEMCG_HIGH
]));
6338 seq_printf(m
, "max %lu\n", atomic_long_read(&events
[MEMCG_MAX
]));
6339 seq_printf(m
, "oom %lu\n", atomic_long_read(&events
[MEMCG_OOM
]));
6340 seq_printf(m
, "oom_kill %lu\n",
6341 atomic_long_read(&events
[MEMCG_OOM_KILL
]));
6344 static int memory_events_show(struct seq_file
*m
, void *v
)
6346 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6348 __memory_events_show(m
, memcg
->memory_events
);
6352 static int memory_events_local_show(struct seq_file
*m
, void *v
)
6354 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6356 __memory_events_show(m
, memcg
->memory_events_local
);
6360 static int memory_stat_show(struct seq_file
*m
, void *v
)
6362 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6365 buf
= memory_stat_format(memcg
);
6373 static int memory_oom_group_show(struct seq_file
*m
, void *v
)
6375 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6377 seq_printf(m
, "%d\n", memcg
->oom_group
);
6382 static ssize_t
memory_oom_group_write(struct kernfs_open_file
*of
,
6383 char *buf
, size_t nbytes
, loff_t off
)
6385 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6388 buf
= strstrip(buf
);
6392 ret
= kstrtoint(buf
, 0, &oom_group
);
6396 if (oom_group
!= 0 && oom_group
!= 1)
6399 memcg
->oom_group
= oom_group
;
6404 static struct cftype memory_files
[] = {
6407 .flags
= CFTYPE_NOT_ON_ROOT
,
6408 .read_u64
= memory_current_read
,
6412 .flags
= CFTYPE_NOT_ON_ROOT
,
6413 .seq_show
= memory_min_show
,
6414 .write
= memory_min_write
,
6418 .flags
= CFTYPE_NOT_ON_ROOT
,
6419 .seq_show
= memory_low_show
,
6420 .write
= memory_low_write
,
6424 .flags
= CFTYPE_NOT_ON_ROOT
,
6425 .seq_show
= memory_high_show
,
6426 .write
= memory_high_write
,
6430 .flags
= CFTYPE_NOT_ON_ROOT
,
6431 .seq_show
= memory_max_show
,
6432 .write
= memory_max_write
,
6436 .flags
= CFTYPE_NOT_ON_ROOT
,
6437 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
6438 .seq_show
= memory_events_show
,
6441 .name
= "events.local",
6442 .flags
= CFTYPE_NOT_ON_ROOT
,
6443 .file_offset
= offsetof(struct mem_cgroup
, events_local_file
),
6444 .seq_show
= memory_events_local_show
,
6448 .seq_show
= memory_stat_show
,
6451 .name
= "oom.group",
6452 .flags
= CFTYPE_NOT_ON_ROOT
| CFTYPE_NS_DELEGATABLE
,
6453 .seq_show
= memory_oom_group_show
,
6454 .write
= memory_oom_group_write
,
6459 struct cgroup_subsys memory_cgrp_subsys
= {
6460 .css_alloc
= mem_cgroup_css_alloc
,
6461 .css_online
= mem_cgroup_css_online
,
6462 .css_offline
= mem_cgroup_css_offline
,
6463 .css_released
= mem_cgroup_css_released
,
6464 .css_free
= mem_cgroup_css_free
,
6465 .css_reset
= mem_cgroup_css_reset
,
6466 .can_attach
= mem_cgroup_can_attach
,
6467 .cancel_attach
= mem_cgroup_cancel_attach
,
6468 .post_attach
= mem_cgroup_move_task
,
6469 .bind
= mem_cgroup_bind
,
6470 .dfl_cftypes
= memory_files
,
6471 .legacy_cftypes
= mem_cgroup_legacy_files
,
6476 * This function calculates an individual cgroup's effective
6477 * protection which is derived from its own memory.min/low, its
6478 * parent's and siblings' settings, as well as the actual memory
6479 * distribution in the tree.
6481 * The following rules apply to the effective protection values:
6483 * 1. At the first level of reclaim, effective protection is equal to
6484 * the declared protection in memory.min and memory.low.
6486 * 2. To enable safe delegation of the protection configuration, at
6487 * subsequent levels the effective protection is capped to the
6488 * parent's effective protection.
6490 * 3. To make complex and dynamic subtrees easier to configure, the
6491 * user is allowed to overcommit the declared protection at a given
6492 * level. If that is the case, the parent's effective protection is
6493 * distributed to the children in proportion to how much protection
6494 * they have declared and how much of it they are utilizing.
6496 * This makes distribution proportional, but also work-conserving:
6497 * if one cgroup claims much more protection than it uses memory,
6498 * the unused remainder is available to its siblings.
6500 * 4. Conversely, when the declared protection is undercommitted at a
6501 * given level, the distribution of the larger parental protection
6502 * budget is NOT proportional. A cgroup's protection from a sibling
6503 * is capped to its own memory.min/low setting.
6505 * 5. However, to allow protecting recursive subtrees from each other
6506 * without having to declare each individual cgroup's fixed share
6507 * of the ancestor's claim to protection, any unutilized -
6508 * "floating" - protection from up the tree is distributed in
6509 * proportion to each cgroup's *usage*. This makes the protection
6510 * neutral wrt sibling cgroups and lets them compete freely over
6511 * the shared parental protection budget, but it protects the
6512 * subtree as a whole from neighboring subtrees.
6514 * Note that 4. and 5. are not in conflict: 4. is about protecting
6515 * against immediate siblings whereas 5. is about protecting against
6516 * neighboring subtrees.
6518 static unsigned long effective_protection(unsigned long usage
,
6519 unsigned long parent_usage
,
6520 unsigned long setting
,
6521 unsigned long parent_effective
,
6522 unsigned long siblings_protected
)
6524 unsigned long protected;
6527 protected = min(usage
, setting
);
6529 * If all cgroups at this level combined claim and use more
6530 * protection then what the parent affords them, distribute
6531 * shares in proportion to utilization.
6533 * We are using actual utilization rather than the statically
6534 * claimed protection in order to be work-conserving: claimed
6535 * but unused protection is available to siblings that would
6536 * otherwise get a smaller chunk than what they claimed.
6538 if (siblings_protected
> parent_effective
)
6539 return protected * parent_effective
/ siblings_protected
;
6542 * Ok, utilized protection of all children is within what the
6543 * parent affords them, so we know whatever this child claims
6544 * and utilizes is effectively protected.
6546 * If there is unprotected usage beyond this value, reclaim
6547 * will apply pressure in proportion to that amount.
6549 * If there is unutilized protection, the cgroup will be fully
6550 * shielded from reclaim, but we do return a smaller value for
6551 * protection than what the group could enjoy in theory. This
6552 * is okay. With the overcommit distribution above, effective
6553 * protection is always dependent on how memory is actually
6554 * consumed among the siblings anyway.
6559 * If the children aren't claiming (all of) the protection
6560 * afforded to them by the parent, distribute the remainder in
6561 * proportion to the (unprotected) memory of each cgroup. That
6562 * way, cgroups that aren't explicitly prioritized wrt each
6563 * other compete freely over the allowance, but they are
6564 * collectively protected from neighboring trees.
6566 * We're using unprotected memory for the weight so that if
6567 * some cgroups DO claim explicit protection, we don't protect
6568 * the same bytes twice.
6570 * Check both usage and parent_usage against the respective
6571 * protected values. One should imply the other, but they
6572 * aren't read atomically - make sure the division is sane.
6574 if (!(cgrp_dfl_root
.flags
& CGRP_ROOT_MEMORY_RECURSIVE_PROT
))
6576 if (parent_effective
> siblings_protected
&&
6577 parent_usage
> siblings_protected
&&
6578 usage
> protected) {
6579 unsigned long unclaimed
;
6581 unclaimed
= parent_effective
- siblings_protected
;
6582 unclaimed
*= usage
- protected;
6583 unclaimed
/= parent_usage
- siblings_protected
;
6592 * mem_cgroup_protected - check if memory consumption is in the normal range
6593 * @root: the top ancestor of the sub-tree being checked
6594 * @memcg: the memory cgroup to check
6596 * WARNING: This function is not stateless! It can only be used as part
6597 * of a top-down tree iteration, not for isolated queries.
6599 void mem_cgroup_calculate_protection(struct mem_cgroup
*root
,
6600 struct mem_cgroup
*memcg
)
6602 unsigned long usage
, parent_usage
;
6603 struct mem_cgroup
*parent
;
6605 if (mem_cgroup_disabled())
6609 root
= root_mem_cgroup
;
6612 * Effective values of the reclaim targets are ignored so they
6613 * can be stale. Have a look at mem_cgroup_protection for more
6615 * TODO: calculation should be more robust so that we do not need
6616 * that special casing.
6621 usage
= page_counter_read(&memcg
->memory
);
6625 parent
= parent_mem_cgroup(memcg
);
6626 /* No parent means a non-hierarchical mode on v1 memcg */
6630 if (parent
== root
) {
6631 memcg
->memory
.emin
= READ_ONCE(memcg
->memory
.min
);
6632 memcg
->memory
.elow
= READ_ONCE(memcg
->memory
.low
);
6636 parent_usage
= page_counter_read(&parent
->memory
);
6638 WRITE_ONCE(memcg
->memory
.emin
, effective_protection(usage
, parent_usage
,
6639 READ_ONCE(memcg
->memory
.min
),
6640 READ_ONCE(parent
->memory
.emin
),
6641 atomic_long_read(&parent
->memory
.children_min_usage
)));
6643 WRITE_ONCE(memcg
->memory
.elow
, effective_protection(usage
, parent_usage
,
6644 READ_ONCE(memcg
->memory
.low
),
6645 READ_ONCE(parent
->memory
.elow
),
6646 atomic_long_read(&parent
->memory
.children_low_usage
)));
6650 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6651 * @page: page to charge
6652 * @mm: mm context of the victim
6653 * @gfp_mask: reclaim mode
6655 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6656 * pages according to @gfp_mask if necessary.
6658 * Returns 0 on success. Otherwise, an error code is returned.
6660 int mem_cgroup_charge(struct page
*page
, struct mm_struct
*mm
, gfp_t gfp_mask
)
6662 unsigned int nr_pages
= hpage_nr_pages(page
);
6663 struct mem_cgroup
*memcg
= NULL
;
6666 if (mem_cgroup_disabled())
6669 if (PageSwapCache(page
)) {
6670 swp_entry_t ent
= { .val
= page_private(page
), };
6674 * Every swap fault against a single page tries to charge the
6675 * page, bail as early as possible. shmem_unuse() encounters
6676 * already charged pages, too. page->mem_cgroup is protected
6677 * by the page lock, which serializes swap cache removal, which
6678 * in turn serializes uncharging.
6680 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
6681 if (compound_head(page
)->mem_cgroup
)
6684 id
= lookup_swap_cgroup_id(ent
);
6686 memcg
= mem_cgroup_from_id(id
);
6687 if (memcg
&& !css_tryget_online(&memcg
->css
))
6693 memcg
= get_mem_cgroup_from_mm(mm
);
6695 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
6699 css_get(&memcg
->css
);
6700 commit_charge(page
, memcg
);
6702 local_irq_disable();
6703 mem_cgroup_charge_statistics(memcg
, page
, nr_pages
);
6704 memcg_check_events(memcg
, page
);
6707 if (PageSwapCache(page
)) {
6708 swp_entry_t entry
= { .val
= page_private(page
) };
6710 * The swap entry might not get freed for a long time,
6711 * let's not wait for it. The page already received a
6712 * memory+swap charge, drop the swap entry duplicate.
6714 mem_cgroup_uncharge_swap(entry
, nr_pages
);
6718 css_put(&memcg
->css
);
6723 struct uncharge_gather
{
6724 struct mem_cgroup
*memcg
;
6725 unsigned long nr_pages
;
6726 unsigned long pgpgout
;
6727 unsigned long nr_kmem
;
6728 struct page
*dummy_page
;
6731 static inline void uncharge_gather_clear(struct uncharge_gather
*ug
)
6733 memset(ug
, 0, sizeof(*ug
));
6736 static void uncharge_batch(const struct uncharge_gather
*ug
)
6738 unsigned long flags
;
6740 if (!mem_cgroup_is_root(ug
->memcg
)) {
6741 page_counter_uncharge(&ug
->memcg
->memory
, ug
->nr_pages
);
6742 if (do_memsw_account())
6743 page_counter_uncharge(&ug
->memcg
->memsw
, ug
->nr_pages
);
6744 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && ug
->nr_kmem
)
6745 page_counter_uncharge(&ug
->memcg
->kmem
, ug
->nr_kmem
);
6746 memcg_oom_recover(ug
->memcg
);
6749 local_irq_save(flags
);
6750 __count_memcg_events(ug
->memcg
, PGPGOUT
, ug
->pgpgout
);
6751 __this_cpu_add(ug
->memcg
->vmstats_percpu
->nr_page_events
, ug
->nr_pages
);
6752 memcg_check_events(ug
->memcg
, ug
->dummy_page
);
6753 local_irq_restore(flags
);
6756 static void uncharge_page(struct page
*page
, struct uncharge_gather
*ug
)
6758 unsigned long nr_pages
;
6760 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6762 if (!page
->mem_cgroup
)
6766 * Nobody should be changing or seriously looking at
6767 * page->mem_cgroup at this point, we have fully
6768 * exclusive access to the page.
6771 if (ug
->memcg
!= page
->mem_cgroup
) {
6774 uncharge_gather_clear(ug
);
6776 ug
->memcg
= page
->mem_cgroup
;
6779 nr_pages
= compound_nr(page
);
6780 ug
->nr_pages
+= nr_pages
;
6782 if (!PageKmemcg(page
)) {
6785 ug
->nr_kmem
+= nr_pages
;
6786 __ClearPageKmemcg(page
);
6789 ug
->dummy_page
= page
;
6790 page
->mem_cgroup
= NULL
;
6791 css_put(&ug
->memcg
->css
);
6794 static void uncharge_list(struct list_head
*page_list
)
6796 struct uncharge_gather ug
;
6797 struct list_head
*next
;
6799 uncharge_gather_clear(&ug
);
6802 * Note that the list can be a single page->lru; hence the
6803 * do-while loop instead of a simple list_for_each_entry().
6805 next
= page_list
->next
;
6809 page
= list_entry(next
, struct page
, lru
);
6810 next
= page
->lru
.next
;
6812 uncharge_page(page
, &ug
);
6813 } while (next
!= page_list
);
6816 uncharge_batch(&ug
);
6820 * mem_cgroup_uncharge - uncharge a page
6821 * @page: page to uncharge
6823 * Uncharge a page previously charged with mem_cgroup_charge().
6825 void mem_cgroup_uncharge(struct page
*page
)
6827 struct uncharge_gather ug
;
6829 if (mem_cgroup_disabled())
6832 /* Don't touch page->lru of any random page, pre-check: */
6833 if (!page
->mem_cgroup
)
6836 uncharge_gather_clear(&ug
);
6837 uncharge_page(page
, &ug
);
6838 uncharge_batch(&ug
);
6842 * mem_cgroup_uncharge_list - uncharge a list of page
6843 * @page_list: list of pages to uncharge
6845 * Uncharge a list of pages previously charged with
6846 * mem_cgroup_charge().
6848 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
6850 if (mem_cgroup_disabled())
6853 if (!list_empty(page_list
))
6854 uncharge_list(page_list
);
6858 * mem_cgroup_migrate - charge a page's replacement
6859 * @oldpage: currently circulating page
6860 * @newpage: replacement page
6862 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6863 * be uncharged upon free.
6865 * Both pages must be locked, @newpage->mapping must be set up.
6867 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
6869 struct mem_cgroup
*memcg
;
6870 unsigned int nr_pages
;
6871 unsigned long flags
;
6873 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
6874 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
6875 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
6876 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
6879 if (mem_cgroup_disabled())
6882 /* Page cache replacement: new page already charged? */
6883 if (newpage
->mem_cgroup
)
6886 /* Swapcache readahead pages can get replaced before being charged */
6887 memcg
= oldpage
->mem_cgroup
;
6891 /* Force-charge the new page. The old one will be freed soon */
6892 nr_pages
= hpage_nr_pages(newpage
);
6894 page_counter_charge(&memcg
->memory
, nr_pages
);
6895 if (do_memsw_account())
6896 page_counter_charge(&memcg
->memsw
, nr_pages
);
6898 css_get(&memcg
->css
);
6899 commit_charge(newpage
, memcg
);
6901 local_irq_save(flags
);
6902 mem_cgroup_charge_statistics(memcg
, newpage
, nr_pages
);
6903 memcg_check_events(memcg
, newpage
);
6904 local_irq_restore(flags
);
6907 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
6908 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
6910 void mem_cgroup_sk_alloc(struct sock
*sk
)
6912 struct mem_cgroup
*memcg
;
6914 if (!mem_cgroup_sockets_enabled
)
6917 /* Do not associate the sock with unrelated interrupted task's memcg. */
6922 memcg
= mem_cgroup_from_task(current
);
6923 if (memcg
== root_mem_cgroup
)
6925 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
6927 if (css_tryget(&memcg
->css
))
6928 sk
->sk_memcg
= memcg
;
6933 void mem_cgroup_sk_free(struct sock
*sk
)
6936 css_put(&sk
->sk_memcg
->css
);
6940 * mem_cgroup_charge_skmem - charge socket memory
6941 * @memcg: memcg to charge
6942 * @nr_pages: number of pages to charge
6944 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6945 * @memcg's configured limit, %false if the charge had to be forced.
6947 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
6949 gfp_t gfp_mask
= GFP_KERNEL
;
6951 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
6952 struct page_counter
*fail
;
6954 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
6955 memcg
->tcpmem_pressure
= 0;
6958 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
6959 memcg
->tcpmem_pressure
= 1;
6963 /* Don't block in the packet receive path */
6965 gfp_mask
= GFP_NOWAIT
;
6967 mod_memcg_state(memcg
, MEMCG_SOCK
, nr_pages
);
6969 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
6972 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
6977 * mem_cgroup_uncharge_skmem - uncharge socket memory
6978 * @memcg: memcg to uncharge
6979 * @nr_pages: number of pages to uncharge
6981 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
6983 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
6984 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
6988 mod_memcg_state(memcg
, MEMCG_SOCK
, -nr_pages
);
6990 refill_stock(memcg
, nr_pages
);
6993 static int __init
cgroup_memory(char *s
)
6997 while ((token
= strsep(&s
, ",")) != NULL
) {
7000 if (!strcmp(token
, "nosocket"))
7001 cgroup_memory_nosocket
= true;
7002 if (!strcmp(token
, "nokmem"))
7003 cgroup_memory_nokmem
= true;
7007 __setup("cgroup.memory=", cgroup_memory
);
7010 * subsys_initcall() for memory controller.
7012 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7013 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7014 * basically everything that doesn't depend on a specific mem_cgroup structure
7015 * should be initialized from here.
7017 static int __init
mem_cgroup_init(void)
7021 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD
, "mm/memctrl:dead", NULL
,
7022 memcg_hotplug_cpu_dead
);
7024 for_each_possible_cpu(cpu
)
7025 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
7028 for_each_node(node
) {
7029 struct mem_cgroup_tree_per_node
*rtpn
;
7031 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
7032 node_online(node
) ? node
: NUMA_NO_NODE
);
7034 rtpn
->rb_root
= RB_ROOT
;
7035 rtpn
->rb_rightmost
= NULL
;
7036 spin_lock_init(&rtpn
->lock
);
7037 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
7042 subsys_initcall(mem_cgroup_init
);
7044 #ifdef CONFIG_MEMCG_SWAP
7045 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
7047 while (!refcount_inc_not_zero(&memcg
->id
.ref
)) {
7049 * The root cgroup cannot be destroyed, so it's refcount must
7052 if (WARN_ON_ONCE(memcg
== root_mem_cgroup
)) {
7056 memcg
= parent_mem_cgroup(memcg
);
7058 memcg
= root_mem_cgroup
;
7064 * mem_cgroup_swapout - transfer a memsw charge to swap
7065 * @page: page whose memsw charge to transfer
7066 * @entry: swap entry to move the charge to
7068 * Transfer the memsw charge of @page to @entry.
7070 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
7072 struct mem_cgroup
*memcg
, *swap_memcg
;
7073 unsigned int nr_entries
;
7074 unsigned short oldid
;
7076 VM_BUG_ON_PAGE(PageLRU(page
), page
);
7077 VM_BUG_ON_PAGE(page_count(page
), page
);
7079 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7082 memcg
= page
->mem_cgroup
;
7084 /* Readahead page, never charged */
7089 * In case the memcg owning these pages has been offlined and doesn't
7090 * have an ID allocated to it anymore, charge the closest online
7091 * ancestor for the swap instead and transfer the memory+swap charge.
7093 swap_memcg
= mem_cgroup_id_get_online(memcg
);
7094 nr_entries
= hpage_nr_pages(page
);
7095 /* Get references for the tail pages, too */
7097 mem_cgroup_id_get_many(swap_memcg
, nr_entries
- 1);
7098 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
),
7100 VM_BUG_ON_PAGE(oldid
, page
);
7101 mod_memcg_state(swap_memcg
, MEMCG_SWAP
, nr_entries
);
7103 page
->mem_cgroup
= NULL
;
7105 if (!mem_cgroup_is_root(memcg
))
7106 page_counter_uncharge(&memcg
->memory
, nr_entries
);
7108 if (!cgroup_memory_noswap
&& memcg
!= swap_memcg
) {
7109 if (!mem_cgroup_is_root(swap_memcg
))
7110 page_counter_charge(&swap_memcg
->memsw
, nr_entries
);
7111 page_counter_uncharge(&memcg
->memsw
, nr_entries
);
7115 * Interrupts should be disabled here because the caller holds the
7116 * i_pages lock which is taken with interrupts-off. It is
7117 * important here to have the interrupts disabled because it is the
7118 * only synchronisation we have for updating the per-CPU variables.
7120 VM_BUG_ON(!irqs_disabled());
7121 mem_cgroup_charge_statistics(memcg
, page
, -nr_entries
);
7122 memcg_check_events(memcg
, page
);
7124 css_put(&memcg
->css
);
7128 * mem_cgroup_try_charge_swap - try charging swap space for a page
7129 * @page: page being added to swap
7130 * @entry: swap entry to charge
7132 * Try to charge @page's memcg for the swap space at @entry.
7134 * Returns 0 on success, -ENOMEM on failure.
7136 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
7138 unsigned int nr_pages
= hpage_nr_pages(page
);
7139 struct page_counter
*counter
;
7140 struct mem_cgroup
*memcg
;
7141 unsigned short oldid
;
7143 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7146 memcg
= page
->mem_cgroup
;
7148 /* Readahead page, never charged */
7153 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7157 memcg
= mem_cgroup_id_get_online(memcg
);
7159 if (!cgroup_memory_noswap
&& !mem_cgroup_is_root(memcg
) &&
7160 !page_counter_try_charge(&memcg
->swap
, nr_pages
, &counter
)) {
7161 memcg_memory_event(memcg
, MEMCG_SWAP_MAX
);
7162 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7163 mem_cgroup_id_put(memcg
);
7167 /* Get references for the tail pages, too */
7169 mem_cgroup_id_get_many(memcg
, nr_pages
- 1);
7170 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
), nr_pages
);
7171 VM_BUG_ON_PAGE(oldid
, page
);
7172 mod_memcg_state(memcg
, MEMCG_SWAP
, nr_pages
);
7178 * mem_cgroup_uncharge_swap - uncharge swap space
7179 * @entry: swap entry to uncharge
7180 * @nr_pages: the amount of swap space to uncharge
7182 void mem_cgroup_uncharge_swap(swp_entry_t entry
, unsigned int nr_pages
)
7184 struct mem_cgroup
*memcg
;
7187 id
= swap_cgroup_record(entry
, 0, nr_pages
);
7189 memcg
= mem_cgroup_from_id(id
);
7191 if (!cgroup_memory_noswap
&& !mem_cgroup_is_root(memcg
)) {
7192 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7193 page_counter_uncharge(&memcg
->swap
, nr_pages
);
7195 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
7197 mod_memcg_state(memcg
, MEMCG_SWAP
, -nr_pages
);
7198 mem_cgroup_id_put_many(memcg
, nr_pages
);
7203 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
7205 long nr_swap_pages
= get_nr_swap_pages();
7207 if (cgroup_memory_noswap
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7208 return nr_swap_pages
;
7209 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
7210 nr_swap_pages
= min_t(long, nr_swap_pages
,
7211 READ_ONCE(memcg
->swap
.max
) -
7212 page_counter_read(&memcg
->swap
));
7213 return nr_swap_pages
;
7216 bool mem_cgroup_swap_full(struct page
*page
)
7218 struct mem_cgroup
*memcg
;
7220 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
7224 if (cgroup_memory_noswap
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7227 memcg
= page
->mem_cgroup
;
7231 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
)) {
7232 unsigned long usage
= page_counter_read(&memcg
->swap
);
7234 if (usage
* 2 >= READ_ONCE(memcg
->swap
.high
) ||
7235 usage
* 2 >= READ_ONCE(memcg
->swap
.max
))
7242 static int __init
setup_swap_account(char *s
)
7244 if (!strcmp(s
, "1"))
7245 cgroup_memory_noswap
= 0;
7246 else if (!strcmp(s
, "0"))
7247 cgroup_memory_noswap
= 1;
7250 __setup("swapaccount=", setup_swap_account
);
7252 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
7255 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
7257 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
7260 static int swap_high_show(struct seq_file
*m
, void *v
)
7262 return seq_puts_memcg_tunable(m
,
7263 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.high
));
7266 static ssize_t
swap_high_write(struct kernfs_open_file
*of
,
7267 char *buf
, size_t nbytes
, loff_t off
)
7269 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7273 buf
= strstrip(buf
);
7274 err
= page_counter_memparse(buf
, "max", &high
);
7278 page_counter_set_high(&memcg
->swap
, high
);
7283 static int swap_max_show(struct seq_file
*m
, void *v
)
7285 return seq_puts_memcg_tunable(m
,
7286 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.max
));
7289 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
7290 char *buf
, size_t nbytes
, loff_t off
)
7292 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7296 buf
= strstrip(buf
);
7297 err
= page_counter_memparse(buf
, "max", &max
);
7301 xchg(&memcg
->swap
.max
, max
);
7306 static int swap_events_show(struct seq_file
*m
, void *v
)
7308 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
7310 seq_printf(m
, "high %lu\n",
7311 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_HIGH
]));
7312 seq_printf(m
, "max %lu\n",
7313 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_MAX
]));
7314 seq_printf(m
, "fail %lu\n",
7315 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_FAIL
]));
7320 static struct cftype swap_files
[] = {
7322 .name
= "swap.current",
7323 .flags
= CFTYPE_NOT_ON_ROOT
,
7324 .read_u64
= swap_current_read
,
7327 .name
= "swap.high",
7328 .flags
= CFTYPE_NOT_ON_ROOT
,
7329 .seq_show
= swap_high_show
,
7330 .write
= swap_high_write
,
7334 .flags
= CFTYPE_NOT_ON_ROOT
,
7335 .seq_show
= swap_max_show
,
7336 .write
= swap_max_write
,
7339 .name
= "swap.events",
7340 .flags
= CFTYPE_NOT_ON_ROOT
,
7341 .file_offset
= offsetof(struct mem_cgroup
, swap_events_file
),
7342 .seq_show
= swap_events_show
,
7347 static struct cftype memsw_files
[] = {
7349 .name
= "memsw.usage_in_bytes",
7350 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
7351 .read_u64
= mem_cgroup_read_u64
,
7354 .name
= "memsw.max_usage_in_bytes",
7355 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
7356 .write
= mem_cgroup_reset
,
7357 .read_u64
= mem_cgroup_read_u64
,
7360 .name
= "memsw.limit_in_bytes",
7361 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
7362 .write
= mem_cgroup_write
,
7363 .read_u64
= mem_cgroup_read_u64
,
7366 .name
= "memsw.failcnt",
7367 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
7368 .write
= mem_cgroup_reset
,
7369 .read_u64
= mem_cgroup_read_u64
,
7371 { }, /* terminate */
7375 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7376 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7377 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7378 * boot parameter. This may result in premature OOPS inside
7379 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7381 static int __init
mem_cgroup_swap_init(void)
7383 /* No memory control -> no swap control */
7384 if (mem_cgroup_disabled())
7385 cgroup_memory_noswap
= true;
7387 if (cgroup_memory_noswap
)
7390 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
, swap_files
));
7391 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
, memsw_files
));
7395 core_initcall(mem_cgroup_swap_init
);
7397 #endif /* CONFIG_MEMCG_SWAP */