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 /* Active memory cgroup to use from an interrupt context */
77 DEFINE_PER_CPU(struct mem_cgroup
*, int_active_memcg
);
79 /* Socket memory accounting disabled? */
80 static bool cgroup_memory_nosocket
;
82 /* Kernel memory accounting disabled? */
83 static bool cgroup_memory_nokmem
;
85 /* Whether the swap controller is active */
86 #ifdef CONFIG_MEMCG_SWAP
87 bool cgroup_memory_noswap __read_mostly
;
89 #define cgroup_memory_noswap 1
92 #ifdef CONFIG_CGROUP_WRITEBACK
93 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq
);
96 /* Whether legacy memory+swap accounting is active */
97 static bool do_memsw_account(void)
99 return !cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_noswap
;
102 #define THRESHOLDS_EVENTS_TARGET 128
103 #define SOFTLIMIT_EVENTS_TARGET 1024
106 * Cgroups above their limits are maintained in a RB-Tree, independent of
107 * their hierarchy representation
110 struct mem_cgroup_tree_per_node
{
111 struct rb_root rb_root
;
112 struct rb_node
*rb_rightmost
;
116 struct mem_cgroup_tree
{
117 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
120 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
123 struct mem_cgroup_eventfd_list
{
124 struct list_head list
;
125 struct eventfd_ctx
*eventfd
;
129 * cgroup_event represents events which userspace want to receive.
131 struct mem_cgroup_event
{
133 * memcg which the event belongs to.
135 struct mem_cgroup
*memcg
;
137 * eventfd to signal userspace about the event.
139 struct eventfd_ctx
*eventfd
;
141 * Each of these stored in a list by the cgroup.
143 struct list_head list
;
145 * register_event() callback will be used to add new userspace
146 * waiter for changes related to this event. Use eventfd_signal()
147 * on eventfd to send notification to userspace.
149 int (*register_event
)(struct mem_cgroup
*memcg
,
150 struct eventfd_ctx
*eventfd
, const char *args
);
152 * unregister_event() callback will be called when userspace closes
153 * the eventfd or on cgroup removing. This callback must be set,
154 * if you want provide notification functionality.
156 void (*unregister_event
)(struct mem_cgroup
*memcg
,
157 struct eventfd_ctx
*eventfd
);
159 * All fields below needed to unregister event when
160 * userspace closes eventfd.
163 wait_queue_head_t
*wqh
;
164 wait_queue_entry_t wait
;
165 struct work_struct remove
;
168 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
169 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
171 /* Stuffs for move charges at task migration. */
173 * Types of charges to be moved.
175 #define MOVE_ANON 0x1U
176 #define MOVE_FILE 0x2U
177 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
179 /* "mc" and its members are protected by cgroup_mutex */
180 static struct move_charge_struct
{
181 spinlock_t lock
; /* for from, to */
182 struct mm_struct
*mm
;
183 struct mem_cgroup
*from
;
184 struct mem_cgroup
*to
;
186 unsigned long precharge
;
187 unsigned long moved_charge
;
188 unsigned long moved_swap
;
189 struct task_struct
*moving_task
; /* a task moving charges */
190 wait_queue_head_t waitq
; /* a waitq for other context */
192 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
193 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
197 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
198 * limit reclaim to prevent infinite loops, if they ever occur.
200 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
201 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
203 /* for encoding cft->private value on file */
212 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
213 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
214 #define MEMFILE_ATTR(val) ((val) & 0xffff)
215 /* Used for OOM nofiier */
216 #define OOM_CONTROL (0)
219 * Iteration constructs for visiting all cgroups (under a tree). If
220 * loops are exited prematurely (break), mem_cgroup_iter_break() must
221 * be used for reference counting.
223 #define for_each_mem_cgroup_tree(iter, root) \
224 for (iter = mem_cgroup_iter(root, NULL, NULL); \
226 iter = mem_cgroup_iter(root, iter, NULL))
228 #define for_each_mem_cgroup(iter) \
229 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
231 iter = mem_cgroup_iter(NULL, iter, NULL))
233 static inline bool should_force_charge(void)
235 return tsk_is_oom_victim(current
) || fatal_signal_pending(current
) ||
236 (current
->flags
& PF_EXITING
);
239 /* Some nice accessors for the vmpressure. */
240 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
243 memcg
= root_mem_cgroup
;
244 return &memcg
->vmpressure
;
247 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
249 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
252 #ifdef CONFIG_MEMCG_KMEM
253 extern spinlock_t css_set_lock
;
255 static void obj_cgroup_release(struct percpu_ref
*ref
)
257 struct obj_cgroup
*objcg
= container_of(ref
, struct obj_cgroup
, refcnt
);
258 struct mem_cgroup
*memcg
;
259 unsigned int nr_bytes
;
260 unsigned int nr_pages
;
264 * At this point all allocated objects are freed, and
265 * objcg->nr_charged_bytes can't have an arbitrary byte value.
266 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
268 * The following sequence can lead to it:
269 * 1) CPU0: objcg == stock->cached_objcg
270 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
271 * PAGE_SIZE bytes are charged
272 * 3) CPU1: a process from another memcg is allocating something,
273 * the stock if flushed,
274 * objcg->nr_charged_bytes = PAGE_SIZE - 92
275 * 5) CPU0: we do release this object,
276 * 92 bytes are added to stock->nr_bytes
277 * 6) CPU0: stock is flushed,
278 * 92 bytes are added to objcg->nr_charged_bytes
280 * In the result, nr_charged_bytes == PAGE_SIZE.
281 * This page will be uncharged in obj_cgroup_release().
283 nr_bytes
= atomic_read(&objcg
->nr_charged_bytes
);
284 WARN_ON_ONCE(nr_bytes
& (PAGE_SIZE
- 1));
285 nr_pages
= nr_bytes
>> PAGE_SHIFT
;
287 spin_lock_irqsave(&css_set_lock
, flags
);
288 memcg
= obj_cgroup_memcg(objcg
);
290 __memcg_kmem_uncharge(memcg
, nr_pages
);
291 list_del(&objcg
->list
);
292 mem_cgroup_put(memcg
);
293 spin_unlock_irqrestore(&css_set_lock
, flags
);
295 percpu_ref_exit(ref
);
296 kfree_rcu(objcg
, rcu
);
299 static struct obj_cgroup
*obj_cgroup_alloc(void)
301 struct obj_cgroup
*objcg
;
304 objcg
= kzalloc(sizeof(struct obj_cgroup
), GFP_KERNEL
);
308 ret
= percpu_ref_init(&objcg
->refcnt
, obj_cgroup_release
, 0,
314 INIT_LIST_HEAD(&objcg
->list
);
318 static void memcg_reparent_objcgs(struct mem_cgroup
*memcg
,
319 struct mem_cgroup
*parent
)
321 struct obj_cgroup
*objcg
, *iter
;
323 objcg
= rcu_replace_pointer(memcg
->objcg
, NULL
, true);
325 spin_lock_irq(&css_set_lock
);
327 /* Move active objcg to the parent's list */
328 xchg(&objcg
->memcg
, parent
);
329 css_get(&parent
->css
);
330 list_add(&objcg
->list
, &parent
->objcg_list
);
332 /* Move already reparented objcgs to the parent's list */
333 list_for_each_entry(iter
, &memcg
->objcg_list
, list
) {
334 css_get(&parent
->css
);
335 xchg(&iter
->memcg
, parent
);
336 css_put(&memcg
->css
);
338 list_splice(&memcg
->objcg_list
, &parent
->objcg_list
);
340 spin_unlock_irq(&css_set_lock
);
342 percpu_ref_kill(&objcg
->refcnt
);
346 * This will be used as a shrinker list's index.
347 * The main reason for not using cgroup id for this:
348 * this works better in sparse environments, where we have a lot of memcgs,
349 * but only a few kmem-limited. Or also, if we have, for instance, 200
350 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
351 * 200 entry array for that.
353 * The current size of the caches array is stored in memcg_nr_cache_ids. It
354 * will double each time we have to increase it.
356 static DEFINE_IDA(memcg_cache_ida
);
357 int memcg_nr_cache_ids
;
359 /* Protects memcg_nr_cache_ids */
360 static DECLARE_RWSEM(memcg_cache_ids_sem
);
362 void memcg_get_cache_ids(void)
364 down_read(&memcg_cache_ids_sem
);
367 void memcg_put_cache_ids(void)
369 up_read(&memcg_cache_ids_sem
);
373 * MIN_SIZE is different than 1, because we would like to avoid going through
374 * the alloc/free process all the time. In a small machine, 4 kmem-limited
375 * cgroups is a reasonable guess. In the future, it could be a parameter or
376 * tunable, but that is strictly not necessary.
378 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
379 * this constant directly from cgroup, but it is understandable that this is
380 * better kept as an internal representation in cgroup.c. In any case, the
381 * cgrp_id space is not getting any smaller, and we don't have to necessarily
382 * increase ours as well if it increases.
384 #define MEMCG_CACHES_MIN_SIZE 4
385 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
388 * A lot of the calls to the cache allocation functions are expected to be
389 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
390 * conditional to this static branch, we'll have to allow modules that does
391 * kmem_cache_alloc and the such to see this symbol as well
393 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
394 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
397 static int memcg_shrinker_map_size
;
398 static DEFINE_MUTEX(memcg_shrinker_map_mutex
);
400 static void memcg_free_shrinker_map_rcu(struct rcu_head
*head
)
402 kvfree(container_of(head
, struct memcg_shrinker_map
, rcu
));
405 static int memcg_expand_one_shrinker_map(struct mem_cgroup
*memcg
,
406 int size
, int old_size
)
408 struct memcg_shrinker_map
*new, *old
;
411 lockdep_assert_held(&memcg_shrinker_map_mutex
);
414 old
= rcu_dereference_protected(
415 mem_cgroup_nodeinfo(memcg
, nid
)->shrinker_map
, true);
416 /* Not yet online memcg */
420 new = kvmalloc_node(sizeof(*new) + size
, GFP_KERNEL
, nid
);
424 /* Set all old bits, clear all new bits */
425 memset(new->map
, (int)0xff, old_size
);
426 memset((void *)new->map
+ old_size
, 0, size
- old_size
);
428 rcu_assign_pointer(memcg
->nodeinfo
[nid
]->shrinker_map
, new);
429 call_rcu(&old
->rcu
, memcg_free_shrinker_map_rcu
);
435 static void memcg_free_shrinker_maps(struct mem_cgroup
*memcg
)
437 struct mem_cgroup_per_node
*pn
;
438 struct memcg_shrinker_map
*map
;
441 if (mem_cgroup_is_root(memcg
))
445 pn
= mem_cgroup_nodeinfo(memcg
, nid
);
446 map
= rcu_dereference_protected(pn
->shrinker_map
, true);
449 rcu_assign_pointer(pn
->shrinker_map
, NULL
);
453 static int memcg_alloc_shrinker_maps(struct mem_cgroup
*memcg
)
455 struct memcg_shrinker_map
*map
;
456 int nid
, size
, ret
= 0;
458 if (mem_cgroup_is_root(memcg
))
461 mutex_lock(&memcg_shrinker_map_mutex
);
462 size
= memcg_shrinker_map_size
;
464 map
= kvzalloc_node(sizeof(*map
) + size
, GFP_KERNEL
, nid
);
466 memcg_free_shrinker_maps(memcg
);
470 rcu_assign_pointer(memcg
->nodeinfo
[nid
]->shrinker_map
, map
);
472 mutex_unlock(&memcg_shrinker_map_mutex
);
477 int memcg_expand_shrinker_maps(int new_id
)
479 int size
, old_size
, ret
= 0;
480 struct mem_cgroup
*memcg
;
482 size
= DIV_ROUND_UP(new_id
+ 1, BITS_PER_LONG
) * sizeof(unsigned long);
483 old_size
= memcg_shrinker_map_size
;
484 if (size
<= old_size
)
487 mutex_lock(&memcg_shrinker_map_mutex
);
488 if (!root_mem_cgroup
)
491 for_each_mem_cgroup(memcg
) {
492 if (mem_cgroup_is_root(memcg
))
494 ret
= memcg_expand_one_shrinker_map(memcg
, size
, old_size
);
496 mem_cgroup_iter_break(NULL
, memcg
);
502 memcg_shrinker_map_size
= size
;
503 mutex_unlock(&memcg_shrinker_map_mutex
);
507 void memcg_set_shrinker_bit(struct mem_cgroup
*memcg
, int nid
, int shrinker_id
)
509 if (shrinker_id
>= 0 && memcg
&& !mem_cgroup_is_root(memcg
)) {
510 struct memcg_shrinker_map
*map
;
513 map
= rcu_dereference(memcg
->nodeinfo
[nid
]->shrinker_map
);
514 /* Pairs with smp mb in shrink_slab() */
515 smp_mb__before_atomic();
516 set_bit(shrinker_id
, map
->map
);
522 * mem_cgroup_css_from_page - css of the memcg associated with a page
523 * @page: page of interest
525 * If memcg is bound to the default hierarchy, css of the memcg associated
526 * with @page is returned. The returned css remains associated with @page
527 * until it is released.
529 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
532 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
534 struct mem_cgroup
*memcg
;
536 memcg
= page
->mem_cgroup
;
538 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
539 memcg
= root_mem_cgroup
;
545 * page_cgroup_ino - return inode number of the memcg a page is charged to
548 * Look up the closest online ancestor of the memory cgroup @page is charged to
549 * and return its inode number or 0 if @page is not charged to any cgroup. It
550 * is safe to call this function without holding a reference to @page.
552 * Note, this function is inherently racy, because there is nothing to prevent
553 * the cgroup inode from getting torn down and potentially reallocated a moment
554 * after page_cgroup_ino() returns, so it only should be used by callers that
555 * do not care (such as procfs interfaces).
557 ino_t
page_cgroup_ino(struct page
*page
)
559 struct mem_cgroup
*memcg
;
560 unsigned long ino
= 0;
563 memcg
= page
->mem_cgroup
;
566 * The lowest bit set means that memcg isn't a valid
567 * memcg pointer, but a obj_cgroups pointer.
568 * In this case the page is shared and doesn't belong
569 * to any specific memory cgroup.
571 if ((unsigned long) memcg
& 0x1UL
)
574 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
575 memcg
= parent_mem_cgroup(memcg
);
577 ino
= cgroup_ino(memcg
->css
.cgroup
);
582 static struct mem_cgroup_per_node
*
583 mem_cgroup_page_nodeinfo(struct mem_cgroup
*memcg
, struct page
*page
)
585 int nid
= page_to_nid(page
);
587 return memcg
->nodeinfo
[nid
];
590 static struct mem_cgroup_tree_per_node
*
591 soft_limit_tree_node(int nid
)
593 return soft_limit_tree
.rb_tree_per_node
[nid
];
596 static struct mem_cgroup_tree_per_node
*
597 soft_limit_tree_from_page(struct page
*page
)
599 int nid
= page_to_nid(page
);
601 return soft_limit_tree
.rb_tree_per_node
[nid
];
604 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node
*mz
,
605 struct mem_cgroup_tree_per_node
*mctz
,
606 unsigned long new_usage_in_excess
)
608 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
609 struct rb_node
*parent
= NULL
;
610 struct mem_cgroup_per_node
*mz_node
;
611 bool rightmost
= true;
616 mz
->usage_in_excess
= new_usage_in_excess
;
617 if (!mz
->usage_in_excess
)
621 mz_node
= rb_entry(parent
, struct mem_cgroup_per_node
,
623 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
) {
629 * We can't avoid mem cgroups that are over their soft
630 * limit by the same amount
632 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
637 mctz
->rb_rightmost
= &mz
->tree_node
;
639 rb_link_node(&mz
->tree_node
, parent
, p
);
640 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
644 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
645 struct mem_cgroup_tree_per_node
*mctz
)
650 if (&mz
->tree_node
== mctz
->rb_rightmost
)
651 mctz
->rb_rightmost
= rb_prev(&mz
->tree_node
);
653 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
657 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
658 struct mem_cgroup_tree_per_node
*mctz
)
662 spin_lock_irqsave(&mctz
->lock
, flags
);
663 __mem_cgroup_remove_exceeded(mz
, mctz
);
664 spin_unlock_irqrestore(&mctz
->lock
, flags
);
667 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
669 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
670 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
671 unsigned long excess
= 0;
673 if (nr_pages
> soft_limit
)
674 excess
= nr_pages
- soft_limit
;
679 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
681 unsigned long excess
;
682 struct mem_cgroup_per_node
*mz
;
683 struct mem_cgroup_tree_per_node
*mctz
;
685 mctz
= soft_limit_tree_from_page(page
);
689 * Necessary to update all ancestors when hierarchy is used.
690 * because their event counter is not touched.
692 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
693 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
694 excess
= soft_limit_excess(memcg
);
696 * We have to update the tree if mz is on RB-tree or
697 * mem is over its softlimit.
699 if (excess
|| mz
->on_tree
) {
702 spin_lock_irqsave(&mctz
->lock
, flags
);
703 /* if on-tree, remove it */
705 __mem_cgroup_remove_exceeded(mz
, mctz
);
707 * Insert again. mz->usage_in_excess will be updated.
708 * If excess is 0, no tree ops.
710 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
711 spin_unlock_irqrestore(&mctz
->lock
, flags
);
716 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
718 struct mem_cgroup_tree_per_node
*mctz
;
719 struct mem_cgroup_per_node
*mz
;
723 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
724 mctz
= soft_limit_tree_node(nid
);
726 mem_cgroup_remove_exceeded(mz
, mctz
);
730 static struct mem_cgroup_per_node
*
731 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
733 struct mem_cgroup_per_node
*mz
;
737 if (!mctz
->rb_rightmost
)
738 goto done
; /* Nothing to reclaim from */
740 mz
= rb_entry(mctz
->rb_rightmost
,
741 struct mem_cgroup_per_node
, tree_node
);
743 * Remove the node now but someone else can add it back,
744 * we will to add it back at the end of reclaim to its correct
745 * position in the tree.
747 __mem_cgroup_remove_exceeded(mz
, mctz
);
748 if (!soft_limit_excess(mz
->memcg
) ||
749 !css_tryget(&mz
->memcg
->css
))
755 static struct mem_cgroup_per_node
*
756 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
758 struct mem_cgroup_per_node
*mz
;
760 spin_lock_irq(&mctz
->lock
);
761 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
762 spin_unlock_irq(&mctz
->lock
);
767 * __mod_memcg_state - update cgroup memory statistics
768 * @memcg: the memory cgroup
769 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
770 * @val: delta to add to the counter, can be negative
772 void __mod_memcg_state(struct mem_cgroup
*memcg
, int idx
, int val
)
774 long x
, threshold
= MEMCG_CHARGE_BATCH
;
776 if (mem_cgroup_disabled())
779 if (memcg_stat_item_in_bytes(idx
))
780 threshold
<<= PAGE_SHIFT
;
782 x
= val
+ __this_cpu_read(memcg
->vmstats_percpu
->stat
[idx
]);
783 if (unlikely(abs(x
) > threshold
)) {
784 struct mem_cgroup
*mi
;
787 * Batch local counters to keep them in sync with
788 * the hierarchical ones.
790 __this_cpu_add(memcg
->vmstats_local
->stat
[idx
], x
);
791 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
792 atomic_long_add(x
, &mi
->vmstats
[idx
]);
795 __this_cpu_write(memcg
->vmstats_percpu
->stat
[idx
], x
);
798 static struct mem_cgroup_per_node
*
799 parent_nodeinfo(struct mem_cgroup_per_node
*pn
, int nid
)
801 struct mem_cgroup
*parent
;
803 parent
= parent_mem_cgroup(pn
->memcg
);
806 return mem_cgroup_nodeinfo(parent
, nid
);
809 void __mod_memcg_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
812 struct mem_cgroup_per_node
*pn
;
813 struct mem_cgroup
*memcg
;
814 long x
, threshold
= MEMCG_CHARGE_BATCH
;
816 pn
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
820 __mod_memcg_state(memcg
, idx
, val
);
823 __this_cpu_add(pn
->lruvec_stat_local
->count
[idx
], val
);
825 if (vmstat_item_in_bytes(idx
))
826 threshold
<<= PAGE_SHIFT
;
828 x
= val
+ __this_cpu_read(pn
->lruvec_stat_cpu
->count
[idx
]);
829 if (unlikely(abs(x
) > threshold
)) {
830 pg_data_t
*pgdat
= lruvec_pgdat(lruvec
);
831 struct mem_cgroup_per_node
*pi
;
833 for (pi
= pn
; pi
; pi
= parent_nodeinfo(pi
, pgdat
->node_id
))
834 atomic_long_add(x
, &pi
->lruvec_stat
[idx
]);
837 __this_cpu_write(pn
->lruvec_stat_cpu
->count
[idx
], x
);
841 * __mod_lruvec_state - update lruvec memory statistics
842 * @lruvec: the lruvec
843 * @idx: the stat item
844 * @val: delta to add to the counter, can be negative
846 * The lruvec is the intersection of the NUMA node and a cgroup. This
847 * function updates the all three counters that are affected by a
848 * change of state at this level: per-node, per-cgroup, per-lruvec.
850 void __mod_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
854 __mod_node_page_state(lruvec_pgdat(lruvec
), idx
, val
);
856 /* Update memcg and lruvec */
857 if (!mem_cgroup_disabled())
858 __mod_memcg_lruvec_state(lruvec
, idx
, val
);
861 void __mod_lruvec_slab_state(void *p
, enum node_stat_item idx
, int val
)
863 pg_data_t
*pgdat
= page_pgdat(virt_to_page(p
));
864 struct mem_cgroup
*memcg
;
865 struct lruvec
*lruvec
;
868 memcg
= mem_cgroup_from_obj(p
);
871 * Untracked pages have no memcg, no lruvec. Update only the
872 * node. If we reparent the slab objects to the root memcg,
873 * when we free the slab object, we need to update the per-memcg
874 * vmstats to keep it correct for the root memcg.
877 __mod_node_page_state(pgdat
, idx
, val
);
879 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
880 __mod_lruvec_state(lruvec
, idx
, val
);
886 * __count_memcg_events - account VM events in a cgroup
887 * @memcg: the memory cgroup
888 * @idx: the event item
889 * @count: the number of events that occured
891 void __count_memcg_events(struct mem_cgroup
*memcg
, enum vm_event_item idx
,
896 if (mem_cgroup_disabled())
899 x
= count
+ __this_cpu_read(memcg
->vmstats_percpu
->events
[idx
]);
900 if (unlikely(x
> MEMCG_CHARGE_BATCH
)) {
901 struct mem_cgroup
*mi
;
904 * Batch local counters to keep them in sync with
905 * the hierarchical ones.
907 __this_cpu_add(memcg
->vmstats_local
->events
[idx
], x
);
908 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
909 atomic_long_add(x
, &mi
->vmevents
[idx
]);
912 __this_cpu_write(memcg
->vmstats_percpu
->events
[idx
], x
);
915 static unsigned long memcg_events(struct mem_cgroup
*memcg
, int event
)
917 return atomic_long_read(&memcg
->vmevents
[event
]);
920 static unsigned long memcg_events_local(struct mem_cgroup
*memcg
, int event
)
925 for_each_possible_cpu(cpu
)
926 x
+= per_cpu(memcg
->vmstats_local
->events
[event
], cpu
);
930 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
934 /* pagein of a big page is an event. So, ignore page size */
936 __count_memcg_events(memcg
, PGPGIN
, 1);
938 __count_memcg_events(memcg
, PGPGOUT
, 1);
939 nr_pages
= -nr_pages
; /* for event */
942 __this_cpu_add(memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
945 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
946 enum mem_cgroup_events_target target
)
948 unsigned long val
, next
;
950 val
= __this_cpu_read(memcg
->vmstats_percpu
->nr_page_events
);
951 next
= __this_cpu_read(memcg
->vmstats_percpu
->targets
[target
]);
952 /* from time_after() in jiffies.h */
953 if ((long)(next
- val
) < 0) {
955 case MEM_CGROUP_TARGET_THRESH
:
956 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
958 case MEM_CGROUP_TARGET_SOFTLIMIT
:
959 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
964 __this_cpu_write(memcg
->vmstats_percpu
->targets
[target
], next
);
971 * Check events in order.
974 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
976 /* threshold event is triggered in finer grain than soft limit */
977 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
978 MEM_CGROUP_TARGET_THRESH
))) {
981 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
982 MEM_CGROUP_TARGET_SOFTLIMIT
);
983 mem_cgroup_threshold(memcg
);
984 if (unlikely(do_softlimit
))
985 mem_cgroup_update_tree(memcg
, page
);
989 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
992 * mm_update_next_owner() may clear mm->owner to NULL
993 * if it races with swapoff, page migration, etc.
994 * So this can be called with p == NULL.
999 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
1001 EXPORT_SYMBOL(mem_cgroup_from_task
);
1004 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1005 * @mm: mm from which memcg should be extracted. It can be NULL.
1007 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1008 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1011 struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1013 struct mem_cgroup
*memcg
;
1015 if (mem_cgroup_disabled())
1021 * Page cache insertions can happen withou an
1022 * actual mm context, e.g. during disk probing
1023 * on boot, loopback IO, acct() writes etc.
1026 memcg
= root_mem_cgroup
;
1028 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1029 if (unlikely(!memcg
))
1030 memcg
= root_mem_cgroup
;
1032 } while (!css_tryget(&memcg
->css
));
1036 EXPORT_SYMBOL(get_mem_cgroup_from_mm
);
1039 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1040 * @page: page from which memcg should be extracted.
1042 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1043 * root_mem_cgroup is returned.
1045 struct mem_cgroup
*get_mem_cgroup_from_page(struct page
*page
)
1047 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
1049 if (mem_cgroup_disabled())
1053 /* Page should not get uncharged and freed memcg under us. */
1054 if (!memcg
|| WARN_ON_ONCE(!css_tryget(&memcg
->css
)))
1055 memcg
= root_mem_cgroup
;
1059 EXPORT_SYMBOL(get_mem_cgroup_from_page
);
1061 static __always_inline
struct mem_cgroup
*active_memcg(void)
1064 return this_cpu_read(int_active_memcg
);
1066 return current
->active_memcg
;
1069 static __always_inline
struct mem_cgroup
*get_active_memcg(void)
1071 struct mem_cgroup
*memcg
;
1074 memcg
= active_memcg();
1076 /* current->active_memcg must hold a ref. */
1077 if (WARN_ON_ONCE(!css_tryget(&memcg
->css
)))
1078 memcg
= root_mem_cgroup
;
1080 memcg
= current
->active_memcg
;
1087 static __always_inline
bool memcg_kmem_bypass(void)
1089 /* Allow remote memcg charging from any context. */
1090 if (unlikely(active_memcg()))
1093 /* Memcg to charge can't be determined. */
1094 if (in_interrupt() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
1101 * If active memcg is set, do not fallback to current->mm->memcg.
1103 static __always_inline
struct mem_cgroup
*get_mem_cgroup_from_current(void)
1105 if (memcg_kmem_bypass())
1108 if (unlikely(active_memcg()))
1109 return get_active_memcg();
1111 return get_mem_cgroup_from_mm(current
->mm
);
1115 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1116 * @root: hierarchy root
1117 * @prev: previously returned memcg, NULL on first invocation
1118 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1120 * Returns references to children of the hierarchy below @root, or
1121 * @root itself, or %NULL after a full round-trip.
1123 * Caller must pass the return value in @prev on subsequent
1124 * invocations for reference counting, or use mem_cgroup_iter_break()
1125 * to cancel a hierarchy walk before the round-trip is complete.
1127 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1128 * in the hierarchy among all concurrent reclaimers operating on the
1131 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1132 struct mem_cgroup
*prev
,
1133 struct mem_cgroup_reclaim_cookie
*reclaim
)
1135 struct mem_cgroup_reclaim_iter
*iter
;
1136 struct cgroup_subsys_state
*css
= NULL
;
1137 struct mem_cgroup
*memcg
= NULL
;
1138 struct mem_cgroup
*pos
= NULL
;
1140 if (mem_cgroup_disabled())
1144 root
= root_mem_cgroup
;
1146 if (prev
&& !reclaim
)
1149 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1158 struct mem_cgroup_per_node
*mz
;
1160 mz
= mem_cgroup_nodeinfo(root
, reclaim
->pgdat
->node_id
);
1163 if (prev
&& reclaim
->generation
!= iter
->generation
)
1167 pos
= READ_ONCE(iter
->position
);
1168 if (!pos
|| css_tryget(&pos
->css
))
1171 * css reference reached zero, so iter->position will
1172 * be cleared by ->css_released. However, we should not
1173 * rely on this happening soon, because ->css_released
1174 * is called from a work queue, and by busy-waiting we
1175 * might block it. So we clear iter->position right
1178 (void)cmpxchg(&iter
->position
, pos
, NULL
);
1186 css
= css_next_descendant_pre(css
, &root
->css
);
1189 * Reclaimers share the hierarchy walk, and a
1190 * new one might jump in right at the end of
1191 * the hierarchy - make sure they see at least
1192 * one group and restart from the beginning.
1200 * Verify the css and acquire a reference. The root
1201 * is provided by the caller, so we know it's alive
1202 * and kicking, and don't take an extra reference.
1204 memcg
= mem_cgroup_from_css(css
);
1206 if (css
== &root
->css
)
1209 if (css_tryget(css
))
1217 * The position could have already been updated by a competing
1218 * thread, so check that the value hasn't changed since we read
1219 * it to avoid reclaiming from the same cgroup twice.
1221 (void)cmpxchg(&iter
->position
, pos
, memcg
);
1229 reclaim
->generation
= iter
->generation
;
1235 if (prev
&& prev
!= root
)
1236 css_put(&prev
->css
);
1242 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1243 * @root: hierarchy root
1244 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1246 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1247 struct mem_cgroup
*prev
)
1250 root
= root_mem_cgroup
;
1251 if (prev
&& prev
!= root
)
1252 css_put(&prev
->css
);
1255 static void __invalidate_reclaim_iterators(struct mem_cgroup
*from
,
1256 struct mem_cgroup
*dead_memcg
)
1258 struct mem_cgroup_reclaim_iter
*iter
;
1259 struct mem_cgroup_per_node
*mz
;
1262 for_each_node(nid
) {
1263 mz
= mem_cgroup_nodeinfo(from
, nid
);
1265 cmpxchg(&iter
->position
, dead_memcg
, NULL
);
1269 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
1271 struct mem_cgroup
*memcg
= dead_memcg
;
1272 struct mem_cgroup
*last
;
1275 __invalidate_reclaim_iterators(memcg
, dead_memcg
);
1277 } while ((memcg
= parent_mem_cgroup(memcg
)));
1280 * When cgruop1 non-hierarchy mode is used,
1281 * parent_mem_cgroup() does not walk all the way up to the
1282 * cgroup root (root_mem_cgroup). So we have to handle
1283 * dead_memcg from cgroup root separately.
1285 if (last
!= root_mem_cgroup
)
1286 __invalidate_reclaim_iterators(root_mem_cgroup
,
1291 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1292 * @memcg: hierarchy root
1293 * @fn: function to call for each task
1294 * @arg: argument passed to @fn
1296 * This function iterates over tasks attached to @memcg or to any of its
1297 * descendants and calls @fn for each task. If @fn returns a non-zero
1298 * value, the function breaks the iteration loop and returns the value.
1299 * Otherwise, it will iterate over all tasks and return 0.
1301 * This function must not be called for the root memory cgroup.
1303 int mem_cgroup_scan_tasks(struct mem_cgroup
*memcg
,
1304 int (*fn
)(struct task_struct
*, void *), void *arg
)
1306 struct mem_cgroup
*iter
;
1309 BUG_ON(memcg
== root_mem_cgroup
);
1311 for_each_mem_cgroup_tree(iter
, memcg
) {
1312 struct css_task_iter it
;
1313 struct task_struct
*task
;
1315 css_task_iter_start(&iter
->css
, CSS_TASK_ITER_PROCS
, &it
);
1316 while (!ret
&& (task
= css_task_iter_next(&it
)))
1317 ret
= fn(task
, arg
);
1318 css_task_iter_end(&it
);
1320 mem_cgroup_iter_break(memcg
, iter
);
1328 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1330 * @pgdat: pgdat of the page
1332 * This function relies on page->mem_cgroup being stable - see the
1333 * access rules in commit_charge().
1335 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct pglist_data
*pgdat
)
1337 struct mem_cgroup_per_node
*mz
;
1338 struct mem_cgroup
*memcg
;
1339 struct lruvec
*lruvec
;
1341 if (mem_cgroup_disabled()) {
1342 lruvec
= &pgdat
->__lruvec
;
1346 memcg
= page
->mem_cgroup
;
1348 * Swapcache readahead pages are added to the LRU - and
1349 * possibly migrated - before they are charged.
1352 memcg
= root_mem_cgroup
;
1354 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
1355 lruvec
= &mz
->lruvec
;
1358 * Since a node can be onlined after the mem_cgroup was created,
1359 * we have to be prepared to initialize lruvec->zone here;
1360 * and if offlined then reonlined, we need to reinitialize it.
1362 if (unlikely(lruvec
->pgdat
!= pgdat
))
1363 lruvec
->pgdat
= pgdat
;
1368 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1369 * @lruvec: mem_cgroup per zone lru vector
1370 * @lru: index of lru list the page is sitting on
1371 * @zid: zone id of the accounted pages
1372 * @nr_pages: positive when adding or negative when removing
1374 * This function must be called under lru_lock, just before a page is added
1375 * to or just after a page is removed from an lru list (that ordering being
1376 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1378 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1379 int zid
, int nr_pages
)
1381 struct mem_cgroup_per_node
*mz
;
1382 unsigned long *lru_size
;
1385 if (mem_cgroup_disabled())
1388 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
1389 lru_size
= &mz
->lru_zone_size
[zid
][lru
];
1392 *lru_size
+= nr_pages
;
1395 if (WARN_ONCE(size
< 0,
1396 "%s(%p, %d, %d): lru_size %ld\n",
1397 __func__
, lruvec
, lru
, nr_pages
, size
)) {
1403 *lru_size
+= nr_pages
;
1407 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1408 * @memcg: the memory cgroup
1410 * Returns the maximum amount of memory @mem can be charged with, in
1413 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1415 unsigned long margin
= 0;
1416 unsigned long count
;
1417 unsigned long limit
;
1419 count
= page_counter_read(&memcg
->memory
);
1420 limit
= READ_ONCE(memcg
->memory
.max
);
1422 margin
= limit
- count
;
1424 if (do_memsw_account()) {
1425 count
= page_counter_read(&memcg
->memsw
);
1426 limit
= READ_ONCE(memcg
->memsw
.max
);
1428 margin
= min(margin
, limit
- count
);
1437 * A routine for checking "mem" is under move_account() or not.
1439 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1440 * moving cgroups. This is for waiting at high-memory pressure
1443 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1445 struct mem_cgroup
*from
;
1446 struct mem_cgroup
*to
;
1449 * Unlike task_move routines, we access mc.to, mc.from not under
1450 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1452 spin_lock(&mc
.lock
);
1458 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1459 mem_cgroup_is_descendant(to
, memcg
);
1461 spin_unlock(&mc
.lock
);
1465 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1467 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1468 if (mem_cgroup_under_move(memcg
)) {
1470 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1471 /* moving charge context might have finished. */
1474 finish_wait(&mc
.waitq
, &wait
);
1481 struct memory_stat
{
1487 static struct memory_stat memory_stats
[] = {
1488 { "anon", PAGE_SIZE
, NR_ANON_MAPPED
},
1489 { "file", PAGE_SIZE
, NR_FILE_PAGES
},
1490 { "kernel_stack", 1024, NR_KERNEL_STACK_KB
},
1491 { "percpu", 1, MEMCG_PERCPU_B
},
1492 { "sock", PAGE_SIZE
, MEMCG_SOCK
},
1493 { "shmem", PAGE_SIZE
, NR_SHMEM
},
1494 { "file_mapped", PAGE_SIZE
, NR_FILE_MAPPED
},
1495 { "file_dirty", PAGE_SIZE
, NR_FILE_DIRTY
},
1496 { "file_writeback", PAGE_SIZE
, NR_WRITEBACK
},
1497 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1499 * The ratio will be initialized in memory_stats_init(). Because
1500 * on some architectures, the macro of HPAGE_PMD_SIZE is not
1501 * constant(e.g. powerpc).
1503 { "anon_thp", 0, NR_ANON_THPS
},
1504 { "file_thp", 0, NR_FILE_THPS
},
1505 { "shmem_thp", 0, NR_SHMEM_THPS
},
1507 { "inactive_anon", PAGE_SIZE
, NR_INACTIVE_ANON
},
1508 { "active_anon", PAGE_SIZE
, NR_ACTIVE_ANON
},
1509 { "inactive_file", PAGE_SIZE
, NR_INACTIVE_FILE
},
1510 { "active_file", PAGE_SIZE
, NR_ACTIVE_FILE
},
1511 { "unevictable", PAGE_SIZE
, NR_UNEVICTABLE
},
1514 * Note: The slab_reclaimable and slab_unreclaimable must be
1515 * together and slab_reclaimable must be in front.
1517 { "slab_reclaimable", 1, NR_SLAB_RECLAIMABLE_B
},
1518 { "slab_unreclaimable", 1, NR_SLAB_UNRECLAIMABLE_B
},
1520 /* The memory events */
1521 { "workingset_refault_anon", 1, WORKINGSET_REFAULT_ANON
},
1522 { "workingset_refault_file", 1, WORKINGSET_REFAULT_FILE
},
1523 { "workingset_activate_anon", 1, WORKINGSET_ACTIVATE_ANON
},
1524 { "workingset_activate_file", 1, WORKINGSET_ACTIVATE_FILE
},
1525 { "workingset_restore_anon", 1, WORKINGSET_RESTORE_ANON
},
1526 { "workingset_restore_file", 1, WORKINGSET_RESTORE_FILE
},
1527 { "workingset_nodereclaim", 1, WORKINGSET_NODERECLAIM
},
1530 static int __init
memory_stats_init(void)
1534 for (i
= 0; i
< ARRAY_SIZE(memory_stats
); i
++) {
1535 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1536 if (memory_stats
[i
].idx
== NR_ANON_THPS
||
1537 memory_stats
[i
].idx
== NR_FILE_THPS
||
1538 memory_stats
[i
].idx
== NR_SHMEM_THPS
)
1539 memory_stats
[i
].ratio
= HPAGE_PMD_SIZE
;
1541 VM_BUG_ON(!memory_stats
[i
].ratio
);
1542 VM_BUG_ON(memory_stats
[i
].idx
>= MEMCG_NR_STAT
);
1547 pure_initcall(memory_stats_init
);
1549 static char *memory_stat_format(struct mem_cgroup
*memcg
)
1554 seq_buf_init(&s
, kmalloc(PAGE_SIZE
, GFP_KERNEL
), PAGE_SIZE
);
1559 * Provide statistics on the state of the memory subsystem as
1560 * well as cumulative event counters that show past behavior.
1562 * This list is ordered following a combination of these gradients:
1563 * 1) generic big picture -> specifics and details
1564 * 2) reflecting userspace activity -> reflecting kernel heuristics
1566 * Current memory state:
1569 for (i
= 0; i
< ARRAY_SIZE(memory_stats
); i
++) {
1572 size
= memcg_page_state(memcg
, memory_stats
[i
].idx
);
1573 size
*= memory_stats
[i
].ratio
;
1574 seq_buf_printf(&s
, "%s %llu\n", memory_stats
[i
].name
, size
);
1576 if (unlikely(memory_stats
[i
].idx
== NR_SLAB_UNRECLAIMABLE_B
)) {
1577 size
= memcg_page_state(memcg
, NR_SLAB_RECLAIMABLE_B
) +
1578 memcg_page_state(memcg
, NR_SLAB_UNRECLAIMABLE_B
);
1579 seq_buf_printf(&s
, "slab %llu\n", size
);
1583 /* Accumulated memory events */
1585 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGFAULT
),
1586 memcg_events(memcg
, PGFAULT
));
1587 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGMAJFAULT
),
1588 memcg_events(memcg
, PGMAJFAULT
));
1589 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGREFILL
),
1590 memcg_events(memcg
, PGREFILL
));
1591 seq_buf_printf(&s
, "pgscan %lu\n",
1592 memcg_events(memcg
, PGSCAN_KSWAPD
) +
1593 memcg_events(memcg
, PGSCAN_DIRECT
));
1594 seq_buf_printf(&s
, "pgsteal %lu\n",
1595 memcg_events(memcg
, PGSTEAL_KSWAPD
) +
1596 memcg_events(memcg
, PGSTEAL_DIRECT
));
1597 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGACTIVATE
),
1598 memcg_events(memcg
, PGACTIVATE
));
1599 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGDEACTIVATE
),
1600 memcg_events(memcg
, PGDEACTIVATE
));
1601 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREE
),
1602 memcg_events(memcg
, PGLAZYFREE
));
1603 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREED
),
1604 memcg_events(memcg
, PGLAZYFREED
));
1606 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1607 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC
),
1608 memcg_events(memcg
, THP_FAULT_ALLOC
));
1609 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC
),
1610 memcg_events(memcg
, THP_COLLAPSE_ALLOC
));
1611 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1613 /* The above should easily fit into one page */
1614 WARN_ON_ONCE(seq_buf_has_overflowed(&s
));
1619 #define K(x) ((x) << (PAGE_SHIFT-10))
1621 * mem_cgroup_print_oom_context: Print OOM information relevant to
1622 * memory controller.
1623 * @memcg: The memory cgroup that went over limit
1624 * @p: Task that is going to be killed
1626 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1629 void mem_cgroup_print_oom_context(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1634 pr_cont(",oom_memcg=");
1635 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1637 pr_cont(",global_oom");
1639 pr_cont(",task_memcg=");
1640 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1646 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1647 * memory controller.
1648 * @memcg: The memory cgroup that went over limit
1650 void mem_cgroup_print_oom_meminfo(struct mem_cgroup
*memcg
)
1654 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1655 K((u64
)page_counter_read(&memcg
->memory
)),
1656 K((u64
)READ_ONCE(memcg
->memory
.max
)), memcg
->memory
.failcnt
);
1657 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1658 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1659 K((u64
)page_counter_read(&memcg
->swap
)),
1660 K((u64
)READ_ONCE(memcg
->swap
.max
)), memcg
->swap
.failcnt
);
1662 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1663 K((u64
)page_counter_read(&memcg
->memsw
)),
1664 K((u64
)memcg
->memsw
.max
), memcg
->memsw
.failcnt
);
1665 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1666 K((u64
)page_counter_read(&memcg
->kmem
)),
1667 K((u64
)memcg
->kmem
.max
), memcg
->kmem
.failcnt
);
1670 pr_info("Memory cgroup stats for ");
1671 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1673 buf
= memory_stat_format(memcg
);
1681 * Return the memory (and swap, if configured) limit for a memcg.
1683 unsigned long mem_cgroup_get_max(struct mem_cgroup
*memcg
)
1685 unsigned long max
= READ_ONCE(memcg
->memory
.max
);
1687 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
1688 if (mem_cgroup_swappiness(memcg
))
1689 max
+= min(READ_ONCE(memcg
->swap
.max
),
1690 (unsigned long)total_swap_pages
);
1692 if (mem_cgroup_swappiness(memcg
)) {
1693 /* Calculate swap excess capacity from memsw limit */
1694 unsigned long swap
= READ_ONCE(memcg
->memsw
.max
) - max
;
1696 max
+= min(swap
, (unsigned long)total_swap_pages
);
1702 unsigned long mem_cgroup_size(struct mem_cgroup
*memcg
)
1704 return page_counter_read(&memcg
->memory
);
1707 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1710 struct oom_control oc
= {
1714 .gfp_mask
= gfp_mask
,
1719 if (mutex_lock_killable(&oom_lock
))
1722 if (mem_cgroup_margin(memcg
) >= (1 << order
))
1726 * A few threads which were not waiting at mutex_lock_killable() can
1727 * fail to bail out. Therefore, check again after holding oom_lock.
1729 ret
= should_force_charge() || out_of_memory(&oc
);
1732 mutex_unlock(&oom_lock
);
1736 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1739 unsigned long *total_scanned
)
1741 struct mem_cgroup
*victim
= NULL
;
1744 unsigned long excess
;
1745 unsigned long nr_scanned
;
1746 struct mem_cgroup_reclaim_cookie reclaim
= {
1750 excess
= soft_limit_excess(root_memcg
);
1753 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1758 * If we have not been able to reclaim
1759 * anything, it might because there are
1760 * no reclaimable pages under this hierarchy
1765 * We want to do more targeted reclaim.
1766 * excess >> 2 is not to excessive so as to
1767 * reclaim too much, nor too less that we keep
1768 * coming back to reclaim from this cgroup
1770 if (total
>= (excess
>> 2) ||
1771 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1776 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1777 pgdat
, &nr_scanned
);
1778 *total_scanned
+= nr_scanned
;
1779 if (!soft_limit_excess(root_memcg
))
1782 mem_cgroup_iter_break(root_memcg
, victim
);
1786 #ifdef CONFIG_LOCKDEP
1787 static struct lockdep_map memcg_oom_lock_dep_map
= {
1788 .name
= "memcg_oom_lock",
1792 static DEFINE_SPINLOCK(memcg_oom_lock
);
1795 * Check OOM-Killer is already running under our hierarchy.
1796 * If someone is running, return false.
1798 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1800 struct mem_cgroup
*iter
, *failed
= NULL
;
1802 spin_lock(&memcg_oom_lock
);
1804 for_each_mem_cgroup_tree(iter
, memcg
) {
1805 if (iter
->oom_lock
) {
1807 * this subtree of our hierarchy is already locked
1808 * so we cannot give a lock.
1811 mem_cgroup_iter_break(memcg
, iter
);
1814 iter
->oom_lock
= true;
1819 * OK, we failed to lock the whole subtree so we have
1820 * to clean up what we set up to the failing subtree
1822 for_each_mem_cgroup_tree(iter
, memcg
) {
1823 if (iter
== failed
) {
1824 mem_cgroup_iter_break(memcg
, iter
);
1827 iter
->oom_lock
= false;
1830 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1832 spin_unlock(&memcg_oom_lock
);
1837 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1839 struct mem_cgroup
*iter
;
1841 spin_lock(&memcg_oom_lock
);
1842 mutex_release(&memcg_oom_lock_dep_map
, _RET_IP_
);
1843 for_each_mem_cgroup_tree(iter
, memcg
)
1844 iter
->oom_lock
= false;
1845 spin_unlock(&memcg_oom_lock
);
1848 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1850 struct mem_cgroup
*iter
;
1852 spin_lock(&memcg_oom_lock
);
1853 for_each_mem_cgroup_tree(iter
, memcg
)
1855 spin_unlock(&memcg_oom_lock
);
1858 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1860 struct mem_cgroup
*iter
;
1863 * Be careful about under_oom underflows becase a child memcg
1864 * could have been added after mem_cgroup_mark_under_oom.
1866 spin_lock(&memcg_oom_lock
);
1867 for_each_mem_cgroup_tree(iter
, memcg
)
1868 if (iter
->under_oom
> 0)
1870 spin_unlock(&memcg_oom_lock
);
1873 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1875 struct oom_wait_info
{
1876 struct mem_cgroup
*memcg
;
1877 wait_queue_entry_t wait
;
1880 static int memcg_oom_wake_function(wait_queue_entry_t
*wait
,
1881 unsigned mode
, int sync
, void *arg
)
1883 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1884 struct mem_cgroup
*oom_wait_memcg
;
1885 struct oom_wait_info
*oom_wait_info
;
1887 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1888 oom_wait_memcg
= oom_wait_info
->memcg
;
1890 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1891 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1893 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1896 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1899 * For the following lockless ->under_oom test, the only required
1900 * guarantee is that it must see the state asserted by an OOM when
1901 * this function is called as a result of userland actions
1902 * triggered by the notification of the OOM. This is trivially
1903 * achieved by invoking mem_cgroup_mark_under_oom() before
1904 * triggering notification.
1906 if (memcg
&& memcg
->under_oom
)
1907 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1917 static enum oom_status
mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1919 enum oom_status ret
;
1922 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1925 memcg_memory_event(memcg
, MEMCG_OOM
);
1928 * We are in the middle of the charge context here, so we
1929 * don't want to block when potentially sitting on a callstack
1930 * that holds all kinds of filesystem and mm locks.
1932 * cgroup1 allows disabling the OOM killer and waiting for outside
1933 * handling until the charge can succeed; remember the context and put
1934 * the task to sleep at the end of the page fault when all locks are
1937 * On the other hand, in-kernel OOM killer allows for an async victim
1938 * memory reclaim (oom_reaper) and that means that we are not solely
1939 * relying on the oom victim to make a forward progress and we can
1940 * invoke the oom killer here.
1942 * Please note that mem_cgroup_out_of_memory might fail to find a
1943 * victim and then we have to bail out from the charge path.
1945 if (memcg
->oom_kill_disable
) {
1946 if (!current
->in_user_fault
)
1948 css_get(&memcg
->css
);
1949 current
->memcg_in_oom
= memcg
;
1950 current
->memcg_oom_gfp_mask
= mask
;
1951 current
->memcg_oom_order
= order
;
1956 mem_cgroup_mark_under_oom(memcg
);
1958 locked
= mem_cgroup_oom_trylock(memcg
);
1961 mem_cgroup_oom_notify(memcg
);
1963 mem_cgroup_unmark_under_oom(memcg
);
1964 if (mem_cgroup_out_of_memory(memcg
, mask
, order
))
1970 mem_cgroup_oom_unlock(memcg
);
1976 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1977 * @handle: actually kill/wait or just clean up the OOM state
1979 * This has to be called at the end of a page fault if the memcg OOM
1980 * handler was enabled.
1982 * Memcg supports userspace OOM handling where failed allocations must
1983 * sleep on a waitqueue until the userspace task resolves the
1984 * situation. Sleeping directly in the charge context with all kinds
1985 * of locks held is not a good idea, instead we remember an OOM state
1986 * in the task and mem_cgroup_oom_synchronize() has to be called at
1987 * the end of the page fault to complete the OOM handling.
1989 * Returns %true if an ongoing memcg OOM situation was detected and
1990 * completed, %false otherwise.
1992 bool mem_cgroup_oom_synchronize(bool handle
)
1994 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1995 struct oom_wait_info owait
;
1998 /* OOM is global, do not handle */
2005 owait
.memcg
= memcg
;
2006 owait
.wait
.flags
= 0;
2007 owait
.wait
.func
= memcg_oom_wake_function
;
2008 owait
.wait
.private = current
;
2009 INIT_LIST_HEAD(&owait
.wait
.entry
);
2011 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2012 mem_cgroup_mark_under_oom(memcg
);
2014 locked
= mem_cgroup_oom_trylock(memcg
);
2017 mem_cgroup_oom_notify(memcg
);
2019 if (locked
&& !memcg
->oom_kill_disable
) {
2020 mem_cgroup_unmark_under_oom(memcg
);
2021 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2022 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
2023 current
->memcg_oom_order
);
2026 mem_cgroup_unmark_under_oom(memcg
);
2027 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2031 mem_cgroup_oom_unlock(memcg
);
2033 * There is no guarantee that an OOM-lock contender
2034 * sees the wakeups triggered by the OOM kill
2035 * uncharges. Wake any sleepers explicitely.
2037 memcg_oom_recover(memcg
);
2040 current
->memcg_in_oom
= NULL
;
2041 css_put(&memcg
->css
);
2046 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2047 * @victim: task to be killed by the OOM killer
2048 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2050 * Returns a pointer to a memory cgroup, which has to be cleaned up
2051 * by killing all belonging OOM-killable tasks.
2053 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2055 struct mem_cgroup
*mem_cgroup_get_oom_group(struct task_struct
*victim
,
2056 struct mem_cgroup
*oom_domain
)
2058 struct mem_cgroup
*oom_group
= NULL
;
2059 struct mem_cgroup
*memcg
;
2061 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2065 oom_domain
= root_mem_cgroup
;
2069 memcg
= mem_cgroup_from_task(victim
);
2070 if (memcg
== root_mem_cgroup
)
2074 * If the victim task has been asynchronously moved to a different
2075 * memory cgroup, we might end up killing tasks outside oom_domain.
2076 * In this case it's better to ignore memory.group.oom.
2078 if (unlikely(!mem_cgroup_is_descendant(memcg
, oom_domain
)))
2082 * Traverse the memory cgroup hierarchy from the victim task's
2083 * cgroup up to the OOMing cgroup (or root) to find the
2084 * highest-level memory cgroup with oom.group set.
2086 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
2087 if (memcg
->oom_group
)
2090 if (memcg
== oom_domain
)
2095 css_get(&oom_group
->css
);
2102 void mem_cgroup_print_oom_group(struct mem_cgroup
*memcg
)
2104 pr_info("Tasks in ");
2105 pr_cont_cgroup_path(memcg
->css
.cgroup
);
2106 pr_cont(" are going to be killed due to memory.oom.group set\n");
2110 * lock_page_memcg - lock a page->mem_cgroup binding
2113 * This function protects unlocked LRU pages from being moved to
2116 * It ensures lifetime of the returned memcg. Caller is responsible
2117 * for the lifetime of the page; __unlock_page_memcg() is available
2118 * when @page might get freed inside the locked section.
2120 struct mem_cgroup
*lock_page_memcg(struct page
*page
)
2122 struct page
*head
= compound_head(page
); /* rmap on tail pages */
2123 struct mem_cgroup
*memcg
;
2124 unsigned long flags
;
2127 * The RCU lock is held throughout the transaction. The fast
2128 * path can get away without acquiring the memcg->move_lock
2129 * because page moving starts with an RCU grace period.
2131 * The RCU lock also protects the memcg from being freed when
2132 * the page state that is going to change is the only thing
2133 * preventing the page itself from being freed. E.g. writeback
2134 * doesn't hold a page reference and relies on PG_writeback to
2135 * keep off truncation, migration and so forth.
2139 if (mem_cgroup_disabled())
2142 memcg
= head
->mem_cgroup
;
2143 if (unlikely(!memcg
))
2146 if (atomic_read(&memcg
->moving_account
) <= 0)
2149 spin_lock_irqsave(&memcg
->move_lock
, flags
);
2150 if (memcg
!= head
->mem_cgroup
) {
2151 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2156 * When charge migration first begins, we can have locked and
2157 * unlocked page stat updates happening concurrently. Track
2158 * the task who has the lock for unlock_page_memcg().
2160 memcg
->move_lock_task
= current
;
2161 memcg
->move_lock_flags
= flags
;
2165 EXPORT_SYMBOL(lock_page_memcg
);
2168 * __unlock_page_memcg - unlock and unpin a memcg
2171 * Unlock and unpin a memcg returned by lock_page_memcg().
2173 void __unlock_page_memcg(struct mem_cgroup
*memcg
)
2175 if (memcg
&& memcg
->move_lock_task
== current
) {
2176 unsigned long flags
= memcg
->move_lock_flags
;
2178 memcg
->move_lock_task
= NULL
;
2179 memcg
->move_lock_flags
= 0;
2181 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2188 * unlock_page_memcg - unlock a page->mem_cgroup binding
2191 void unlock_page_memcg(struct page
*page
)
2193 struct page
*head
= compound_head(page
);
2195 __unlock_page_memcg(head
->mem_cgroup
);
2197 EXPORT_SYMBOL(unlock_page_memcg
);
2199 struct memcg_stock_pcp
{
2200 struct mem_cgroup
*cached
; /* this never be root cgroup */
2201 unsigned int nr_pages
;
2203 #ifdef CONFIG_MEMCG_KMEM
2204 struct obj_cgroup
*cached_objcg
;
2205 unsigned int nr_bytes
;
2208 struct work_struct work
;
2209 unsigned long flags
;
2210 #define FLUSHING_CACHED_CHARGE 0
2212 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2213 static DEFINE_MUTEX(percpu_charge_mutex
);
2215 #ifdef CONFIG_MEMCG_KMEM
2216 static void drain_obj_stock(struct memcg_stock_pcp
*stock
);
2217 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
2218 struct mem_cgroup
*root_memcg
);
2221 static inline void drain_obj_stock(struct memcg_stock_pcp
*stock
)
2224 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
2225 struct mem_cgroup
*root_memcg
)
2232 * consume_stock: Try to consume stocked charge on this cpu.
2233 * @memcg: memcg to consume from.
2234 * @nr_pages: how many pages to charge.
2236 * The charges will only happen if @memcg matches the current cpu's memcg
2237 * stock, and at least @nr_pages are available in that stock. Failure to
2238 * service an allocation will refill the stock.
2240 * returns true if successful, false otherwise.
2242 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2244 struct memcg_stock_pcp
*stock
;
2245 unsigned long flags
;
2248 if (nr_pages
> MEMCG_CHARGE_BATCH
)
2251 local_irq_save(flags
);
2253 stock
= this_cpu_ptr(&memcg_stock
);
2254 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
2255 stock
->nr_pages
-= nr_pages
;
2259 local_irq_restore(flags
);
2265 * Returns stocks cached in percpu and reset cached information.
2267 static void drain_stock(struct memcg_stock_pcp
*stock
)
2269 struct mem_cgroup
*old
= stock
->cached
;
2274 if (stock
->nr_pages
) {
2275 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2276 if (do_memsw_account())
2277 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2278 stock
->nr_pages
= 0;
2282 stock
->cached
= NULL
;
2285 static void drain_local_stock(struct work_struct
*dummy
)
2287 struct memcg_stock_pcp
*stock
;
2288 unsigned long flags
;
2291 * The only protection from memory hotplug vs. drain_stock races is
2292 * that we always operate on local CPU stock here with IRQ disabled
2294 local_irq_save(flags
);
2296 stock
= this_cpu_ptr(&memcg_stock
);
2297 drain_obj_stock(stock
);
2299 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2301 local_irq_restore(flags
);
2305 * Cache charges(val) to local per_cpu area.
2306 * This will be consumed by consume_stock() function, later.
2308 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2310 struct memcg_stock_pcp
*stock
;
2311 unsigned long flags
;
2313 local_irq_save(flags
);
2315 stock
= this_cpu_ptr(&memcg_stock
);
2316 if (stock
->cached
!= memcg
) { /* reset if necessary */
2318 css_get(&memcg
->css
);
2319 stock
->cached
= memcg
;
2321 stock
->nr_pages
+= nr_pages
;
2323 if (stock
->nr_pages
> MEMCG_CHARGE_BATCH
)
2326 local_irq_restore(flags
);
2330 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2331 * of the hierarchy under it.
2333 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2337 /* If someone's already draining, avoid adding running more workers. */
2338 if (!mutex_trylock(&percpu_charge_mutex
))
2341 * Notify other cpus that system-wide "drain" is running
2342 * We do not care about races with the cpu hotplug because cpu down
2343 * as well as workers from this path always operate on the local
2344 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2347 for_each_online_cpu(cpu
) {
2348 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2349 struct mem_cgroup
*memcg
;
2353 memcg
= stock
->cached
;
2354 if (memcg
&& stock
->nr_pages
&&
2355 mem_cgroup_is_descendant(memcg
, root_memcg
))
2357 if (obj_stock_flush_required(stock
, root_memcg
))
2362 !test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2364 drain_local_stock(&stock
->work
);
2366 schedule_work_on(cpu
, &stock
->work
);
2370 mutex_unlock(&percpu_charge_mutex
);
2373 static int memcg_hotplug_cpu_dead(unsigned int cpu
)
2375 struct memcg_stock_pcp
*stock
;
2376 struct mem_cgroup
*memcg
, *mi
;
2378 stock
= &per_cpu(memcg_stock
, cpu
);
2381 for_each_mem_cgroup(memcg
) {
2384 for (i
= 0; i
< MEMCG_NR_STAT
; i
++) {
2388 x
= this_cpu_xchg(memcg
->vmstats_percpu
->stat
[i
], 0);
2390 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2391 atomic_long_add(x
, &memcg
->vmstats
[i
]);
2393 if (i
>= NR_VM_NODE_STAT_ITEMS
)
2396 for_each_node(nid
) {
2397 struct mem_cgroup_per_node
*pn
;
2399 pn
= mem_cgroup_nodeinfo(memcg
, nid
);
2400 x
= this_cpu_xchg(pn
->lruvec_stat_cpu
->count
[i
], 0);
2403 atomic_long_add(x
, &pn
->lruvec_stat
[i
]);
2404 } while ((pn
= parent_nodeinfo(pn
, nid
)));
2408 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++) {
2411 x
= this_cpu_xchg(memcg
->vmstats_percpu
->events
[i
], 0);
2413 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2414 atomic_long_add(x
, &memcg
->vmevents
[i
]);
2421 static unsigned long reclaim_high(struct mem_cgroup
*memcg
,
2422 unsigned int nr_pages
,
2425 unsigned long nr_reclaimed
= 0;
2428 unsigned long pflags
;
2430 if (page_counter_read(&memcg
->memory
) <=
2431 READ_ONCE(memcg
->memory
.high
))
2434 memcg_memory_event(memcg
, MEMCG_HIGH
);
2436 psi_memstall_enter(&pflags
);
2437 nr_reclaimed
+= try_to_free_mem_cgroup_pages(memcg
, nr_pages
,
2439 psi_memstall_leave(&pflags
);
2440 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2441 !mem_cgroup_is_root(memcg
));
2443 return nr_reclaimed
;
2446 static void high_work_func(struct work_struct
*work
)
2448 struct mem_cgroup
*memcg
;
2450 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
2451 reclaim_high(memcg
, MEMCG_CHARGE_BATCH
, GFP_KERNEL
);
2455 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2456 * enough to still cause a significant slowdown in most cases, while still
2457 * allowing diagnostics and tracing to proceed without becoming stuck.
2459 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2462 * When calculating the delay, we use these either side of the exponentiation to
2463 * maintain precision and scale to a reasonable number of jiffies (see the table
2466 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2467 * overage ratio to a delay.
2468 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2469 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2470 * to produce a reasonable delay curve.
2472 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2473 * reasonable delay curve compared to precision-adjusted overage, not
2474 * penalising heavily at first, but still making sure that growth beyond the
2475 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2476 * example, with a high of 100 megabytes:
2478 * +-------+------------------------+
2479 * | usage | time to allocate in ms |
2480 * +-------+------------------------+
2502 * +-------+------------------------+
2504 #define MEMCG_DELAY_PRECISION_SHIFT 20
2505 #define MEMCG_DELAY_SCALING_SHIFT 14
2507 static u64
calculate_overage(unsigned long usage
, unsigned long high
)
2515 * Prevent division by 0 in overage calculation by acting as if
2516 * it was a threshold of 1 page
2518 high
= max(high
, 1UL);
2520 overage
= usage
- high
;
2521 overage
<<= MEMCG_DELAY_PRECISION_SHIFT
;
2522 return div64_u64(overage
, high
);
2525 static u64
mem_find_max_overage(struct mem_cgroup
*memcg
)
2527 u64 overage
, max_overage
= 0;
2530 overage
= calculate_overage(page_counter_read(&memcg
->memory
),
2531 READ_ONCE(memcg
->memory
.high
));
2532 max_overage
= max(overage
, max_overage
);
2533 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2534 !mem_cgroup_is_root(memcg
));
2539 static u64
swap_find_max_overage(struct mem_cgroup
*memcg
)
2541 u64 overage
, max_overage
= 0;
2544 overage
= calculate_overage(page_counter_read(&memcg
->swap
),
2545 READ_ONCE(memcg
->swap
.high
));
2547 memcg_memory_event(memcg
, MEMCG_SWAP_HIGH
);
2548 max_overage
= max(overage
, max_overage
);
2549 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2550 !mem_cgroup_is_root(memcg
));
2556 * Get the number of jiffies that we should penalise a mischievous cgroup which
2557 * is exceeding its memory.high by checking both it and its ancestors.
2559 static unsigned long calculate_high_delay(struct mem_cgroup
*memcg
,
2560 unsigned int nr_pages
,
2563 unsigned long penalty_jiffies
;
2569 * We use overage compared to memory.high to calculate the number of
2570 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2571 * fairly lenient on small overages, and increasingly harsh when the
2572 * memcg in question makes it clear that it has no intention of stopping
2573 * its crazy behaviour, so we exponentially increase the delay based on
2576 penalty_jiffies
= max_overage
* max_overage
* HZ
;
2577 penalty_jiffies
>>= MEMCG_DELAY_PRECISION_SHIFT
;
2578 penalty_jiffies
>>= MEMCG_DELAY_SCALING_SHIFT
;
2581 * Factor in the task's own contribution to the overage, such that four
2582 * N-sized allocations are throttled approximately the same as one
2583 * 4N-sized allocation.
2585 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2586 * larger the current charge patch is than that.
2588 return penalty_jiffies
* nr_pages
/ MEMCG_CHARGE_BATCH
;
2592 * Scheduled by try_charge() to be executed from the userland return path
2593 * and reclaims memory over the high limit.
2595 void mem_cgroup_handle_over_high(void)
2597 unsigned long penalty_jiffies
;
2598 unsigned long pflags
;
2599 unsigned long nr_reclaimed
;
2600 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
2601 int nr_retries
= MAX_RECLAIM_RETRIES
;
2602 struct mem_cgroup
*memcg
;
2603 bool in_retry
= false;
2605 if (likely(!nr_pages
))
2608 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2609 current
->memcg_nr_pages_over_high
= 0;
2613 * The allocating task should reclaim at least the batch size, but for
2614 * subsequent retries we only want to do what's necessary to prevent oom
2615 * or breaching resource isolation.
2617 * This is distinct from memory.max or page allocator behaviour because
2618 * memory.high is currently batched, whereas memory.max and the page
2619 * allocator run every time an allocation is made.
2621 nr_reclaimed
= reclaim_high(memcg
,
2622 in_retry
? SWAP_CLUSTER_MAX
: nr_pages
,
2626 * memory.high is breached and reclaim is unable to keep up. Throttle
2627 * allocators proactively to slow down excessive growth.
2629 penalty_jiffies
= calculate_high_delay(memcg
, nr_pages
,
2630 mem_find_max_overage(memcg
));
2632 penalty_jiffies
+= calculate_high_delay(memcg
, nr_pages
,
2633 swap_find_max_overage(memcg
));
2636 * Clamp the max delay per usermode return so as to still keep the
2637 * application moving forwards and also permit diagnostics, albeit
2640 penalty_jiffies
= min(penalty_jiffies
, MEMCG_MAX_HIGH_DELAY_JIFFIES
);
2643 * Don't sleep if the amount of jiffies this memcg owes us is so low
2644 * that it's not even worth doing, in an attempt to be nice to those who
2645 * go only a small amount over their memory.high value and maybe haven't
2646 * been aggressively reclaimed enough yet.
2648 if (penalty_jiffies
<= HZ
/ 100)
2652 * If reclaim is making forward progress but we're still over
2653 * memory.high, we want to encourage that rather than doing allocator
2656 if (nr_reclaimed
|| nr_retries
--) {
2662 * If we exit early, we're guaranteed to die (since
2663 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2664 * need to account for any ill-begotten jiffies to pay them off later.
2666 psi_memstall_enter(&pflags
);
2667 schedule_timeout_killable(penalty_jiffies
);
2668 psi_memstall_leave(&pflags
);
2671 css_put(&memcg
->css
);
2674 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2675 unsigned int nr_pages
)
2677 unsigned int batch
= max(MEMCG_CHARGE_BATCH
, nr_pages
);
2678 int nr_retries
= MAX_RECLAIM_RETRIES
;
2679 struct mem_cgroup
*mem_over_limit
;
2680 struct page_counter
*counter
;
2681 enum oom_status oom_status
;
2682 unsigned long nr_reclaimed
;
2683 bool may_swap
= true;
2684 bool drained
= false;
2685 unsigned long pflags
;
2687 if (mem_cgroup_is_root(memcg
))
2690 if (consume_stock(memcg
, nr_pages
))
2693 if (!do_memsw_account() ||
2694 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2695 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2697 if (do_memsw_account())
2698 page_counter_uncharge(&memcg
->memsw
, batch
);
2699 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2701 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2705 if (batch
> nr_pages
) {
2711 * Memcg doesn't have a dedicated reserve for atomic
2712 * allocations. But like the global atomic pool, we need to
2713 * put the burden of reclaim on regular allocation requests
2714 * and let these go through as privileged allocations.
2716 if (gfp_mask
& __GFP_ATOMIC
)
2720 * Unlike in global OOM situations, memcg is not in a physical
2721 * memory shortage. Allow dying and OOM-killed tasks to
2722 * bypass the last charges so that they can exit quickly and
2723 * free their memory.
2725 if (unlikely(should_force_charge()))
2729 * Prevent unbounded recursion when reclaim operations need to
2730 * allocate memory. This might exceed the limits temporarily,
2731 * but we prefer facilitating memory reclaim and getting back
2732 * under the limit over triggering OOM kills in these cases.
2734 if (unlikely(current
->flags
& PF_MEMALLOC
))
2737 if (unlikely(task_in_memcg_oom(current
)))
2740 if (!gfpflags_allow_blocking(gfp_mask
))
2743 memcg_memory_event(mem_over_limit
, MEMCG_MAX
);
2745 psi_memstall_enter(&pflags
);
2746 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2747 gfp_mask
, may_swap
);
2748 psi_memstall_leave(&pflags
);
2750 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2754 drain_all_stock(mem_over_limit
);
2759 if (gfp_mask
& __GFP_NORETRY
)
2762 * Even though the limit is exceeded at this point, reclaim
2763 * may have been able to free some pages. Retry the charge
2764 * before killing the task.
2766 * Only for regular pages, though: huge pages are rather
2767 * unlikely to succeed so close to the limit, and we fall back
2768 * to regular pages anyway in case of failure.
2770 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2773 * At task move, charge accounts can be doubly counted. So, it's
2774 * better to wait until the end of task_move if something is going on.
2776 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2782 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
2785 if (gfp_mask
& __GFP_NOFAIL
)
2788 if (fatal_signal_pending(current
))
2792 * keep retrying as long as the memcg oom killer is able to make
2793 * a forward progress or bypass the charge if the oom killer
2794 * couldn't make any progress.
2796 oom_status
= mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2797 get_order(nr_pages
* PAGE_SIZE
));
2798 switch (oom_status
) {
2800 nr_retries
= MAX_RECLAIM_RETRIES
;
2808 if (!(gfp_mask
& __GFP_NOFAIL
))
2812 * The allocation either can't fail or will lead to more memory
2813 * being freed very soon. Allow memory usage go over the limit
2814 * temporarily by force charging it.
2816 page_counter_charge(&memcg
->memory
, nr_pages
);
2817 if (do_memsw_account())
2818 page_counter_charge(&memcg
->memsw
, nr_pages
);
2823 if (batch
> nr_pages
)
2824 refill_stock(memcg
, batch
- nr_pages
);
2827 * If the hierarchy is above the normal consumption range, schedule
2828 * reclaim on returning to userland. We can perform reclaim here
2829 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2830 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2831 * not recorded as it most likely matches current's and won't
2832 * change in the meantime. As high limit is checked again before
2833 * reclaim, the cost of mismatch is negligible.
2836 bool mem_high
, swap_high
;
2838 mem_high
= page_counter_read(&memcg
->memory
) >
2839 READ_ONCE(memcg
->memory
.high
);
2840 swap_high
= page_counter_read(&memcg
->swap
) >
2841 READ_ONCE(memcg
->swap
.high
);
2843 /* Don't bother a random interrupted task */
2844 if (in_interrupt()) {
2846 schedule_work(&memcg
->high_work
);
2852 if (mem_high
|| swap_high
) {
2854 * The allocating tasks in this cgroup will need to do
2855 * reclaim or be throttled to prevent further growth
2856 * of the memory or swap footprints.
2858 * Target some best-effort fairness between the tasks,
2859 * and distribute reclaim work and delay penalties
2860 * based on how much each task is actually allocating.
2862 current
->memcg_nr_pages_over_high
+= batch
;
2863 set_notify_resume(current
);
2866 } while ((memcg
= parent_mem_cgroup(memcg
)));
2871 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2872 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2874 if (mem_cgroup_is_root(memcg
))
2877 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2878 if (do_memsw_account())
2879 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2883 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
)
2885 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2887 * Any of the following ensures page->mem_cgroup stability:
2891 * - lock_page_memcg()
2892 * - exclusive reference
2894 page
->mem_cgroup
= memcg
;
2897 #ifdef CONFIG_MEMCG_KMEM
2898 int memcg_alloc_page_obj_cgroups(struct page
*page
, struct kmem_cache
*s
,
2901 unsigned int objects
= objs_per_slab_page(s
, page
);
2904 vec
= kcalloc_node(objects
, sizeof(struct obj_cgroup
*), gfp
,
2909 if (cmpxchg(&page
->obj_cgroups
, NULL
,
2910 (struct obj_cgroup
**) ((unsigned long)vec
| 0x1UL
)))
2913 kmemleak_not_leak(vec
);
2919 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2921 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2922 * cgroup_mutex, etc.
2924 struct mem_cgroup
*mem_cgroup_from_obj(void *p
)
2928 if (mem_cgroup_disabled())
2931 page
= virt_to_head_page(p
);
2934 * If page->mem_cgroup is set, it's either a simple mem_cgroup pointer
2935 * or a pointer to obj_cgroup vector. In the latter case the lowest
2936 * bit of the pointer is set.
2937 * The page->mem_cgroup pointer can be asynchronously changed
2938 * from NULL to (obj_cgroup_vec | 0x1UL), but can't be changed
2939 * from a valid memcg pointer to objcg vector or back.
2941 if (!page
->mem_cgroup
)
2945 * Slab objects are accounted individually, not per-page.
2946 * Memcg membership data for each individual object is saved in
2947 * the page->obj_cgroups.
2949 if (page_has_obj_cgroups(page
)) {
2950 struct obj_cgroup
*objcg
;
2953 off
= obj_to_index(page
->slab_cache
, page
, p
);
2954 objcg
= page_obj_cgroups(page
)[off
];
2956 return obj_cgroup_memcg(objcg
);
2961 /* All other pages use page->mem_cgroup */
2962 return page
->mem_cgroup
;
2965 __always_inline
struct obj_cgroup
*get_obj_cgroup_from_current(void)
2967 struct obj_cgroup
*objcg
= NULL
;
2968 struct mem_cgroup
*memcg
;
2970 if (memcg_kmem_bypass())
2974 if (unlikely(active_memcg()))
2975 memcg
= active_memcg();
2977 memcg
= mem_cgroup_from_task(current
);
2979 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
)) {
2980 objcg
= rcu_dereference(memcg
->objcg
);
2981 if (objcg
&& obj_cgroup_tryget(objcg
))
2989 static int memcg_alloc_cache_id(void)
2994 id
= ida_simple_get(&memcg_cache_ida
,
2995 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2999 if (id
< memcg_nr_cache_ids
)
3003 * There's no space for the new id in memcg_caches arrays,
3004 * so we have to grow them.
3006 down_write(&memcg_cache_ids_sem
);
3008 size
= 2 * (id
+ 1);
3009 if (size
< MEMCG_CACHES_MIN_SIZE
)
3010 size
= MEMCG_CACHES_MIN_SIZE
;
3011 else if (size
> MEMCG_CACHES_MAX_SIZE
)
3012 size
= MEMCG_CACHES_MAX_SIZE
;
3014 err
= memcg_update_all_list_lrus(size
);
3016 memcg_nr_cache_ids
= size
;
3018 up_write(&memcg_cache_ids_sem
);
3021 ida_simple_remove(&memcg_cache_ida
, id
);
3027 static void memcg_free_cache_id(int id
)
3029 ida_simple_remove(&memcg_cache_ida
, id
);
3033 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3034 * @memcg: memory cgroup to charge
3035 * @gfp: reclaim mode
3036 * @nr_pages: number of pages to charge
3038 * Returns 0 on success, an error code on failure.
3040 int __memcg_kmem_charge(struct mem_cgroup
*memcg
, gfp_t gfp
,
3041 unsigned int nr_pages
)
3043 struct page_counter
*counter
;
3046 ret
= try_charge(memcg
, gfp
, nr_pages
);
3050 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
3051 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
3054 * Enforce __GFP_NOFAIL allocation because callers are not
3055 * prepared to see failures and likely do not have any failure
3058 if (gfp
& __GFP_NOFAIL
) {
3059 page_counter_charge(&memcg
->kmem
, nr_pages
);
3062 cancel_charge(memcg
, nr_pages
);
3069 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3070 * @memcg: memcg to uncharge
3071 * @nr_pages: number of pages to uncharge
3073 void __memcg_kmem_uncharge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
3075 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
3076 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
3078 page_counter_uncharge(&memcg
->memory
, nr_pages
);
3079 if (do_memsw_account())
3080 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
3084 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3085 * @page: page to charge
3086 * @gfp: reclaim mode
3087 * @order: allocation order
3089 * Returns 0 on success, an error code on failure.
3091 int __memcg_kmem_charge_page(struct page
*page
, gfp_t gfp
, int order
)
3093 struct mem_cgroup
*memcg
;
3096 memcg
= get_mem_cgroup_from_current();
3097 if (memcg
&& !mem_cgroup_is_root(memcg
)) {
3098 ret
= __memcg_kmem_charge(memcg
, gfp
, 1 << order
);
3100 page
->mem_cgroup
= memcg
;
3101 __SetPageKmemcg(page
);
3104 css_put(&memcg
->css
);
3110 * __memcg_kmem_uncharge_page: uncharge a kmem page
3111 * @page: page to uncharge
3112 * @order: allocation order
3114 void __memcg_kmem_uncharge_page(struct page
*page
, int order
)
3116 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
3117 unsigned int nr_pages
= 1 << order
;
3122 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
3123 __memcg_kmem_uncharge(memcg
, nr_pages
);
3124 page
->mem_cgroup
= NULL
;
3125 css_put(&memcg
->css
);
3127 /* slab pages do not have PageKmemcg flag set */
3128 if (PageKmemcg(page
))
3129 __ClearPageKmemcg(page
);
3132 static bool consume_obj_stock(struct obj_cgroup
*objcg
, unsigned int nr_bytes
)
3134 struct memcg_stock_pcp
*stock
;
3135 unsigned long flags
;
3138 local_irq_save(flags
);
3140 stock
= this_cpu_ptr(&memcg_stock
);
3141 if (objcg
== stock
->cached_objcg
&& stock
->nr_bytes
>= nr_bytes
) {
3142 stock
->nr_bytes
-= nr_bytes
;
3146 local_irq_restore(flags
);
3151 static void drain_obj_stock(struct memcg_stock_pcp
*stock
)
3153 struct obj_cgroup
*old
= stock
->cached_objcg
;
3158 if (stock
->nr_bytes
) {
3159 unsigned int nr_pages
= stock
->nr_bytes
>> PAGE_SHIFT
;
3160 unsigned int nr_bytes
= stock
->nr_bytes
& (PAGE_SIZE
- 1);
3164 __memcg_kmem_uncharge(obj_cgroup_memcg(old
), nr_pages
);
3169 * The leftover is flushed to the centralized per-memcg value.
3170 * On the next attempt to refill obj stock it will be moved
3171 * to a per-cpu stock (probably, on an other CPU), see
3172 * refill_obj_stock().
3174 * How often it's flushed is a trade-off between the memory
3175 * limit enforcement accuracy and potential CPU contention,
3176 * so it might be changed in the future.
3178 atomic_add(nr_bytes
, &old
->nr_charged_bytes
);
3179 stock
->nr_bytes
= 0;
3182 obj_cgroup_put(old
);
3183 stock
->cached_objcg
= NULL
;
3186 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
3187 struct mem_cgroup
*root_memcg
)
3189 struct mem_cgroup
*memcg
;
3191 if (stock
->cached_objcg
) {
3192 memcg
= obj_cgroup_memcg(stock
->cached_objcg
);
3193 if (memcg
&& mem_cgroup_is_descendant(memcg
, root_memcg
))
3200 static void refill_obj_stock(struct obj_cgroup
*objcg
, unsigned int nr_bytes
)
3202 struct memcg_stock_pcp
*stock
;
3203 unsigned long flags
;
3205 local_irq_save(flags
);
3207 stock
= this_cpu_ptr(&memcg_stock
);
3208 if (stock
->cached_objcg
!= objcg
) { /* reset if necessary */
3209 drain_obj_stock(stock
);
3210 obj_cgroup_get(objcg
);
3211 stock
->cached_objcg
= objcg
;
3212 stock
->nr_bytes
= atomic_xchg(&objcg
->nr_charged_bytes
, 0);
3214 stock
->nr_bytes
+= nr_bytes
;
3216 if (stock
->nr_bytes
> PAGE_SIZE
)
3217 drain_obj_stock(stock
);
3219 local_irq_restore(flags
);
3222 int obj_cgroup_charge(struct obj_cgroup
*objcg
, gfp_t gfp
, size_t size
)
3224 struct mem_cgroup
*memcg
;
3225 unsigned int nr_pages
, nr_bytes
;
3228 if (consume_obj_stock(objcg
, size
))
3232 * In theory, memcg->nr_charged_bytes can have enough
3233 * pre-charged bytes to satisfy the allocation. However,
3234 * flushing memcg->nr_charged_bytes requires two atomic
3235 * operations, and memcg->nr_charged_bytes can't be big,
3236 * so it's better to ignore it and try grab some new pages.
3237 * memcg->nr_charged_bytes will be flushed in
3238 * refill_obj_stock(), called from this function or
3239 * independently later.
3242 memcg
= obj_cgroup_memcg(objcg
);
3243 css_get(&memcg
->css
);
3246 nr_pages
= size
>> PAGE_SHIFT
;
3247 nr_bytes
= size
& (PAGE_SIZE
- 1);
3252 ret
= __memcg_kmem_charge(memcg
, gfp
, nr_pages
);
3253 if (!ret
&& nr_bytes
)
3254 refill_obj_stock(objcg
, PAGE_SIZE
- nr_bytes
);
3256 css_put(&memcg
->css
);
3260 void obj_cgroup_uncharge(struct obj_cgroup
*objcg
, size_t size
)
3262 refill_obj_stock(objcg
, size
);
3265 #endif /* CONFIG_MEMCG_KMEM */
3267 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3270 * Because tail pages are not marked as "used", set it. We're under
3271 * pgdat->lru_lock and migration entries setup in all page mappings.
3273 void mem_cgroup_split_huge_fixup(struct page
*head
)
3275 struct mem_cgroup
*memcg
= head
->mem_cgroup
;
3278 if (mem_cgroup_disabled())
3281 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3282 css_get(&memcg
->css
);
3283 head
[i
].mem_cgroup
= memcg
;
3286 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3288 #ifdef CONFIG_MEMCG_SWAP
3290 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3291 * @entry: swap entry to be moved
3292 * @from: mem_cgroup which the entry is moved from
3293 * @to: mem_cgroup which the entry is moved to
3295 * It succeeds only when the swap_cgroup's record for this entry is the same
3296 * as the mem_cgroup's id of @from.
3298 * Returns 0 on success, -EINVAL on failure.
3300 * The caller must have charged to @to, IOW, called page_counter_charge() about
3301 * both res and memsw, and called css_get().
3303 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
3304 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3306 unsigned short old_id
, new_id
;
3308 old_id
= mem_cgroup_id(from
);
3309 new_id
= mem_cgroup_id(to
);
3311 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
3312 mod_memcg_state(from
, MEMCG_SWAP
, -1);
3313 mod_memcg_state(to
, MEMCG_SWAP
, 1);
3319 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
3320 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3326 static DEFINE_MUTEX(memcg_max_mutex
);
3328 static int mem_cgroup_resize_max(struct mem_cgroup
*memcg
,
3329 unsigned long max
, bool memsw
)
3331 bool enlarge
= false;
3332 bool drained
= false;
3334 bool limits_invariant
;
3335 struct page_counter
*counter
= memsw
? &memcg
->memsw
: &memcg
->memory
;
3338 if (signal_pending(current
)) {
3343 mutex_lock(&memcg_max_mutex
);
3345 * Make sure that the new limit (memsw or memory limit) doesn't
3346 * break our basic invariant rule memory.max <= memsw.max.
3348 limits_invariant
= memsw
? max
>= READ_ONCE(memcg
->memory
.max
) :
3349 max
<= memcg
->memsw
.max
;
3350 if (!limits_invariant
) {
3351 mutex_unlock(&memcg_max_mutex
);
3355 if (max
> counter
->max
)
3357 ret
= page_counter_set_max(counter
, max
);
3358 mutex_unlock(&memcg_max_mutex
);
3364 drain_all_stock(memcg
);
3369 if (!try_to_free_mem_cgroup_pages(memcg
, 1,
3370 GFP_KERNEL
, !memsw
)) {
3376 if (!ret
&& enlarge
)
3377 memcg_oom_recover(memcg
);
3382 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
3384 unsigned long *total_scanned
)
3386 unsigned long nr_reclaimed
= 0;
3387 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
3388 unsigned long reclaimed
;
3390 struct mem_cgroup_tree_per_node
*mctz
;
3391 unsigned long excess
;
3392 unsigned long nr_scanned
;
3397 mctz
= soft_limit_tree_node(pgdat
->node_id
);
3400 * Do not even bother to check the largest node if the root
3401 * is empty. Do it lockless to prevent lock bouncing. Races
3402 * are acceptable as soft limit is best effort anyway.
3404 if (!mctz
|| RB_EMPTY_ROOT(&mctz
->rb_root
))
3408 * This loop can run a while, specially if mem_cgroup's continuously
3409 * keep exceeding their soft limit and putting the system under
3416 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3421 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
3422 gfp_mask
, &nr_scanned
);
3423 nr_reclaimed
+= reclaimed
;
3424 *total_scanned
+= nr_scanned
;
3425 spin_lock_irq(&mctz
->lock
);
3426 __mem_cgroup_remove_exceeded(mz
, mctz
);
3429 * If we failed to reclaim anything from this memory cgroup
3430 * it is time to move on to the next cgroup
3434 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
3436 excess
= soft_limit_excess(mz
->memcg
);
3438 * One school of thought says that we should not add
3439 * back the node to the tree if reclaim returns 0.
3440 * But our reclaim could return 0, simply because due
3441 * to priority we are exposing a smaller subset of
3442 * memory to reclaim from. Consider this as a longer
3445 /* If excess == 0, no tree ops */
3446 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
3447 spin_unlock_irq(&mctz
->lock
);
3448 css_put(&mz
->memcg
->css
);
3451 * Could not reclaim anything and there are no more
3452 * mem cgroups to try or we seem to be looping without
3453 * reclaiming anything.
3455 if (!nr_reclaimed
&&
3457 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3459 } while (!nr_reclaimed
);
3461 css_put(&next_mz
->memcg
->css
);
3462 return nr_reclaimed
;
3466 * Test whether @memcg has children, dead or alive. Note that this
3467 * function doesn't care whether @memcg has use_hierarchy enabled and
3468 * returns %true if there are child csses according to the cgroup
3469 * hierarchy. Testing use_hierarchy is the caller's responsibility.
3471 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
3476 ret
= css_next_child(NULL
, &memcg
->css
);
3482 * Reclaims as many pages from the given memcg as possible.
3484 * Caller is responsible for holding css reference for memcg.
3486 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
3488 int nr_retries
= MAX_RECLAIM_RETRIES
;
3490 /* we call try-to-free pages for make this cgroup empty */
3491 lru_add_drain_all();
3493 drain_all_stock(memcg
);
3495 /* try to free all pages in this cgroup */
3496 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
3499 if (signal_pending(current
))
3502 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
3506 /* maybe some writeback is necessary */
3507 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3515 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3516 char *buf
, size_t nbytes
,
3519 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3521 if (mem_cgroup_is_root(memcg
))
3523 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3526 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3529 return mem_cgroup_from_css(css
)->use_hierarchy
;
3532 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3533 struct cftype
*cft
, u64 val
)
3536 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3537 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
3539 if (memcg
->use_hierarchy
== val
)
3543 * If parent's use_hierarchy is set, we can't make any modifications
3544 * in the child subtrees. If it is unset, then the change can
3545 * occur, provided the current cgroup has no children.
3547 * For the root cgroup, parent_mem is NULL, we allow value to be
3548 * set if there are no children.
3550 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
3551 (val
== 1 || val
== 0)) {
3552 if (!memcg_has_children(memcg
))
3553 memcg
->use_hierarchy
= val
;
3562 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3566 if (mem_cgroup_is_root(memcg
)) {
3567 val
= memcg_page_state(memcg
, NR_FILE_PAGES
) +
3568 memcg_page_state(memcg
, NR_ANON_MAPPED
);
3570 val
+= memcg_page_state(memcg
, MEMCG_SWAP
);
3573 val
= page_counter_read(&memcg
->memory
);
3575 val
= page_counter_read(&memcg
->memsw
);
3588 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3591 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3592 struct page_counter
*counter
;
3594 switch (MEMFILE_TYPE(cft
->private)) {
3596 counter
= &memcg
->memory
;
3599 counter
= &memcg
->memsw
;
3602 counter
= &memcg
->kmem
;
3605 counter
= &memcg
->tcpmem
;
3611 switch (MEMFILE_ATTR(cft
->private)) {
3613 if (counter
== &memcg
->memory
)
3614 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
3615 if (counter
== &memcg
->memsw
)
3616 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
3617 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3619 return (u64
)counter
->max
* PAGE_SIZE
;
3621 return (u64
)counter
->watermark
* PAGE_SIZE
;
3623 return counter
->failcnt
;
3624 case RES_SOFT_LIMIT
:
3625 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
3631 static void memcg_flush_percpu_vmstats(struct mem_cgroup
*memcg
)
3633 unsigned long stat
[MEMCG_NR_STAT
] = {0};
3634 struct mem_cgroup
*mi
;
3637 for_each_online_cpu(cpu
)
3638 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
3639 stat
[i
] += per_cpu(memcg
->vmstats_percpu
->stat
[i
], cpu
);
3641 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
3642 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
3643 atomic_long_add(stat
[i
], &mi
->vmstats
[i
]);
3645 for_each_node(node
) {
3646 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
3647 struct mem_cgroup_per_node
*pi
;
3649 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3652 for_each_online_cpu(cpu
)
3653 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3655 pn
->lruvec_stat_cpu
->count
[i
], cpu
);
3657 for (pi
= pn
; pi
; pi
= parent_nodeinfo(pi
, node
))
3658 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3659 atomic_long_add(stat
[i
], &pi
->lruvec_stat
[i
]);
3663 static void memcg_flush_percpu_vmevents(struct mem_cgroup
*memcg
)
3665 unsigned long events
[NR_VM_EVENT_ITEMS
];
3666 struct mem_cgroup
*mi
;
3669 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3672 for_each_online_cpu(cpu
)
3673 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3674 events
[i
] += per_cpu(memcg
->vmstats_percpu
->events
[i
],
3677 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
3678 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3679 atomic_long_add(events
[i
], &mi
->vmevents
[i
]);
3682 #ifdef CONFIG_MEMCG_KMEM
3683 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3685 struct obj_cgroup
*objcg
;
3688 if (cgroup_memory_nokmem
)
3691 BUG_ON(memcg
->kmemcg_id
>= 0);
3692 BUG_ON(memcg
->kmem_state
);
3694 memcg_id
= memcg_alloc_cache_id();
3698 objcg
= obj_cgroup_alloc();
3700 memcg_free_cache_id(memcg_id
);
3703 objcg
->memcg
= memcg
;
3704 rcu_assign_pointer(memcg
->objcg
, objcg
);
3706 static_branch_enable(&memcg_kmem_enabled_key
);
3709 * A memory cgroup is considered kmem-online as soon as it gets
3710 * kmemcg_id. Setting the id after enabling static branching will
3711 * guarantee no one starts accounting before all call sites are
3714 memcg
->kmemcg_id
= memcg_id
;
3715 memcg
->kmem_state
= KMEM_ONLINE
;
3720 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3722 struct cgroup_subsys_state
*css
;
3723 struct mem_cgroup
*parent
, *child
;
3726 if (memcg
->kmem_state
!= KMEM_ONLINE
)
3729 memcg
->kmem_state
= KMEM_ALLOCATED
;
3731 parent
= parent_mem_cgroup(memcg
);
3733 parent
= root_mem_cgroup
;
3735 memcg_reparent_objcgs(memcg
, parent
);
3737 kmemcg_id
= memcg
->kmemcg_id
;
3738 BUG_ON(kmemcg_id
< 0);
3741 * Change kmemcg_id of this cgroup and all its descendants to the
3742 * parent's id, and then move all entries from this cgroup's list_lrus
3743 * to ones of the parent. After we have finished, all list_lrus
3744 * corresponding to this cgroup are guaranteed to remain empty. The
3745 * ordering is imposed by list_lru_node->lock taken by
3746 * memcg_drain_all_list_lrus().
3748 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3749 css_for_each_descendant_pre(css
, &memcg
->css
) {
3750 child
= mem_cgroup_from_css(css
);
3751 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
3752 child
->kmemcg_id
= parent
->kmemcg_id
;
3753 if (!memcg
->use_hierarchy
)
3758 memcg_drain_all_list_lrus(kmemcg_id
, parent
);
3760 memcg_free_cache_id(kmemcg_id
);
3763 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3765 /* css_alloc() failed, offlining didn't happen */
3766 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
3767 memcg_offline_kmem(memcg
);
3770 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3774 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3777 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3780 #endif /* CONFIG_MEMCG_KMEM */
3782 static int memcg_update_kmem_max(struct mem_cgroup
*memcg
,
3787 mutex_lock(&memcg_max_mutex
);
3788 ret
= page_counter_set_max(&memcg
->kmem
, max
);
3789 mutex_unlock(&memcg_max_mutex
);
3793 static int memcg_update_tcp_max(struct mem_cgroup
*memcg
, unsigned long max
)
3797 mutex_lock(&memcg_max_mutex
);
3799 ret
= page_counter_set_max(&memcg
->tcpmem
, max
);
3803 if (!memcg
->tcpmem_active
) {
3805 * The active flag needs to be written after the static_key
3806 * update. This is what guarantees that the socket activation
3807 * function is the last one to run. See mem_cgroup_sk_alloc()
3808 * for details, and note that we don't mark any socket as
3809 * belonging to this memcg until that flag is up.
3811 * We need to do this, because static_keys will span multiple
3812 * sites, but we can't control their order. If we mark a socket
3813 * as accounted, but the accounting functions are not patched in
3814 * yet, we'll lose accounting.
3816 * We never race with the readers in mem_cgroup_sk_alloc(),
3817 * because when this value change, the code to process it is not
3820 static_branch_inc(&memcg_sockets_enabled_key
);
3821 memcg
->tcpmem_active
= true;
3824 mutex_unlock(&memcg_max_mutex
);
3829 * The user of this function is...
3832 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3833 char *buf
, size_t nbytes
, loff_t off
)
3835 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3836 unsigned long nr_pages
;
3839 buf
= strstrip(buf
);
3840 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3844 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3846 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3850 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3852 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, false);
3855 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, true);
3858 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3859 "Please report your usecase to linux-mm@kvack.org if you "
3860 "depend on this functionality.\n");
3861 ret
= memcg_update_kmem_max(memcg
, nr_pages
);
3864 ret
= memcg_update_tcp_max(memcg
, nr_pages
);
3868 case RES_SOFT_LIMIT
:
3869 memcg
->soft_limit
= nr_pages
;
3873 return ret
?: nbytes
;
3876 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3877 size_t nbytes
, loff_t off
)
3879 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3880 struct page_counter
*counter
;
3882 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3884 counter
= &memcg
->memory
;
3887 counter
= &memcg
->memsw
;
3890 counter
= &memcg
->kmem
;
3893 counter
= &memcg
->tcpmem
;
3899 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3901 page_counter_reset_watermark(counter
);
3904 counter
->failcnt
= 0;
3913 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3916 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3920 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3921 struct cftype
*cft
, u64 val
)
3923 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3925 if (val
& ~MOVE_MASK
)
3929 * No kind of locking is needed in here, because ->can_attach() will
3930 * check this value once in the beginning of the process, and then carry
3931 * on with stale data. This means that changes to this value will only
3932 * affect task migrations starting after the change.
3934 memcg
->move_charge_at_immigrate
= val
;
3938 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3939 struct cftype
*cft
, u64 val
)
3947 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3948 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3949 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3951 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
3952 int nid
, unsigned int lru_mask
, bool tree
)
3954 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, NODE_DATA(nid
));
3955 unsigned long nr
= 0;
3958 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
3961 if (!(BIT(lru
) & lru_mask
))
3964 nr
+= lruvec_page_state(lruvec
, NR_LRU_BASE
+ lru
);
3966 nr
+= lruvec_page_state_local(lruvec
, NR_LRU_BASE
+ lru
);
3971 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
3972 unsigned int lru_mask
,
3975 unsigned long nr
= 0;
3979 if (!(BIT(lru
) & lru_mask
))
3982 nr
+= memcg_page_state(memcg
, NR_LRU_BASE
+ lru
);
3984 nr
+= memcg_page_state_local(memcg
, NR_LRU_BASE
+ lru
);
3989 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3993 unsigned int lru_mask
;
3996 static const struct numa_stat stats
[] = {
3997 { "total", LRU_ALL
},
3998 { "file", LRU_ALL_FILE
},
3999 { "anon", LRU_ALL_ANON
},
4000 { "unevictable", BIT(LRU_UNEVICTABLE
) },
4002 const struct numa_stat
*stat
;
4004 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
4006 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
4007 seq_printf(m
, "%s=%lu", stat
->name
,
4008 mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
,
4010 for_each_node_state(nid
, N_MEMORY
)
4011 seq_printf(m
, " N%d=%lu", nid
,
4012 mem_cgroup_node_nr_lru_pages(memcg
, nid
,
4013 stat
->lru_mask
, false));
4017 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
4019 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
,
4020 mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
,
4022 for_each_node_state(nid
, N_MEMORY
)
4023 seq_printf(m
, " N%d=%lu", nid
,
4024 mem_cgroup_node_nr_lru_pages(memcg
, nid
,
4025 stat
->lru_mask
, true));
4031 #endif /* CONFIG_NUMA */
4033 static const unsigned int memcg1_stats
[] = {
4036 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4046 static const char *const memcg1_stat_names
[] = {
4049 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4059 /* Universal VM events cgroup1 shows, original sort order */
4060 static const unsigned int memcg1_events
[] = {
4067 static int memcg_stat_show(struct seq_file
*m
, void *v
)
4069 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
4070 unsigned long memory
, memsw
;
4071 struct mem_cgroup
*mi
;
4074 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names
) != ARRAY_SIZE(memcg1_stats
));
4076 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
4079 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
4081 nr
= memcg_page_state_local(memcg
, memcg1_stats
[i
]);
4082 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4083 if (memcg1_stats
[i
] == NR_ANON_THPS
)
4086 seq_printf(m
, "%s %lu\n", memcg1_stat_names
[i
], nr
* PAGE_SIZE
);
4089 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
4090 seq_printf(m
, "%s %lu\n", vm_event_name(memcg1_events
[i
]),
4091 memcg_events_local(memcg
, memcg1_events
[i
]));
4093 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
4094 seq_printf(m
, "%s %lu\n", lru_list_name(i
),
4095 memcg_page_state_local(memcg
, NR_LRU_BASE
+ i
) *
4098 /* Hierarchical information */
4099 memory
= memsw
= PAGE_COUNTER_MAX
;
4100 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
4101 memory
= min(memory
, READ_ONCE(mi
->memory
.max
));
4102 memsw
= min(memsw
, READ_ONCE(mi
->memsw
.max
));
4104 seq_printf(m
, "hierarchical_memory_limit %llu\n",
4105 (u64
)memory
* PAGE_SIZE
);
4106 if (do_memsw_account())
4107 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
4108 (u64
)memsw
* PAGE_SIZE
);
4110 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
4113 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
4115 nr
= memcg_page_state(memcg
, memcg1_stats
[i
]);
4116 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4117 if (memcg1_stats
[i
] == NR_ANON_THPS
)
4120 seq_printf(m
, "total_%s %llu\n", memcg1_stat_names
[i
],
4121 (u64
)nr
* PAGE_SIZE
);
4124 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
4125 seq_printf(m
, "total_%s %llu\n",
4126 vm_event_name(memcg1_events
[i
]),
4127 (u64
)memcg_events(memcg
, memcg1_events
[i
]));
4129 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
4130 seq_printf(m
, "total_%s %llu\n", lru_list_name(i
),
4131 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
4134 #ifdef CONFIG_DEBUG_VM
4137 struct mem_cgroup_per_node
*mz
;
4138 unsigned long anon_cost
= 0;
4139 unsigned long file_cost
= 0;
4141 for_each_online_pgdat(pgdat
) {
4142 mz
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
4144 anon_cost
+= mz
->lruvec
.anon_cost
;
4145 file_cost
+= mz
->lruvec
.file_cost
;
4147 seq_printf(m
, "anon_cost %lu\n", anon_cost
);
4148 seq_printf(m
, "file_cost %lu\n", file_cost
);
4155 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
4158 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4160 return mem_cgroup_swappiness(memcg
);
4163 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
4164 struct cftype
*cft
, u64 val
)
4166 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4172 memcg
->swappiness
= val
;
4174 vm_swappiness
= val
;
4179 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
4181 struct mem_cgroup_threshold_ary
*t
;
4182 unsigned long usage
;
4187 t
= rcu_dereference(memcg
->thresholds
.primary
);
4189 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
4194 usage
= mem_cgroup_usage(memcg
, swap
);
4197 * current_threshold points to threshold just below or equal to usage.
4198 * If it's not true, a threshold was crossed after last
4199 * call of __mem_cgroup_threshold().
4201 i
= t
->current_threshold
;
4204 * Iterate backward over array of thresholds starting from
4205 * current_threshold and check if a threshold is crossed.
4206 * If none of thresholds below usage is crossed, we read
4207 * only one element of the array here.
4209 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
4210 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4212 /* i = current_threshold + 1 */
4216 * Iterate forward over array of thresholds starting from
4217 * current_threshold+1 and check if a threshold is crossed.
4218 * If none of thresholds above usage is crossed, we read
4219 * only one element of the array here.
4221 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
4222 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4224 /* Update current_threshold */
4225 t
->current_threshold
= i
- 1;
4230 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
4233 __mem_cgroup_threshold(memcg
, false);
4234 if (do_memsw_account())
4235 __mem_cgroup_threshold(memcg
, true);
4237 memcg
= parent_mem_cgroup(memcg
);
4241 static int compare_thresholds(const void *a
, const void *b
)
4243 const struct mem_cgroup_threshold
*_a
= a
;
4244 const struct mem_cgroup_threshold
*_b
= b
;
4246 if (_a
->threshold
> _b
->threshold
)
4249 if (_a
->threshold
< _b
->threshold
)
4255 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
4257 struct mem_cgroup_eventfd_list
*ev
;
4259 spin_lock(&memcg_oom_lock
);
4261 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
4262 eventfd_signal(ev
->eventfd
, 1);
4264 spin_unlock(&memcg_oom_lock
);
4268 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
4270 struct mem_cgroup
*iter
;
4272 for_each_mem_cgroup_tree(iter
, memcg
)
4273 mem_cgroup_oom_notify_cb(iter
);
4276 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4277 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
4279 struct mem_cgroup_thresholds
*thresholds
;
4280 struct mem_cgroup_threshold_ary
*new;
4281 unsigned long threshold
;
4282 unsigned long usage
;
4285 ret
= page_counter_memparse(args
, "-1", &threshold
);
4289 mutex_lock(&memcg
->thresholds_lock
);
4292 thresholds
= &memcg
->thresholds
;
4293 usage
= mem_cgroup_usage(memcg
, false);
4294 } else if (type
== _MEMSWAP
) {
4295 thresholds
= &memcg
->memsw_thresholds
;
4296 usage
= mem_cgroup_usage(memcg
, true);
4300 /* Check if a threshold crossed before adding a new one */
4301 if (thresholds
->primary
)
4302 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4304 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
4306 /* Allocate memory for new array of thresholds */
4307 new = kmalloc(struct_size(new, entries
, size
), GFP_KERNEL
);
4314 /* Copy thresholds (if any) to new array */
4315 if (thresholds
->primary
)
4316 memcpy(new->entries
, thresholds
->primary
->entries
,
4317 flex_array_size(new, entries
, size
- 1));
4319 /* Add new threshold */
4320 new->entries
[size
- 1].eventfd
= eventfd
;
4321 new->entries
[size
- 1].threshold
= threshold
;
4323 /* Sort thresholds. Registering of new threshold isn't time-critical */
4324 sort(new->entries
, size
, sizeof(*new->entries
),
4325 compare_thresholds
, NULL
);
4327 /* Find current threshold */
4328 new->current_threshold
= -1;
4329 for (i
= 0; i
< size
; i
++) {
4330 if (new->entries
[i
].threshold
<= usage
) {
4332 * new->current_threshold will not be used until
4333 * rcu_assign_pointer(), so it's safe to increment
4336 ++new->current_threshold
;
4341 /* Free old spare buffer and save old primary buffer as spare */
4342 kfree(thresholds
->spare
);
4343 thresholds
->spare
= thresholds
->primary
;
4345 rcu_assign_pointer(thresholds
->primary
, new);
4347 /* To be sure that nobody uses thresholds */
4351 mutex_unlock(&memcg
->thresholds_lock
);
4356 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4357 struct eventfd_ctx
*eventfd
, const char *args
)
4359 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
4362 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4363 struct eventfd_ctx
*eventfd
, const char *args
)
4365 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
4368 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4369 struct eventfd_ctx
*eventfd
, enum res_type type
)
4371 struct mem_cgroup_thresholds
*thresholds
;
4372 struct mem_cgroup_threshold_ary
*new;
4373 unsigned long usage
;
4374 int i
, j
, size
, entries
;
4376 mutex_lock(&memcg
->thresholds_lock
);
4379 thresholds
= &memcg
->thresholds
;
4380 usage
= mem_cgroup_usage(memcg
, false);
4381 } else if (type
== _MEMSWAP
) {
4382 thresholds
= &memcg
->memsw_thresholds
;
4383 usage
= mem_cgroup_usage(memcg
, true);
4387 if (!thresholds
->primary
)
4390 /* Check if a threshold crossed before removing */
4391 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4393 /* Calculate new number of threshold */
4395 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4396 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4402 new = thresholds
->spare
;
4404 /* If no items related to eventfd have been cleared, nothing to do */
4408 /* Set thresholds array to NULL if we don't have thresholds */
4417 /* Copy thresholds and find current threshold */
4418 new->current_threshold
= -1;
4419 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4420 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4423 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4424 if (new->entries
[j
].threshold
<= usage
) {
4426 * new->current_threshold will not be used
4427 * until rcu_assign_pointer(), so it's safe to increment
4430 ++new->current_threshold
;
4436 /* Swap primary and spare array */
4437 thresholds
->spare
= thresholds
->primary
;
4439 rcu_assign_pointer(thresholds
->primary
, new);
4441 /* To be sure that nobody uses thresholds */
4444 /* If all events are unregistered, free the spare array */
4446 kfree(thresholds
->spare
);
4447 thresholds
->spare
= NULL
;
4450 mutex_unlock(&memcg
->thresholds_lock
);
4453 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4454 struct eventfd_ctx
*eventfd
)
4456 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
4459 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4460 struct eventfd_ctx
*eventfd
)
4462 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
4465 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
4466 struct eventfd_ctx
*eventfd
, const char *args
)
4468 struct mem_cgroup_eventfd_list
*event
;
4470 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4474 spin_lock(&memcg_oom_lock
);
4476 event
->eventfd
= eventfd
;
4477 list_add(&event
->list
, &memcg
->oom_notify
);
4479 /* already in OOM ? */
4480 if (memcg
->under_oom
)
4481 eventfd_signal(eventfd
, 1);
4482 spin_unlock(&memcg_oom_lock
);
4487 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
4488 struct eventfd_ctx
*eventfd
)
4490 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4492 spin_lock(&memcg_oom_lock
);
4494 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
4495 if (ev
->eventfd
== eventfd
) {
4496 list_del(&ev
->list
);
4501 spin_unlock(&memcg_oom_lock
);
4504 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
4506 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(sf
);
4508 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
4509 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
4510 seq_printf(sf
, "oom_kill %lu\n",
4511 atomic_long_read(&memcg
->memory_events
[MEMCG_OOM_KILL
]));
4515 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
4516 struct cftype
*cft
, u64 val
)
4518 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4520 /* cannot set to root cgroup and only 0 and 1 are allowed */
4521 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
4524 memcg
->oom_kill_disable
= val
;
4526 memcg_oom_recover(memcg
);
4531 #ifdef CONFIG_CGROUP_WRITEBACK
4533 #include <trace/events/writeback.h>
4535 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4537 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
4540 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4542 wb_domain_exit(&memcg
->cgwb_domain
);
4545 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4547 wb_domain_size_changed(&memcg
->cgwb_domain
);
4550 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
4552 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4554 if (!memcg
->css
.parent
)
4557 return &memcg
->cgwb_domain
;
4561 * idx can be of type enum memcg_stat_item or node_stat_item.
4562 * Keep in sync with memcg_exact_page().
4564 static unsigned long memcg_exact_page_state(struct mem_cgroup
*memcg
, int idx
)
4566 long x
= atomic_long_read(&memcg
->vmstats
[idx
]);
4569 for_each_online_cpu(cpu
)
4570 x
+= per_cpu_ptr(memcg
->vmstats_percpu
, cpu
)->stat
[idx
];
4577 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4578 * @wb: bdi_writeback in question
4579 * @pfilepages: out parameter for number of file pages
4580 * @pheadroom: out parameter for number of allocatable pages according to memcg
4581 * @pdirty: out parameter for number of dirty pages
4582 * @pwriteback: out parameter for number of pages under writeback
4584 * Determine the numbers of file, headroom, dirty, and writeback pages in
4585 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4586 * is a bit more involved.
4588 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4589 * headroom is calculated as the lowest headroom of itself and the
4590 * ancestors. Note that this doesn't consider the actual amount of
4591 * available memory in the system. The caller should further cap
4592 * *@pheadroom accordingly.
4594 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
4595 unsigned long *pheadroom
, unsigned long *pdirty
,
4596 unsigned long *pwriteback
)
4598 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4599 struct mem_cgroup
*parent
;
4601 *pdirty
= memcg_exact_page_state(memcg
, NR_FILE_DIRTY
);
4603 *pwriteback
= memcg_exact_page_state(memcg
, NR_WRITEBACK
);
4604 *pfilepages
= memcg_exact_page_state(memcg
, NR_INACTIVE_FILE
) +
4605 memcg_exact_page_state(memcg
, NR_ACTIVE_FILE
);
4606 *pheadroom
= PAGE_COUNTER_MAX
;
4608 while ((parent
= parent_mem_cgroup(memcg
))) {
4609 unsigned long ceiling
= min(READ_ONCE(memcg
->memory
.max
),
4610 READ_ONCE(memcg
->memory
.high
));
4611 unsigned long used
= page_counter_read(&memcg
->memory
);
4613 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
4619 * Foreign dirty flushing
4621 * There's an inherent mismatch between memcg and writeback. The former
4622 * trackes ownership per-page while the latter per-inode. This was a
4623 * deliberate design decision because honoring per-page ownership in the
4624 * writeback path is complicated, may lead to higher CPU and IO overheads
4625 * and deemed unnecessary given that write-sharing an inode across
4626 * different cgroups isn't a common use-case.
4628 * Combined with inode majority-writer ownership switching, this works well
4629 * enough in most cases but there are some pathological cases. For
4630 * example, let's say there are two cgroups A and B which keep writing to
4631 * different but confined parts of the same inode. B owns the inode and
4632 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4633 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4634 * triggering background writeback. A will be slowed down without a way to
4635 * make writeback of the dirty pages happen.
4637 * Conditions like the above can lead to a cgroup getting repatedly and
4638 * severely throttled after making some progress after each
4639 * dirty_expire_interval while the underyling IO device is almost
4642 * Solving this problem completely requires matching the ownership tracking
4643 * granularities between memcg and writeback in either direction. However,
4644 * the more egregious behaviors can be avoided by simply remembering the
4645 * most recent foreign dirtying events and initiating remote flushes on
4646 * them when local writeback isn't enough to keep the memory clean enough.
4648 * The following two functions implement such mechanism. When a foreign
4649 * page - a page whose memcg and writeback ownerships don't match - is
4650 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4651 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4652 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4653 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4654 * foreign bdi_writebacks which haven't expired. Both the numbers of
4655 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4656 * limited to MEMCG_CGWB_FRN_CNT.
4658 * The mechanism only remembers IDs and doesn't hold any object references.
4659 * As being wrong occasionally doesn't matter, updates and accesses to the
4660 * records are lockless and racy.
4662 void mem_cgroup_track_foreign_dirty_slowpath(struct page
*page
,
4663 struct bdi_writeback
*wb
)
4665 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
4666 struct memcg_cgwb_frn
*frn
;
4667 u64 now
= get_jiffies_64();
4668 u64 oldest_at
= now
;
4672 trace_track_foreign_dirty(page
, wb
);
4675 * Pick the slot to use. If there is already a slot for @wb, keep
4676 * using it. If not replace the oldest one which isn't being
4679 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4680 frn
= &memcg
->cgwb_frn
[i
];
4681 if (frn
->bdi_id
== wb
->bdi
->id
&&
4682 frn
->memcg_id
== wb
->memcg_css
->id
)
4684 if (time_before64(frn
->at
, oldest_at
) &&
4685 atomic_read(&frn
->done
.cnt
) == 1) {
4687 oldest_at
= frn
->at
;
4691 if (i
< MEMCG_CGWB_FRN_CNT
) {
4693 * Re-using an existing one. Update timestamp lazily to
4694 * avoid making the cacheline hot. We want them to be
4695 * reasonably up-to-date and significantly shorter than
4696 * dirty_expire_interval as that's what expires the record.
4697 * Use the shorter of 1s and dirty_expire_interval / 8.
4699 unsigned long update_intv
=
4700 min_t(unsigned long, HZ
,
4701 msecs_to_jiffies(dirty_expire_interval
* 10) / 8);
4703 if (time_before64(frn
->at
, now
- update_intv
))
4705 } else if (oldest
>= 0) {
4706 /* replace the oldest free one */
4707 frn
= &memcg
->cgwb_frn
[oldest
];
4708 frn
->bdi_id
= wb
->bdi
->id
;
4709 frn
->memcg_id
= wb
->memcg_css
->id
;
4714 /* issue foreign writeback flushes for recorded foreign dirtying events */
4715 void mem_cgroup_flush_foreign(struct bdi_writeback
*wb
)
4717 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4718 unsigned long intv
= msecs_to_jiffies(dirty_expire_interval
* 10);
4719 u64 now
= jiffies_64
;
4722 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4723 struct memcg_cgwb_frn
*frn
= &memcg
->cgwb_frn
[i
];
4726 * If the record is older than dirty_expire_interval,
4727 * writeback on it has already started. No need to kick it
4728 * off again. Also, don't start a new one if there's
4729 * already one in flight.
4731 if (time_after64(frn
->at
, now
- intv
) &&
4732 atomic_read(&frn
->done
.cnt
) == 1) {
4734 trace_flush_foreign(wb
, frn
->bdi_id
, frn
->memcg_id
);
4735 cgroup_writeback_by_id(frn
->bdi_id
, frn
->memcg_id
, 0,
4736 WB_REASON_FOREIGN_FLUSH
,
4742 #else /* CONFIG_CGROUP_WRITEBACK */
4744 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4749 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4753 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4757 #endif /* CONFIG_CGROUP_WRITEBACK */
4760 * DO NOT USE IN NEW FILES.
4762 * "cgroup.event_control" implementation.
4764 * This is way over-engineered. It tries to support fully configurable
4765 * events for each user. Such level of flexibility is completely
4766 * unnecessary especially in the light of the planned unified hierarchy.
4768 * Please deprecate this and replace with something simpler if at all
4773 * Unregister event and free resources.
4775 * Gets called from workqueue.
4777 static void memcg_event_remove(struct work_struct
*work
)
4779 struct mem_cgroup_event
*event
=
4780 container_of(work
, struct mem_cgroup_event
, remove
);
4781 struct mem_cgroup
*memcg
= event
->memcg
;
4783 remove_wait_queue(event
->wqh
, &event
->wait
);
4785 event
->unregister_event(memcg
, event
->eventfd
);
4787 /* Notify userspace the event is going away. */
4788 eventfd_signal(event
->eventfd
, 1);
4790 eventfd_ctx_put(event
->eventfd
);
4792 css_put(&memcg
->css
);
4796 * Gets called on EPOLLHUP on eventfd when user closes it.
4798 * Called with wqh->lock held and interrupts disabled.
4800 static int memcg_event_wake(wait_queue_entry_t
*wait
, unsigned mode
,
4801 int sync
, void *key
)
4803 struct mem_cgroup_event
*event
=
4804 container_of(wait
, struct mem_cgroup_event
, wait
);
4805 struct mem_cgroup
*memcg
= event
->memcg
;
4806 __poll_t flags
= key_to_poll(key
);
4808 if (flags
& EPOLLHUP
) {
4810 * If the event has been detached at cgroup removal, we
4811 * can simply return knowing the other side will cleanup
4814 * We can't race against event freeing since the other
4815 * side will require wqh->lock via remove_wait_queue(),
4818 spin_lock(&memcg
->event_list_lock
);
4819 if (!list_empty(&event
->list
)) {
4820 list_del_init(&event
->list
);
4822 * We are in atomic context, but cgroup_event_remove()
4823 * may sleep, so we have to call it in workqueue.
4825 schedule_work(&event
->remove
);
4827 spin_unlock(&memcg
->event_list_lock
);
4833 static void memcg_event_ptable_queue_proc(struct file
*file
,
4834 wait_queue_head_t
*wqh
, poll_table
*pt
)
4836 struct mem_cgroup_event
*event
=
4837 container_of(pt
, struct mem_cgroup_event
, pt
);
4840 add_wait_queue(wqh
, &event
->wait
);
4844 * DO NOT USE IN NEW FILES.
4846 * Parse input and register new cgroup event handler.
4848 * Input must be in format '<event_fd> <control_fd> <args>'.
4849 * Interpretation of args is defined by control file implementation.
4851 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4852 char *buf
, size_t nbytes
, loff_t off
)
4854 struct cgroup_subsys_state
*css
= of_css(of
);
4855 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4856 struct mem_cgroup_event
*event
;
4857 struct cgroup_subsys_state
*cfile_css
;
4858 unsigned int efd
, cfd
;
4865 buf
= strstrip(buf
);
4867 efd
= simple_strtoul(buf
, &endp
, 10);
4872 cfd
= simple_strtoul(buf
, &endp
, 10);
4873 if ((*endp
!= ' ') && (*endp
!= '\0'))
4877 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
4881 event
->memcg
= memcg
;
4882 INIT_LIST_HEAD(&event
->list
);
4883 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
4884 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
4885 INIT_WORK(&event
->remove
, memcg_event_remove
);
4893 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
4894 if (IS_ERR(event
->eventfd
)) {
4895 ret
= PTR_ERR(event
->eventfd
);
4902 goto out_put_eventfd
;
4905 /* the process need read permission on control file */
4906 /* AV: shouldn't we check that it's been opened for read instead? */
4907 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
4912 * Determine the event callbacks and set them in @event. This used
4913 * to be done via struct cftype but cgroup core no longer knows
4914 * about these events. The following is crude but the whole thing
4915 * is for compatibility anyway.
4917 * DO NOT ADD NEW FILES.
4919 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
4921 if (!strcmp(name
, "memory.usage_in_bytes")) {
4922 event
->register_event
= mem_cgroup_usage_register_event
;
4923 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
4924 } else if (!strcmp(name
, "memory.oom_control")) {
4925 event
->register_event
= mem_cgroup_oom_register_event
;
4926 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
4927 } else if (!strcmp(name
, "memory.pressure_level")) {
4928 event
->register_event
= vmpressure_register_event
;
4929 event
->unregister_event
= vmpressure_unregister_event
;
4930 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
4931 event
->register_event
= memsw_cgroup_usage_register_event
;
4932 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
4939 * Verify @cfile should belong to @css. Also, remaining events are
4940 * automatically removed on cgroup destruction but the removal is
4941 * asynchronous, so take an extra ref on @css.
4943 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
4944 &memory_cgrp_subsys
);
4946 if (IS_ERR(cfile_css
))
4948 if (cfile_css
!= css
) {
4953 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
4957 vfs_poll(efile
.file
, &event
->pt
);
4959 spin_lock(&memcg
->event_list_lock
);
4960 list_add(&event
->list
, &memcg
->event_list
);
4961 spin_unlock(&memcg
->event_list_lock
);
4973 eventfd_ctx_put(event
->eventfd
);
4982 static struct cftype mem_cgroup_legacy_files
[] = {
4984 .name
= "usage_in_bytes",
4985 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4986 .read_u64
= mem_cgroup_read_u64
,
4989 .name
= "max_usage_in_bytes",
4990 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4991 .write
= mem_cgroup_reset
,
4992 .read_u64
= mem_cgroup_read_u64
,
4995 .name
= "limit_in_bytes",
4996 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4997 .write
= mem_cgroup_write
,
4998 .read_u64
= mem_cgroup_read_u64
,
5001 .name
= "soft_limit_in_bytes",
5002 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5003 .write
= mem_cgroup_write
,
5004 .read_u64
= mem_cgroup_read_u64
,
5008 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5009 .write
= mem_cgroup_reset
,
5010 .read_u64
= mem_cgroup_read_u64
,
5014 .seq_show
= memcg_stat_show
,
5017 .name
= "force_empty",
5018 .write
= mem_cgroup_force_empty_write
,
5021 .name
= "use_hierarchy",
5022 .write_u64
= mem_cgroup_hierarchy_write
,
5023 .read_u64
= mem_cgroup_hierarchy_read
,
5026 .name
= "cgroup.event_control", /* XXX: for compat */
5027 .write
= memcg_write_event_control
,
5028 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
5031 .name
= "swappiness",
5032 .read_u64
= mem_cgroup_swappiness_read
,
5033 .write_u64
= mem_cgroup_swappiness_write
,
5036 .name
= "move_charge_at_immigrate",
5037 .read_u64
= mem_cgroup_move_charge_read
,
5038 .write_u64
= mem_cgroup_move_charge_write
,
5041 .name
= "oom_control",
5042 .seq_show
= mem_cgroup_oom_control_read
,
5043 .write_u64
= mem_cgroup_oom_control_write
,
5044 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
5047 .name
= "pressure_level",
5051 .name
= "numa_stat",
5052 .seq_show
= memcg_numa_stat_show
,
5056 .name
= "kmem.limit_in_bytes",
5057 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
5058 .write
= mem_cgroup_write
,
5059 .read_u64
= mem_cgroup_read_u64
,
5062 .name
= "kmem.usage_in_bytes",
5063 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
5064 .read_u64
= mem_cgroup_read_u64
,
5067 .name
= "kmem.failcnt",
5068 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
5069 .write
= mem_cgroup_reset
,
5070 .read_u64
= mem_cgroup_read_u64
,
5073 .name
= "kmem.max_usage_in_bytes",
5074 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
5075 .write
= mem_cgroup_reset
,
5076 .read_u64
= mem_cgroup_read_u64
,
5078 #if defined(CONFIG_MEMCG_KMEM) && \
5079 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5081 .name
= "kmem.slabinfo",
5082 .seq_show
= memcg_slab_show
,
5086 .name
= "kmem.tcp.limit_in_bytes",
5087 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
5088 .write
= mem_cgroup_write
,
5089 .read_u64
= mem_cgroup_read_u64
,
5092 .name
= "kmem.tcp.usage_in_bytes",
5093 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
5094 .read_u64
= mem_cgroup_read_u64
,
5097 .name
= "kmem.tcp.failcnt",
5098 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
5099 .write
= mem_cgroup_reset
,
5100 .read_u64
= mem_cgroup_read_u64
,
5103 .name
= "kmem.tcp.max_usage_in_bytes",
5104 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
5105 .write
= mem_cgroup_reset
,
5106 .read_u64
= mem_cgroup_read_u64
,
5108 { }, /* terminate */
5112 * Private memory cgroup IDR
5114 * Swap-out records and page cache shadow entries need to store memcg
5115 * references in constrained space, so we maintain an ID space that is
5116 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5117 * memory-controlled cgroups to 64k.
5119 * However, there usually are many references to the offline CSS after
5120 * the cgroup has been destroyed, such as page cache or reclaimable
5121 * slab objects, that don't need to hang on to the ID. We want to keep
5122 * those dead CSS from occupying IDs, or we might quickly exhaust the
5123 * relatively small ID space and prevent the creation of new cgroups
5124 * even when there are much fewer than 64k cgroups - possibly none.
5126 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5127 * be freed and recycled when it's no longer needed, which is usually
5128 * when the CSS is offlined.
5130 * The only exception to that are records of swapped out tmpfs/shmem
5131 * pages that need to be attributed to live ancestors on swapin. But
5132 * those references are manageable from userspace.
5135 static DEFINE_IDR(mem_cgroup_idr
);
5137 static void mem_cgroup_id_remove(struct mem_cgroup
*memcg
)
5139 if (memcg
->id
.id
> 0) {
5140 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
5145 static void __maybe_unused
mem_cgroup_id_get_many(struct mem_cgroup
*memcg
,
5148 refcount_add(n
, &memcg
->id
.ref
);
5151 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
5153 if (refcount_sub_and_test(n
, &memcg
->id
.ref
)) {
5154 mem_cgroup_id_remove(memcg
);
5156 /* Memcg ID pins CSS */
5157 css_put(&memcg
->css
);
5161 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
5163 mem_cgroup_id_put_many(memcg
, 1);
5167 * mem_cgroup_from_id - look up a memcg from a memcg id
5168 * @id: the memcg id to look up
5170 * Caller must hold rcu_read_lock().
5172 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
5174 WARN_ON_ONCE(!rcu_read_lock_held());
5175 return idr_find(&mem_cgroup_idr
, id
);
5178 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
5180 struct mem_cgroup_per_node
*pn
;
5183 * This routine is called against possible nodes.
5184 * But it's BUG to call kmalloc() against offline node.
5186 * TODO: this routine can waste much memory for nodes which will
5187 * never be onlined. It's better to use memory hotplug callback
5190 if (!node_state(node
, N_NORMAL_MEMORY
))
5192 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
5196 pn
->lruvec_stat_local
= alloc_percpu_gfp(struct lruvec_stat
,
5197 GFP_KERNEL_ACCOUNT
);
5198 if (!pn
->lruvec_stat_local
) {
5203 pn
->lruvec_stat_cpu
= alloc_percpu_gfp(struct lruvec_stat
,
5204 GFP_KERNEL_ACCOUNT
);
5205 if (!pn
->lruvec_stat_cpu
) {
5206 free_percpu(pn
->lruvec_stat_local
);
5211 lruvec_init(&pn
->lruvec
);
5212 pn
->usage_in_excess
= 0;
5213 pn
->on_tree
= false;
5216 memcg
->nodeinfo
[node
] = pn
;
5220 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
5222 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
5227 free_percpu(pn
->lruvec_stat_cpu
);
5228 free_percpu(pn
->lruvec_stat_local
);
5232 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5237 free_mem_cgroup_per_node_info(memcg
, node
);
5238 free_percpu(memcg
->vmstats_percpu
);
5239 free_percpu(memcg
->vmstats_local
);
5243 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
5245 memcg_wb_domain_exit(memcg
);
5247 * Flush percpu vmstats and vmevents to guarantee the value correctness
5248 * on parent's and all ancestor levels.
5250 memcg_flush_percpu_vmstats(memcg
);
5251 memcg_flush_percpu_vmevents(memcg
);
5252 __mem_cgroup_free(memcg
);
5255 static struct mem_cgroup
*mem_cgroup_alloc(void)
5257 struct mem_cgroup
*memcg
;
5260 int __maybe_unused i
;
5261 long error
= -ENOMEM
;
5263 size
= sizeof(struct mem_cgroup
);
5264 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
5266 memcg
= kzalloc(size
, GFP_KERNEL
);
5268 return ERR_PTR(error
);
5270 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
5271 1, MEM_CGROUP_ID_MAX
,
5273 if (memcg
->id
.id
< 0) {
5274 error
= memcg
->id
.id
;
5278 memcg
->vmstats_local
= alloc_percpu_gfp(struct memcg_vmstats_percpu
,
5279 GFP_KERNEL_ACCOUNT
);
5280 if (!memcg
->vmstats_local
)
5283 memcg
->vmstats_percpu
= alloc_percpu_gfp(struct memcg_vmstats_percpu
,
5284 GFP_KERNEL_ACCOUNT
);
5285 if (!memcg
->vmstats_percpu
)
5289 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
5292 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
5295 INIT_WORK(&memcg
->high_work
, high_work_func
);
5296 INIT_LIST_HEAD(&memcg
->oom_notify
);
5297 mutex_init(&memcg
->thresholds_lock
);
5298 spin_lock_init(&memcg
->move_lock
);
5299 vmpressure_init(&memcg
->vmpressure
);
5300 INIT_LIST_HEAD(&memcg
->event_list
);
5301 spin_lock_init(&memcg
->event_list_lock
);
5302 memcg
->socket_pressure
= jiffies
;
5303 #ifdef CONFIG_MEMCG_KMEM
5304 memcg
->kmemcg_id
= -1;
5305 INIT_LIST_HEAD(&memcg
->objcg_list
);
5307 #ifdef CONFIG_CGROUP_WRITEBACK
5308 INIT_LIST_HEAD(&memcg
->cgwb_list
);
5309 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5310 memcg
->cgwb_frn
[i
].done
=
5311 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq
);
5313 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5314 spin_lock_init(&memcg
->deferred_split_queue
.split_queue_lock
);
5315 INIT_LIST_HEAD(&memcg
->deferred_split_queue
.split_queue
);
5316 memcg
->deferred_split_queue
.split_queue_len
= 0;
5318 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
5321 mem_cgroup_id_remove(memcg
);
5322 __mem_cgroup_free(memcg
);
5323 return ERR_PTR(error
);
5326 static struct cgroup_subsys_state
* __ref
5327 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
5329 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
5330 struct mem_cgroup
*memcg
, *old_memcg
;
5331 long error
= -ENOMEM
;
5333 old_memcg
= set_active_memcg(parent
);
5334 memcg
= mem_cgroup_alloc();
5335 set_active_memcg(old_memcg
);
5337 return ERR_CAST(memcg
);
5339 page_counter_set_high(&memcg
->memory
, PAGE_COUNTER_MAX
);
5340 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5341 page_counter_set_high(&memcg
->swap
, PAGE_COUNTER_MAX
);
5343 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
5344 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
5347 page_counter_init(&memcg
->memory
, NULL
);
5348 page_counter_init(&memcg
->swap
, NULL
);
5349 page_counter_init(&memcg
->kmem
, NULL
);
5350 page_counter_init(&memcg
->tcpmem
, NULL
);
5351 } else if (parent
->use_hierarchy
) {
5352 memcg
->use_hierarchy
= true;
5353 page_counter_init(&memcg
->memory
, &parent
->memory
);
5354 page_counter_init(&memcg
->swap
, &parent
->swap
);
5355 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
5356 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
5358 page_counter_init(&memcg
->memory
, &root_mem_cgroup
->memory
);
5359 page_counter_init(&memcg
->swap
, &root_mem_cgroup
->swap
);
5360 page_counter_init(&memcg
->kmem
, &root_mem_cgroup
->kmem
);
5361 page_counter_init(&memcg
->tcpmem
, &root_mem_cgroup
->tcpmem
);
5363 * Deeper hierachy with use_hierarchy == false doesn't make
5364 * much sense so let cgroup subsystem know about this
5365 * unfortunate state in our controller.
5367 if (parent
!= root_mem_cgroup
)
5368 memory_cgrp_subsys
.broken_hierarchy
= true;
5371 /* The following stuff does not apply to the root */
5373 root_mem_cgroup
= memcg
;
5377 error
= memcg_online_kmem(memcg
);
5381 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5382 static_branch_inc(&memcg_sockets_enabled_key
);
5386 mem_cgroup_id_remove(memcg
);
5387 mem_cgroup_free(memcg
);
5388 return ERR_PTR(error
);
5391 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
5393 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5396 * A memcg must be visible for memcg_expand_shrinker_maps()
5397 * by the time the maps are allocated. So, we allocate maps
5398 * here, when for_each_mem_cgroup() can't skip it.
5400 if (memcg_alloc_shrinker_maps(memcg
)) {
5401 mem_cgroup_id_remove(memcg
);
5405 /* Online state pins memcg ID, memcg ID pins CSS */
5406 refcount_set(&memcg
->id
.ref
, 1);
5411 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
5413 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5414 struct mem_cgroup_event
*event
, *tmp
;
5417 * Unregister events and notify userspace.
5418 * Notify userspace about cgroup removing only after rmdir of cgroup
5419 * directory to avoid race between userspace and kernelspace.
5421 spin_lock(&memcg
->event_list_lock
);
5422 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
5423 list_del_init(&event
->list
);
5424 schedule_work(&event
->remove
);
5426 spin_unlock(&memcg
->event_list_lock
);
5428 page_counter_set_min(&memcg
->memory
, 0);
5429 page_counter_set_low(&memcg
->memory
, 0);
5431 memcg_offline_kmem(memcg
);
5432 wb_memcg_offline(memcg
);
5434 drain_all_stock(memcg
);
5436 mem_cgroup_id_put(memcg
);
5439 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
5441 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5443 invalidate_reclaim_iterators(memcg
);
5446 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
5448 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5449 int __maybe_unused i
;
5451 #ifdef CONFIG_CGROUP_WRITEBACK
5452 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5453 wb_wait_for_completion(&memcg
->cgwb_frn
[i
].done
);
5455 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5456 static_branch_dec(&memcg_sockets_enabled_key
);
5458 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
5459 static_branch_dec(&memcg_sockets_enabled_key
);
5461 vmpressure_cleanup(&memcg
->vmpressure
);
5462 cancel_work_sync(&memcg
->high_work
);
5463 mem_cgroup_remove_from_trees(memcg
);
5464 memcg_free_shrinker_maps(memcg
);
5465 memcg_free_kmem(memcg
);
5466 mem_cgroup_free(memcg
);
5470 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5471 * @css: the target css
5473 * Reset the states of the mem_cgroup associated with @css. This is
5474 * invoked when the userland requests disabling on the default hierarchy
5475 * but the memcg is pinned through dependency. The memcg should stop
5476 * applying policies and should revert to the vanilla state as it may be
5477 * made visible again.
5479 * The current implementation only resets the essential configurations.
5480 * This needs to be expanded to cover all the visible parts.
5482 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
5484 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5486 page_counter_set_max(&memcg
->memory
, PAGE_COUNTER_MAX
);
5487 page_counter_set_max(&memcg
->swap
, PAGE_COUNTER_MAX
);
5488 page_counter_set_max(&memcg
->kmem
, PAGE_COUNTER_MAX
);
5489 page_counter_set_max(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
5490 page_counter_set_min(&memcg
->memory
, 0);
5491 page_counter_set_low(&memcg
->memory
, 0);
5492 page_counter_set_high(&memcg
->memory
, PAGE_COUNTER_MAX
);
5493 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5494 page_counter_set_high(&memcg
->swap
, PAGE_COUNTER_MAX
);
5495 memcg_wb_domain_size_changed(memcg
);
5499 /* Handlers for move charge at task migration. */
5500 static int mem_cgroup_do_precharge(unsigned long count
)
5504 /* Try a single bulk charge without reclaim first, kswapd may wake */
5505 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
5507 mc
.precharge
+= count
;
5511 /* Try charges one by one with reclaim, but do not retry */
5513 ret
= try_charge(mc
.to
, GFP_KERNEL
| __GFP_NORETRY
, 1);
5527 enum mc_target_type
{
5534 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
5535 unsigned long addr
, pte_t ptent
)
5537 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
5539 if (!page
|| !page_mapped(page
))
5541 if (PageAnon(page
)) {
5542 if (!(mc
.flags
& MOVE_ANON
))
5545 if (!(mc
.flags
& MOVE_FILE
))
5548 if (!get_page_unless_zero(page
))
5554 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5555 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5556 pte_t ptent
, swp_entry_t
*entry
)
5558 struct page
*page
= NULL
;
5559 swp_entry_t ent
= pte_to_swp_entry(ptent
);
5561 if (!(mc
.flags
& MOVE_ANON
))
5565 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5566 * a device and because they are not accessible by CPU they are store
5567 * as special swap entry in the CPU page table.
5569 if (is_device_private_entry(ent
)) {
5570 page
= device_private_entry_to_page(ent
);
5572 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5573 * a refcount of 1 when free (unlike normal page)
5575 if (!page_ref_add_unless(page
, 1, 1))
5580 if (non_swap_entry(ent
))
5584 * Because lookup_swap_cache() updates some statistics counter,
5585 * we call find_get_page() with swapper_space directly.
5587 page
= find_get_page(swap_address_space(ent
), swp_offset(ent
));
5588 entry
->val
= ent
.val
;
5593 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5594 pte_t ptent
, swp_entry_t
*entry
)
5600 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
5601 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5603 if (!vma
->vm_file
) /* anonymous vma */
5605 if (!(mc
.flags
& MOVE_FILE
))
5608 /* page is moved even if it's not RSS of this task(page-faulted). */
5609 /* shmem/tmpfs may report page out on swap: account for that too. */
5610 return find_get_incore_page(vma
->vm_file
->f_mapping
,
5611 linear_page_index(vma
, addr
));
5615 * mem_cgroup_move_account - move account of the page
5617 * @compound: charge the page as compound or small page
5618 * @from: mem_cgroup which the page is moved from.
5619 * @to: mem_cgroup which the page is moved to. @from != @to.
5621 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5623 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5626 static int mem_cgroup_move_account(struct page
*page
,
5628 struct mem_cgroup
*from
,
5629 struct mem_cgroup
*to
)
5631 struct lruvec
*from_vec
, *to_vec
;
5632 struct pglist_data
*pgdat
;
5633 unsigned int nr_pages
= compound
? thp_nr_pages(page
) : 1;
5636 VM_BUG_ON(from
== to
);
5637 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5638 VM_BUG_ON(compound
&& !PageTransHuge(page
));
5641 * Prevent mem_cgroup_migrate() from looking at
5642 * page->mem_cgroup of its source page while we change it.
5645 if (!trylock_page(page
))
5649 if (page
->mem_cgroup
!= from
)
5652 pgdat
= page_pgdat(page
);
5653 from_vec
= mem_cgroup_lruvec(from
, pgdat
);
5654 to_vec
= mem_cgroup_lruvec(to
, pgdat
);
5656 lock_page_memcg(page
);
5658 if (PageAnon(page
)) {
5659 if (page_mapped(page
)) {
5660 __mod_lruvec_state(from_vec
, NR_ANON_MAPPED
, -nr_pages
);
5661 __mod_lruvec_state(to_vec
, NR_ANON_MAPPED
, nr_pages
);
5662 if (PageTransHuge(page
)) {
5663 __mod_lruvec_state(from_vec
, NR_ANON_THPS
,
5665 __mod_lruvec_state(to_vec
, NR_ANON_THPS
,
5671 __mod_lruvec_state(from_vec
, NR_FILE_PAGES
, -nr_pages
);
5672 __mod_lruvec_state(to_vec
, NR_FILE_PAGES
, nr_pages
);
5674 if (PageSwapBacked(page
)) {
5675 __mod_lruvec_state(from_vec
, NR_SHMEM
, -nr_pages
);
5676 __mod_lruvec_state(to_vec
, NR_SHMEM
, nr_pages
);
5679 if (page_mapped(page
)) {
5680 __mod_lruvec_state(from_vec
, NR_FILE_MAPPED
, -nr_pages
);
5681 __mod_lruvec_state(to_vec
, NR_FILE_MAPPED
, nr_pages
);
5684 if (PageDirty(page
)) {
5685 struct address_space
*mapping
= page_mapping(page
);
5687 if (mapping_can_writeback(mapping
)) {
5688 __mod_lruvec_state(from_vec
, NR_FILE_DIRTY
,
5690 __mod_lruvec_state(to_vec
, NR_FILE_DIRTY
,
5696 if (PageWriteback(page
)) {
5697 __mod_lruvec_state(from_vec
, NR_WRITEBACK
, -nr_pages
);
5698 __mod_lruvec_state(to_vec
, NR_WRITEBACK
, nr_pages
);
5702 * All state has been migrated, let's switch to the new memcg.
5704 * It is safe to change page->mem_cgroup here because the page
5705 * is referenced, charged, isolated, and locked: we can't race
5706 * with (un)charging, migration, LRU putback, or anything else
5707 * that would rely on a stable page->mem_cgroup.
5709 * Note that lock_page_memcg is a memcg lock, not a page lock,
5710 * to save space. As soon as we switch page->mem_cgroup to a
5711 * new memcg that isn't locked, the above state can change
5712 * concurrently again. Make sure we're truly done with it.
5717 css_put(&from
->css
);
5719 page
->mem_cgroup
= to
;
5721 __unlock_page_memcg(from
);
5725 local_irq_disable();
5726 mem_cgroup_charge_statistics(to
, page
, nr_pages
);
5727 memcg_check_events(to
, page
);
5728 mem_cgroup_charge_statistics(from
, page
, -nr_pages
);
5729 memcg_check_events(from
, page
);
5738 * get_mctgt_type - get target type of moving charge
5739 * @vma: the vma the pte to be checked belongs
5740 * @addr: the address corresponding to the pte to be checked
5741 * @ptent: the pte to be checked
5742 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5745 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5746 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5747 * move charge. if @target is not NULL, the page is stored in target->page
5748 * with extra refcnt got(Callers should handle it).
5749 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5750 * target for charge migration. if @target is not NULL, the entry is stored
5752 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5753 * (so ZONE_DEVICE page and thus not on the lru).
5754 * For now we such page is charge like a regular page would be as for all
5755 * intent and purposes it is just special memory taking the place of a
5758 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5760 * Called with pte lock held.
5763 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
5764 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
5766 struct page
*page
= NULL
;
5767 enum mc_target_type ret
= MC_TARGET_NONE
;
5768 swp_entry_t ent
= { .val
= 0 };
5770 if (pte_present(ptent
))
5771 page
= mc_handle_present_pte(vma
, addr
, ptent
);
5772 else if (is_swap_pte(ptent
))
5773 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
5774 else if (pte_none(ptent
))
5775 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
5777 if (!page
&& !ent
.val
)
5781 * Do only loose check w/o serialization.
5782 * mem_cgroup_move_account() checks the page is valid or
5783 * not under LRU exclusion.
5785 if (page
->mem_cgroup
== mc
.from
) {
5786 ret
= MC_TARGET_PAGE
;
5787 if (is_device_private_page(page
))
5788 ret
= MC_TARGET_DEVICE
;
5790 target
->page
= page
;
5792 if (!ret
|| !target
)
5796 * There is a swap entry and a page doesn't exist or isn't charged.
5797 * But we cannot move a tail-page in a THP.
5799 if (ent
.val
&& !ret
&& (!page
|| !PageTransCompound(page
)) &&
5800 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
5801 ret
= MC_TARGET_SWAP
;
5808 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5810 * We don't consider PMD mapped swapping or file mapped pages because THP does
5811 * not support them for now.
5812 * Caller should make sure that pmd_trans_huge(pmd) is true.
5814 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5815 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5817 struct page
*page
= NULL
;
5818 enum mc_target_type ret
= MC_TARGET_NONE
;
5820 if (unlikely(is_swap_pmd(pmd
))) {
5821 VM_BUG_ON(thp_migration_supported() &&
5822 !is_pmd_migration_entry(pmd
));
5825 page
= pmd_page(pmd
);
5826 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
5827 if (!(mc
.flags
& MOVE_ANON
))
5829 if (page
->mem_cgroup
== mc
.from
) {
5830 ret
= MC_TARGET_PAGE
;
5833 target
->page
= page
;
5839 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5840 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5842 return MC_TARGET_NONE
;
5846 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5847 unsigned long addr
, unsigned long end
,
5848 struct mm_walk
*walk
)
5850 struct vm_area_struct
*vma
= walk
->vma
;
5854 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5857 * Note their can not be MC_TARGET_DEVICE for now as we do not
5858 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5859 * this might change.
5861 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5862 mc
.precharge
+= HPAGE_PMD_NR
;
5867 if (pmd_trans_unstable(pmd
))
5869 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5870 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5871 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5872 mc
.precharge
++; /* increment precharge temporarily */
5873 pte_unmap_unlock(pte
- 1, ptl
);
5879 static const struct mm_walk_ops precharge_walk_ops
= {
5880 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
5883 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5885 unsigned long precharge
;
5888 walk_page_range(mm
, 0, mm
->highest_vm_end
, &precharge_walk_ops
, NULL
);
5889 mmap_read_unlock(mm
);
5891 precharge
= mc
.precharge
;
5897 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
5899 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
5901 VM_BUG_ON(mc
.moving_task
);
5902 mc
.moving_task
= current
;
5903 return mem_cgroup_do_precharge(precharge
);
5906 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5907 static void __mem_cgroup_clear_mc(void)
5909 struct mem_cgroup
*from
= mc
.from
;
5910 struct mem_cgroup
*to
= mc
.to
;
5912 /* we must uncharge all the leftover precharges from mc.to */
5914 cancel_charge(mc
.to
, mc
.precharge
);
5918 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5919 * we must uncharge here.
5921 if (mc
.moved_charge
) {
5922 cancel_charge(mc
.from
, mc
.moved_charge
);
5923 mc
.moved_charge
= 0;
5925 /* we must fixup refcnts and charges */
5926 if (mc
.moved_swap
) {
5927 /* uncharge swap account from the old cgroup */
5928 if (!mem_cgroup_is_root(mc
.from
))
5929 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
5931 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
5934 * we charged both to->memory and to->memsw, so we
5935 * should uncharge to->memory.
5937 if (!mem_cgroup_is_root(mc
.to
))
5938 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
5942 memcg_oom_recover(from
);
5943 memcg_oom_recover(to
);
5944 wake_up_all(&mc
.waitq
);
5947 static void mem_cgroup_clear_mc(void)
5949 struct mm_struct
*mm
= mc
.mm
;
5952 * we must clear moving_task before waking up waiters at the end of
5955 mc
.moving_task
= NULL
;
5956 __mem_cgroup_clear_mc();
5957 spin_lock(&mc
.lock
);
5961 spin_unlock(&mc
.lock
);
5966 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5968 struct cgroup_subsys_state
*css
;
5969 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
5970 struct mem_cgroup
*from
;
5971 struct task_struct
*leader
, *p
;
5972 struct mm_struct
*mm
;
5973 unsigned long move_flags
;
5976 /* charge immigration isn't supported on the default hierarchy */
5977 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5981 * Multi-process migrations only happen on the default hierarchy
5982 * where charge immigration is not used. Perform charge
5983 * immigration if @tset contains a leader and whine if there are
5987 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
5990 memcg
= mem_cgroup_from_css(css
);
5996 * We are now commited to this value whatever it is. Changes in this
5997 * tunable will only affect upcoming migrations, not the current one.
5998 * So we need to save it, and keep it going.
6000 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
6004 from
= mem_cgroup_from_task(p
);
6006 VM_BUG_ON(from
== memcg
);
6008 mm
= get_task_mm(p
);
6011 /* We move charges only when we move a owner of the mm */
6012 if (mm
->owner
== p
) {
6015 VM_BUG_ON(mc
.precharge
);
6016 VM_BUG_ON(mc
.moved_charge
);
6017 VM_BUG_ON(mc
.moved_swap
);
6019 spin_lock(&mc
.lock
);
6023 mc
.flags
= move_flags
;
6024 spin_unlock(&mc
.lock
);
6025 /* We set mc.moving_task later */
6027 ret
= mem_cgroup_precharge_mc(mm
);
6029 mem_cgroup_clear_mc();
6036 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
6039 mem_cgroup_clear_mc();
6042 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6043 unsigned long addr
, unsigned long end
,
6044 struct mm_walk
*walk
)
6047 struct vm_area_struct
*vma
= walk
->vma
;
6050 enum mc_target_type target_type
;
6051 union mc_target target
;
6054 ptl
= pmd_trans_huge_lock(pmd
, vma
);
6056 if (mc
.precharge
< HPAGE_PMD_NR
) {
6060 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6061 if (target_type
== MC_TARGET_PAGE
) {
6063 if (!isolate_lru_page(page
)) {
6064 if (!mem_cgroup_move_account(page
, true,
6066 mc
.precharge
-= HPAGE_PMD_NR
;
6067 mc
.moved_charge
+= HPAGE_PMD_NR
;
6069 putback_lru_page(page
);
6072 } else if (target_type
== MC_TARGET_DEVICE
) {
6074 if (!mem_cgroup_move_account(page
, true,
6076 mc
.precharge
-= HPAGE_PMD_NR
;
6077 mc
.moved_charge
+= HPAGE_PMD_NR
;
6085 if (pmd_trans_unstable(pmd
))
6088 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6089 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6090 pte_t ptent
= *(pte
++);
6091 bool device
= false;
6097 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6098 case MC_TARGET_DEVICE
:
6101 case MC_TARGET_PAGE
:
6104 * We can have a part of the split pmd here. Moving it
6105 * can be done but it would be too convoluted so simply
6106 * ignore such a partial THP and keep it in original
6107 * memcg. There should be somebody mapping the head.
6109 if (PageTransCompound(page
))
6111 if (!device
&& isolate_lru_page(page
))
6113 if (!mem_cgroup_move_account(page
, false,
6116 /* we uncharge from mc.from later. */
6120 putback_lru_page(page
);
6121 put
: /* get_mctgt_type() gets the page */
6124 case MC_TARGET_SWAP
:
6126 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6128 mem_cgroup_id_get_many(mc
.to
, 1);
6129 /* we fixup other refcnts and charges later. */
6137 pte_unmap_unlock(pte
- 1, ptl
);
6142 * We have consumed all precharges we got in can_attach().
6143 * We try charge one by one, but don't do any additional
6144 * charges to mc.to if we have failed in charge once in attach()
6147 ret
= mem_cgroup_do_precharge(1);
6155 static const struct mm_walk_ops charge_walk_ops
= {
6156 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6159 static void mem_cgroup_move_charge(void)
6161 lru_add_drain_all();
6163 * Signal lock_page_memcg() to take the memcg's move_lock
6164 * while we're moving its pages to another memcg. Then wait
6165 * for already started RCU-only updates to finish.
6167 atomic_inc(&mc
.from
->moving_account
);
6170 if (unlikely(!mmap_read_trylock(mc
.mm
))) {
6172 * Someone who are holding the mmap_lock might be waiting in
6173 * waitq. So we cancel all extra charges, wake up all waiters,
6174 * and retry. Because we cancel precharges, we might not be able
6175 * to move enough charges, but moving charge is a best-effort
6176 * feature anyway, so it wouldn't be a big problem.
6178 __mem_cgroup_clear_mc();
6183 * When we have consumed all precharges and failed in doing
6184 * additional charge, the page walk just aborts.
6186 walk_page_range(mc
.mm
, 0, mc
.mm
->highest_vm_end
, &charge_walk_ops
,
6189 mmap_read_unlock(mc
.mm
);
6190 atomic_dec(&mc
.from
->moving_account
);
6193 static void mem_cgroup_move_task(void)
6196 mem_cgroup_move_charge();
6197 mem_cgroup_clear_mc();
6200 #else /* !CONFIG_MMU */
6201 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
6205 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
6208 static void mem_cgroup_move_task(void)
6214 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6215 * to verify whether we're attached to the default hierarchy on each mount
6218 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
6221 * use_hierarchy is forced on the default hierarchy. cgroup core
6222 * guarantees that @root doesn't have any children, so turning it
6223 * on for the root memcg is enough.
6225 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
6226 root_mem_cgroup
->use_hierarchy
= true;
6228 root_mem_cgroup
->use_hierarchy
= false;
6231 static int seq_puts_memcg_tunable(struct seq_file
*m
, unsigned long value
)
6233 if (value
== PAGE_COUNTER_MAX
)
6234 seq_puts(m
, "max\n");
6236 seq_printf(m
, "%llu\n", (u64
)value
* PAGE_SIZE
);
6241 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
6244 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6246 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
6249 static int memory_min_show(struct seq_file
*m
, void *v
)
6251 return seq_puts_memcg_tunable(m
,
6252 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.min
));
6255 static ssize_t
memory_min_write(struct kernfs_open_file
*of
,
6256 char *buf
, size_t nbytes
, loff_t off
)
6258 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6262 buf
= strstrip(buf
);
6263 err
= page_counter_memparse(buf
, "max", &min
);
6267 page_counter_set_min(&memcg
->memory
, min
);
6272 static int memory_low_show(struct seq_file
*m
, void *v
)
6274 return seq_puts_memcg_tunable(m
,
6275 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.low
));
6278 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
6279 char *buf
, size_t nbytes
, loff_t off
)
6281 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6285 buf
= strstrip(buf
);
6286 err
= page_counter_memparse(buf
, "max", &low
);
6290 page_counter_set_low(&memcg
->memory
, low
);
6295 static int memory_high_show(struct seq_file
*m
, void *v
)
6297 return seq_puts_memcg_tunable(m
,
6298 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.high
));
6301 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
6302 char *buf
, size_t nbytes
, loff_t off
)
6304 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6305 unsigned int nr_retries
= MAX_RECLAIM_RETRIES
;
6306 bool drained
= false;
6310 buf
= strstrip(buf
);
6311 err
= page_counter_memparse(buf
, "max", &high
);
6316 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6317 unsigned long reclaimed
;
6319 if (nr_pages
<= high
)
6322 if (signal_pending(current
))
6326 drain_all_stock(memcg
);
6331 reclaimed
= try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
6334 if (!reclaimed
&& !nr_retries
--)
6338 page_counter_set_high(&memcg
->memory
, high
);
6340 memcg_wb_domain_size_changed(memcg
);
6345 static int memory_max_show(struct seq_file
*m
, void *v
)
6347 return seq_puts_memcg_tunable(m
,
6348 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.max
));
6351 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
6352 char *buf
, size_t nbytes
, loff_t off
)
6354 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6355 unsigned int nr_reclaims
= MAX_RECLAIM_RETRIES
;
6356 bool drained
= false;
6360 buf
= strstrip(buf
);
6361 err
= page_counter_memparse(buf
, "max", &max
);
6365 xchg(&memcg
->memory
.max
, max
);
6368 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6370 if (nr_pages
<= max
)
6373 if (signal_pending(current
))
6377 drain_all_stock(memcg
);
6383 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
6389 memcg_memory_event(memcg
, MEMCG_OOM
);
6390 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
6394 memcg_wb_domain_size_changed(memcg
);
6398 static void __memory_events_show(struct seq_file
*m
, atomic_long_t
*events
)
6400 seq_printf(m
, "low %lu\n", atomic_long_read(&events
[MEMCG_LOW
]));
6401 seq_printf(m
, "high %lu\n", atomic_long_read(&events
[MEMCG_HIGH
]));
6402 seq_printf(m
, "max %lu\n", atomic_long_read(&events
[MEMCG_MAX
]));
6403 seq_printf(m
, "oom %lu\n", atomic_long_read(&events
[MEMCG_OOM
]));
6404 seq_printf(m
, "oom_kill %lu\n",
6405 atomic_long_read(&events
[MEMCG_OOM_KILL
]));
6408 static int memory_events_show(struct seq_file
*m
, void *v
)
6410 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6412 __memory_events_show(m
, memcg
->memory_events
);
6416 static int memory_events_local_show(struct seq_file
*m
, void *v
)
6418 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6420 __memory_events_show(m
, memcg
->memory_events_local
);
6424 static int memory_stat_show(struct seq_file
*m
, void *v
)
6426 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6429 buf
= memory_stat_format(memcg
);
6438 static int memory_numa_stat_show(struct seq_file
*m
, void *v
)
6441 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6443 for (i
= 0; i
< ARRAY_SIZE(memory_stats
); i
++) {
6446 if (memory_stats
[i
].idx
>= NR_VM_NODE_STAT_ITEMS
)
6449 seq_printf(m
, "%s", memory_stats
[i
].name
);
6450 for_each_node_state(nid
, N_MEMORY
) {
6452 struct lruvec
*lruvec
;
6454 lruvec
= mem_cgroup_lruvec(memcg
, NODE_DATA(nid
));
6455 size
= lruvec_page_state(lruvec
, memory_stats
[i
].idx
);
6456 size
*= memory_stats
[i
].ratio
;
6457 seq_printf(m
, " N%d=%llu", nid
, size
);
6466 static int memory_oom_group_show(struct seq_file
*m
, void *v
)
6468 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6470 seq_printf(m
, "%d\n", memcg
->oom_group
);
6475 static ssize_t
memory_oom_group_write(struct kernfs_open_file
*of
,
6476 char *buf
, size_t nbytes
, loff_t off
)
6478 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6481 buf
= strstrip(buf
);
6485 ret
= kstrtoint(buf
, 0, &oom_group
);
6489 if (oom_group
!= 0 && oom_group
!= 1)
6492 memcg
->oom_group
= oom_group
;
6497 static struct cftype memory_files
[] = {
6500 .flags
= CFTYPE_NOT_ON_ROOT
,
6501 .read_u64
= memory_current_read
,
6505 .flags
= CFTYPE_NOT_ON_ROOT
,
6506 .seq_show
= memory_min_show
,
6507 .write
= memory_min_write
,
6511 .flags
= CFTYPE_NOT_ON_ROOT
,
6512 .seq_show
= memory_low_show
,
6513 .write
= memory_low_write
,
6517 .flags
= CFTYPE_NOT_ON_ROOT
,
6518 .seq_show
= memory_high_show
,
6519 .write
= memory_high_write
,
6523 .flags
= CFTYPE_NOT_ON_ROOT
,
6524 .seq_show
= memory_max_show
,
6525 .write
= memory_max_write
,
6529 .flags
= CFTYPE_NOT_ON_ROOT
,
6530 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
6531 .seq_show
= memory_events_show
,
6534 .name
= "events.local",
6535 .flags
= CFTYPE_NOT_ON_ROOT
,
6536 .file_offset
= offsetof(struct mem_cgroup
, events_local_file
),
6537 .seq_show
= memory_events_local_show
,
6541 .seq_show
= memory_stat_show
,
6545 .name
= "numa_stat",
6546 .seq_show
= memory_numa_stat_show
,
6550 .name
= "oom.group",
6551 .flags
= CFTYPE_NOT_ON_ROOT
| CFTYPE_NS_DELEGATABLE
,
6552 .seq_show
= memory_oom_group_show
,
6553 .write
= memory_oom_group_write
,
6558 struct cgroup_subsys memory_cgrp_subsys
= {
6559 .css_alloc
= mem_cgroup_css_alloc
,
6560 .css_online
= mem_cgroup_css_online
,
6561 .css_offline
= mem_cgroup_css_offline
,
6562 .css_released
= mem_cgroup_css_released
,
6563 .css_free
= mem_cgroup_css_free
,
6564 .css_reset
= mem_cgroup_css_reset
,
6565 .can_attach
= mem_cgroup_can_attach
,
6566 .cancel_attach
= mem_cgroup_cancel_attach
,
6567 .post_attach
= mem_cgroup_move_task
,
6568 .bind
= mem_cgroup_bind
,
6569 .dfl_cftypes
= memory_files
,
6570 .legacy_cftypes
= mem_cgroup_legacy_files
,
6575 * This function calculates an individual cgroup's effective
6576 * protection which is derived from its own memory.min/low, its
6577 * parent's and siblings' settings, as well as the actual memory
6578 * distribution in the tree.
6580 * The following rules apply to the effective protection values:
6582 * 1. At the first level of reclaim, effective protection is equal to
6583 * the declared protection in memory.min and memory.low.
6585 * 2. To enable safe delegation of the protection configuration, at
6586 * subsequent levels the effective protection is capped to the
6587 * parent's effective protection.
6589 * 3. To make complex and dynamic subtrees easier to configure, the
6590 * user is allowed to overcommit the declared protection at a given
6591 * level. If that is the case, the parent's effective protection is
6592 * distributed to the children in proportion to how much protection
6593 * they have declared and how much of it they are utilizing.
6595 * This makes distribution proportional, but also work-conserving:
6596 * if one cgroup claims much more protection than it uses memory,
6597 * the unused remainder is available to its siblings.
6599 * 4. Conversely, when the declared protection is undercommitted at a
6600 * given level, the distribution of the larger parental protection
6601 * budget is NOT proportional. A cgroup's protection from a sibling
6602 * is capped to its own memory.min/low setting.
6604 * 5. However, to allow protecting recursive subtrees from each other
6605 * without having to declare each individual cgroup's fixed share
6606 * of the ancestor's claim to protection, any unutilized -
6607 * "floating" - protection from up the tree is distributed in
6608 * proportion to each cgroup's *usage*. This makes the protection
6609 * neutral wrt sibling cgroups and lets them compete freely over
6610 * the shared parental protection budget, but it protects the
6611 * subtree as a whole from neighboring subtrees.
6613 * Note that 4. and 5. are not in conflict: 4. is about protecting
6614 * against immediate siblings whereas 5. is about protecting against
6615 * neighboring subtrees.
6617 static unsigned long effective_protection(unsigned long usage
,
6618 unsigned long parent_usage
,
6619 unsigned long setting
,
6620 unsigned long parent_effective
,
6621 unsigned long siblings_protected
)
6623 unsigned long protected;
6626 protected = min(usage
, setting
);
6628 * If all cgroups at this level combined claim and use more
6629 * protection then what the parent affords them, distribute
6630 * shares in proportion to utilization.
6632 * We are using actual utilization rather than the statically
6633 * claimed protection in order to be work-conserving: claimed
6634 * but unused protection is available to siblings that would
6635 * otherwise get a smaller chunk than what they claimed.
6637 if (siblings_protected
> parent_effective
)
6638 return protected * parent_effective
/ siblings_protected
;
6641 * Ok, utilized protection of all children is within what the
6642 * parent affords them, so we know whatever this child claims
6643 * and utilizes is effectively protected.
6645 * If there is unprotected usage beyond this value, reclaim
6646 * will apply pressure in proportion to that amount.
6648 * If there is unutilized protection, the cgroup will be fully
6649 * shielded from reclaim, but we do return a smaller value for
6650 * protection than what the group could enjoy in theory. This
6651 * is okay. With the overcommit distribution above, effective
6652 * protection is always dependent on how memory is actually
6653 * consumed among the siblings anyway.
6658 * If the children aren't claiming (all of) the protection
6659 * afforded to them by the parent, distribute the remainder in
6660 * proportion to the (unprotected) memory of each cgroup. That
6661 * way, cgroups that aren't explicitly prioritized wrt each
6662 * other compete freely over the allowance, but they are
6663 * collectively protected from neighboring trees.
6665 * We're using unprotected memory for the weight so that if
6666 * some cgroups DO claim explicit protection, we don't protect
6667 * the same bytes twice.
6669 * Check both usage and parent_usage against the respective
6670 * protected values. One should imply the other, but they
6671 * aren't read atomically - make sure the division is sane.
6673 if (!(cgrp_dfl_root
.flags
& CGRP_ROOT_MEMORY_RECURSIVE_PROT
))
6675 if (parent_effective
> siblings_protected
&&
6676 parent_usage
> siblings_protected
&&
6677 usage
> protected) {
6678 unsigned long unclaimed
;
6680 unclaimed
= parent_effective
- siblings_protected
;
6681 unclaimed
*= usage
- protected;
6682 unclaimed
/= parent_usage
- siblings_protected
;
6691 * mem_cgroup_protected - check if memory consumption is in the normal range
6692 * @root: the top ancestor of the sub-tree being checked
6693 * @memcg: the memory cgroup to check
6695 * WARNING: This function is not stateless! It can only be used as part
6696 * of a top-down tree iteration, not for isolated queries.
6698 void mem_cgroup_calculate_protection(struct mem_cgroup
*root
,
6699 struct mem_cgroup
*memcg
)
6701 unsigned long usage
, parent_usage
;
6702 struct mem_cgroup
*parent
;
6704 if (mem_cgroup_disabled())
6708 root
= root_mem_cgroup
;
6711 * Effective values of the reclaim targets are ignored so they
6712 * can be stale. Have a look at mem_cgroup_protection for more
6714 * TODO: calculation should be more robust so that we do not need
6715 * that special casing.
6720 usage
= page_counter_read(&memcg
->memory
);
6724 parent
= parent_mem_cgroup(memcg
);
6725 /* No parent means a non-hierarchical mode on v1 memcg */
6729 if (parent
== root
) {
6730 memcg
->memory
.emin
= READ_ONCE(memcg
->memory
.min
);
6731 memcg
->memory
.elow
= READ_ONCE(memcg
->memory
.low
);
6735 parent_usage
= page_counter_read(&parent
->memory
);
6737 WRITE_ONCE(memcg
->memory
.emin
, effective_protection(usage
, parent_usage
,
6738 READ_ONCE(memcg
->memory
.min
),
6739 READ_ONCE(parent
->memory
.emin
),
6740 atomic_long_read(&parent
->memory
.children_min_usage
)));
6742 WRITE_ONCE(memcg
->memory
.elow
, effective_protection(usage
, parent_usage
,
6743 READ_ONCE(memcg
->memory
.low
),
6744 READ_ONCE(parent
->memory
.elow
),
6745 atomic_long_read(&parent
->memory
.children_low_usage
)));
6749 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6750 * @page: page to charge
6751 * @mm: mm context of the victim
6752 * @gfp_mask: reclaim mode
6754 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6755 * pages according to @gfp_mask if necessary.
6757 * Returns 0 on success. Otherwise, an error code is returned.
6759 int mem_cgroup_charge(struct page
*page
, struct mm_struct
*mm
, gfp_t gfp_mask
)
6761 unsigned int nr_pages
= thp_nr_pages(page
);
6762 struct mem_cgroup
*memcg
= NULL
;
6765 if (mem_cgroup_disabled())
6768 if (PageSwapCache(page
)) {
6769 swp_entry_t ent
= { .val
= page_private(page
), };
6773 * Every swap fault against a single page tries to charge the
6774 * page, bail as early as possible. shmem_unuse() encounters
6775 * already charged pages, too. page->mem_cgroup is protected
6776 * by the page lock, which serializes swap cache removal, which
6777 * in turn serializes uncharging.
6779 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
6780 if (compound_head(page
)->mem_cgroup
)
6783 id
= lookup_swap_cgroup_id(ent
);
6785 memcg
= mem_cgroup_from_id(id
);
6786 if (memcg
&& !css_tryget_online(&memcg
->css
))
6792 memcg
= get_mem_cgroup_from_mm(mm
);
6794 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
6798 css_get(&memcg
->css
);
6799 commit_charge(page
, memcg
);
6801 local_irq_disable();
6802 mem_cgroup_charge_statistics(memcg
, page
, nr_pages
);
6803 memcg_check_events(memcg
, page
);
6806 if (PageSwapCache(page
)) {
6807 swp_entry_t entry
= { .val
= page_private(page
) };
6809 * The swap entry might not get freed for a long time,
6810 * let's not wait for it. The page already received a
6811 * memory+swap charge, drop the swap entry duplicate.
6813 mem_cgroup_uncharge_swap(entry
, nr_pages
);
6817 css_put(&memcg
->css
);
6822 struct uncharge_gather
{
6823 struct mem_cgroup
*memcg
;
6824 unsigned long nr_pages
;
6825 unsigned long pgpgout
;
6826 unsigned long nr_kmem
;
6827 struct page
*dummy_page
;
6830 static inline void uncharge_gather_clear(struct uncharge_gather
*ug
)
6832 memset(ug
, 0, sizeof(*ug
));
6835 static void uncharge_batch(const struct uncharge_gather
*ug
)
6837 unsigned long flags
;
6839 if (!mem_cgroup_is_root(ug
->memcg
)) {
6840 page_counter_uncharge(&ug
->memcg
->memory
, ug
->nr_pages
);
6841 if (do_memsw_account())
6842 page_counter_uncharge(&ug
->memcg
->memsw
, ug
->nr_pages
);
6843 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && ug
->nr_kmem
)
6844 page_counter_uncharge(&ug
->memcg
->kmem
, ug
->nr_kmem
);
6845 memcg_oom_recover(ug
->memcg
);
6848 local_irq_save(flags
);
6849 __count_memcg_events(ug
->memcg
, PGPGOUT
, ug
->pgpgout
);
6850 __this_cpu_add(ug
->memcg
->vmstats_percpu
->nr_page_events
, ug
->nr_pages
);
6851 memcg_check_events(ug
->memcg
, ug
->dummy_page
);
6852 local_irq_restore(flags
);
6854 /* drop reference from uncharge_page */
6855 css_put(&ug
->memcg
->css
);
6858 static void uncharge_page(struct page
*page
, struct uncharge_gather
*ug
)
6860 unsigned long nr_pages
;
6862 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6864 if (!page
->mem_cgroup
)
6868 * Nobody should be changing or seriously looking at
6869 * page->mem_cgroup at this point, we have fully
6870 * exclusive access to the page.
6873 if (ug
->memcg
!= page
->mem_cgroup
) {
6876 uncharge_gather_clear(ug
);
6878 ug
->memcg
= page
->mem_cgroup
;
6880 /* pairs with css_put in uncharge_batch */
6881 css_get(&ug
->memcg
->css
);
6884 nr_pages
= compound_nr(page
);
6885 ug
->nr_pages
+= nr_pages
;
6887 if (!PageKmemcg(page
)) {
6890 ug
->nr_kmem
+= nr_pages
;
6891 __ClearPageKmemcg(page
);
6894 ug
->dummy_page
= page
;
6895 page
->mem_cgroup
= NULL
;
6896 css_put(&ug
->memcg
->css
);
6899 static void uncharge_list(struct list_head
*page_list
)
6901 struct uncharge_gather ug
;
6902 struct list_head
*next
;
6904 uncharge_gather_clear(&ug
);
6907 * Note that the list can be a single page->lru; hence the
6908 * do-while loop instead of a simple list_for_each_entry().
6910 next
= page_list
->next
;
6914 page
= list_entry(next
, struct page
, lru
);
6915 next
= page
->lru
.next
;
6917 uncharge_page(page
, &ug
);
6918 } while (next
!= page_list
);
6921 uncharge_batch(&ug
);
6925 * mem_cgroup_uncharge - uncharge a page
6926 * @page: page to uncharge
6928 * Uncharge a page previously charged with mem_cgroup_charge().
6930 void mem_cgroup_uncharge(struct page
*page
)
6932 struct uncharge_gather ug
;
6934 if (mem_cgroup_disabled())
6937 /* Don't touch page->lru of any random page, pre-check: */
6938 if (!page
->mem_cgroup
)
6941 uncharge_gather_clear(&ug
);
6942 uncharge_page(page
, &ug
);
6943 uncharge_batch(&ug
);
6947 * mem_cgroup_uncharge_list - uncharge a list of page
6948 * @page_list: list of pages to uncharge
6950 * Uncharge a list of pages previously charged with
6951 * mem_cgroup_charge().
6953 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
6955 if (mem_cgroup_disabled())
6958 if (!list_empty(page_list
))
6959 uncharge_list(page_list
);
6963 * mem_cgroup_migrate - charge a page's replacement
6964 * @oldpage: currently circulating page
6965 * @newpage: replacement page
6967 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6968 * be uncharged upon free.
6970 * Both pages must be locked, @newpage->mapping must be set up.
6972 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
6974 struct mem_cgroup
*memcg
;
6975 unsigned int nr_pages
;
6976 unsigned long flags
;
6978 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
6979 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
6980 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
6981 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
6984 if (mem_cgroup_disabled())
6987 /* Page cache replacement: new page already charged? */
6988 if (newpage
->mem_cgroup
)
6991 /* Swapcache readahead pages can get replaced before being charged */
6992 memcg
= oldpage
->mem_cgroup
;
6996 /* Force-charge the new page. The old one will be freed soon */
6997 nr_pages
= thp_nr_pages(newpage
);
6999 page_counter_charge(&memcg
->memory
, nr_pages
);
7000 if (do_memsw_account())
7001 page_counter_charge(&memcg
->memsw
, nr_pages
);
7003 css_get(&memcg
->css
);
7004 commit_charge(newpage
, memcg
);
7006 local_irq_save(flags
);
7007 mem_cgroup_charge_statistics(memcg
, newpage
, nr_pages
);
7008 memcg_check_events(memcg
, newpage
);
7009 local_irq_restore(flags
);
7012 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
7013 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
7015 void mem_cgroup_sk_alloc(struct sock
*sk
)
7017 struct mem_cgroup
*memcg
;
7019 if (!mem_cgroup_sockets_enabled
)
7022 /* Do not associate the sock with unrelated interrupted task's memcg. */
7027 memcg
= mem_cgroup_from_task(current
);
7028 if (memcg
== root_mem_cgroup
)
7030 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
7032 if (css_tryget(&memcg
->css
))
7033 sk
->sk_memcg
= memcg
;
7038 void mem_cgroup_sk_free(struct sock
*sk
)
7041 css_put(&sk
->sk_memcg
->css
);
7045 * mem_cgroup_charge_skmem - charge socket memory
7046 * @memcg: memcg to charge
7047 * @nr_pages: number of pages to charge
7049 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7050 * @memcg's configured limit, %false if the charge had to be forced.
7052 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
7054 gfp_t gfp_mask
= GFP_KERNEL
;
7056 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
7057 struct page_counter
*fail
;
7059 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
7060 memcg
->tcpmem_pressure
= 0;
7063 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
7064 memcg
->tcpmem_pressure
= 1;
7068 /* Don't block in the packet receive path */
7070 gfp_mask
= GFP_NOWAIT
;
7072 mod_memcg_state(memcg
, MEMCG_SOCK
, nr_pages
);
7074 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
7077 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
7082 * mem_cgroup_uncharge_skmem - uncharge socket memory
7083 * @memcg: memcg to uncharge
7084 * @nr_pages: number of pages to uncharge
7086 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
7088 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
7089 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
7093 mod_memcg_state(memcg
, MEMCG_SOCK
, -nr_pages
);
7095 refill_stock(memcg
, nr_pages
);
7098 static int __init
cgroup_memory(char *s
)
7102 while ((token
= strsep(&s
, ",")) != NULL
) {
7105 if (!strcmp(token
, "nosocket"))
7106 cgroup_memory_nosocket
= true;
7107 if (!strcmp(token
, "nokmem"))
7108 cgroup_memory_nokmem
= true;
7112 __setup("cgroup.memory=", cgroup_memory
);
7115 * subsys_initcall() for memory controller.
7117 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7118 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7119 * basically everything that doesn't depend on a specific mem_cgroup structure
7120 * should be initialized from here.
7122 static int __init
mem_cgroup_init(void)
7126 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD
, "mm/memctrl:dead", NULL
,
7127 memcg_hotplug_cpu_dead
);
7129 for_each_possible_cpu(cpu
)
7130 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
7133 for_each_node(node
) {
7134 struct mem_cgroup_tree_per_node
*rtpn
;
7136 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
7137 node_online(node
) ? node
: NUMA_NO_NODE
);
7139 rtpn
->rb_root
= RB_ROOT
;
7140 rtpn
->rb_rightmost
= NULL
;
7141 spin_lock_init(&rtpn
->lock
);
7142 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
7147 subsys_initcall(mem_cgroup_init
);
7149 #ifdef CONFIG_MEMCG_SWAP
7150 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
7152 while (!refcount_inc_not_zero(&memcg
->id
.ref
)) {
7154 * The root cgroup cannot be destroyed, so it's refcount must
7157 if (WARN_ON_ONCE(memcg
== root_mem_cgroup
)) {
7161 memcg
= parent_mem_cgroup(memcg
);
7163 memcg
= root_mem_cgroup
;
7169 * mem_cgroup_swapout - transfer a memsw charge to swap
7170 * @page: page whose memsw charge to transfer
7171 * @entry: swap entry to move the charge to
7173 * Transfer the memsw charge of @page to @entry.
7175 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
7177 struct mem_cgroup
*memcg
, *swap_memcg
;
7178 unsigned int nr_entries
;
7179 unsigned short oldid
;
7181 VM_BUG_ON_PAGE(PageLRU(page
), page
);
7182 VM_BUG_ON_PAGE(page_count(page
), page
);
7184 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7187 memcg
= page
->mem_cgroup
;
7189 /* Readahead page, never charged */
7194 * In case the memcg owning these pages has been offlined and doesn't
7195 * have an ID allocated to it anymore, charge the closest online
7196 * ancestor for the swap instead and transfer the memory+swap charge.
7198 swap_memcg
= mem_cgroup_id_get_online(memcg
);
7199 nr_entries
= thp_nr_pages(page
);
7200 /* Get references for the tail pages, too */
7202 mem_cgroup_id_get_many(swap_memcg
, nr_entries
- 1);
7203 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
),
7205 VM_BUG_ON_PAGE(oldid
, page
);
7206 mod_memcg_state(swap_memcg
, MEMCG_SWAP
, nr_entries
);
7208 page
->mem_cgroup
= NULL
;
7210 if (!mem_cgroup_is_root(memcg
))
7211 page_counter_uncharge(&memcg
->memory
, nr_entries
);
7213 if (!cgroup_memory_noswap
&& memcg
!= swap_memcg
) {
7214 if (!mem_cgroup_is_root(swap_memcg
))
7215 page_counter_charge(&swap_memcg
->memsw
, nr_entries
);
7216 page_counter_uncharge(&memcg
->memsw
, nr_entries
);
7220 * Interrupts should be disabled here because the caller holds the
7221 * i_pages lock which is taken with interrupts-off. It is
7222 * important here to have the interrupts disabled because it is the
7223 * only synchronisation we have for updating the per-CPU variables.
7225 VM_BUG_ON(!irqs_disabled());
7226 mem_cgroup_charge_statistics(memcg
, page
, -nr_entries
);
7227 memcg_check_events(memcg
, page
);
7229 css_put(&memcg
->css
);
7233 * mem_cgroup_try_charge_swap - try charging swap space for a page
7234 * @page: page being added to swap
7235 * @entry: swap entry to charge
7237 * Try to charge @page's memcg for the swap space at @entry.
7239 * Returns 0 on success, -ENOMEM on failure.
7241 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
7243 unsigned int nr_pages
= thp_nr_pages(page
);
7244 struct page_counter
*counter
;
7245 struct mem_cgroup
*memcg
;
7246 unsigned short oldid
;
7248 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7251 memcg
= page
->mem_cgroup
;
7253 /* Readahead page, never charged */
7258 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7262 memcg
= mem_cgroup_id_get_online(memcg
);
7264 if (!cgroup_memory_noswap
&& !mem_cgroup_is_root(memcg
) &&
7265 !page_counter_try_charge(&memcg
->swap
, nr_pages
, &counter
)) {
7266 memcg_memory_event(memcg
, MEMCG_SWAP_MAX
);
7267 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7268 mem_cgroup_id_put(memcg
);
7272 /* Get references for the tail pages, too */
7274 mem_cgroup_id_get_many(memcg
, nr_pages
- 1);
7275 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
), nr_pages
);
7276 VM_BUG_ON_PAGE(oldid
, page
);
7277 mod_memcg_state(memcg
, MEMCG_SWAP
, nr_pages
);
7283 * mem_cgroup_uncharge_swap - uncharge swap space
7284 * @entry: swap entry to uncharge
7285 * @nr_pages: the amount of swap space to uncharge
7287 void mem_cgroup_uncharge_swap(swp_entry_t entry
, unsigned int nr_pages
)
7289 struct mem_cgroup
*memcg
;
7292 id
= swap_cgroup_record(entry
, 0, nr_pages
);
7294 memcg
= mem_cgroup_from_id(id
);
7296 if (!cgroup_memory_noswap
&& !mem_cgroup_is_root(memcg
)) {
7297 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7298 page_counter_uncharge(&memcg
->swap
, nr_pages
);
7300 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
7302 mod_memcg_state(memcg
, MEMCG_SWAP
, -nr_pages
);
7303 mem_cgroup_id_put_many(memcg
, nr_pages
);
7308 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
7310 long nr_swap_pages
= get_nr_swap_pages();
7312 if (cgroup_memory_noswap
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7313 return nr_swap_pages
;
7314 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
7315 nr_swap_pages
= min_t(long, nr_swap_pages
,
7316 READ_ONCE(memcg
->swap
.max
) -
7317 page_counter_read(&memcg
->swap
));
7318 return nr_swap_pages
;
7321 bool mem_cgroup_swap_full(struct page
*page
)
7323 struct mem_cgroup
*memcg
;
7325 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
7329 if (cgroup_memory_noswap
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7332 memcg
= page
->mem_cgroup
;
7336 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
)) {
7337 unsigned long usage
= page_counter_read(&memcg
->swap
);
7339 if (usage
* 2 >= READ_ONCE(memcg
->swap
.high
) ||
7340 usage
* 2 >= READ_ONCE(memcg
->swap
.max
))
7347 static int __init
setup_swap_account(char *s
)
7349 if (!strcmp(s
, "1"))
7350 cgroup_memory_noswap
= 0;
7351 else if (!strcmp(s
, "0"))
7352 cgroup_memory_noswap
= 1;
7355 __setup("swapaccount=", setup_swap_account
);
7357 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
7360 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
7362 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
7365 static int swap_high_show(struct seq_file
*m
, void *v
)
7367 return seq_puts_memcg_tunable(m
,
7368 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.high
));
7371 static ssize_t
swap_high_write(struct kernfs_open_file
*of
,
7372 char *buf
, size_t nbytes
, loff_t off
)
7374 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7378 buf
= strstrip(buf
);
7379 err
= page_counter_memparse(buf
, "max", &high
);
7383 page_counter_set_high(&memcg
->swap
, high
);
7388 static int swap_max_show(struct seq_file
*m
, void *v
)
7390 return seq_puts_memcg_tunable(m
,
7391 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.max
));
7394 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
7395 char *buf
, size_t nbytes
, loff_t off
)
7397 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7401 buf
= strstrip(buf
);
7402 err
= page_counter_memparse(buf
, "max", &max
);
7406 xchg(&memcg
->swap
.max
, max
);
7411 static int swap_events_show(struct seq_file
*m
, void *v
)
7413 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
7415 seq_printf(m
, "high %lu\n",
7416 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_HIGH
]));
7417 seq_printf(m
, "max %lu\n",
7418 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_MAX
]));
7419 seq_printf(m
, "fail %lu\n",
7420 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_FAIL
]));
7425 static struct cftype swap_files
[] = {
7427 .name
= "swap.current",
7428 .flags
= CFTYPE_NOT_ON_ROOT
,
7429 .read_u64
= swap_current_read
,
7432 .name
= "swap.high",
7433 .flags
= CFTYPE_NOT_ON_ROOT
,
7434 .seq_show
= swap_high_show
,
7435 .write
= swap_high_write
,
7439 .flags
= CFTYPE_NOT_ON_ROOT
,
7440 .seq_show
= swap_max_show
,
7441 .write
= swap_max_write
,
7444 .name
= "swap.events",
7445 .flags
= CFTYPE_NOT_ON_ROOT
,
7446 .file_offset
= offsetof(struct mem_cgroup
, swap_events_file
),
7447 .seq_show
= swap_events_show
,
7452 static struct cftype memsw_files
[] = {
7454 .name
= "memsw.usage_in_bytes",
7455 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
7456 .read_u64
= mem_cgroup_read_u64
,
7459 .name
= "memsw.max_usage_in_bytes",
7460 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
7461 .write
= mem_cgroup_reset
,
7462 .read_u64
= mem_cgroup_read_u64
,
7465 .name
= "memsw.limit_in_bytes",
7466 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
7467 .write
= mem_cgroup_write
,
7468 .read_u64
= mem_cgroup_read_u64
,
7471 .name
= "memsw.failcnt",
7472 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
7473 .write
= mem_cgroup_reset
,
7474 .read_u64
= mem_cgroup_read_u64
,
7476 { }, /* terminate */
7480 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7481 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7482 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7483 * boot parameter. This may result in premature OOPS inside
7484 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7486 static int __init
mem_cgroup_swap_init(void)
7488 /* No memory control -> no swap control */
7489 if (mem_cgroup_disabled())
7490 cgroup_memory_noswap
= true;
7492 if (cgroup_memory_noswap
)
7495 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
, swap_files
));
7496 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
, memsw_files
));
7500 core_initcall(mem_cgroup_swap_init
);
7502 #endif /* CONFIG_MEMCG_SWAP */