1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
24 * Per memcg lru locking
25 * Copyright (C) 2020 Alibaba, Inc, Alex Shi
28 #include <linux/page_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
31 #include <linux/pagewalk.h>
32 #include <linux/sched/mm.h>
33 #include <linux/shmem_fs.h>
34 #include <linux/hugetlb.h>
35 #include <linux/pagemap.h>
36 #include <linux/vm_event_item.h>
37 #include <linux/smp.h>
38 #include <linux/page-flags.h>
39 #include <linux/backing-dev.h>
40 #include <linux/bit_spinlock.h>
41 #include <linux/rcupdate.h>
42 #include <linux/limits.h>
43 #include <linux/export.h>
44 #include <linux/mutex.h>
45 #include <linux/rbtree.h>
46 #include <linux/slab.h>
47 #include <linux/swap.h>
48 #include <linux/swapops.h>
49 #include <linux/spinlock.h>
50 #include <linux/eventfd.h>
51 #include <linux/poll.h>
52 #include <linux/sort.h>
54 #include <linux/seq_file.h>
55 #include <linux/vmpressure.h>
56 #include <linux/mm_inline.h>
57 #include <linux/swap_cgroup.h>
58 #include <linux/cpu.h>
59 #include <linux/oom.h>
60 #include <linux/lockdep.h>
61 #include <linux/file.h>
62 #include <linux/tracehook.h>
63 #include <linux/psi.h>
64 #include <linux/seq_buf.h>
70 #include <linux/uaccess.h>
72 #include <trace/events/vmscan.h>
74 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
75 EXPORT_SYMBOL(memory_cgrp_subsys
);
77 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
79 /* Active memory cgroup to use from an interrupt context */
80 DEFINE_PER_CPU(struct mem_cgroup
*, int_active_memcg
);
82 /* Socket memory accounting disabled? */
83 static bool cgroup_memory_nosocket
;
85 /* Kernel memory accounting disabled? */
86 static bool cgroup_memory_nokmem
;
88 /* Whether the swap controller is active */
89 #ifdef CONFIG_MEMCG_SWAP
90 bool cgroup_memory_noswap __read_mostly
;
92 #define cgroup_memory_noswap 1
95 #ifdef CONFIG_CGROUP_WRITEBACK
96 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq
);
99 /* Whether legacy memory+swap accounting is active */
100 static bool do_memsw_account(void)
102 return !cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_noswap
;
105 #define THRESHOLDS_EVENTS_TARGET 128
106 #define SOFTLIMIT_EVENTS_TARGET 1024
109 * Cgroups above their limits are maintained in a RB-Tree, independent of
110 * their hierarchy representation
113 struct mem_cgroup_tree_per_node
{
114 struct rb_root rb_root
;
115 struct rb_node
*rb_rightmost
;
119 struct mem_cgroup_tree
{
120 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
123 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
126 struct mem_cgroup_eventfd_list
{
127 struct list_head list
;
128 struct eventfd_ctx
*eventfd
;
132 * cgroup_event represents events which userspace want to receive.
134 struct mem_cgroup_event
{
136 * memcg which the event belongs to.
138 struct mem_cgroup
*memcg
;
140 * eventfd to signal userspace about the event.
142 struct eventfd_ctx
*eventfd
;
144 * Each of these stored in a list by the cgroup.
146 struct list_head list
;
148 * register_event() callback will be used to add new userspace
149 * waiter for changes related to this event. Use eventfd_signal()
150 * on eventfd to send notification to userspace.
152 int (*register_event
)(struct mem_cgroup
*memcg
,
153 struct eventfd_ctx
*eventfd
, const char *args
);
155 * unregister_event() callback will be called when userspace closes
156 * the eventfd or on cgroup removing. This callback must be set,
157 * if you want provide notification functionality.
159 void (*unregister_event
)(struct mem_cgroup
*memcg
,
160 struct eventfd_ctx
*eventfd
);
162 * All fields below needed to unregister event when
163 * userspace closes eventfd.
166 wait_queue_head_t
*wqh
;
167 wait_queue_entry_t wait
;
168 struct work_struct remove
;
171 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
172 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
174 /* Stuffs for move charges at task migration. */
176 * Types of charges to be moved.
178 #define MOVE_ANON 0x1U
179 #define MOVE_FILE 0x2U
180 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
182 /* "mc" and its members are protected by cgroup_mutex */
183 static struct move_charge_struct
{
184 spinlock_t lock
; /* for from, to */
185 struct mm_struct
*mm
;
186 struct mem_cgroup
*from
;
187 struct mem_cgroup
*to
;
189 unsigned long precharge
;
190 unsigned long moved_charge
;
191 unsigned long moved_swap
;
192 struct task_struct
*moving_task
; /* a task moving charges */
193 wait_queue_head_t waitq
; /* a waitq for other context */
195 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
196 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
200 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
201 * limit reclaim to prevent infinite loops, if they ever occur.
203 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
204 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
206 /* for encoding cft->private value on file */
215 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
216 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
217 #define MEMFILE_ATTR(val) ((val) & 0xffff)
218 /* Used for OOM nofiier */
219 #define OOM_CONTROL (0)
222 * Iteration constructs for visiting all cgroups (under a tree). If
223 * loops are exited prematurely (break), mem_cgroup_iter_break() must
224 * be used for reference counting.
226 #define for_each_mem_cgroup_tree(iter, root) \
227 for (iter = mem_cgroup_iter(root, NULL, NULL); \
229 iter = mem_cgroup_iter(root, iter, NULL))
231 #define for_each_mem_cgroup(iter) \
232 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
234 iter = mem_cgroup_iter(NULL, iter, NULL))
236 static inline bool should_force_charge(void)
238 return tsk_is_oom_victim(current
) || fatal_signal_pending(current
) ||
239 (current
->flags
& PF_EXITING
);
242 /* Some nice accessors for the vmpressure. */
243 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
246 memcg
= root_mem_cgroup
;
247 return &memcg
->vmpressure
;
250 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
252 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
255 #ifdef CONFIG_MEMCG_KMEM
256 extern spinlock_t css_set_lock
;
258 static int __memcg_kmem_charge(struct mem_cgroup
*memcg
, gfp_t gfp
,
259 unsigned int nr_pages
);
260 static void __memcg_kmem_uncharge(struct mem_cgroup
*memcg
,
261 unsigned int nr_pages
);
263 static void obj_cgroup_release(struct percpu_ref
*ref
)
265 struct obj_cgroup
*objcg
= container_of(ref
, struct obj_cgroup
, refcnt
);
266 struct mem_cgroup
*memcg
;
267 unsigned int nr_bytes
;
268 unsigned int nr_pages
;
272 * At this point all allocated objects are freed, and
273 * objcg->nr_charged_bytes can't have an arbitrary byte value.
274 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
276 * The following sequence can lead to it:
277 * 1) CPU0: objcg == stock->cached_objcg
278 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
279 * PAGE_SIZE bytes are charged
280 * 3) CPU1: a process from another memcg is allocating something,
281 * the stock if flushed,
282 * objcg->nr_charged_bytes = PAGE_SIZE - 92
283 * 5) CPU0: we do release this object,
284 * 92 bytes are added to stock->nr_bytes
285 * 6) CPU0: stock is flushed,
286 * 92 bytes are added to objcg->nr_charged_bytes
288 * In the result, nr_charged_bytes == PAGE_SIZE.
289 * This page will be uncharged in obj_cgroup_release().
291 nr_bytes
= atomic_read(&objcg
->nr_charged_bytes
);
292 WARN_ON_ONCE(nr_bytes
& (PAGE_SIZE
- 1));
293 nr_pages
= nr_bytes
>> PAGE_SHIFT
;
295 spin_lock_irqsave(&css_set_lock
, flags
);
296 memcg
= obj_cgroup_memcg(objcg
);
298 __memcg_kmem_uncharge(memcg
, nr_pages
);
299 list_del(&objcg
->list
);
300 mem_cgroup_put(memcg
);
301 spin_unlock_irqrestore(&css_set_lock
, flags
);
303 percpu_ref_exit(ref
);
304 kfree_rcu(objcg
, rcu
);
307 static struct obj_cgroup
*obj_cgroup_alloc(void)
309 struct obj_cgroup
*objcg
;
312 objcg
= kzalloc(sizeof(struct obj_cgroup
), GFP_KERNEL
);
316 ret
= percpu_ref_init(&objcg
->refcnt
, obj_cgroup_release
, 0,
322 INIT_LIST_HEAD(&objcg
->list
);
326 static void memcg_reparent_objcgs(struct mem_cgroup
*memcg
,
327 struct mem_cgroup
*parent
)
329 struct obj_cgroup
*objcg
, *iter
;
331 objcg
= rcu_replace_pointer(memcg
->objcg
, NULL
, true);
333 spin_lock_irq(&css_set_lock
);
335 /* Move active objcg to the parent's list */
336 xchg(&objcg
->memcg
, parent
);
337 css_get(&parent
->css
);
338 list_add(&objcg
->list
, &parent
->objcg_list
);
340 /* Move already reparented objcgs to the parent's list */
341 list_for_each_entry(iter
, &memcg
->objcg_list
, list
) {
342 css_get(&parent
->css
);
343 xchg(&iter
->memcg
, parent
);
344 css_put(&memcg
->css
);
346 list_splice(&memcg
->objcg_list
, &parent
->objcg_list
);
348 spin_unlock_irq(&css_set_lock
);
350 percpu_ref_kill(&objcg
->refcnt
);
354 * This will be used as a shrinker list's index.
355 * The main reason for not using cgroup id for this:
356 * this works better in sparse environments, where we have a lot of memcgs,
357 * but only a few kmem-limited. Or also, if we have, for instance, 200
358 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
359 * 200 entry array for that.
361 * The current size of the caches array is stored in memcg_nr_cache_ids. It
362 * will double each time we have to increase it.
364 static DEFINE_IDA(memcg_cache_ida
);
365 int memcg_nr_cache_ids
;
367 /* Protects memcg_nr_cache_ids */
368 static DECLARE_RWSEM(memcg_cache_ids_sem
);
370 void memcg_get_cache_ids(void)
372 down_read(&memcg_cache_ids_sem
);
375 void memcg_put_cache_ids(void)
377 up_read(&memcg_cache_ids_sem
);
381 * MIN_SIZE is different than 1, because we would like to avoid going through
382 * the alloc/free process all the time. In a small machine, 4 kmem-limited
383 * cgroups is a reasonable guess. In the future, it could be a parameter or
384 * tunable, but that is strictly not necessary.
386 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
387 * this constant directly from cgroup, but it is understandable that this is
388 * better kept as an internal representation in cgroup.c. In any case, the
389 * cgrp_id space is not getting any smaller, and we don't have to necessarily
390 * increase ours as well if it increases.
392 #define MEMCG_CACHES_MIN_SIZE 4
393 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
396 * A lot of the calls to the cache allocation functions are expected to be
397 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
398 * conditional to this static branch, we'll have to allow modules that does
399 * kmem_cache_alloc and the such to see this symbol as well
401 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
402 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
405 static int memcg_shrinker_map_size
;
406 static DEFINE_MUTEX(memcg_shrinker_map_mutex
);
408 static void memcg_free_shrinker_map_rcu(struct rcu_head
*head
)
410 kvfree(container_of(head
, struct memcg_shrinker_map
, rcu
));
413 static int memcg_expand_one_shrinker_map(struct mem_cgroup
*memcg
,
414 int size
, int old_size
)
416 struct memcg_shrinker_map
*new, *old
;
419 lockdep_assert_held(&memcg_shrinker_map_mutex
);
422 old
= rcu_dereference_protected(
423 mem_cgroup_nodeinfo(memcg
, nid
)->shrinker_map
, true);
424 /* Not yet online memcg */
428 new = kvmalloc_node(sizeof(*new) + size
, GFP_KERNEL
, nid
);
432 /* Set all old bits, clear all new bits */
433 memset(new->map
, (int)0xff, old_size
);
434 memset((void *)new->map
+ old_size
, 0, size
- old_size
);
436 rcu_assign_pointer(memcg
->nodeinfo
[nid
]->shrinker_map
, new);
437 call_rcu(&old
->rcu
, memcg_free_shrinker_map_rcu
);
443 static void memcg_free_shrinker_maps(struct mem_cgroup
*memcg
)
445 struct mem_cgroup_per_node
*pn
;
446 struct memcg_shrinker_map
*map
;
449 if (mem_cgroup_is_root(memcg
))
453 pn
= mem_cgroup_nodeinfo(memcg
, nid
);
454 map
= rcu_dereference_protected(pn
->shrinker_map
, true);
456 rcu_assign_pointer(pn
->shrinker_map
, NULL
);
460 static int memcg_alloc_shrinker_maps(struct mem_cgroup
*memcg
)
462 struct memcg_shrinker_map
*map
;
463 int nid
, size
, ret
= 0;
465 if (mem_cgroup_is_root(memcg
))
468 mutex_lock(&memcg_shrinker_map_mutex
);
469 size
= memcg_shrinker_map_size
;
471 map
= kvzalloc_node(sizeof(*map
) + size
, GFP_KERNEL
, nid
);
473 memcg_free_shrinker_maps(memcg
);
477 rcu_assign_pointer(memcg
->nodeinfo
[nid
]->shrinker_map
, map
);
479 mutex_unlock(&memcg_shrinker_map_mutex
);
484 int memcg_expand_shrinker_maps(int new_id
)
486 int size
, old_size
, ret
= 0;
487 struct mem_cgroup
*memcg
;
489 size
= DIV_ROUND_UP(new_id
+ 1, BITS_PER_LONG
) * sizeof(unsigned long);
490 old_size
= memcg_shrinker_map_size
;
491 if (size
<= old_size
)
494 mutex_lock(&memcg_shrinker_map_mutex
);
495 if (!root_mem_cgroup
)
498 for_each_mem_cgroup(memcg
) {
499 if (mem_cgroup_is_root(memcg
))
501 ret
= memcg_expand_one_shrinker_map(memcg
, size
, old_size
);
503 mem_cgroup_iter_break(NULL
, memcg
);
509 memcg_shrinker_map_size
= size
;
510 mutex_unlock(&memcg_shrinker_map_mutex
);
514 void memcg_set_shrinker_bit(struct mem_cgroup
*memcg
, int nid
, int shrinker_id
)
516 if (shrinker_id
>= 0 && memcg
&& !mem_cgroup_is_root(memcg
)) {
517 struct memcg_shrinker_map
*map
;
520 map
= rcu_dereference(memcg
->nodeinfo
[nid
]->shrinker_map
);
521 /* Pairs with smp mb in shrink_slab() */
522 smp_mb__before_atomic();
523 set_bit(shrinker_id
, map
->map
);
529 * mem_cgroup_css_from_page - css of the memcg associated with a page
530 * @page: page of interest
532 * If memcg is bound to the default hierarchy, css of the memcg associated
533 * with @page is returned. The returned css remains associated with @page
534 * until it is released.
536 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
539 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
541 struct mem_cgroup
*memcg
;
543 memcg
= page_memcg(page
);
545 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
546 memcg
= root_mem_cgroup
;
552 * page_cgroup_ino - return inode number of the memcg a page is charged to
555 * Look up the closest online ancestor of the memory cgroup @page is charged to
556 * and return its inode number or 0 if @page is not charged to any cgroup. It
557 * is safe to call this function without holding a reference to @page.
559 * Note, this function is inherently racy, because there is nothing to prevent
560 * the cgroup inode from getting torn down and potentially reallocated a moment
561 * after page_cgroup_ino() returns, so it only should be used by callers that
562 * do not care (such as procfs interfaces).
564 ino_t
page_cgroup_ino(struct page
*page
)
566 struct mem_cgroup
*memcg
;
567 unsigned long ino
= 0;
570 memcg
= page_memcg_check(page
);
572 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
573 memcg
= parent_mem_cgroup(memcg
);
575 ino
= cgroup_ino(memcg
->css
.cgroup
);
580 static struct mem_cgroup_per_node
*
581 mem_cgroup_page_nodeinfo(struct mem_cgroup
*memcg
, struct page
*page
)
583 int nid
= page_to_nid(page
);
585 return memcg
->nodeinfo
[nid
];
588 static struct mem_cgroup_tree_per_node
*
589 soft_limit_tree_node(int nid
)
591 return soft_limit_tree
.rb_tree_per_node
[nid
];
594 static struct mem_cgroup_tree_per_node
*
595 soft_limit_tree_from_page(struct page
*page
)
597 int nid
= page_to_nid(page
);
599 return soft_limit_tree
.rb_tree_per_node
[nid
];
602 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node
*mz
,
603 struct mem_cgroup_tree_per_node
*mctz
,
604 unsigned long new_usage_in_excess
)
606 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
607 struct rb_node
*parent
= NULL
;
608 struct mem_cgroup_per_node
*mz_node
;
609 bool rightmost
= true;
614 mz
->usage_in_excess
= new_usage_in_excess
;
615 if (!mz
->usage_in_excess
)
619 mz_node
= rb_entry(parent
, struct mem_cgroup_per_node
,
621 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
) {
630 mctz
->rb_rightmost
= &mz
->tree_node
;
632 rb_link_node(&mz
->tree_node
, parent
, p
);
633 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
637 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
638 struct mem_cgroup_tree_per_node
*mctz
)
643 if (&mz
->tree_node
== mctz
->rb_rightmost
)
644 mctz
->rb_rightmost
= rb_prev(&mz
->tree_node
);
646 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
650 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
651 struct mem_cgroup_tree_per_node
*mctz
)
655 spin_lock_irqsave(&mctz
->lock
, flags
);
656 __mem_cgroup_remove_exceeded(mz
, mctz
);
657 spin_unlock_irqrestore(&mctz
->lock
, flags
);
660 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
662 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
663 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
664 unsigned long excess
= 0;
666 if (nr_pages
> soft_limit
)
667 excess
= nr_pages
- soft_limit
;
672 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
674 unsigned long excess
;
675 struct mem_cgroup_per_node
*mz
;
676 struct mem_cgroup_tree_per_node
*mctz
;
678 mctz
= soft_limit_tree_from_page(page
);
682 * Necessary to update all ancestors when hierarchy is used.
683 * because their event counter is not touched.
685 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
686 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
687 excess
= soft_limit_excess(memcg
);
689 * We have to update the tree if mz is on RB-tree or
690 * mem is over its softlimit.
692 if (excess
|| mz
->on_tree
) {
695 spin_lock_irqsave(&mctz
->lock
, flags
);
696 /* if on-tree, remove it */
698 __mem_cgroup_remove_exceeded(mz
, mctz
);
700 * Insert again. mz->usage_in_excess will be updated.
701 * If excess is 0, no tree ops.
703 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
704 spin_unlock_irqrestore(&mctz
->lock
, flags
);
709 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
711 struct mem_cgroup_tree_per_node
*mctz
;
712 struct mem_cgroup_per_node
*mz
;
716 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
717 mctz
= soft_limit_tree_node(nid
);
719 mem_cgroup_remove_exceeded(mz
, mctz
);
723 static struct mem_cgroup_per_node
*
724 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
726 struct mem_cgroup_per_node
*mz
;
730 if (!mctz
->rb_rightmost
)
731 goto done
; /* Nothing to reclaim from */
733 mz
= rb_entry(mctz
->rb_rightmost
,
734 struct mem_cgroup_per_node
, tree_node
);
736 * Remove the node now but someone else can add it back,
737 * we will to add it back at the end of reclaim to its correct
738 * position in the tree.
740 __mem_cgroup_remove_exceeded(mz
, mctz
);
741 if (!soft_limit_excess(mz
->memcg
) ||
742 !css_tryget(&mz
->memcg
->css
))
748 static struct mem_cgroup_per_node
*
749 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
751 struct mem_cgroup_per_node
*mz
;
753 spin_lock_irq(&mctz
->lock
);
754 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
755 spin_unlock_irq(&mctz
->lock
);
760 * __mod_memcg_state - update cgroup memory statistics
761 * @memcg: the memory cgroup
762 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
763 * @val: delta to add to the counter, can be negative
765 void __mod_memcg_state(struct mem_cgroup
*memcg
, int idx
, int val
)
767 long x
, threshold
= MEMCG_CHARGE_BATCH
;
769 if (mem_cgroup_disabled())
772 if (memcg_stat_item_in_bytes(idx
))
773 threshold
<<= PAGE_SHIFT
;
775 x
= val
+ __this_cpu_read(memcg
->vmstats_percpu
->stat
[idx
]);
776 if (unlikely(abs(x
) > threshold
)) {
777 struct mem_cgroup
*mi
;
780 * Batch local counters to keep them in sync with
781 * the hierarchical ones.
783 __this_cpu_add(memcg
->vmstats_local
->stat
[idx
], x
);
784 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
785 atomic_long_add(x
, &mi
->vmstats
[idx
]);
788 __this_cpu_write(memcg
->vmstats_percpu
->stat
[idx
], x
);
791 static struct mem_cgroup_per_node
*
792 parent_nodeinfo(struct mem_cgroup_per_node
*pn
, int nid
)
794 struct mem_cgroup
*parent
;
796 parent
= parent_mem_cgroup(pn
->memcg
);
799 return mem_cgroup_nodeinfo(parent
, nid
);
802 void __mod_memcg_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
805 struct mem_cgroup_per_node
*pn
;
806 struct mem_cgroup
*memcg
;
807 long x
, threshold
= MEMCG_CHARGE_BATCH
;
809 pn
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
813 __mod_memcg_state(memcg
, idx
, val
);
816 __this_cpu_add(pn
->lruvec_stat_local
->count
[idx
], val
);
818 if (vmstat_item_in_bytes(idx
))
819 threshold
<<= PAGE_SHIFT
;
821 x
= val
+ __this_cpu_read(pn
->lruvec_stat_cpu
->count
[idx
]);
822 if (unlikely(abs(x
) > threshold
)) {
823 pg_data_t
*pgdat
= lruvec_pgdat(lruvec
);
824 struct mem_cgroup_per_node
*pi
;
826 for (pi
= pn
; pi
; pi
= parent_nodeinfo(pi
, pgdat
->node_id
))
827 atomic_long_add(x
, &pi
->lruvec_stat
[idx
]);
830 __this_cpu_write(pn
->lruvec_stat_cpu
->count
[idx
], x
);
834 * __mod_lruvec_state - update lruvec memory statistics
835 * @lruvec: the lruvec
836 * @idx: the stat item
837 * @val: delta to add to the counter, can be negative
839 * The lruvec is the intersection of the NUMA node and a cgroup. This
840 * function updates the all three counters that are affected by a
841 * change of state at this level: per-node, per-cgroup, per-lruvec.
843 void __mod_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
847 __mod_node_page_state(lruvec_pgdat(lruvec
), idx
, val
);
849 /* Update memcg and lruvec */
850 if (!mem_cgroup_disabled())
851 __mod_memcg_lruvec_state(lruvec
, idx
, val
);
854 void __mod_lruvec_page_state(struct page
*page
, enum node_stat_item idx
,
857 struct page
*head
= compound_head(page
); /* rmap on tail pages */
858 struct mem_cgroup
*memcg
= page_memcg(head
);
859 pg_data_t
*pgdat
= page_pgdat(page
);
860 struct lruvec
*lruvec
;
862 /* Untracked pages have no memcg, no lruvec. Update only the node */
864 __mod_node_page_state(pgdat
, idx
, val
);
868 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
869 __mod_lruvec_state(lruvec
, idx
, val
);
871 EXPORT_SYMBOL(__mod_lruvec_page_state
);
873 void __mod_lruvec_kmem_state(void *p
, enum node_stat_item idx
, int val
)
875 pg_data_t
*pgdat
= page_pgdat(virt_to_page(p
));
876 struct mem_cgroup
*memcg
;
877 struct lruvec
*lruvec
;
880 memcg
= mem_cgroup_from_obj(p
);
883 * Untracked pages have no memcg, no lruvec. Update only the
884 * node. If we reparent the slab objects to the root memcg,
885 * when we free the slab object, we need to update the per-memcg
886 * vmstats to keep it correct for the root memcg.
889 __mod_node_page_state(pgdat
, idx
, val
);
891 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
892 __mod_lruvec_state(lruvec
, idx
, val
);
898 * __count_memcg_events - account VM events in a cgroup
899 * @memcg: the memory cgroup
900 * @idx: the event item
901 * @count: the number of events that occured
903 void __count_memcg_events(struct mem_cgroup
*memcg
, enum vm_event_item idx
,
908 if (mem_cgroup_disabled())
911 x
= count
+ __this_cpu_read(memcg
->vmstats_percpu
->events
[idx
]);
912 if (unlikely(x
> MEMCG_CHARGE_BATCH
)) {
913 struct mem_cgroup
*mi
;
916 * Batch local counters to keep them in sync with
917 * the hierarchical ones.
919 __this_cpu_add(memcg
->vmstats_local
->events
[idx
], x
);
920 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
921 atomic_long_add(x
, &mi
->vmevents
[idx
]);
924 __this_cpu_write(memcg
->vmstats_percpu
->events
[idx
], x
);
927 static unsigned long memcg_events(struct mem_cgroup
*memcg
, int event
)
929 return atomic_long_read(&memcg
->vmevents
[event
]);
932 static unsigned long memcg_events_local(struct mem_cgroup
*memcg
, int event
)
937 for_each_possible_cpu(cpu
)
938 x
+= per_cpu(memcg
->vmstats_local
->events
[event
], cpu
);
942 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
946 /* pagein of a big page is an event. So, ignore page size */
948 __count_memcg_events(memcg
, PGPGIN
, 1);
950 __count_memcg_events(memcg
, PGPGOUT
, 1);
951 nr_pages
= -nr_pages
; /* for event */
954 __this_cpu_add(memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
957 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
958 enum mem_cgroup_events_target target
)
960 unsigned long val
, next
;
962 val
= __this_cpu_read(memcg
->vmstats_percpu
->nr_page_events
);
963 next
= __this_cpu_read(memcg
->vmstats_percpu
->targets
[target
]);
964 /* from time_after() in jiffies.h */
965 if ((long)(next
- val
) < 0) {
967 case MEM_CGROUP_TARGET_THRESH
:
968 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
970 case MEM_CGROUP_TARGET_SOFTLIMIT
:
971 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
976 __this_cpu_write(memcg
->vmstats_percpu
->targets
[target
], next
);
983 * Check events in order.
986 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
988 /* threshold event is triggered in finer grain than soft limit */
989 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
990 MEM_CGROUP_TARGET_THRESH
))) {
993 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
994 MEM_CGROUP_TARGET_SOFTLIMIT
);
995 mem_cgroup_threshold(memcg
);
996 if (unlikely(do_softlimit
))
997 mem_cgroup_update_tree(memcg
, page
);
1001 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1004 * mm_update_next_owner() may clear mm->owner to NULL
1005 * if it races with swapoff, page migration, etc.
1006 * So this can be called with p == NULL.
1011 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
1013 EXPORT_SYMBOL(mem_cgroup_from_task
);
1016 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1017 * @mm: mm from which memcg should be extracted. It can be NULL.
1019 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1020 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1023 struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1025 struct mem_cgroup
*memcg
;
1027 if (mem_cgroup_disabled())
1033 * Page cache insertions can happen withou an
1034 * actual mm context, e.g. during disk probing
1035 * on boot, loopback IO, acct() writes etc.
1038 memcg
= root_mem_cgroup
;
1040 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1041 if (unlikely(!memcg
))
1042 memcg
= root_mem_cgroup
;
1044 } while (!css_tryget(&memcg
->css
));
1048 EXPORT_SYMBOL(get_mem_cgroup_from_mm
);
1050 static __always_inline
struct mem_cgroup
*active_memcg(void)
1053 return this_cpu_read(int_active_memcg
);
1055 return current
->active_memcg
;
1058 static __always_inline
struct mem_cgroup
*get_active_memcg(void)
1060 struct mem_cgroup
*memcg
;
1063 memcg
= active_memcg();
1064 /* remote memcg must hold a ref. */
1065 if (memcg
&& WARN_ON_ONCE(!css_tryget(&memcg
->css
)))
1066 memcg
= root_mem_cgroup
;
1072 static __always_inline
bool memcg_kmem_bypass(void)
1074 /* Allow remote memcg charging from any context. */
1075 if (unlikely(active_memcg()))
1078 /* Memcg to charge can't be determined. */
1079 if (in_interrupt() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
1086 * If active memcg is set, do not fallback to current->mm->memcg.
1088 static __always_inline
struct mem_cgroup
*get_mem_cgroup_from_current(void)
1090 if (memcg_kmem_bypass())
1093 if (unlikely(active_memcg()))
1094 return get_active_memcg();
1096 return get_mem_cgroup_from_mm(current
->mm
);
1100 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1101 * @root: hierarchy root
1102 * @prev: previously returned memcg, NULL on first invocation
1103 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1105 * Returns references to children of the hierarchy below @root, or
1106 * @root itself, or %NULL after a full round-trip.
1108 * Caller must pass the return value in @prev on subsequent
1109 * invocations for reference counting, or use mem_cgroup_iter_break()
1110 * to cancel a hierarchy walk before the round-trip is complete.
1112 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1113 * in the hierarchy among all concurrent reclaimers operating on the
1116 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1117 struct mem_cgroup
*prev
,
1118 struct mem_cgroup_reclaim_cookie
*reclaim
)
1120 struct mem_cgroup_reclaim_iter
*iter
;
1121 struct cgroup_subsys_state
*css
= NULL
;
1122 struct mem_cgroup
*memcg
= NULL
;
1123 struct mem_cgroup
*pos
= NULL
;
1125 if (mem_cgroup_disabled())
1129 root
= root_mem_cgroup
;
1131 if (prev
&& !reclaim
)
1137 struct mem_cgroup_per_node
*mz
;
1139 mz
= mem_cgroup_nodeinfo(root
, reclaim
->pgdat
->node_id
);
1142 if (prev
&& reclaim
->generation
!= iter
->generation
)
1146 pos
= READ_ONCE(iter
->position
);
1147 if (!pos
|| css_tryget(&pos
->css
))
1150 * css reference reached zero, so iter->position will
1151 * be cleared by ->css_released. However, we should not
1152 * rely on this happening soon, because ->css_released
1153 * is called from a work queue, and by busy-waiting we
1154 * might block it. So we clear iter->position right
1157 (void)cmpxchg(&iter
->position
, pos
, NULL
);
1165 css
= css_next_descendant_pre(css
, &root
->css
);
1168 * Reclaimers share the hierarchy walk, and a
1169 * new one might jump in right at the end of
1170 * the hierarchy - make sure they see at least
1171 * one group and restart from the beginning.
1179 * Verify the css and acquire a reference. The root
1180 * is provided by the caller, so we know it's alive
1181 * and kicking, and don't take an extra reference.
1183 memcg
= mem_cgroup_from_css(css
);
1185 if (css
== &root
->css
)
1188 if (css_tryget(css
))
1196 * The position could have already been updated by a competing
1197 * thread, so check that the value hasn't changed since we read
1198 * it to avoid reclaiming from the same cgroup twice.
1200 (void)cmpxchg(&iter
->position
, pos
, memcg
);
1208 reclaim
->generation
= iter
->generation
;
1213 if (prev
&& prev
!= root
)
1214 css_put(&prev
->css
);
1220 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1221 * @root: hierarchy root
1222 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1224 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1225 struct mem_cgroup
*prev
)
1228 root
= root_mem_cgroup
;
1229 if (prev
&& prev
!= root
)
1230 css_put(&prev
->css
);
1233 static void __invalidate_reclaim_iterators(struct mem_cgroup
*from
,
1234 struct mem_cgroup
*dead_memcg
)
1236 struct mem_cgroup_reclaim_iter
*iter
;
1237 struct mem_cgroup_per_node
*mz
;
1240 for_each_node(nid
) {
1241 mz
= mem_cgroup_nodeinfo(from
, nid
);
1243 cmpxchg(&iter
->position
, dead_memcg
, NULL
);
1247 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
1249 struct mem_cgroup
*memcg
= dead_memcg
;
1250 struct mem_cgroup
*last
;
1253 __invalidate_reclaim_iterators(memcg
, dead_memcg
);
1255 } while ((memcg
= parent_mem_cgroup(memcg
)));
1258 * When cgruop1 non-hierarchy mode is used,
1259 * parent_mem_cgroup() does not walk all the way up to the
1260 * cgroup root (root_mem_cgroup). So we have to handle
1261 * dead_memcg from cgroup root separately.
1263 if (last
!= root_mem_cgroup
)
1264 __invalidate_reclaim_iterators(root_mem_cgroup
,
1269 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1270 * @memcg: hierarchy root
1271 * @fn: function to call for each task
1272 * @arg: argument passed to @fn
1274 * This function iterates over tasks attached to @memcg or to any of its
1275 * descendants and calls @fn for each task. If @fn returns a non-zero
1276 * value, the function breaks the iteration loop and returns the value.
1277 * Otherwise, it will iterate over all tasks and return 0.
1279 * This function must not be called for the root memory cgroup.
1281 int mem_cgroup_scan_tasks(struct mem_cgroup
*memcg
,
1282 int (*fn
)(struct task_struct
*, void *), void *arg
)
1284 struct mem_cgroup
*iter
;
1287 BUG_ON(memcg
== root_mem_cgroup
);
1289 for_each_mem_cgroup_tree(iter
, memcg
) {
1290 struct css_task_iter it
;
1291 struct task_struct
*task
;
1293 css_task_iter_start(&iter
->css
, CSS_TASK_ITER_PROCS
, &it
);
1294 while (!ret
&& (task
= css_task_iter_next(&it
)))
1295 ret
= fn(task
, arg
);
1296 css_task_iter_end(&it
);
1298 mem_cgroup_iter_break(memcg
, iter
);
1305 #ifdef CONFIG_DEBUG_VM
1306 void lruvec_memcg_debug(struct lruvec
*lruvec
, struct page
*page
)
1308 struct mem_cgroup
*memcg
;
1310 if (mem_cgroup_disabled())
1313 memcg
= page_memcg(page
);
1316 VM_BUG_ON_PAGE(lruvec_memcg(lruvec
) != root_mem_cgroup
, page
);
1318 VM_BUG_ON_PAGE(lruvec_memcg(lruvec
) != memcg
, page
);
1323 * lock_page_lruvec - lock and return lruvec for a given page.
1326 * These functions are safe to use under any of the following conditions:
1329 * - lock_page_memcg()
1330 * - page->_refcount is zero
1332 struct lruvec
*lock_page_lruvec(struct page
*page
)
1334 struct lruvec
*lruvec
;
1335 struct pglist_data
*pgdat
= page_pgdat(page
);
1337 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1338 spin_lock(&lruvec
->lru_lock
);
1340 lruvec_memcg_debug(lruvec
, page
);
1345 struct lruvec
*lock_page_lruvec_irq(struct page
*page
)
1347 struct lruvec
*lruvec
;
1348 struct pglist_data
*pgdat
= page_pgdat(page
);
1350 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1351 spin_lock_irq(&lruvec
->lru_lock
);
1353 lruvec_memcg_debug(lruvec
, page
);
1358 struct lruvec
*lock_page_lruvec_irqsave(struct page
*page
, unsigned long *flags
)
1360 struct lruvec
*lruvec
;
1361 struct pglist_data
*pgdat
= page_pgdat(page
);
1363 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1364 spin_lock_irqsave(&lruvec
->lru_lock
, *flags
);
1366 lruvec_memcg_debug(lruvec
, page
);
1372 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1373 * @lruvec: mem_cgroup per zone lru vector
1374 * @lru: index of lru list the page is sitting on
1375 * @zid: zone id of the accounted pages
1376 * @nr_pages: positive when adding or negative when removing
1378 * This function must be called under lru_lock, just before a page is added
1379 * to or just after a page is removed from an lru list (that ordering being
1380 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1382 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1383 int zid
, int nr_pages
)
1385 struct mem_cgroup_per_node
*mz
;
1386 unsigned long *lru_size
;
1389 if (mem_cgroup_disabled())
1392 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
1393 lru_size
= &mz
->lru_zone_size
[zid
][lru
];
1396 *lru_size
+= nr_pages
;
1399 if (WARN_ONCE(size
< 0,
1400 "%s(%p, %d, %d): lru_size %ld\n",
1401 __func__
, lruvec
, lru
, nr_pages
, size
)) {
1407 *lru_size
+= nr_pages
;
1411 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1412 * @memcg: the memory cgroup
1414 * Returns the maximum amount of memory @mem can be charged with, in
1417 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1419 unsigned long margin
= 0;
1420 unsigned long count
;
1421 unsigned long limit
;
1423 count
= page_counter_read(&memcg
->memory
);
1424 limit
= READ_ONCE(memcg
->memory
.max
);
1426 margin
= limit
- count
;
1428 if (do_memsw_account()) {
1429 count
= page_counter_read(&memcg
->memsw
);
1430 limit
= READ_ONCE(memcg
->memsw
.max
);
1432 margin
= min(margin
, limit
- count
);
1441 * A routine for checking "mem" is under move_account() or not.
1443 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1444 * moving cgroups. This is for waiting at high-memory pressure
1447 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1449 struct mem_cgroup
*from
;
1450 struct mem_cgroup
*to
;
1453 * Unlike task_move routines, we access mc.to, mc.from not under
1454 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1456 spin_lock(&mc
.lock
);
1462 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1463 mem_cgroup_is_descendant(to
, memcg
);
1465 spin_unlock(&mc
.lock
);
1469 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1471 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1472 if (mem_cgroup_under_move(memcg
)) {
1474 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1475 /* moving charge context might have finished. */
1478 finish_wait(&mc
.waitq
, &wait
);
1485 struct memory_stat
{
1490 static const struct memory_stat memory_stats
[] = {
1491 { "anon", NR_ANON_MAPPED
},
1492 { "file", NR_FILE_PAGES
},
1493 { "kernel_stack", NR_KERNEL_STACK_KB
},
1494 { "pagetables", NR_PAGETABLE
},
1495 { "percpu", MEMCG_PERCPU_B
},
1496 { "sock", MEMCG_SOCK
},
1497 { "shmem", NR_SHMEM
},
1498 { "file_mapped", NR_FILE_MAPPED
},
1499 { "file_dirty", NR_FILE_DIRTY
},
1500 { "file_writeback", NR_WRITEBACK
},
1502 { "swapcached", NR_SWAPCACHE
},
1504 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1505 { "anon_thp", NR_ANON_THPS
},
1506 { "file_thp", NR_FILE_THPS
},
1507 { "shmem_thp", NR_SHMEM_THPS
},
1509 { "inactive_anon", NR_INACTIVE_ANON
},
1510 { "active_anon", NR_ACTIVE_ANON
},
1511 { "inactive_file", NR_INACTIVE_FILE
},
1512 { "active_file", NR_ACTIVE_FILE
},
1513 { "unevictable", NR_UNEVICTABLE
},
1514 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B
},
1515 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B
},
1517 /* The memory events */
1518 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON
},
1519 { "workingset_refault_file", WORKINGSET_REFAULT_FILE
},
1520 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON
},
1521 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE
},
1522 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON
},
1523 { "workingset_restore_file", WORKINGSET_RESTORE_FILE
},
1524 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM
},
1527 /* Translate stat items to the correct unit for memory.stat output */
1528 static int memcg_page_state_unit(int item
)
1531 case MEMCG_PERCPU_B
:
1532 case NR_SLAB_RECLAIMABLE_B
:
1533 case NR_SLAB_UNRECLAIMABLE_B
:
1534 case WORKINGSET_REFAULT_ANON
:
1535 case WORKINGSET_REFAULT_FILE
:
1536 case WORKINGSET_ACTIVATE_ANON
:
1537 case WORKINGSET_ACTIVATE_FILE
:
1538 case WORKINGSET_RESTORE_ANON
:
1539 case WORKINGSET_RESTORE_FILE
:
1540 case WORKINGSET_NODERECLAIM
:
1542 case NR_KERNEL_STACK_KB
:
1549 static inline unsigned long memcg_page_state_output(struct mem_cgroup
*memcg
,
1552 return memcg_page_state(memcg
, item
) * memcg_page_state_unit(item
);
1555 static char *memory_stat_format(struct mem_cgroup
*memcg
)
1560 seq_buf_init(&s
, kmalloc(PAGE_SIZE
, GFP_KERNEL
), PAGE_SIZE
);
1565 * Provide statistics on the state of the memory subsystem as
1566 * well as cumulative event counters that show past behavior.
1568 * This list is ordered following a combination of these gradients:
1569 * 1) generic big picture -> specifics and details
1570 * 2) reflecting userspace activity -> reflecting kernel heuristics
1572 * Current memory state:
1575 for (i
= 0; i
< ARRAY_SIZE(memory_stats
); i
++) {
1578 size
= memcg_page_state_output(memcg
, memory_stats
[i
].idx
);
1579 seq_buf_printf(&s
, "%s %llu\n", memory_stats
[i
].name
, size
);
1581 if (unlikely(memory_stats
[i
].idx
== NR_SLAB_UNRECLAIMABLE_B
)) {
1582 size
+= memcg_page_state_output(memcg
,
1583 NR_SLAB_RECLAIMABLE_B
);
1584 seq_buf_printf(&s
, "slab %llu\n", size
);
1588 /* Accumulated memory events */
1590 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGFAULT
),
1591 memcg_events(memcg
, PGFAULT
));
1592 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGMAJFAULT
),
1593 memcg_events(memcg
, PGMAJFAULT
));
1594 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGREFILL
),
1595 memcg_events(memcg
, PGREFILL
));
1596 seq_buf_printf(&s
, "pgscan %lu\n",
1597 memcg_events(memcg
, PGSCAN_KSWAPD
) +
1598 memcg_events(memcg
, PGSCAN_DIRECT
));
1599 seq_buf_printf(&s
, "pgsteal %lu\n",
1600 memcg_events(memcg
, PGSTEAL_KSWAPD
) +
1601 memcg_events(memcg
, PGSTEAL_DIRECT
));
1602 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGACTIVATE
),
1603 memcg_events(memcg
, PGACTIVATE
));
1604 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGDEACTIVATE
),
1605 memcg_events(memcg
, PGDEACTIVATE
));
1606 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREE
),
1607 memcg_events(memcg
, PGLAZYFREE
));
1608 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREED
),
1609 memcg_events(memcg
, PGLAZYFREED
));
1611 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1612 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC
),
1613 memcg_events(memcg
, THP_FAULT_ALLOC
));
1614 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC
),
1615 memcg_events(memcg
, THP_COLLAPSE_ALLOC
));
1616 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1618 /* The above should easily fit into one page */
1619 WARN_ON_ONCE(seq_buf_has_overflowed(&s
));
1624 #define K(x) ((x) << (PAGE_SHIFT-10))
1626 * mem_cgroup_print_oom_context: Print OOM information relevant to
1627 * memory controller.
1628 * @memcg: The memory cgroup that went over limit
1629 * @p: Task that is going to be killed
1631 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1634 void mem_cgroup_print_oom_context(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1639 pr_cont(",oom_memcg=");
1640 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1642 pr_cont(",global_oom");
1644 pr_cont(",task_memcg=");
1645 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1651 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1652 * memory controller.
1653 * @memcg: The memory cgroup that went over limit
1655 void mem_cgroup_print_oom_meminfo(struct mem_cgroup
*memcg
)
1659 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1660 K((u64
)page_counter_read(&memcg
->memory
)),
1661 K((u64
)READ_ONCE(memcg
->memory
.max
)), memcg
->memory
.failcnt
);
1662 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1663 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1664 K((u64
)page_counter_read(&memcg
->swap
)),
1665 K((u64
)READ_ONCE(memcg
->swap
.max
)), memcg
->swap
.failcnt
);
1667 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1668 K((u64
)page_counter_read(&memcg
->memsw
)),
1669 K((u64
)memcg
->memsw
.max
), memcg
->memsw
.failcnt
);
1670 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1671 K((u64
)page_counter_read(&memcg
->kmem
)),
1672 K((u64
)memcg
->kmem
.max
), memcg
->kmem
.failcnt
);
1675 pr_info("Memory cgroup stats for ");
1676 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1678 buf
= memory_stat_format(memcg
);
1686 * Return the memory (and swap, if configured) limit for a memcg.
1688 unsigned long mem_cgroup_get_max(struct mem_cgroup
*memcg
)
1690 unsigned long max
= READ_ONCE(memcg
->memory
.max
);
1692 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
1693 if (mem_cgroup_swappiness(memcg
))
1694 max
+= min(READ_ONCE(memcg
->swap
.max
),
1695 (unsigned long)total_swap_pages
);
1697 if (mem_cgroup_swappiness(memcg
)) {
1698 /* Calculate swap excess capacity from memsw limit */
1699 unsigned long swap
= READ_ONCE(memcg
->memsw
.max
) - max
;
1701 max
+= min(swap
, (unsigned long)total_swap_pages
);
1707 unsigned long mem_cgroup_size(struct mem_cgroup
*memcg
)
1709 return page_counter_read(&memcg
->memory
);
1712 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1715 struct oom_control oc
= {
1719 .gfp_mask
= gfp_mask
,
1724 if (mutex_lock_killable(&oom_lock
))
1727 if (mem_cgroup_margin(memcg
) >= (1 << order
))
1731 * A few threads which were not waiting at mutex_lock_killable() can
1732 * fail to bail out. Therefore, check again after holding oom_lock.
1734 ret
= should_force_charge() || out_of_memory(&oc
);
1737 mutex_unlock(&oom_lock
);
1741 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1744 unsigned long *total_scanned
)
1746 struct mem_cgroup
*victim
= NULL
;
1749 unsigned long excess
;
1750 unsigned long nr_scanned
;
1751 struct mem_cgroup_reclaim_cookie reclaim
= {
1755 excess
= soft_limit_excess(root_memcg
);
1758 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1763 * If we have not been able to reclaim
1764 * anything, it might because there are
1765 * no reclaimable pages under this hierarchy
1770 * We want to do more targeted reclaim.
1771 * excess >> 2 is not to excessive so as to
1772 * reclaim too much, nor too less that we keep
1773 * coming back to reclaim from this cgroup
1775 if (total
>= (excess
>> 2) ||
1776 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1781 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1782 pgdat
, &nr_scanned
);
1783 *total_scanned
+= nr_scanned
;
1784 if (!soft_limit_excess(root_memcg
))
1787 mem_cgroup_iter_break(root_memcg
, victim
);
1791 #ifdef CONFIG_LOCKDEP
1792 static struct lockdep_map memcg_oom_lock_dep_map
= {
1793 .name
= "memcg_oom_lock",
1797 static DEFINE_SPINLOCK(memcg_oom_lock
);
1800 * Check OOM-Killer is already running under our hierarchy.
1801 * If someone is running, return false.
1803 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1805 struct mem_cgroup
*iter
, *failed
= NULL
;
1807 spin_lock(&memcg_oom_lock
);
1809 for_each_mem_cgroup_tree(iter
, memcg
) {
1810 if (iter
->oom_lock
) {
1812 * this subtree of our hierarchy is already locked
1813 * so we cannot give a lock.
1816 mem_cgroup_iter_break(memcg
, iter
);
1819 iter
->oom_lock
= true;
1824 * OK, we failed to lock the whole subtree so we have
1825 * to clean up what we set up to the failing subtree
1827 for_each_mem_cgroup_tree(iter
, memcg
) {
1828 if (iter
== failed
) {
1829 mem_cgroup_iter_break(memcg
, iter
);
1832 iter
->oom_lock
= false;
1835 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1837 spin_unlock(&memcg_oom_lock
);
1842 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1844 struct mem_cgroup
*iter
;
1846 spin_lock(&memcg_oom_lock
);
1847 mutex_release(&memcg_oom_lock_dep_map
, _RET_IP_
);
1848 for_each_mem_cgroup_tree(iter
, memcg
)
1849 iter
->oom_lock
= false;
1850 spin_unlock(&memcg_oom_lock
);
1853 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1855 struct mem_cgroup
*iter
;
1857 spin_lock(&memcg_oom_lock
);
1858 for_each_mem_cgroup_tree(iter
, memcg
)
1860 spin_unlock(&memcg_oom_lock
);
1863 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1865 struct mem_cgroup
*iter
;
1868 * Be careful about under_oom underflows becase a child memcg
1869 * could have been added after mem_cgroup_mark_under_oom.
1871 spin_lock(&memcg_oom_lock
);
1872 for_each_mem_cgroup_tree(iter
, memcg
)
1873 if (iter
->under_oom
> 0)
1875 spin_unlock(&memcg_oom_lock
);
1878 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1880 struct oom_wait_info
{
1881 struct mem_cgroup
*memcg
;
1882 wait_queue_entry_t wait
;
1885 static int memcg_oom_wake_function(wait_queue_entry_t
*wait
,
1886 unsigned mode
, int sync
, void *arg
)
1888 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1889 struct mem_cgroup
*oom_wait_memcg
;
1890 struct oom_wait_info
*oom_wait_info
;
1892 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1893 oom_wait_memcg
= oom_wait_info
->memcg
;
1895 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1896 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1898 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1901 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1904 * For the following lockless ->under_oom test, the only required
1905 * guarantee is that it must see the state asserted by an OOM when
1906 * this function is called as a result of userland actions
1907 * triggered by the notification of the OOM. This is trivially
1908 * achieved by invoking mem_cgroup_mark_under_oom() before
1909 * triggering notification.
1911 if (memcg
&& memcg
->under_oom
)
1912 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1922 static enum oom_status
mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1924 enum oom_status ret
;
1927 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1930 memcg_memory_event(memcg
, MEMCG_OOM
);
1933 * We are in the middle of the charge context here, so we
1934 * don't want to block when potentially sitting on a callstack
1935 * that holds all kinds of filesystem and mm locks.
1937 * cgroup1 allows disabling the OOM killer and waiting for outside
1938 * handling until the charge can succeed; remember the context and put
1939 * the task to sleep at the end of the page fault when all locks are
1942 * On the other hand, in-kernel OOM killer allows for an async victim
1943 * memory reclaim (oom_reaper) and that means that we are not solely
1944 * relying on the oom victim to make a forward progress and we can
1945 * invoke the oom killer here.
1947 * Please note that mem_cgroup_out_of_memory might fail to find a
1948 * victim and then we have to bail out from the charge path.
1950 if (memcg
->oom_kill_disable
) {
1951 if (!current
->in_user_fault
)
1953 css_get(&memcg
->css
);
1954 current
->memcg_in_oom
= memcg
;
1955 current
->memcg_oom_gfp_mask
= mask
;
1956 current
->memcg_oom_order
= order
;
1961 mem_cgroup_mark_under_oom(memcg
);
1963 locked
= mem_cgroup_oom_trylock(memcg
);
1966 mem_cgroup_oom_notify(memcg
);
1968 mem_cgroup_unmark_under_oom(memcg
);
1969 if (mem_cgroup_out_of_memory(memcg
, mask
, order
))
1975 mem_cgroup_oom_unlock(memcg
);
1981 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1982 * @handle: actually kill/wait or just clean up the OOM state
1984 * This has to be called at the end of a page fault if the memcg OOM
1985 * handler was enabled.
1987 * Memcg supports userspace OOM handling where failed allocations must
1988 * sleep on a waitqueue until the userspace task resolves the
1989 * situation. Sleeping directly in the charge context with all kinds
1990 * of locks held is not a good idea, instead we remember an OOM state
1991 * in the task and mem_cgroup_oom_synchronize() has to be called at
1992 * the end of the page fault to complete the OOM handling.
1994 * Returns %true if an ongoing memcg OOM situation was detected and
1995 * completed, %false otherwise.
1997 bool mem_cgroup_oom_synchronize(bool handle
)
1999 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
2000 struct oom_wait_info owait
;
2003 /* OOM is global, do not handle */
2010 owait
.memcg
= memcg
;
2011 owait
.wait
.flags
= 0;
2012 owait
.wait
.func
= memcg_oom_wake_function
;
2013 owait
.wait
.private = current
;
2014 INIT_LIST_HEAD(&owait
.wait
.entry
);
2016 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2017 mem_cgroup_mark_under_oom(memcg
);
2019 locked
= mem_cgroup_oom_trylock(memcg
);
2022 mem_cgroup_oom_notify(memcg
);
2024 if (locked
&& !memcg
->oom_kill_disable
) {
2025 mem_cgroup_unmark_under_oom(memcg
);
2026 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2027 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
2028 current
->memcg_oom_order
);
2031 mem_cgroup_unmark_under_oom(memcg
);
2032 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2036 mem_cgroup_oom_unlock(memcg
);
2038 * There is no guarantee that an OOM-lock contender
2039 * sees the wakeups triggered by the OOM kill
2040 * uncharges. Wake any sleepers explicitely.
2042 memcg_oom_recover(memcg
);
2045 current
->memcg_in_oom
= NULL
;
2046 css_put(&memcg
->css
);
2051 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2052 * @victim: task to be killed by the OOM killer
2053 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2055 * Returns a pointer to a memory cgroup, which has to be cleaned up
2056 * by killing all belonging OOM-killable tasks.
2058 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2060 struct mem_cgroup
*mem_cgroup_get_oom_group(struct task_struct
*victim
,
2061 struct mem_cgroup
*oom_domain
)
2063 struct mem_cgroup
*oom_group
= NULL
;
2064 struct mem_cgroup
*memcg
;
2066 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2070 oom_domain
= root_mem_cgroup
;
2074 memcg
= mem_cgroup_from_task(victim
);
2075 if (memcg
== root_mem_cgroup
)
2079 * If the victim task has been asynchronously moved to a different
2080 * memory cgroup, we might end up killing tasks outside oom_domain.
2081 * In this case it's better to ignore memory.group.oom.
2083 if (unlikely(!mem_cgroup_is_descendant(memcg
, oom_domain
)))
2087 * Traverse the memory cgroup hierarchy from the victim task's
2088 * cgroup up to the OOMing cgroup (or root) to find the
2089 * highest-level memory cgroup with oom.group set.
2091 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
2092 if (memcg
->oom_group
)
2095 if (memcg
== oom_domain
)
2100 css_get(&oom_group
->css
);
2107 void mem_cgroup_print_oom_group(struct mem_cgroup
*memcg
)
2109 pr_info("Tasks in ");
2110 pr_cont_cgroup_path(memcg
->css
.cgroup
);
2111 pr_cont(" are going to be killed due to memory.oom.group set\n");
2115 * lock_page_memcg - lock a page and memcg binding
2118 * This function protects unlocked LRU pages from being moved to
2121 * It ensures lifetime of the returned memcg. Caller is responsible
2122 * for the lifetime of the page; __unlock_page_memcg() is available
2123 * when @page might get freed inside the locked section.
2125 struct mem_cgroup
*lock_page_memcg(struct page
*page
)
2127 struct page
*head
= compound_head(page
); /* rmap on tail pages */
2128 struct mem_cgroup
*memcg
;
2129 unsigned long flags
;
2132 * The RCU lock is held throughout the transaction. The fast
2133 * path can get away without acquiring the memcg->move_lock
2134 * because page moving starts with an RCU grace period.
2136 * The RCU lock also protects the memcg from being freed when
2137 * the page state that is going to change is the only thing
2138 * preventing the page itself from being freed. E.g. writeback
2139 * doesn't hold a page reference and relies on PG_writeback to
2140 * keep off truncation, migration and so forth.
2144 if (mem_cgroup_disabled())
2147 memcg
= page_memcg(head
);
2148 if (unlikely(!memcg
))
2151 #ifdef CONFIG_PROVE_LOCKING
2152 local_irq_save(flags
);
2153 might_lock(&memcg
->move_lock
);
2154 local_irq_restore(flags
);
2157 if (atomic_read(&memcg
->moving_account
) <= 0)
2160 spin_lock_irqsave(&memcg
->move_lock
, flags
);
2161 if (memcg
!= page_memcg(head
)) {
2162 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2167 * When charge migration first begins, we can have locked and
2168 * unlocked page stat updates happening concurrently. Track
2169 * the task who has the lock for unlock_page_memcg().
2171 memcg
->move_lock_task
= current
;
2172 memcg
->move_lock_flags
= flags
;
2176 EXPORT_SYMBOL(lock_page_memcg
);
2179 * __unlock_page_memcg - unlock and unpin a memcg
2182 * Unlock and unpin a memcg returned by lock_page_memcg().
2184 void __unlock_page_memcg(struct mem_cgroup
*memcg
)
2186 if (memcg
&& memcg
->move_lock_task
== current
) {
2187 unsigned long flags
= memcg
->move_lock_flags
;
2189 memcg
->move_lock_task
= NULL
;
2190 memcg
->move_lock_flags
= 0;
2192 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2199 * unlock_page_memcg - unlock a page and memcg binding
2202 void unlock_page_memcg(struct page
*page
)
2204 struct page
*head
= compound_head(page
);
2206 __unlock_page_memcg(page_memcg(head
));
2208 EXPORT_SYMBOL(unlock_page_memcg
);
2210 struct memcg_stock_pcp
{
2211 struct mem_cgroup
*cached
; /* this never be root cgroup */
2212 unsigned int nr_pages
;
2214 #ifdef CONFIG_MEMCG_KMEM
2215 struct obj_cgroup
*cached_objcg
;
2216 unsigned int nr_bytes
;
2219 struct work_struct work
;
2220 unsigned long flags
;
2221 #define FLUSHING_CACHED_CHARGE 0
2223 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2224 static DEFINE_MUTEX(percpu_charge_mutex
);
2226 #ifdef CONFIG_MEMCG_KMEM
2227 static void drain_obj_stock(struct memcg_stock_pcp
*stock
);
2228 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
2229 struct mem_cgroup
*root_memcg
);
2232 static inline void drain_obj_stock(struct memcg_stock_pcp
*stock
)
2235 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
2236 struct mem_cgroup
*root_memcg
)
2243 * consume_stock: Try to consume stocked charge on this cpu.
2244 * @memcg: memcg to consume from.
2245 * @nr_pages: how many pages to charge.
2247 * The charges will only happen if @memcg matches the current cpu's memcg
2248 * stock, and at least @nr_pages are available in that stock. Failure to
2249 * service an allocation will refill the stock.
2251 * returns true if successful, false otherwise.
2253 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2255 struct memcg_stock_pcp
*stock
;
2256 unsigned long flags
;
2259 if (nr_pages
> MEMCG_CHARGE_BATCH
)
2262 local_irq_save(flags
);
2264 stock
= this_cpu_ptr(&memcg_stock
);
2265 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
2266 stock
->nr_pages
-= nr_pages
;
2270 local_irq_restore(flags
);
2276 * Returns stocks cached in percpu and reset cached information.
2278 static void drain_stock(struct memcg_stock_pcp
*stock
)
2280 struct mem_cgroup
*old
= stock
->cached
;
2285 if (stock
->nr_pages
) {
2286 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2287 if (do_memsw_account())
2288 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2289 stock
->nr_pages
= 0;
2293 stock
->cached
= NULL
;
2296 static void drain_local_stock(struct work_struct
*dummy
)
2298 struct memcg_stock_pcp
*stock
;
2299 unsigned long flags
;
2302 * The only protection from memory hotplug vs. drain_stock races is
2303 * that we always operate on local CPU stock here with IRQ disabled
2305 local_irq_save(flags
);
2307 stock
= this_cpu_ptr(&memcg_stock
);
2308 drain_obj_stock(stock
);
2310 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2312 local_irq_restore(flags
);
2316 * Cache charges(val) to local per_cpu area.
2317 * This will be consumed by consume_stock() function, later.
2319 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2321 struct memcg_stock_pcp
*stock
;
2322 unsigned long flags
;
2324 local_irq_save(flags
);
2326 stock
= this_cpu_ptr(&memcg_stock
);
2327 if (stock
->cached
!= memcg
) { /* reset if necessary */
2329 css_get(&memcg
->css
);
2330 stock
->cached
= memcg
;
2332 stock
->nr_pages
+= nr_pages
;
2334 if (stock
->nr_pages
> MEMCG_CHARGE_BATCH
)
2337 local_irq_restore(flags
);
2341 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2342 * of the hierarchy under it.
2344 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2348 /* If someone's already draining, avoid adding running more workers. */
2349 if (!mutex_trylock(&percpu_charge_mutex
))
2352 * Notify other cpus that system-wide "drain" is running
2353 * We do not care about races with the cpu hotplug because cpu down
2354 * as well as workers from this path always operate on the local
2355 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2358 for_each_online_cpu(cpu
) {
2359 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2360 struct mem_cgroup
*memcg
;
2364 memcg
= stock
->cached
;
2365 if (memcg
&& stock
->nr_pages
&&
2366 mem_cgroup_is_descendant(memcg
, root_memcg
))
2368 if (obj_stock_flush_required(stock
, root_memcg
))
2373 !test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2375 drain_local_stock(&stock
->work
);
2377 schedule_work_on(cpu
, &stock
->work
);
2381 mutex_unlock(&percpu_charge_mutex
);
2384 static int memcg_hotplug_cpu_dead(unsigned int cpu
)
2386 struct memcg_stock_pcp
*stock
;
2387 struct mem_cgroup
*memcg
, *mi
;
2389 stock
= &per_cpu(memcg_stock
, cpu
);
2392 for_each_mem_cgroup(memcg
) {
2395 for (i
= 0; i
< MEMCG_NR_STAT
; i
++) {
2399 x
= this_cpu_xchg(memcg
->vmstats_percpu
->stat
[i
], 0);
2401 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2402 atomic_long_add(x
, &memcg
->vmstats
[i
]);
2404 if (i
>= NR_VM_NODE_STAT_ITEMS
)
2407 for_each_node(nid
) {
2408 struct mem_cgroup_per_node
*pn
;
2410 pn
= mem_cgroup_nodeinfo(memcg
, nid
);
2411 x
= this_cpu_xchg(pn
->lruvec_stat_cpu
->count
[i
], 0);
2414 atomic_long_add(x
, &pn
->lruvec_stat
[i
]);
2415 } while ((pn
= parent_nodeinfo(pn
, nid
)));
2419 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++) {
2422 x
= this_cpu_xchg(memcg
->vmstats_percpu
->events
[i
], 0);
2424 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2425 atomic_long_add(x
, &memcg
->vmevents
[i
]);
2432 static unsigned long reclaim_high(struct mem_cgroup
*memcg
,
2433 unsigned int nr_pages
,
2436 unsigned long nr_reclaimed
= 0;
2439 unsigned long pflags
;
2441 if (page_counter_read(&memcg
->memory
) <=
2442 READ_ONCE(memcg
->memory
.high
))
2445 memcg_memory_event(memcg
, MEMCG_HIGH
);
2447 psi_memstall_enter(&pflags
);
2448 nr_reclaimed
+= try_to_free_mem_cgroup_pages(memcg
, nr_pages
,
2450 psi_memstall_leave(&pflags
);
2451 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2452 !mem_cgroup_is_root(memcg
));
2454 return nr_reclaimed
;
2457 static void high_work_func(struct work_struct
*work
)
2459 struct mem_cgroup
*memcg
;
2461 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
2462 reclaim_high(memcg
, MEMCG_CHARGE_BATCH
, GFP_KERNEL
);
2466 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2467 * enough to still cause a significant slowdown in most cases, while still
2468 * allowing diagnostics and tracing to proceed without becoming stuck.
2470 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2473 * When calculating the delay, we use these either side of the exponentiation to
2474 * maintain precision and scale to a reasonable number of jiffies (see the table
2477 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2478 * overage ratio to a delay.
2479 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2480 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2481 * to produce a reasonable delay curve.
2483 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2484 * reasonable delay curve compared to precision-adjusted overage, not
2485 * penalising heavily at first, but still making sure that growth beyond the
2486 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2487 * example, with a high of 100 megabytes:
2489 * +-------+------------------------+
2490 * | usage | time to allocate in ms |
2491 * +-------+------------------------+
2513 * +-------+------------------------+
2515 #define MEMCG_DELAY_PRECISION_SHIFT 20
2516 #define MEMCG_DELAY_SCALING_SHIFT 14
2518 static u64
calculate_overage(unsigned long usage
, unsigned long high
)
2526 * Prevent division by 0 in overage calculation by acting as if
2527 * it was a threshold of 1 page
2529 high
= max(high
, 1UL);
2531 overage
= usage
- high
;
2532 overage
<<= MEMCG_DELAY_PRECISION_SHIFT
;
2533 return div64_u64(overage
, high
);
2536 static u64
mem_find_max_overage(struct mem_cgroup
*memcg
)
2538 u64 overage
, max_overage
= 0;
2541 overage
= calculate_overage(page_counter_read(&memcg
->memory
),
2542 READ_ONCE(memcg
->memory
.high
));
2543 max_overage
= max(overage
, max_overage
);
2544 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2545 !mem_cgroup_is_root(memcg
));
2550 static u64
swap_find_max_overage(struct mem_cgroup
*memcg
)
2552 u64 overage
, max_overage
= 0;
2555 overage
= calculate_overage(page_counter_read(&memcg
->swap
),
2556 READ_ONCE(memcg
->swap
.high
));
2558 memcg_memory_event(memcg
, MEMCG_SWAP_HIGH
);
2559 max_overage
= max(overage
, max_overage
);
2560 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2561 !mem_cgroup_is_root(memcg
));
2567 * Get the number of jiffies that we should penalise a mischievous cgroup which
2568 * is exceeding its memory.high by checking both it and its ancestors.
2570 static unsigned long calculate_high_delay(struct mem_cgroup
*memcg
,
2571 unsigned int nr_pages
,
2574 unsigned long penalty_jiffies
;
2580 * We use overage compared to memory.high to calculate the number of
2581 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2582 * fairly lenient on small overages, and increasingly harsh when the
2583 * memcg in question makes it clear that it has no intention of stopping
2584 * its crazy behaviour, so we exponentially increase the delay based on
2587 penalty_jiffies
= max_overage
* max_overage
* HZ
;
2588 penalty_jiffies
>>= MEMCG_DELAY_PRECISION_SHIFT
;
2589 penalty_jiffies
>>= MEMCG_DELAY_SCALING_SHIFT
;
2592 * Factor in the task's own contribution to the overage, such that four
2593 * N-sized allocations are throttled approximately the same as one
2594 * 4N-sized allocation.
2596 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2597 * larger the current charge patch is than that.
2599 return penalty_jiffies
* nr_pages
/ MEMCG_CHARGE_BATCH
;
2603 * Scheduled by try_charge() to be executed from the userland return path
2604 * and reclaims memory over the high limit.
2606 void mem_cgroup_handle_over_high(void)
2608 unsigned long penalty_jiffies
;
2609 unsigned long pflags
;
2610 unsigned long nr_reclaimed
;
2611 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
2612 int nr_retries
= MAX_RECLAIM_RETRIES
;
2613 struct mem_cgroup
*memcg
;
2614 bool in_retry
= false;
2616 if (likely(!nr_pages
))
2619 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2620 current
->memcg_nr_pages_over_high
= 0;
2624 * The allocating task should reclaim at least the batch size, but for
2625 * subsequent retries we only want to do what's necessary to prevent oom
2626 * or breaching resource isolation.
2628 * This is distinct from memory.max or page allocator behaviour because
2629 * memory.high is currently batched, whereas memory.max and the page
2630 * allocator run every time an allocation is made.
2632 nr_reclaimed
= reclaim_high(memcg
,
2633 in_retry
? SWAP_CLUSTER_MAX
: nr_pages
,
2637 * memory.high is breached and reclaim is unable to keep up. Throttle
2638 * allocators proactively to slow down excessive growth.
2640 penalty_jiffies
= calculate_high_delay(memcg
, nr_pages
,
2641 mem_find_max_overage(memcg
));
2643 penalty_jiffies
+= calculate_high_delay(memcg
, nr_pages
,
2644 swap_find_max_overage(memcg
));
2647 * Clamp the max delay per usermode return so as to still keep the
2648 * application moving forwards and also permit diagnostics, albeit
2651 penalty_jiffies
= min(penalty_jiffies
, MEMCG_MAX_HIGH_DELAY_JIFFIES
);
2654 * Don't sleep if the amount of jiffies this memcg owes us is so low
2655 * that it's not even worth doing, in an attempt to be nice to those who
2656 * go only a small amount over their memory.high value and maybe haven't
2657 * been aggressively reclaimed enough yet.
2659 if (penalty_jiffies
<= HZ
/ 100)
2663 * If reclaim is making forward progress but we're still over
2664 * memory.high, we want to encourage that rather than doing allocator
2667 if (nr_reclaimed
|| nr_retries
--) {
2673 * If we exit early, we're guaranteed to die (since
2674 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2675 * need to account for any ill-begotten jiffies to pay them off later.
2677 psi_memstall_enter(&pflags
);
2678 schedule_timeout_killable(penalty_jiffies
);
2679 psi_memstall_leave(&pflags
);
2682 css_put(&memcg
->css
);
2685 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2686 unsigned int nr_pages
)
2688 unsigned int batch
= max(MEMCG_CHARGE_BATCH
, nr_pages
);
2689 int nr_retries
= MAX_RECLAIM_RETRIES
;
2690 struct mem_cgroup
*mem_over_limit
;
2691 struct page_counter
*counter
;
2692 enum oom_status oom_status
;
2693 unsigned long nr_reclaimed
;
2694 bool may_swap
= true;
2695 bool drained
= false;
2696 unsigned long pflags
;
2698 if (mem_cgroup_is_root(memcg
))
2701 if (consume_stock(memcg
, nr_pages
))
2704 if (!do_memsw_account() ||
2705 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2706 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2708 if (do_memsw_account())
2709 page_counter_uncharge(&memcg
->memsw
, batch
);
2710 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2712 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2716 if (batch
> nr_pages
) {
2722 * Memcg doesn't have a dedicated reserve for atomic
2723 * allocations. But like the global atomic pool, we need to
2724 * put the burden of reclaim on regular allocation requests
2725 * and let these go through as privileged allocations.
2727 if (gfp_mask
& __GFP_ATOMIC
)
2731 * Unlike in global OOM situations, memcg is not in a physical
2732 * memory shortage. Allow dying and OOM-killed tasks to
2733 * bypass the last charges so that they can exit quickly and
2734 * free their memory.
2736 if (unlikely(should_force_charge()))
2740 * Prevent unbounded recursion when reclaim operations need to
2741 * allocate memory. This might exceed the limits temporarily,
2742 * but we prefer facilitating memory reclaim and getting back
2743 * under the limit over triggering OOM kills in these cases.
2745 if (unlikely(current
->flags
& PF_MEMALLOC
))
2748 if (unlikely(task_in_memcg_oom(current
)))
2751 if (!gfpflags_allow_blocking(gfp_mask
))
2754 memcg_memory_event(mem_over_limit
, MEMCG_MAX
);
2756 psi_memstall_enter(&pflags
);
2757 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2758 gfp_mask
, may_swap
);
2759 psi_memstall_leave(&pflags
);
2761 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2765 drain_all_stock(mem_over_limit
);
2770 if (gfp_mask
& __GFP_NORETRY
)
2773 * Even though the limit is exceeded at this point, reclaim
2774 * may have been able to free some pages. Retry the charge
2775 * before killing the task.
2777 * Only for regular pages, though: huge pages are rather
2778 * unlikely to succeed so close to the limit, and we fall back
2779 * to regular pages anyway in case of failure.
2781 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2784 * At task move, charge accounts can be doubly counted. So, it's
2785 * better to wait until the end of task_move if something is going on.
2787 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2793 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
2796 if (gfp_mask
& __GFP_NOFAIL
)
2799 if (fatal_signal_pending(current
))
2803 * keep retrying as long as the memcg oom killer is able to make
2804 * a forward progress or bypass the charge if the oom killer
2805 * couldn't make any progress.
2807 oom_status
= mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2808 get_order(nr_pages
* PAGE_SIZE
));
2809 switch (oom_status
) {
2811 nr_retries
= MAX_RECLAIM_RETRIES
;
2819 if (!(gfp_mask
& __GFP_NOFAIL
))
2823 * The allocation either can't fail or will lead to more memory
2824 * being freed very soon. Allow memory usage go over the limit
2825 * temporarily by force charging it.
2827 page_counter_charge(&memcg
->memory
, nr_pages
);
2828 if (do_memsw_account())
2829 page_counter_charge(&memcg
->memsw
, nr_pages
);
2834 if (batch
> nr_pages
)
2835 refill_stock(memcg
, batch
- nr_pages
);
2838 * If the hierarchy is above the normal consumption range, schedule
2839 * reclaim on returning to userland. We can perform reclaim here
2840 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2841 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2842 * not recorded as it most likely matches current's and won't
2843 * change in the meantime. As high limit is checked again before
2844 * reclaim, the cost of mismatch is negligible.
2847 bool mem_high
, swap_high
;
2849 mem_high
= page_counter_read(&memcg
->memory
) >
2850 READ_ONCE(memcg
->memory
.high
);
2851 swap_high
= page_counter_read(&memcg
->swap
) >
2852 READ_ONCE(memcg
->swap
.high
);
2854 /* Don't bother a random interrupted task */
2855 if (in_interrupt()) {
2857 schedule_work(&memcg
->high_work
);
2863 if (mem_high
|| swap_high
) {
2865 * The allocating tasks in this cgroup will need to do
2866 * reclaim or be throttled to prevent further growth
2867 * of the memory or swap footprints.
2869 * Target some best-effort fairness between the tasks,
2870 * and distribute reclaim work and delay penalties
2871 * based on how much each task is actually allocating.
2873 current
->memcg_nr_pages_over_high
+= batch
;
2874 set_notify_resume(current
);
2877 } while ((memcg
= parent_mem_cgroup(memcg
)));
2882 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2883 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2885 if (mem_cgroup_is_root(memcg
))
2888 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2889 if (do_memsw_account())
2890 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2894 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
)
2896 VM_BUG_ON_PAGE(page_memcg(page
), page
);
2898 * Any of the following ensures page's memcg stability:
2902 * - lock_page_memcg()
2903 * - exclusive reference
2905 page
->memcg_data
= (unsigned long)memcg
;
2908 #ifdef CONFIG_MEMCG_KMEM
2909 int memcg_alloc_page_obj_cgroups(struct page
*page
, struct kmem_cache
*s
,
2910 gfp_t gfp
, bool new_page
)
2912 unsigned int objects
= objs_per_slab_page(s
, page
);
2913 unsigned long memcg_data
;
2916 vec
= kcalloc_node(objects
, sizeof(struct obj_cgroup
*), gfp
,
2921 memcg_data
= (unsigned long) vec
| MEMCG_DATA_OBJCGS
;
2924 * If the slab page is brand new and nobody can yet access
2925 * it's memcg_data, no synchronization is required and
2926 * memcg_data can be simply assigned.
2928 page
->memcg_data
= memcg_data
;
2929 } else if (cmpxchg(&page
->memcg_data
, 0, memcg_data
)) {
2931 * If the slab page is already in use, somebody can allocate
2932 * and assign obj_cgroups in parallel. In this case the existing
2933 * objcg vector should be reused.
2939 kmemleak_not_leak(vec
);
2944 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2946 * A passed kernel object can be a slab object or a generic kernel page, so
2947 * different mechanisms for getting the memory cgroup pointer should be used.
2948 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2949 * can not know for sure how the kernel object is implemented.
2950 * mem_cgroup_from_obj() can be safely used in such cases.
2952 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2953 * cgroup_mutex, etc.
2955 struct mem_cgroup
*mem_cgroup_from_obj(void *p
)
2959 if (mem_cgroup_disabled())
2962 page
= virt_to_head_page(p
);
2965 * Slab objects are accounted individually, not per-page.
2966 * Memcg membership data for each individual object is saved in
2967 * the page->obj_cgroups.
2969 if (page_objcgs_check(page
)) {
2970 struct obj_cgroup
*objcg
;
2973 off
= obj_to_index(page
->slab_cache
, page
, p
);
2974 objcg
= page_objcgs(page
)[off
];
2976 return obj_cgroup_memcg(objcg
);
2982 * page_memcg_check() is used here, because page_has_obj_cgroups()
2983 * check above could fail because the object cgroups vector wasn't set
2984 * at that moment, but it can be set concurrently.
2985 * page_memcg_check(page) will guarantee that a proper memory
2986 * cgroup pointer or NULL will be returned.
2988 return page_memcg_check(page
);
2991 __always_inline
struct obj_cgroup
*get_obj_cgroup_from_current(void)
2993 struct obj_cgroup
*objcg
= NULL
;
2994 struct mem_cgroup
*memcg
;
2996 if (memcg_kmem_bypass())
3000 if (unlikely(active_memcg()))
3001 memcg
= active_memcg();
3003 memcg
= mem_cgroup_from_task(current
);
3005 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
)) {
3006 objcg
= rcu_dereference(memcg
->objcg
);
3007 if (objcg
&& obj_cgroup_tryget(objcg
))
3016 static int memcg_alloc_cache_id(void)
3021 id
= ida_simple_get(&memcg_cache_ida
,
3022 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
3026 if (id
< memcg_nr_cache_ids
)
3030 * There's no space for the new id in memcg_caches arrays,
3031 * so we have to grow them.
3033 down_write(&memcg_cache_ids_sem
);
3035 size
= 2 * (id
+ 1);
3036 if (size
< MEMCG_CACHES_MIN_SIZE
)
3037 size
= MEMCG_CACHES_MIN_SIZE
;
3038 else if (size
> MEMCG_CACHES_MAX_SIZE
)
3039 size
= MEMCG_CACHES_MAX_SIZE
;
3041 err
= memcg_update_all_list_lrus(size
);
3043 memcg_nr_cache_ids
= size
;
3045 up_write(&memcg_cache_ids_sem
);
3048 ida_simple_remove(&memcg_cache_ida
, id
);
3054 static void memcg_free_cache_id(int id
)
3056 ida_simple_remove(&memcg_cache_ida
, id
);
3060 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3061 * @memcg: memory cgroup to charge
3062 * @gfp: reclaim mode
3063 * @nr_pages: number of pages to charge
3065 * Returns 0 on success, an error code on failure.
3067 static int __memcg_kmem_charge(struct mem_cgroup
*memcg
, gfp_t gfp
,
3068 unsigned int nr_pages
)
3070 struct page_counter
*counter
;
3073 ret
= try_charge(memcg
, gfp
, nr_pages
);
3077 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
3078 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
3081 * Enforce __GFP_NOFAIL allocation because callers are not
3082 * prepared to see failures and likely do not have any failure
3085 if (gfp
& __GFP_NOFAIL
) {
3086 page_counter_charge(&memcg
->kmem
, nr_pages
);
3089 cancel_charge(memcg
, nr_pages
);
3096 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3097 * @memcg: memcg to uncharge
3098 * @nr_pages: number of pages to uncharge
3100 static void __memcg_kmem_uncharge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
3102 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
3103 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
3105 refill_stock(memcg
, nr_pages
);
3109 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3110 * @page: page to charge
3111 * @gfp: reclaim mode
3112 * @order: allocation order
3114 * Returns 0 on success, an error code on failure.
3116 int __memcg_kmem_charge_page(struct page
*page
, gfp_t gfp
, int order
)
3118 struct mem_cgroup
*memcg
;
3121 memcg
= get_mem_cgroup_from_current();
3122 if (memcg
&& !mem_cgroup_is_root(memcg
)) {
3123 ret
= __memcg_kmem_charge(memcg
, gfp
, 1 << order
);
3125 page
->memcg_data
= (unsigned long)memcg
|
3129 css_put(&memcg
->css
);
3135 * __memcg_kmem_uncharge_page: uncharge a kmem page
3136 * @page: page to uncharge
3137 * @order: allocation order
3139 void __memcg_kmem_uncharge_page(struct page
*page
, int order
)
3141 struct mem_cgroup
*memcg
= page_memcg(page
);
3142 unsigned int nr_pages
= 1 << order
;
3147 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
3148 __memcg_kmem_uncharge(memcg
, nr_pages
);
3149 page
->memcg_data
= 0;
3150 css_put(&memcg
->css
);
3153 static bool consume_obj_stock(struct obj_cgroup
*objcg
, unsigned int nr_bytes
)
3155 struct memcg_stock_pcp
*stock
;
3156 unsigned long flags
;
3159 local_irq_save(flags
);
3161 stock
= this_cpu_ptr(&memcg_stock
);
3162 if (objcg
== stock
->cached_objcg
&& stock
->nr_bytes
>= nr_bytes
) {
3163 stock
->nr_bytes
-= nr_bytes
;
3167 local_irq_restore(flags
);
3172 static void drain_obj_stock(struct memcg_stock_pcp
*stock
)
3174 struct obj_cgroup
*old
= stock
->cached_objcg
;
3179 if (stock
->nr_bytes
) {
3180 unsigned int nr_pages
= stock
->nr_bytes
>> PAGE_SHIFT
;
3181 unsigned int nr_bytes
= stock
->nr_bytes
& (PAGE_SIZE
- 1);
3185 __memcg_kmem_uncharge(obj_cgroup_memcg(old
), nr_pages
);
3190 * The leftover is flushed to the centralized per-memcg value.
3191 * On the next attempt to refill obj stock it will be moved
3192 * to a per-cpu stock (probably, on an other CPU), see
3193 * refill_obj_stock().
3195 * How often it's flushed is a trade-off between the memory
3196 * limit enforcement accuracy and potential CPU contention,
3197 * so it might be changed in the future.
3199 atomic_add(nr_bytes
, &old
->nr_charged_bytes
);
3200 stock
->nr_bytes
= 0;
3203 obj_cgroup_put(old
);
3204 stock
->cached_objcg
= NULL
;
3207 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
3208 struct mem_cgroup
*root_memcg
)
3210 struct mem_cgroup
*memcg
;
3212 if (stock
->cached_objcg
) {
3213 memcg
= obj_cgroup_memcg(stock
->cached_objcg
);
3214 if (memcg
&& mem_cgroup_is_descendant(memcg
, root_memcg
))
3221 static void refill_obj_stock(struct obj_cgroup
*objcg
, unsigned int nr_bytes
)
3223 struct memcg_stock_pcp
*stock
;
3224 unsigned long flags
;
3226 local_irq_save(flags
);
3228 stock
= this_cpu_ptr(&memcg_stock
);
3229 if (stock
->cached_objcg
!= objcg
) { /* reset if necessary */
3230 drain_obj_stock(stock
);
3231 obj_cgroup_get(objcg
);
3232 stock
->cached_objcg
= objcg
;
3233 stock
->nr_bytes
= atomic_xchg(&objcg
->nr_charged_bytes
, 0);
3235 stock
->nr_bytes
+= nr_bytes
;
3237 if (stock
->nr_bytes
> PAGE_SIZE
)
3238 drain_obj_stock(stock
);
3240 local_irq_restore(flags
);
3243 int obj_cgroup_charge(struct obj_cgroup
*objcg
, gfp_t gfp
, size_t size
)
3245 struct mem_cgroup
*memcg
;
3246 unsigned int nr_pages
, nr_bytes
;
3249 if (consume_obj_stock(objcg
, size
))
3253 * In theory, memcg->nr_charged_bytes can have enough
3254 * pre-charged bytes to satisfy the allocation. However,
3255 * flushing memcg->nr_charged_bytes requires two atomic
3256 * operations, and memcg->nr_charged_bytes can't be big,
3257 * so it's better to ignore it and try grab some new pages.
3258 * memcg->nr_charged_bytes will be flushed in
3259 * refill_obj_stock(), called from this function or
3260 * independently later.
3264 memcg
= obj_cgroup_memcg(objcg
);
3265 if (unlikely(!css_tryget(&memcg
->css
)))
3269 nr_pages
= size
>> PAGE_SHIFT
;
3270 nr_bytes
= size
& (PAGE_SIZE
- 1);
3275 ret
= __memcg_kmem_charge(memcg
, gfp
, nr_pages
);
3276 if (!ret
&& nr_bytes
)
3277 refill_obj_stock(objcg
, PAGE_SIZE
- nr_bytes
);
3279 css_put(&memcg
->css
);
3283 void obj_cgroup_uncharge(struct obj_cgroup
*objcg
, size_t size
)
3285 refill_obj_stock(objcg
, size
);
3288 #endif /* CONFIG_MEMCG_KMEM */
3290 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3292 * Because page_memcg(head) is not set on compound tails, set it now.
3294 void mem_cgroup_split_huge_fixup(struct page
*head
)
3296 struct mem_cgroup
*memcg
= page_memcg(head
);
3299 if (mem_cgroup_disabled())
3302 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3303 css_get(&memcg
->css
);
3304 head
[i
].memcg_data
= (unsigned long)memcg
;
3307 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3309 #ifdef CONFIG_MEMCG_SWAP
3311 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3312 * @entry: swap entry to be moved
3313 * @from: mem_cgroup which the entry is moved from
3314 * @to: mem_cgroup which the entry is moved to
3316 * It succeeds only when the swap_cgroup's record for this entry is the same
3317 * as the mem_cgroup's id of @from.
3319 * Returns 0 on success, -EINVAL on failure.
3321 * The caller must have charged to @to, IOW, called page_counter_charge() about
3322 * both res and memsw, and called css_get().
3324 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
3325 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3327 unsigned short old_id
, new_id
;
3329 old_id
= mem_cgroup_id(from
);
3330 new_id
= mem_cgroup_id(to
);
3332 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
3333 mod_memcg_state(from
, MEMCG_SWAP
, -1);
3334 mod_memcg_state(to
, MEMCG_SWAP
, 1);
3340 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
3341 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3347 static DEFINE_MUTEX(memcg_max_mutex
);
3349 static int mem_cgroup_resize_max(struct mem_cgroup
*memcg
,
3350 unsigned long max
, bool memsw
)
3352 bool enlarge
= false;
3353 bool drained
= false;
3355 bool limits_invariant
;
3356 struct page_counter
*counter
= memsw
? &memcg
->memsw
: &memcg
->memory
;
3359 if (signal_pending(current
)) {
3364 mutex_lock(&memcg_max_mutex
);
3366 * Make sure that the new limit (memsw or memory limit) doesn't
3367 * break our basic invariant rule memory.max <= memsw.max.
3369 limits_invariant
= memsw
? max
>= READ_ONCE(memcg
->memory
.max
) :
3370 max
<= memcg
->memsw
.max
;
3371 if (!limits_invariant
) {
3372 mutex_unlock(&memcg_max_mutex
);
3376 if (max
> counter
->max
)
3378 ret
= page_counter_set_max(counter
, max
);
3379 mutex_unlock(&memcg_max_mutex
);
3385 drain_all_stock(memcg
);
3390 if (!try_to_free_mem_cgroup_pages(memcg
, 1,
3391 GFP_KERNEL
, !memsw
)) {
3397 if (!ret
&& enlarge
)
3398 memcg_oom_recover(memcg
);
3403 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
3405 unsigned long *total_scanned
)
3407 unsigned long nr_reclaimed
= 0;
3408 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
3409 unsigned long reclaimed
;
3411 struct mem_cgroup_tree_per_node
*mctz
;
3412 unsigned long excess
;
3413 unsigned long nr_scanned
;
3418 mctz
= soft_limit_tree_node(pgdat
->node_id
);
3421 * Do not even bother to check the largest node if the root
3422 * is empty. Do it lockless to prevent lock bouncing. Races
3423 * are acceptable as soft limit is best effort anyway.
3425 if (!mctz
|| RB_EMPTY_ROOT(&mctz
->rb_root
))
3429 * This loop can run a while, specially if mem_cgroup's continuously
3430 * keep exceeding their soft limit and putting the system under
3437 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3442 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
3443 gfp_mask
, &nr_scanned
);
3444 nr_reclaimed
+= reclaimed
;
3445 *total_scanned
+= nr_scanned
;
3446 spin_lock_irq(&mctz
->lock
);
3447 __mem_cgroup_remove_exceeded(mz
, mctz
);
3450 * If we failed to reclaim anything from this memory cgroup
3451 * it is time to move on to the next cgroup
3455 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
3457 excess
= soft_limit_excess(mz
->memcg
);
3459 * One school of thought says that we should not add
3460 * back the node to the tree if reclaim returns 0.
3461 * But our reclaim could return 0, simply because due
3462 * to priority we are exposing a smaller subset of
3463 * memory to reclaim from. Consider this as a longer
3466 /* If excess == 0, no tree ops */
3467 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
3468 spin_unlock_irq(&mctz
->lock
);
3469 css_put(&mz
->memcg
->css
);
3472 * Could not reclaim anything and there are no more
3473 * mem cgroups to try or we seem to be looping without
3474 * reclaiming anything.
3476 if (!nr_reclaimed
&&
3478 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3480 } while (!nr_reclaimed
);
3482 css_put(&next_mz
->memcg
->css
);
3483 return nr_reclaimed
;
3487 * Reclaims as many pages from the given memcg as possible.
3489 * Caller is responsible for holding css reference for memcg.
3491 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
3493 int nr_retries
= MAX_RECLAIM_RETRIES
;
3495 /* we call try-to-free pages for make this cgroup empty */
3496 lru_add_drain_all();
3498 drain_all_stock(memcg
);
3500 /* try to free all pages in this cgroup */
3501 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
3504 if (signal_pending(current
))
3507 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
3511 /* maybe some writeback is necessary */
3512 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3520 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3521 char *buf
, size_t nbytes
,
3524 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3526 if (mem_cgroup_is_root(memcg
))
3528 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3531 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3537 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3538 struct cftype
*cft
, u64 val
)
3543 pr_warn_once("Non-hierarchical mode is deprecated. "
3544 "Please report your usecase to linux-mm@kvack.org if you "
3545 "depend on this functionality.\n");
3550 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3554 if (mem_cgroup_is_root(memcg
)) {
3555 val
= memcg_page_state(memcg
, NR_FILE_PAGES
) +
3556 memcg_page_state(memcg
, NR_ANON_MAPPED
);
3558 val
+= memcg_page_state(memcg
, MEMCG_SWAP
);
3561 val
= page_counter_read(&memcg
->memory
);
3563 val
= page_counter_read(&memcg
->memsw
);
3576 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3579 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3580 struct page_counter
*counter
;
3582 switch (MEMFILE_TYPE(cft
->private)) {
3584 counter
= &memcg
->memory
;
3587 counter
= &memcg
->memsw
;
3590 counter
= &memcg
->kmem
;
3593 counter
= &memcg
->tcpmem
;
3599 switch (MEMFILE_ATTR(cft
->private)) {
3601 if (counter
== &memcg
->memory
)
3602 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
3603 if (counter
== &memcg
->memsw
)
3604 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
3605 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3607 return (u64
)counter
->max
* PAGE_SIZE
;
3609 return (u64
)counter
->watermark
* PAGE_SIZE
;
3611 return counter
->failcnt
;
3612 case RES_SOFT_LIMIT
:
3613 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
3619 static void memcg_flush_percpu_vmstats(struct mem_cgroup
*memcg
)
3621 unsigned long stat
[MEMCG_NR_STAT
] = {0};
3622 struct mem_cgroup
*mi
;
3625 for_each_online_cpu(cpu
)
3626 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
3627 stat
[i
] += per_cpu(memcg
->vmstats_percpu
->stat
[i
], cpu
);
3629 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
3630 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
3631 atomic_long_add(stat
[i
], &mi
->vmstats
[i
]);
3633 for_each_node(node
) {
3634 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
3635 struct mem_cgroup_per_node
*pi
;
3637 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3640 for_each_online_cpu(cpu
)
3641 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3643 pn
->lruvec_stat_cpu
->count
[i
], cpu
);
3645 for (pi
= pn
; pi
; pi
= parent_nodeinfo(pi
, node
))
3646 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3647 atomic_long_add(stat
[i
], &pi
->lruvec_stat
[i
]);
3651 static void memcg_flush_percpu_vmevents(struct mem_cgroup
*memcg
)
3653 unsigned long events
[NR_VM_EVENT_ITEMS
];
3654 struct mem_cgroup
*mi
;
3657 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3660 for_each_online_cpu(cpu
)
3661 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3662 events
[i
] += per_cpu(memcg
->vmstats_percpu
->events
[i
],
3665 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
3666 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3667 atomic_long_add(events
[i
], &mi
->vmevents
[i
]);
3670 #ifdef CONFIG_MEMCG_KMEM
3671 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3673 struct obj_cgroup
*objcg
;
3676 if (cgroup_memory_nokmem
)
3679 BUG_ON(memcg
->kmemcg_id
>= 0);
3680 BUG_ON(memcg
->kmem_state
);
3682 memcg_id
= memcg_alloc_cache_id();
3686 objcg
= obj_cgroup_alloc();
3688 memcg_free_cache_id(memcg_id
);
3691 objcg
->memcg
= memcg
;
3692 rcu_assign_pointer(memcg
->objcg
, objcg
);
3694 static_branch_enable(&memcg_kmem_enabled_key
);
3696 memcg
->kmemcg_id
= memcg_id
;
3697 memcg
->kmem_state
= KMEM_ONLINE
;
3702 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3704 struct cgroup_subsys_state
*css
;
3705 struct mem_cgroup
*parent
, *child
;
3708 if (memcg
->kmem_state
!= KMEM_ONLINE
)
3711 memcg
->kmem_state
= KMEM_ALLOCATED
;
3713 parent
= parent_mem_cgroup(memcg
);
3715 parent
= root_mem_cgroup
;
3717 memcg_reparent_objcgs(memcg
, parent
);
3719 kmemcg_id
= memcg
->kmemcg_id
;
3720 BUG_ON(kmemcg_id
< 0);
3723 * Change kmemcg_id of this cgroup and all its descendants to the
3724 * parent's id, and then move all entries from this cgroup's list_lrus
3725 * to ones of the parent. After we have finished, all list_lrus
3726 * corresponding to this cgroup are guaranteed to remain empty. The
3727 * ordering is imposed by list_lru_node->lock taken by
3728 * memcg_drain_all_list_lrus().
3730 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3731 css_for_each_descendant_pre(css
, &memcg
->css
) {
3732 child
= mem_cgroup_from_css(css
);
3733 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
3734 child
->kmemcg_id
= parent
->kmemcg_id
;
3738 memcg_drain_all_list_lrus(kmemcg_id
, parent
);
3740 memcg_free_cache_id(kmemcg_id
);
3743 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3745 /* css_alloc() failed, offlining didn't happen */
3746 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
3747 memcg_offline_kmem(memcg
);
3750 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3754 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3757 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3760 #endif /* CONFIG_MEMCG_KMEM */
3762 static int memcg_update_kmem_max(struct mem_cgroup
*memcg
,
3767 mutex_lock(&memcg_max_mutex
);
3768 ret
= page_counter_set_max(&memcg
->kmem
, max
);
3769 mutex_unlock(&memcg_max_mutex
);
3773 static int memcg_update_tcp_max(struct mem_cgroup
*memcg
, unsigned long max
)
3777 mutex_lock(&memcg_max_mutex
);
3779 ret
= page_counter_set_max(&memcg
->tcpmem
, max
);
3783 if (!memcg
->tcpmem_active
) {
3785 * The active flag needs to be written after the static_key
3786 * update. This is what guarantees that the socket activation
3787 * function is the last one to run. See mem_cgroup_sk_alloc()
3788 * for details, and note that we don't mark any socket as
3789 * belonging to this memcg until that flag is up.
3791 * We need to do this, because static_keys will span multiple
3792 * sites, but we can't control their order. If we mark a socket
3793 * as accounted, but the accounting functions are not patched in
3794 * yet, we'll lose accounting.
3796 * We never race with the readers in mem_cgroup_sk_alloc(),
3797 * because when this value change, the code to process it is not
3800 static_branch_inc(&memcg_sockets_enabled_key
);
3801 memcg
->tcpmem_active
= true;
3804 mutex_unlock(&memcg_max_mutex
);
3809 * The user of this function is...
3812 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3813 char *buf
, size_t nbytes
, loff_t off
)
3815 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3816 unsigned long nr_pages
;
3819 buf
= strstrip(buf
);
3820 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3824 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3826 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3830 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3832 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, false);
3835 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, true);
3838 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3839 "Please report your usecase to linux-mm@kvack.org if you "
3840 "depend on this functionality.\n");
3841 ret
= memcg_update_kmem_max(memcg
, nr_pages
);
3844 ret
= memcg_update_tcp_max(memcg
, nr_pages
);
3848 case RES_SOFT_LIMIT
:
3849 memcg
->soft_limit
= nr_pages
;
3853 return ret
?: nbytes
;
3856 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3857 size_t nbytes
, loff_t off
)
3859 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3860 struct page_counter
*counter
;
3862 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3864 counter
= &memcg
->memory
;
3867 counter
= &memcg
->memsw
;
3870 counter
= &memcg
->kmem
;
3873 counter
= &memcg
->tcpmem
;
3879 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3881 page_counter_reset_watermark(counter
);
3884 counter
->failcnt
= 0;
3893 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3896 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3900 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3901 struct cftype
*cft
, u64 val
)
3903 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3905 if (val
& ~MOVE_MASK
)
3909 * No kind of locking is needed in here, because ->can_attach() will
3910 * check this value once in the beginning of the process, and then carry
3911 * on with stale data. This means that changes to this value will only
3912 * affect task migrations starting after the change.
3914 memcg
->move_charge_at_immigrate
= val
;
3918 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3919 struct cftype
*cft
, u64 val
)
3927 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3928 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3929 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3931 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
3932 int nid
, unsigned int lru_mask
, bool tree
)
3934 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, NODE_DATA(nid
));
3935 unsigned long nr
= 0;
3938 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
3941 if (!(BIT(lru
) & lru_mask
))
3944 nr
+= lruvec_page_state(lruvec
, NR_LRU_BASE
+ lru
);
3946 nr
+= lruvec_page_state_local(lruvec
, NR_LRU_BASE
+ lru
);
3951 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
3952 unsigned int lru_mask
,
3955 unsigned long nr
= 0;
3959 if (!(BIT(lru
) & lru_mask
))
3962 nr
+= memcg_page_state(memcg
, NR_LRU_BASE
+ lru
);
3964 nr
+= memcg_page_state_local(memcg
, NR_LRU_BASE
+ lru
);
3969 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3973 unsigned int lru_mask
;
3976 static const struct numa_stat stats
[] = {
3977 { "total", LRU_ALL
},
3978 { "file", LRU_ALL_FILE
},
3979 { "anon", LRU_ALL_ANON
},
3980 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3982 const struct numa_stat
*stat
;
3984 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
3986 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3987 seq_printf(m
, "%s=%lu", stat
->name
,
3988 mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
,
3990 for_each_node_state(nid
, N_MEMORY
)
3991 seq_printf(m
, " N%d=%lu", nid
,
3992 mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3993 stat
->lru_mask
, false));
3997 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3999 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
,
4000 mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
,
4002 for_each_node_state(nid
, N_MEMORY
)
4003 seq_printf(m
, " N%d=%lu", nid
,
4004 mem_cgroup_node_nr_lru_pages(memcg
, nid
,
4005 stat
->lru_mask
, true));
4011 #endif /* CONFIG_NUMA */
4013 static const unsigned int memcg1_stats
[] = {
4016 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4026 static const char *const memcg1_stat_names
[] = {
4029 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4039 /* Universal VM events cgroup1 shows, original sort order */
4040 static const unsigned int memcg1_events
[] = {
4047 static int memcg_stat_show(struct seq_file
*m
, void *v
)
4049 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
4050 unsigned long memory
, memsw
;
4051 struct mem_cgroup
*mi
;
4054 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names
) != ARRAY_SIZE(memcg1_stats
));
4056 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
4059 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
4061 nr
= memcg_page_state_local(memcg
, memcg1_stats
[i
]);
4062 seq_printf(m
, "%s %lu\n", memcg1_stat_names
[i
], nr
* PAGE_SIZE
);
4065 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
4066 seq_printf(m
, "%s %lu\n", vm_event_name(memcg1_events
[i
]),
4067 memcg_events_local(memcg
, memcg1_events
[i
]));
4069 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
4070 seq_printf(m
, "%s %lu\n", lru_list_name(i
),
4071 memcg_page_state_local(memcg
, NR_LRU_BASE
+ i
) *
4074 /* Hierarchical information */
4075 memory
= memsw
= PAGE_COUNTER_MAX
;
4076 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
4077 memory
= min(memory
, READ_ONCE(mi
->memory
.max
));
4078 memsw
= min(memsw
, READ_ONCE(mi
->memsw
.max
));
4080 seq_printf(m
, "hierarchical_memory_limit %llu\n",
4081 (u64
)memory
* PAGE_SIZE
);
4082 if (do_memsw_account())
4083 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
4084 (u64
)memsw
* PAGE_SIZE
);
4086 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
4089 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
4091 nr
= memcg_page_state(memcg
, memcg1_stats
[i
]);
4092 seq_printf(m
, "total_%s %llu\n", memcg1_stat_names
[i
],
4093 (u64
)nr
* PAGE_SIZE
);
4096 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
4097 seq_printf(m
, "total_%s %llu\n",
4098 vm_event_name(memcg1_events
[i
]),
4099 (u64
)memcg_events(memcg
, memcg1_events
[i
]));
4101 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
4102 seq_printf(m
, "total_%s %llu\n", lru_list_name(i
),
4103 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
4106 #ifdef CONFIG_DEBUG_VM
4109 struct mem_cgroup_per_node
*mz
;
4110 unsigned long anon_cost
= 0;
4111 unsigned long file_cost
= 0;
4113 for_each_online_pgdat(pgdat
) {
4114 mz
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
4116 anon_cost
+= mz
->lruvec
.anon_cost
;
4117 file_cost
+= mz
->lruvec
.file_cost
;
4119 seq_printf(m
, "anon_cost %lu\n", anon_cost
);
4120 seq_printf(m
, "file_cost %lu\n", file_cost
);
4127 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
4130 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4132 return mem_cgroup_swappiness(memcg
);
4135 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
4136 struct cftype
*cft
, u64 val
)
4138 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4144 memcg
->swappiness
= val
;
4146 vm_swappiness
= val
;
4151 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
4153 struct mem_cgroup_threshold_ary
*t
;
4154 unsigned long usage
;
4159 t
= rcu_dereference(memcg
->thresholds
.primary
);
4161 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
4166 usage
= mem_cgroup_usage(memcg
, swap
);
4169 * current_threshold points to threshold just below or equal to usage.
4170 * If it's not true, a threshold was crossed after last
4171 * call of __mem_cgroup_threshold().
4173 i
= t
->current_threshold
;
4176 * Iterate backward over array of thresholds starting from
4177 * current_threshold and check if a threshold is crossed.
4178 * If none of thresholds below usage is crossed, we read
4179 * only one element of the array here.
4181 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
4182 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4184 /* i = current_threshold + 1 */
4188 * Iterate forward over array of thresholds starting from
4189 * current_threshold+1 and check if a threshold is crossed.
4190 * If none of thresholds above usage is crossed, we read
4191 * only one element of the array here.
4193 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
4194 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4196 /* Update current_threshold */
4197 t
->current_threshold
= i
- 1;
4202 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
4205 __mem_cgroup_threshold(memcg
, false);
4206 if (do_memsw_account())
4207 __mem_cgroup_threshold(memcg
, true);
4209 memcg
= parent_mem_cgroup(memcg
);
4213 static int compare_thresholds(const void *a
, const void *b
)
4215 const struct mem_cgroup_threshold
*_a
= a
;
4216 const struct mem_cgroup_threshold
*_b
= b
;
4218 if (_a
->threshold
> _b
->threshold
)
4221 if (_a
->threshold
< _b
->threshold
)
4227 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
4229 struct mem_cgroup_eventfd_list
*ev
;
4231 spin_lock(&memcg_oom_lock
);
4233 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
4234 eventfd_signal(ev
->eventfd
, 1);
4236 spin_unlock(&memcg_oom_lock
);
4240 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
4242 struct mem_cgroup
*iter
;
4244 for_each_mem_cgroup_tree(iter
, memcg
)
4245 mem_cgroup_oom_notify_cb(iter
);
4248 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4249 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
4251 struct mem_cgroup_thresholds
*thresholds
;
4252 struct mem_cgroup_threshold_ary
*new;
4253 unsigned long threshold
;
4254 unsigned long usage
;
4257 ret
= page_counter_memparse(args
, "-1", &threshold
);
4261 mutex_lock(&memcg
->thresholds_lock
);
4264 thresholds
= &memcg
->thresholds
;
4265 usage
= mem_cgroup_usage(memcg
, false);
4266 } else if (type
== _MEMSWAP
) {
4267 thresholds
= &memcg
->memsw_thresholds
;
4268 usage
= mem_cgroup_usage(memcg
, true);
4272 /* Check if a threshold crossed before adding a new one */
4273 if (thresholds
->primary
)
4274 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4276 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
4278 /* Allocate memory for new array of thresholds */
4279 new = kmalloc(struct_size(new, entries
, size
), GFP_KERNEL
);
4286 /* Copy thresholds (if any) to new array */
4287 if (thresholds
->primary
)
4288 memcpy(new->entries
, thresholds
->primary
->entries
,
4289 flex_array_size(new, entries
, size
- 1));
4291 /* Add new threshold */
4292 new->entries
[size
- 1].eventfd
= eventfd
;
4293 new->entries
[size
- 1].threshold
= threshold
;
4295 /* Sort thresholds. Registering of new threshold isn't time-critical */
4296 sort(new->entries
, size
, sizeof(*new->entries
),
4297 compare_thresholds
, NULL
);
4299 /* Find current threshold */
4300 new->current_threshold
= -1;
4301 for (i
= 0; i
< size
; i
++) {
4302 if (new->entries
[i
].threshold
<= usage
) {
4304 * new->current_threshold will not be used until
4305 * rcu_assign_pointer(), so it's safe to increment
4308 ++new->current_threshold
;
4313 /* Free old spare buffer and save old primary buffer as spare */
4314 kfree(thresholds
->spare
);
4315 thresholds
->spare
= thresholds
->primary
;
4317 rcu_assign_pointer(thresholds
->primary
, new);
4319 /* To be sure that nobody uses thresholds */
4323 mutex_unlock(&memcg
->thresholds_lock
);
4328 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4329 struct eventfd_ctx
*eventfd
, const char *args
)
4331 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
4334 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4335 struct eventfd_ctx
*eventfd
, const char *args
)
4337 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
4340 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4341 struct eventfd_ctx
*eventfd
, enum res_type type
)
4343 struct mem_cgroup_thresholds
*thresholds
;
4344 struct mem_cgroup_threshold_ary
*new;
4345 unsigned long usage
;
4346 int i
, j
, size
, entries
;
4348 mutex_lock(&memcg
->thresholds_lock
);
4351 thresholds
= &memcg
->thresholds
;
4352 usage
= mem_cgroup_usage(memcg
, false);
4353 } else if (type
== _MEMSWAP
) {
4354 thresholds
= &memcg
->memsw_thresholds
;
4355 usage
= mem_cgroup_usage(memcg
, true);
4359 if (!thresholds
->primary
)
4362 /* Check if a threshold crossed before removing */
4363 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4365 /* Calculate new number of threshold */
4367 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4368 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4374 new = thresholds
->spare
;
4376 /* If no items related to eventfd have been cleared, nothing to do */
4380 /* Set thresholds array to NULL if we don't have thresholds */
4389 /* Copy thresholds and find current threshold */
4390 new->current_threshold
= -1;
4391 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4392 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4395 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4396 if (new->entries
[j
].threshold
<= usage
) {
4398 * new->current_threshold will not be used
4399 * until rcu_assign_pointer(), so it's safe to increment
4402 ++new->current_threshold
;
4408 /* Swap primary and spare array */
4409 thresholds
->spare
= thresholds
->primary
;
4411 rcu_assign_pointer(thresholds
->primary
, new);
4413 /* To be sure that nobody uses thresholds */
4416 /* If all events are unregistered, free the spare array */
4418 kfree(thresholds
->spare
);
4419 thresholds
->spare
= NULL
;
4422 mutex_unlock(&memcg
->thresholds_lock
);
4425 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4426 struct eventfd_ctx
*eventfd
)
4428 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
4431 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4432 struct eventfd_ctx
*eventfd
)
4434 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
4437 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
4438 struct eventfd_ctx
*eventfd
, const char *args
)
4440 struct mem_cgroup_eventfd_list
*event
;
4442 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4446 spin_lock(&memcg_oom_lock
);
4448 event
->eventfd
= eventfd
;
4449 list_add(&event
->list
, &memcg
->oom_notify
);
4451 /* already in OOM ? */
4452 if (memcg
->under_oom
)
4453 eventfd_signal(eventfd
, 1);
4454 spin_unlock(&memcg_oom_lock
);
4459 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
4460 struct eventfd_ctx
*eventfd
)
4462 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4464 spin_lock(&memcg_oom_lock
);
4466 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
4467 if (ev
->eventfd
== eventfd
) {
4468 list_del(&ev
->list
);
4473 spin_unlock(&memcg_oom_lock
);
4476 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
4478 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(sf
);
4480 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
4481 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
4482 seq_printf(sf
, "oom_kill %lu\n",
4483 atomic_long_read(&memcg
->memory_events
[MEMCG_OOM_KILL
]));
4487 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
4488 struct cftype
*cft
, u64 val
)
4490 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4492 /* cannot set to root cgroup and only 0 and 1 are allowed */
4493 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
4496 memcg
->oom_kill_disable
= val
;
4498 memcg_oom_recover(memcg
);
4503 #ifdef CONFIG_CGROUP_WRITEBACK
4505 #include <trace/events/writeback.h>
4507 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4509 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
4512 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4514 wb_domain_exit(&memcg
->cgwb_domain
);
4517 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4519 wb_domain_size_changed(&memcg
->cgwb_domain
);
4522 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
4524 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4526 if (!memcg
->css
.parent
)
4529 return &memcg
->cgwb_domain
;
4533 * idx can be of type enum memcg_stat_item or node_stat_item.
4534 * Keep in sync with memcg_exact_page().
4536 static unsigned long memcg_exact_page_state(struct mem_cgroup
*memcg
, int idx
)
4538 long x
= atomic_long_read(&memcg
->vmstats
[idx
]);
4541 for_each_online_cpu(cpu
)
4542 x
+= per_cpu_ptr(memcg
->vmstats_percpu
, cpu
)->stat
[idx
];
4549 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4550 * @wb: bdi_writeback in question
4551 * @pfilepages: out parameter for number of file pages
4552 * @pheadroom: out parameter for number of allocatable pages according to memcg
4553 * @pdirty: out parameter for number of dirty pages
4554 * @pwriteback: out parameter for number of pages under writeback
4556 * Determine the numbers of file, headroom, dirty, and writeback pages in
4557 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4558 * is a bit more involved.
4560 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4561 * headroom is calculated as the lowest headroom of itself and the
4562 * ancestors. Note that this doesn't consider the actual amount of
4563 * available memory in the system. The caller should further cap
4564 * *@pheadroom accordingly.
4566 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
4567 unsigned long *pheadroom
, unsigned long *pdirty
,
4568 unsigned long *pwriteback
)
4570 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4571 struct mem_cgroup
*parent
;
4573 *pdirty
= memcg_exact_page_state(memcg
, NR_FILE_DIRTY
);
4575 *pwriteback
= memcg_exact_page_state(memcg
, NR_WRITEBACK
);
4576 *pfilepages
= memcg_exact_page_state(memcg
, NR_INACTIVE_FILE
) +
4577 memcg_exact_page_state(memcg
, NR_ACTIVE_FILE
);
4578 *pheadroom
= PAGE_COUNTER_MAX
;
4580 while ((parent
= parent_mem_cgroup(memcg
))) {
4581 unsigned long ceiling
= min(READ_ONCE(memcg
->memory
.max
),
4582 READ_ONCE(memcg
->memory
.high
));
4583 unsigned long used
= page_counter_read(&memcg
->memory
);
4585 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
4591 * Foreign dirty flushing
4593 * There's an inherent mismatch between memcg and writeback. The former
4594 * trackes ownership per-page while the latter per-inode. This was a
4595 * deliberate design decision because honoring per-page ownership in the
4596 * writeback path is complicated, may lead to higher CPU and IO overheads
4597 * and deemed unnecessary given that write-sharing an inode across
4598 * different cgroups isn't a common use-case.
4600 * Combined with inode majority-writer ownership switching, this works well
4601 * enough in most cases but there are some pathological cases. For
4602 * example, let's say there are two cgroups A and B which keep writing to
4603 * different but confined parts of the same inode. B owns the inode and
4604 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4605 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4606 * triggering background writeback. A will be slowed down without a way to
4607 * make writeback of the dirty pages happen.
4609 * Conditions like the above can lead to a cgroup getting repatedly and
4610 * severely throttled after making some progress after each
4611 * dirty_expire_interval while the underyling IO device is almost
4614 * Solving this problem completely requires matching the ownership tracking
4615 * granularities between memcg and writeback in either direction. However,
4616 * the more egregious behaviors can be avoided by simply remembering the
4617 * most recent foreign dirtying events and initiating remote flushes on
4618 * them when local writeback isn't enough to keep the memory clean enough.
4620 * The following two functions implement such mechanism. When a foreign
4621 * page - a page whose memcg and writeback ownerships don't match - is
4622 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4623 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4624 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4625 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4626 * foreign bdi_writebacks which haven't expired. Both the numbers of
4627 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4628 * limited to MEMCG_CGWB_FRN_CNT.
4630 * The mechanism only remembers IDs and doesn't hold any object references.
4631 * As being wrong occasionally doesn't matter, updates and accesses to the
4632 * records are lockless and racy.
4634 void mem_cgroup_track_foreign_dirty_slowpath(struct page
*page
,
4635 struct bdi_writeback
*wb
)
4637 struct mem_cgroup
*memcg
= page_memcg(page
);
4638 struct memcg_cgwb_frn
*frn
;
4639 u64 now
= get_jiffies_64();
4640 u64 oldest_at
= now
;
4644 trace_track_foreign_dirty(page
, wb
);
4647 * Pick the slot to use. If there is already a slot for @wb, keep
4648 * using it. If not replace the oldest one which isn't being
4651 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4652 frn
= &memcg
->cgwb_frn
[i
];
4653 if (frn
->bdi_id
== wb
->bdi
->id
&&
4654 frn
->memcg_id
== wb
->memcg_css
->id
)
4656 if (time_before64(frn
->at
, oldest_at
) &&
4657 atomic_read(&frn
->done
.cnt
) == 1) {
4659 oldest_at
= frn
->at
;
4663 if (i
< MEMCG_CGWB_FRN_CNT
) {
4665 * Re-using an existing one. Update timestamp lazily to
4666 * avoid making the cacheline hot. We want them to be
4667 * reasonably up-to-date and significantly shorter than
4668 * dirty_expire_interval as that's what expires the record.
4669 * Use the shorter of 1s and dirty_expire_interval / 8.
4671 unsigned long update_intv
=
4672 min_t(unsigned long, HZ
,
4673 msecs_to_jiffies(dirty_expire_interval
* 10) / 8);
4675 if (time_before64(frn
->at
, now
- update_intv
))
4677 } else if (oldest
>= 0) {
4678 /* replace the oldest free one */
4679 frn
= &memcg
->cgwb_frn
[oldest
];
4680 frn
->bdi_id
= wb
->bdi
->id
;
4681 frn
->memcg_id
= wb
->memcg_css
->id
;
4686 /* issue foreign writeback flushes for recorded foreign dirtying events */
4687 void mem_cgroup_flush_foreign(struct bdi_writeback
*wb
)
4689 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4690 unsigned long intv
= msecs_to_jiffies(dirty_expire_interval
* 10);
4691 u64 now
= jiffies_64
;
4694 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4695 struct memcg_cgwb_frn
*frn
= &memcg
->cgwb_frn
[i
];
4698 * If the record is older than dirty_expire_interval,
4699 * writeback on it has already started. No need to kick it
4700 * off again. Also, don't start a new one if there's
4701 * already one in flight.
4703 if (time_after64(frn
->at
, now
- intv
) &&
4704 atomic_read(&frn
->done
.cnt
) == 1) {
4706 trace_flush_foreign(wb
, frn
->bdi_id
, frn
->memcg_id
);
4707 cgroup_writeback_by_id(frn
->bdi_id
, frn
->memcg_id
, 0,
4708 WB_REASON_FOREIGN_FLUSH
,
4714 #else /* CONFIG_CGROUP_WRITEBACK */
4716 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4721 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4725 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4729 #endif /* CONFIG_CGROUP_WRITEBACK */
4732 * DO NOT USE IN NEW FILES.
4734 * "cgroup.event_control" implementation.
4736 * This is way over-engineered. It tries to support fully configurable
4737 * events for each user. Such level of flexibility is completely
4738 * unnecessary especially in the light of the planned unified hierarchy.
4740 * Please deprecate this and replace with something simpler if at all
4745 * Unregister event and free resources.
4747 * Gets called from workqueue.
4749 static void memcg_event_remove(struct work_struct
*work
)
4751 struct mem_cgroup_event
*event
=
4752 container_of(work
, struct mem_cgroup_event
, remove
);
4753 struct mem_cgroup
*memcg
= event
->memcg
;
4755 remove_wait_queue(event
->wqh
, &event
->wait
);
4757 event
->unregister_event(memcg
, event
->eventfd
);
4759 /* Notify userspace the event is going away. */
4760 eventfd_signal(event
->eventfd
, 1);
4762 eventfd_ctx_put(event
->eventfd
);
4764 css_put(&memcg
->css
);
4768 * Gets called on EPOLLHUP on eventfd when user closes it.
4770 * Called with wqh->lock held and interrupts disabled.
4772 static int memcg_event_wake(wait_queue_entry_t
*wait
, unsigned mode
,
4773 int sync
, void *key
)
4775 struct mem_cgroup_event
*event
=
4776 container_of(wait
, struct mem_cgroup_event
, wait
);
4777 struct mem_cgroup
*memcg
= event
->memcg
;
4778 __poll_t flags
= key_to_poll(key
);
4780 if (flags
& EPOLLHUP
) {
4782 * If the event has been detached at cgroup removal, we
4783 * can simply return knowing the other side will cleanup
4786 * We can't race against event freeing since the other
4787 * side will require wqh->lock via remove_wait_queue(),
4790 spin_lock(&memcg
->event_list_lock
);
4791 if (!list_empty(&event
->list
)) {
4792 list_del_init(&event
->list
);
4794 * We are in atomic context, but cgroup_event_remove()
4795 * may sleep, so we have to call it in workqueue.
4797 schedule_work(&event
->remove
);
4799 spin_unlock(&memcg
->event_list_lock
);
4805 static void memcg_event_ptable_queue_proc(struct file
*file
,
4806 wait_queue_head_t
*wqh
, poll_table
*pt
)
4808 struct mem_cgroup_event
*event
=
4809 container_of(pt
, struct mem_cgroup_event
, pt
);
4812 add_wait_queue(wqh
, &event
->wait
);
4816 * DO NOT USE IN NEW FILES.
4818 * Parse input and register new cgroup event handler.
4820 * Input must be in format '<event_fd> <control_fd> <args>'.
4821 * Interpretation of args is defined by control file implementation.
4823 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4824 char *buf
, size_t nbytes
, loff_t off
)
4826 struct cgroup_subsys_state
*css
= of_css(of
);
4827 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4828 struct mem_cgroup_event
*event
;
4829 struct cgroup_subsys_state
*cfile_css
;
4830 unsigned int efd
, cfd
;
4837 buf
= strstrip(buf
);
4839 efd
= simple_strtoul(buf
, &endp
, 10);
4844 cfd
= simple_strtoul(buf
, &endp
, 10);
4845 if ((*endp
!= ' ') && (*endp
!= '\0'))
4849 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
4853 event
->memcg
= memcg
;
4854 INIT_LIST_HEAD(&event
->list
);
4855 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
4856 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
4857 INIT_WORK(&event
->remove
, memcg_event_remove
);
4865 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
4866 if (IS_ERR(event
->eventfd
)) {
4867 ret
= PTR_ERR(event
->eventfd
);
4874 goto out_put_eventfd
;
4877 /* the process need read permission on control file */
4878 /* AV: shouldn't we check that it's been opened for read instead? */
4879 ret
= file_permission(cfile
.file
, MAY_READ
);
4884 * Determine the event callbacks and set them in @event. This used
4885 * to be done via struct cftype but cgroup core no longer knows
4886 * about these events. The following is crude but the whole thing
4887 * is for compatibility anyway.
4889 * DO NOT ADD NEW FILES.
4891 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
4893 if (!strcmp(name
, "memory.usage_in_bytes")) {
4894 event
->register_event
= mem_cgroup_usage_register_event
;
4895 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
4896 } else if (!strcmp(name
, "memory.oom_control")) {
4897 event
->register_event
= mem_cgroup_oom_register_event
;
4898 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
4899 } else if (!strcmp(name
, "memory.pressure_level")) {
4900 event
->register_event
= vmpressure_register_event
;
4901 event
->unregister_event
= vmpressure_unregister_event
;
4902 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
4903 event
->register_event
= memsw_cgroup_usage_register_event
;
4904 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
4911 * Verify @cfile should belong to @css. Also, remaining events are
4912 * automatically removed on cgroup destruction but the removal is
4913 * asynchronous, so take an extra ref on @css.
4915 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
4916 &memory_cgrp_subsys
);
4918 if (IS_ERR(cfile_css
))
4920 if (cfile_css
!= css
) {
4925 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
4929 vfs_poll(efile
.file
, &event
->pt
);
4931 spin_lock(&memcg
->event_list_lock
);
4932 list_add(&event
->list
, &memcg
->event_list
);
4933 spin_unlock(&memcg
->event_list_lock
);
4945 eventfd_ctx_put(event
->eventfd
);
4954 static struct cftype mem_cgroup_legacy_files
[] = {
4956 .name
= "usage_in_bytes",
4957 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4958 .read_u64
= mem_cgroup_read_u64
,
4961 .name
= "max_usage_in_bytes",
4962 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4963 .write
= mem_cgroup_reset
,
4964 .read_u64
= mem_cgroup_read_u64
,
4967 .name
= "limit_in_bytes",
4968 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4969 .write
= mem_cgroup_write
,
4970 .read_u64
= mem_cgroup_read_u64
,
4973 .name
= "soft_limit_in_bytes",
4974 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4975 .write
= mem_cgroup_write
,
4976 .read_u64
= mem_cgroup_read_u64
,
4980 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4981 .write
= mem_cgroup_reset
,
4982 .read_u64
= mem_cgroup_read_u64
,
4986 .seq_show
= memcg_stat_show
,
4989 .name
= "force_empty",
4990 .write
= mem_cgroup_force_empty_write
,
4993 .name
= "use_hierarchy",
4994 .write_u64
= mem_cgroup_hierarchy_write
,
4995 .read_u64
= mem_cgroup_hierarchy_read
,
4998 .name
= "cgroup.event_control", /* XXX: for compat */
4999 .write
= memcg_write_event_control
,
5000 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
5003 .name
= "swappiness",
5004 .read_u64
= mem_cgroup_swappiness_read
,
5005 .write_u64
= mem_cgroup_swappiness_write
,
5008 .name
= "move_charge_at_immigrate",
5009 .read_u64
= mem_cgroup_move_charge_read
,
5010 .write_u64
= mem_cgroup_move_charge_write
,
5013 .name
= "oom_control",
5014 .seq_show
= mem_cgroup_oom_control_read
,
5015 .write_u64
= mem_cgroup_oom_control_write
,
5016 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
5019 .name
= "pressure_level",
5023 .name
= "numa_stat",
5024 .seq_show
= memcg_numa_stat_show
,
5028 .name
= "kmem.limit_in_bytes",
5029 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
5030 .write
= mem_cgroup_write
,
5031 .read_u64
= mem_cgroup_read_u64
,
5034 .name
= "kmem.usage_in_bytes",
5035 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
5036 .read_u64
= mem_cgroup_read_u64
,
5039 .name
= "kmem.failcnt",
5040 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
5041 .write
= mem_cgroup_reset
,
5042 .read_u64
= mem_cgroup_read_u64
,
5045 .name
= "kmem.max_usage_in_bytes",
5046 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
5047 .write
= mem_cgroup_reset
,
5048 .read_u64
= mem_cgroup_read_u64
,
5050 #if defined(CONFIG_MEMCG_KMEM) && \
5051 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5053 .name
= "kmem.slabinfo",
5054 .seq_show
= memcg_slab_show
,
5058 .name
= "kmem.tcp.limit_in_bytes",
5059 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
5060 .write
= mem_cgroup_write
,
5061 .read_u64
= mem_cgroup_read_u64
,
5064 .name
= "kmem.tcp.usage_in_bytes",
5065 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
5066 .read_u64
= mem_cgroup_read_u64
,
5069 .name
= "kmem.tcp.failcnt",
5070 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
5071 .write
= mem_cgroup_reset
,
5072 .read_u64
= mem_cgroup_read_u64
,
5075 .name
= "kmem.tcp.max_usage_in_bytes",
5076 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
5077 .write
= mem_cgroup_reset
,
5078 .read_u64
= mem_cgroup_read_u64
,
5080 { }, /* terminate */
5084 * Private memory cgroup IDR
5086 * Swap-out records and page cache shadow entries need to store memcg
5087 * references in constrained space, so we maintain an ID space that is
5088 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5089 * memory-controlled cgroups to 64k.
5091 * However, there usually are many references to the offline CSS after
5092 * the cgroup has been destroyed, such as page cache or reclaimable
5093 * slab objects, that don't need to hang on to the ID. We want to keep
5094 * those dead CSS from occupying IDs, or we might quickly exhaust the
5095 * relatively small ID space and prevent the creation of new cgroups
5096 * even when there are much fewer than 64k cgroups - possibly none.
5098 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5099 * be freed and recycled when it's no longer needed, which is usually
5100 * when the CSS is offlined.
5102 * The only exception to that are records of swapped out tmpfs/shmem
5103 * pages that need to be attributed to live ancestors on swapin. But
5104 * those references are manageable from userspace.
5107 static DEFINE_IDR(mem_cgroup_idr
);
5109 static void mem_cgroup_id_remove(struct mem_cgroup
*memcg
)
5111 if (memcg
->id
.id
> 0) {
5112 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
5117 static void __maybe_unused
mem_cgroup_id_get_many(struct mem_cgroup
*memcg
,
5120 refcount_add(n
, &memcg
->id
.ref
);
5123 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
5125 if (refcount_sub_and_test(n
, &memcg
->id
.ref
)) {
5126 mem_cgroup_id_remove(memcg
);
5128 /* Memcg ID pins CSS */
5129 css_put(&memcg
->css
);
5133 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
5135 mem_cgroup_id_put_many(memcg
, 1);
5139 * mem_cgroup_from_id - look up a memcg from a memcg id
5140 * @id: the memcg id to look up
5142 * Caller must hold rcu_read_lock().
5144 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
5146 WARN_ON_ONCE(!rcu_read_lock_held());
5147 return idr_find(&mem_cgroup_idr
, id
);
5150 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
5152 struct mem_cgroup_per_node
*pn
;
5155 * This routine is called against possible nodes.
5156 * But it's BUG to call kmalloc() against offline node.
5158 * TODO: this routine can waste much memory for nodes which will
5159 * never be onlined. It's better to use memory hotplug callback
5162 if (!node_state(node
, N_NORMAL_MEMORY
))
5164 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
5168 pn
->lruvec_stat_local
= alloc_percpu_gfp(struct lruvec_stat
,
5169 GFP_KERNEL_ACCOUNT
);
5170 if (!pn
->lruvec_stat_local
) {
5175 pn
->lruvec_stat_cpu
= alloc_percpu_gfp(struct batched_lruvec_stat
,
5176 GFP_KERNEL_ACCOUNT
);
5177 if (!pn
->lruvec_stat_cpu
) {
5178 free_percpu(pn
->lruvec_stat_local
);
5183 lruvec_init(&pn
->lruvec
);
5184 pn
->usage_in_excess
= 0;
5185 pn
->on_tree
= false;
5188 memcg
->nodeinfo
[node
] = pn
;
5192 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
5194 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
5199 free_percpu(pn
->lruvec_stat_cpu
);
5200 free_percpu(pn
->lruvec_stat_local
);
5204 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5209 free_mem_cgroup_per_node_info(memcg
, node
);
5210 free_percpu(memcg
->vmstats_percpu
);
5211 free_percpu(memcg
->vmstats_local
);
5215 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
5217 memcg_wb_domain_exit(memcg
);
5219 * Flush percpu vmstats and vmevents to guarantee the value correctness
5220 * on parent's and all ancestor levels.
5222 memcg_flush_percpu_vmstats(memcg
);
5223 memcg_flush_percpu_vmevents(memcg
);
5224 __mem_cgroup_free(memcg
);
5227 static struct mem_cgroup
*mem_cgroup_alloc(void)
5229 struct mem_cgroup
*memcg
;
5232 int __maybe_unused i
;
5233 long error
= -ENOMEM
;
5235 size
= sizeof(struct mem_cgroup
);
5236 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
5238 memcg
= kzalloc(size
, GFP_KERNEL
);
5240 return ERR_PTR(error
);
5242 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
5243 1, MEM_CGROUP_ID_MAX
,
5245 if (memcg
->id
.id
< 0) {
5246 error
= memcg
->id
.id
;
5250 memcg
->vmstats_local
= alloc_percpu_gfp(struct memcg_vmstats_percpu
,
5251 GFP_KERNEL_ACCOUNT
);
5252 if (!memcg
->vmstats_local
)
5255 memcg
->vmstats_percpu
= alloc_percpu_gfp(struct memcg_vmstats_percpu
,
5256 GFP_KERNEL_ACCOUNT
);
5257 if (!memcg
->vmstats_percpu
)
5261 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
5264 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
5267 INIT_WORK(&memcg
->high_work
, high_work_func
);
5268 INIT_LIST_HEAD(&memcg
->oom_notify
);
5269 mutex_init(&memcg
->thresholds_lock
);
5270 spin_lock_init(&memcg
->move_lock
);
5271 vmpressure_init(&memcg
->vmpressure
);
5272 INIT_LIST_HEAD(&memcg
->event_list
);
5273 spin_lock_init(&memcg
->event_list_lock
);
5274 memcg
->socket_pressure
= jiffies
;
5275 #ifdef CONFIG_MEMCG_KMEM
5276 memcg
->kmemcg_id
= -1;
5277 INIT_LIST_HEAD(&memcg
->objcg_list
);
5279 #ifdef CONFIG_CGROUP_WRITEBACK
5280 INIT_LIST_HEAD(&memcg
->cgwb_list
);
5281 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5282 memcg
->cgwb_frn
[i
].done
=
5283 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq
);
5285 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5286 spin_lock_init(&memcg
->deferred_split_queue
.split_queue_lock
);
5287 INIT_LIST_HEAD(&memcg
->deferred_split_queue
.split_queue
);
5288 memcg
->deferred_split_queue
.split_queue_len
= 0;
5290 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
5293 mem_cgroup_id_remove(memcg
);
5294 __mem_cgroup_free(memcg
);
5295 return ERR_PTR(error
);
5298 static struct cgroup_subsys_state
* __ref
5299 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
5301 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
5302 struct mem_cgroup
*memcg
, *old_memcg
;
5303 long error
= -ENOMEM
;
5305 old_memcg
= set_active_memcg(parent
);
5306 memcg
= mem_cgroup_alloc();
5307 set_active_memcg(old_memcg
);
5309 return ERR_CAST(memcg
);
5311 page_counter_set_high(&memcg
->memory
, PAGE_COUNTER_MAX
);
5312 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5313 page_counter_set_high(&memcg
->swap
, PAGE_COUNTER_MAX
);
5315 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
5316 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
5318 page_counter_init(&memcg
->memory
, &parent
->memory
);
5319 page_counter_init(&memcg
->swap
, &parent
->swap
);
5320 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
5321 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
5323 page_counter_init(&memcg
->memory
, NULL
);
5324 page_counter_init(&memcg
->swap
, NULL
);
5325 page_counter_init(&memcg
->kmem
, NULL
);
5326 page_counter_init(&memcg
->tcpmem
, NULL
);
5328 root_mem_cgroup
= memcg
;
5332 /* The following stuff does not apply to the root */
5333 error
= memcg_online_kmem(memcg
);
5337 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5338 static_branch_inc(&memcg_sockets_enabled_key
);
5342 mem_cgroup_id_remove(memcg
);
5343 mem_cgroup_free(memcg
);
5344 return ERR_PTR(error
);
5347 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
5349 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5352 * A memcg must be visible for memcg_expand_shrinker_maps()
5353 * by the time the maps are allocated. So, we allocate maps
5354 * here, when for_each_mem_cgroup() can't skip it.
5356 if (memcg_alloc_shrinker_maps(memcg
)) {
5357 mem_cgroup_id_remove(memcg
);
5361 /* Online state pins memcg ID, memcg ID pins CSS */
5362 refcount_set(&memcg
->id
.ref
, 1);
5367 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
5369 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5370 struct mem_cgroup_event
*event
, *tmp
;
5373 * Unregister events and notify userspace.
5374 * Notify userspace about cgroup removing only after rmdir of cgroup
5375 * directory to avoid race between userspace and kernelspace.
5377 spin_lock(&memcg
->event_list_lock
);
5378 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
5379 list_del_init(&event
->list
);
5380 schedule_work(&event
->remove
);
5382 spin_unlock(&memcg
->event_list_lock
);
5384 page_counter_set_min(&memcg
->memory
, 0);
5385 page_counter_set_low(&memcg
->memory
, 0);
5387 memcg_offline_kmem(memcg
);
5388 wb_memcg_offline(memcg
);
5390 drain_all_stock(memcg
);
5392 mem_cgroup_id_put(memcg
);
5395 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
5397 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5399 invalidate_reclaim_iterators(memcg
);
5402 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
5404 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5405 int __maybe_unused i
;
5407 #ifdef CONFIG_CGROUP_WRITEBACK
5408 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5409 wb_wait_for_completion(&memcg
->cgwb_frn
[i
].done
);
5411 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5412 static_branch_dec(&memcg_sockets_enabled_key
);
5414 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
5415 static_branch_dec(&memcg_sockets_enabled_key
);
5417 vmpressure_cleanup(&memcg
->vmpressure
);
5418 cancel_work_sync(&memcg
->high_work
);
5419 mem_cgroup_remove_from_trees(memcg
);
5420 memcg_free_shrinker_maps(memcg
);
5421 memcg_free_kmem(memcg
);
5422 mem_cgroup_free(memcg
);
5426 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5427 * @css: the target css
5429 * Reset the states of the mem_cgroup associated with @css. This is
5430 * invoked when the userland requests disabling on the default hierarchy
5431 * but the memcg is pinned through dependency. The memcg should stop
5432 * applying policies and should revert to the vanilla state as it may be
5433 * made visible again.
5435 * The current implementation only resets the essential configurations.
5436 * This needs to be expanded to cover all the visible parts.
5438 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
5440 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5442 page_counter_set_max(&memcg
->memory
, PAGE_COUNTER_MAX
);
5443 page_counter_set_max(&memcg
->swap
, PAGE_COUNTER_MAX
);
5444 page_counter_set_max(&memcg
->kmem
, PAGE_COUNTER_MAX
);
5445 page_counter_set_max(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
5446 page_counter_set_min(&memcg
->memory
, 0);
5447 page_counter_set_low(&memcg
->memory
, 0);
5448 page_counter_set_high(&memcg
->memory
, PAGE_COUNTER_MAX
);
5449 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5450 page_counter_set_high(&memcg
->swap
, PAGE_COUNTER_MAX
);
5451 memcg_wb_domain_size_changed(memcg
);
5455 /* Handlers for move charge at task migration. */
5456 static int mem_cgroup_do_precharge(unsigned long count
)
5460 /* Try a single bulk charge without reclaim first, kswapd may wake */
5461 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
5463 mc
.precharge
+= count
;
5467 /* Try charges one by one with reclaim, but do not retry */
5469 ret
= try_charge(mc
.to
, GFP_KERNEL
| __GFP_NORETRY
, 1);
5483 enum mc_target_type
{
5490 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
5491 unsigned long addr
, pte_t ptent
)
5493 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
5495 if (!page
|| !page_mapped(page
))
5497 if (PageAnon(page
)) {
5498 if (!(mc
.flags
& MOVE_ANON
))
5501 if (!(mc
.flags
& MOVE_FILE
))
5504 if (!get_page_unless_zero(page
))
5510 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5511 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5512 pte_t ptent
, swp_entry_t
*entry
)
5514 struct page
*page
= NULL
;
5515 swp_entry_t ent
= pte_to_swp_entry(ptent
);
5517 if (!(mc
.flags
& MOVE_ANON
))
5521 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5522 * a device and because they are not accessible by CPU they are store
5523 * as special swap entry in the CPU page table.
5525 if (is_device_private_entry(ent
)) {
5526 page
= device_private_entry_to_page(ent
);
5528 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5529 * a refcount of 1 when free (unlike normal page)
5531 if (!page_ref_add_unless(page
, 1, 1))
5536 if (non_swap_entry(ent
))
5540 * Because lookup_swap_cache() updates some statistics counter,
5541 * we call find_get_page() with swapper_space directly.
5543 page
= find_get_page(swap_address_space(ent
), swp_offset(ent
));
5544 entry
->val
= ent
.val
;
5549 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5550 pte_t ptent
, swp_entry_t
*entry
)
5556 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
5557 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5559 if (!vma
->vm_file
) /* anonymous vma */
5561 if (!(mc
.flags
& MOVE_FILE
))
5564 /* page is moved even if it's not RSS of this task(page-faulted). */
5565 /* shmem/tmpfs may report page out on swap: account for that too. */
5566 return find_get_incore_page(vma
->vm_file
->f_mapping
,
5567 linear_page_index(vma
, addr
));
5571 * mem_cgroup_move_account - move account of the page
5573 * @compound: charge the page as compound or small page
5574 * @from: mem_cgroup which the page is moved from.
5575 * @to: mem_cgroup which the page is moved to. @from != @to.
5577 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5579 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5582 static int mem_cgroup_move_account(struct page
*page
,
5584 struct mem_cgroup
*from
,
5585 struct mem_cgroup
*to
)
5587 struct lruvec
*from_vec
, *to_vec
;
5588 struct pglist_data
*pgdat
;
5589 unsigned int nr_pages
= compound
? thp_nr_pages(page
) : 1;
5592 VM_BUG_ON(from
== to
);
5593 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5594 VM_BUG_ON(compound
&& !PageTransHuge(page
));
5597 * Prevent mem_cgroup_migrate() from looking at
5598 * page's memory cgroup of its source page while we change it.
5601 if (!trylock_page(page
))
5605 if (page_memcg(page
) != from
)
5608 pgdat
= page_pgdat(page
);
5609 from_vec
= mem_cgroup_lruvec(from
, pgdat
);
5610 to_vec
= mem_cgroup_lruvec(to
, pgdat
);
5612 lock_page_memcg(page
);
5614 if (PageAnon(page
)) {
5615 if (page_mapped(page
)) {
5616 __mod_lruvec_state(from_vec
, NR_ANON_MAPPED
, -nr_pages
);
5617 __mod_lruvec_state(to_vec
, NR_ANON_MAPPED
, nr_pages
);
5618 if (PageTransHuge(page
)) {
5619 __mod_lruvec_state(from_vec
, NR_ANON_THPS
,
5621 __mod_lruvec_state(to_vec
, NR_ANON_THPS
,
5626 __mod_lruvec_state(from_vec
, NR_FILE_PAGES
, -nr_pages
);
5627 __mod_lruvec_state(to_vec
, NR_FILE_PAGES
, nr_pages
);
5629 if (PageSwapBacked(page
)) {
5630 __mod_lruvec_state(from_vec
, NR_SHMEM
, -nr_pages
);
5631 __mod_lruvec_state(to_vec
, NR_SHMEM
, nr_pages
);
5634 if (page_mapped(page
)) {
5635 __mod_lruvec_state(from_vec
, NR_FILE_MAPPED
, -nr_pages
);
5636 __mod_lruvec_state(to_vec
, NR_FILE_MAPPED
, nr_pages
);
5639 if (PageDirty(page
)) {
5640 struct address_space
*mapping
= page_mapping(page
);
5642 if (mapping_can_writeback(mapping
)) {
5643 __mod_lruvec_state(from_vec
, NR_FILE_DIRTY
,
5645 __mod_lruvec_state(to_vec
, NR_FILE_DIRTY
,
5651 if (PageWriteback(page
)) {
5652 __mod_lruvec_state(from_vec
, NR_WRITEBACK
, -nr_pages
);
5653 __mod_lruvec_state(to_vec
, NR_WRITEBACK
, nr_pages
);
5657 * All state has been migrated, let's switch to the new memcg.
5659 * It is safe to change page's memcg here because the page
5660 * is referenced, charged, isolated, and locked: we can't race
5661 * with (un)charging, migration, LRU putback, or anything else
5662 * that would rely on a stable page's memory cgroup.
5664 * Note that lock_page_memcg is a memcg lock, not a page lock,
5665 * to save space. As soon as we switch page's memory cgroup to a
5666 * new memcg that isn't locked, the above state can change
5667 * concurrently again. Make sure we're truly done with it.
5672 css_put(&from
->css
);
5674 page
->memcg_data
= (unsigned long)to
;
5676 __unlock_page_memcg(from
);
5680 local_irq_disable();
5681 mem_cgroup_charge_statistics(to
, page
, nr_pages
);
5682 memcg_check_events(to
, page
);
5683 mem_cgroup_charge_statistics(from
, page
, -nr_pages
);
5684 memcg_check_events(from
, page
);
5693 * get_mctgt_type - get target type of moving charge
5694 * @vma: the vma the pte to be checked belongs
5695 * @addr: the address corresponding to the pte to be checked
5696 * @ptent: the pte to be checked
5697 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5700 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5701 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5702 * move charge. if @target is not NULL, the page is stored in target->page
5703 * with extra refcnt got(Callers should handle it).
5704 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5705 * target for charge migration. if @target is not NULL, the entry is stored
5707 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5708 * (so ZONE_DEVICE page and thus not on the lru).
5709 * For now we such page is charge like a regular page would be as for all
5710 * intent and purposes it is just special memory taking the place of a
5713 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5715 * Called with pte lock held.
5718 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
5719 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
5721 struct page
*page
= NULL
;
5722 enum mc_target_type ret
= MC_TARGET_NONE
;
5723 swp_entry_t ent
= { .val
= 0 };
5725 if (pte_present(ptent
))
5726 page
= mc_handle_present_pte(vma
, addr
, ptent
);
5727 else if (is_swap_pte(ptent
))
5728 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
5729 else if (pte_none(ptent
))
5730 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
5732 if (!page
&& !ent
.val
)
5736 * Do only loose check w/o serialization.
5737 * mem_cgroup_move_account() checks the page is valid or
5738 * not under LRU exclusion.
5740 if (page_memcg(page
) == mc
.from
) {
5741 ret
= MC_TARGET_PAGE
;
5742 if (is_device_private_page(page
))
5743 ret
= MC_TARGET_DEVICE
;
5745 target
->page
= page
;
5747 if (!ret
|| !target
)
5751 * There is a swap entry and a page doesn't exist or isn't charged.
5752 * But we cannot move a tail-page in a THP.
5754 if (ent
.val
&& !ret
&& (!page
|| !PageTransCompound(page
)) &&
5755 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
5756 ret
= MC_TARGET_SWAP
;
5763 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5765 * We don't consider PMD mapped swapping or file mapped pages because THP does
5766 * not support them for now.
5767 * Caller should make sure that pmd_trans_huge(pmd) is true.
5769 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5770 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5772 struct page
*page
= NULL
;
5773 enum mc_target_type ret
= MC_TARGET_NONE
;
5775 if (unlikely(is_swap_pmd(pmd
))) {
5776 VM_BUG_ON(thp_migration_supported() &&
5777 !is_pmd_migration_entry(pmd
));
5780 page
= pmd_page(pmd
);
5781 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
5782 if (!(mc
.flags
& MOVE_ANON
))
5784 if (page_memcg(page
) == mc
.from
) {
5785 ret
= MC_TARGET_PAGE
;
5788 target
->page
= page
;
5794 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5795 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5797 return MC_TARGET_NONE
;
5801 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5802 unsigned long addr
, unsigned long end
,
5803 struct mm_walk
*walk
)
5805 struct vm_area_struct
*vma
= walk
->vma
;
5809 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5812 * Note their can not be MC_TARGET_DEVICE for now as we do not
5813 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5814 * this might change.
5816 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5817 mc
.precharge
+= HPAGE_PMD_NR
;
5822 if (pmd_trans_unstable(pmd
))
5824 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5825 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5826 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5827 mc
.precharge
++; /* increment precharge temporarily */
5828 pte_unmap_unlock(pte
- 1, ptl
);
5834 static const struct mm_walk_ops precharge_walk_ops
= {
5835 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
5838 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5840 unsigned long precharge
;
5843 walk_page_range(mm
, 0, mm
->highest_vm_end
, &precharge_walk_ops
, NULL
);
5844 mmap_read_unlock(mm
);
5846 precharge
= mc
.precharge
;
5852 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
5854 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
5856 VM_BUG_ON(mc
.moving_task
);
5857 mc
.moving_task
= current
;
5858 return mem_cgroup_do_precharge(precharge
);
5861 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5862 static void __mem_cgroup_clear_mc(void)
5864 struct mem_cgroup
*from
= mc
.from
;
5865 struct mem_cgroup
*to
= mc
.to
;
5867 /* we must uncharge all the leftover precharges from mc.to */
5869 cancel_charge(mc
.to
, mc
.precharge
);
5873 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5874 * we must uncharge here.
5876 if (mc
.moved_charge
) {
5877 cancel_charge(mc
.from
, mc
.moved_charge
);
5878 mc
.moved_charge
= 0;
5880 /* we must fixup refcnts and charges */
5881 if (mc
.moved_swap
) {
5882 /* uncharge swap account from the old cgroup */
5883 if (!mem_cgroup_is_root(mc
.from
))
5884 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
5886 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
5889 * we charged both to->memory and to->memsw, so we
5890 * should uncharge to->memory.
5892 if (!mem_cgroup_is_root(mc
.to
))
5893 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
5897 memcg_oom_recover(from
);
5898 memcg_oom_recover(to
);
5899 wake_up_all(&mc
.waitq
);
5902 static void mem_cgroup_clear_mc(void)
5904 struct mm_struct
*mm
= mc
.mm
;
5907 * we must clear moving_task before waking up waiters at the end of
5910 mc
.moving_task
= NULL
;
5911 __mem_cgroup_clear_mc();
5912 spin_lock(&mc
.lock
);
5916 spin_unlock(&mc
.lock
);
5921 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5923 struct cgroup_subsys_state
*css
;
5924 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
5925 struct mem_cgroup
*from
;
5926 struct task_struct
*leader
, *p
;
5927 struct mm_struct
*mm
;
5928 unsigned long move_flags
;
5931 /* charge immigration isn't supported on the default hierarchy */
5932 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5936 * Multi-process migrations only happen on the default hierarchy
5937 * where charge immigration is not used. Perform charge
5938 * immigration if @tset contains a leader and whine if there are
5942 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
5945 memcg
= mem_cgroup_from_css(css
);
5951 * We are now commited to this value whatever it is. Changes in this
5952 * tunable will only affect upcoming migrations, not the current one.
5953 * So we need to save it, and keep it going.
5955 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
5959 from
= mem_cgroup_from_task(p
);
5961 VM_BUG_ON(from
== memcg
);
5963 mm
= get_task_mm(p
);
5966 /* We move charges only when we move a owner of the mm */
5967 if (mm
->owner
== p
) {
5970 VM_BUG_ON(mc
.precharge
);
5971 VM_BUG_ON(mc
.moved_charge
);
5972 VM_BUG_ON(mc
.moved_swap
);
5974 spin_lock(&mc
.lock
);
5978 mc
.flags
= move_flags
;
5979 spin_unlock(&mc
.lock
);
5980 /* We set mc.moving_task later */
5982 ret
= mem_cgroup_precharge_mc(mm
);
5984 mem_cgroup_clear_mc();
5991 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5994 mem_cgroup_clear_mc();
5997 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
5998 unsigned long addr
, unsigned long end
,
5999 struct mm_walk
*walk
)
6002 struct vm_area_struct
*vma
= walk
->vma
;
6005 enum mc_target_type target_type
;
6006 union mc_target target
;
6009 ptl
= pmd_trans_huge_lock(pmd
, vma
);
6011 if (mc
.precharge
< HPAGE_PMD_NR
) {
6015 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6016 if (target_type
== MC_TARGET_PAGE
) {
6018 if (!isolate_lru_page(page
)) {
6019 if (!mem_cgroup_move_account(page
, true,
6021 mc
.precharge
-= HPAGE_PMD_NR
;
6022 mc
.moved_charge
+= HPAGE_PMD_NR
;
6024 putback_lru_page(page
);
6027 } else if (target_type
== MC_TARGET_DEVICE
) {
6029 if (!mem_cgroup_move_account(page
, true,
6031 mc
.precharge
-= HPAGE_PMD_NR
;
6032 mc
.moved_charge
+= HPAGE_PMD_NR
;
6040 if (pmd_trans_unstable(pmd
))
6043 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6044 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6045 pte_t ptent
= *(pte
++);
6046 bool device
= false;
6052 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6053 case MC_TARGET_DEVICE
:
6056 case MC_TARGET_PAGE
:
6059 * We can have a part of the split pmd here. Moving it
6060 * can be done but it would be too convoluted so simply
6061 * ignore such a partial THP and keep it in original
6062 * memcg. There should be somebody mapping the head.
6064 if (PageTransCompound(page
))
6066 if (!device
&& isolate_lru_page(page
))
6068 if (!mem_cgroup_move_account(page
, false,
6071 /* we uncharge from mc.from later. */
6075 putback_lru_page(page
);
6076 put
: /* get_mctgt_type() gets the page */
6079 case MC_TARGET_SWAP
:
6081 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6083 mem_cgroup_id_get_many(mc
.to
, 1);
6084 /* we fixup other refcnts and charges later. */
6092 pte_unmap_unlock(pte
- 1, ptl
);
6097 * We have consumed all precharges we got in can_attach().
6098 * We try charge one by one, but don't do any additional
6099 * charges to mc.to if we have failed in charge once in attach()
6102 ret
= mem_cgroup_do_precharge(1);
6110 static const struct mm_walk_ops charge_walk_ops
= {
6111 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6114 static void mem_cgroup_move_charge(void)
6116 lru_add_drain_all();
6118 * Signal lock_page_memcg() to take the memcg's move_lock
6119 * while we're moving its pages to another memcg. Then wait
6120 * for already started RCU-only updates to finish.
6122 atomic_inc(&mc
.from
->moving_account
);
6125 if (unlikely(!mmap_read_trylock(mc
.mm
))) {
6127 * Someone who are holding the mmap_lock might be waiting in
6128 * waitq. So we cancel all extra charges, wake up all waiters,
6129 * and retry. Because we cancel precharges, we might not be able
6130 * to move enough charges, but moving charge is a best-effort
6131 * feature anyway, so it wouldn't be a big problem.
6133 __mem_cgroup_clear_mc();
6138 * When we have consumed all precharges and failed in doing
6139 * additional charge, the page walk just aborts.
6141 walk_page_range(mc
.mm
, 0, mc
.mm
->highest_vm_end
, &charge_walk_ops
,
6144 mmap_read_unlock(mc
.mm
);
6145 atomic_dec(&mc
.from
->moving_account
);
6148 static void mem_cgroup_move_task(void)
6151 mem_cgroup_move_charge();
6152 mem_cgroup_clear_mc();
6155 #else /* !CONFIG_MMU */
6156 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
6160 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
6163 static void mem_cgroup_move_task(void)
6168 static int seq_puts_memcg_tunable(struct seq_file
*m
, unsigned long value
)
6170 if (value
== PAGE_COUNTER_MAX
)
6171 seq_puts(m
, "max\n");
6173 seq_printf(m
, "%llu\n", (u64
)value
* PAGE_SIZE
);
6178 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
6181 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6183 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
6186 static int memory_min_show(struct seq_file
*m
, void *v
)
6188 return seq_puts_memcg_tunable(m
,
6189 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.min
));
6192 static ssize_t
memory_min_write(struct kernfs_open_file
*of
,
6193 char *buf
, size_t nbytes
, loff_t off
)
6195 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6199 buf
= strstrip(buf
);
6200 err
= page_counter_memparse(buf
, "max", &min
);
6204 page_counter_set_min(&memcg
->memory
, min
);
6209 static int memory_low_show(struct seq_file
*m
, void *v
)
6211 return seq_puts_memcg_tunable(m
,
6212 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.low
));
6215 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
6216 char *buf
, size_t nbytes
, loff_t off
)
6218 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6222 buf
= strstrip(buf
);
6223 err
= page_counter_memparse(buf
, "max", &low
);
6227 page_counter_set_low(&memcg
->memory
, low
);
6232 static int memory_high_show(struct seq_file
*m
, void *v
)
6234 return seq_puts_memcg_tunable(m
,
6235 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.high
));
6238 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
6239 char *buf
, size_t nbytes
, loff_t off
)
6241 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6242 unsigned int nr_retries
= MAX_RECLAIM_RETRIES
;
6243 bool drained
= false;
6247 buf
= strstrip(buf
);
6248 err
= page_counter_memparse(buf
, "max", &high
);
6252 page_counter_set_high(&memcg
->memory
, high
);
6255 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6256 unsigned long reclaimed
;
6258 if (nr_pages
<= high
)
6261 if (signal_pending(current
))
6265 drain_all_stock(memcg
);
6270 reclaimed
= try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
6273 if (!reclaimed
&& !nr_retries
--)
6277 memcg_wb_domain_size_changed(memcg
);
6281 static int memory_max_show(struct seq_file
*m
, void *v
)
6283 return seq_puts_memcg_tunable(m
,
6284 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.max
));
6287 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
6288 char *buf
, size_t nbytes
, loff_t off
)
6290 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6291 unsigned int nr_reclaims
= MAX_RECLAIM_RETRIES
;
6292 bool drained
= false;
6296 buf
= strstrip(buf
);
6297 err
= page_counter_memparse(buf
, "max", &max
);
6301 xchg(&memcg
->memory
.max
, max
);
6304 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6306 if (nr_pages
<= max
)
6309 if (signal_pending(current
))
6313 drain_all_stock(memcg
);
6319 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
6325 memcg_memory_event(memcg
, MEMCG_OOM
);
6326 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
6330 memcg_wb_domain_size_changed(memcg
);
6334 static void __memory_events_show(struct seq_file
*m
, atomic_long_t
*events
)
6336 seq_printf(m
, "low %lu\n", atomic_long_read(&events
[MEMCG_LOW
]));
6337 seq_printf(m
, "high %lu\n", atomic_long_read(&events
[MEMCG_HIGH
]));
6338 seq_printf(m
, "max %lu\n", atomic_long_read(&events
[MEMCG_MAX
]));
6339 seq_printf(m
, "oom %lu\n", atomic_long_read(&events
[MEMCG_OOM
]));
6340 seq_printf(m
, "oom_kill %lu\n",
6341 atomic_long_read(&events
[MEMCG_OOM_KILL
]));
6344 static int memory_events_show(struct seq_file
*m
, void *v
)
6346 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6348 __memory_events_show(m
, memcg
->memory_events
);
6352 static int memory_events_local_show(struct seq_file
*m
, void *v
)
6354 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6356 __memory_events_show(m
, memcg
->memory_events_local
);
6360 static int memory_stat_show(struct seq_file
*m
, void *v
)
6362 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6365 buf
= memory_stat_format(memcg
);
6374 static inline unsigned long lruvec_page_state_output(struct lruvec
*lruvec
,
6377 return lruvec_page_state(lruvec
, item
) * memcg_page_state_unit(item
);
6380 static int memory_numa_stat_show(struct seq_file
*m
, void *v
)
6383 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6385 for (i
= 0; i
< ARRAY_SIZE(memory_stats
); i
++) {
6388 if (memory_stats
[i
].idx
>= NR_VM_NODE_STAT_ITEMS
)
6391 seq_printf(m
, "%s", memory_stats
[i
].name
);
6392 for_each_node_state(nid
, N_MEMORY
) {
6394 struct lruvec
*lruvec
;
6396 lruvec
= mem_cgroup_lruvec(memcg
, NODE_DATA(nid
));
6397 size
= lruvec_page_state_output(lruvec
,
6398 memory_stats
[i
].idx
);
6399 seq_printf(m
, " N%d=%llu", nid
, size
);
6408 static int memory_oom_group_show(struct seq_file
*m
, void *v
)
6410 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6412 seq_printf(m
, "%d\n", memcg
->oom_group
);
6417 static ssize_t
memory_oom_group_write(struct kernfs_open_file
*of
,
6418 char *buf
, size_t nbytes
, loff_t off
)
6420 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6423 buf
= strstrip(buf
);
6427 ret
= kstrtoint(buf
, 0, &oom_group
);
6431 if (oom_group
!= 0 && oom_group
!= 1)
6434 memcg
->oom_group
= oom_group
;
6439 static struct cftype memory_files
[] = {
6442 .flags
= CFTYPE_NOT_ON_ROOT
,
6443 .read_u64
= memory_current_read
,
6447 .flags
= CFTYPE_NOT_ON_ROOT
,
6448 .seq_show
= memory_min_show
,
6449 .write
= memory_min_write
,
6453 .flags
= CFTYPE_NOT_ON_ROOT
,
6454 .seq_show
= memory_low_show
,
6455 .write
= memory_low_write
,
6459 .flags
= CFTYPE_NOT_ON_ROOT
,
6460 .seq_show
= memory_high_show
,
6461 .write
= memory_high_write
,
6465 .flags
= CFTYPE_NOT_ON_ROOT
,
6466 .seq_show
= memory_max_show
,
6467 .write
= memory_max_write
,
6471 .flags
= CFTYPE_NOT_ON_ROOT
,
6472 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
6473 .seq_show
= memory_events_show
,
6476 .name
= "events.local",
6477 .flags
= CFTYPE_NOT_ON_ROOT
,
6478 .file_offset
= offsetof(struct mem_cgroup
, events_local_file
),
6479 .seq_show
= memory_events_local_show
,
6483 .seq_show
= memory_stat_show
,
6487 .name
= "numa_stat",
6488 .seq_show
= memory_numa_stat_show
,
6492 .name
= "oom.group",
6493 .flags
= CFTYPE_NOT_ON_ROOT
| CFTYPE_NS_DELEGATABLE
,
6494 .seq_show
= memory_oom_group_show
,
6495 .write
= memory_oom_group_write
,
6500 struct cgroup_subsys memory_cgrp_subsys
= {
6501 .css_alloc
= mem_cgroup_css_alloc
,
6502 .css_online
= mem_cgroup_css_online
,
6503 .css_offline
= mem_cgroup_css_offline
,
6504 .css_released
= mem_cgroup_css_released
,
6505 .css_free
= mem_cgroup_css_free
,
6506 .css_reset
= mem_cgroup_css_reset
,
6507 .can_attach
= mem_cgroup_can_attach
,
6508 .cancel_attach
= mem_cgroup_cancel_attach
,
6509 .post_attach
= mem_cgroup_move_task
,
6510 .dfl_cftypes
= memory_files
,
6511 .legacy_cftypes
= mem_cgroup_legacy_files
,
6516 * This function calculates an individual cgroup's effective
6517 * protection which is derived from its own memory.min/low, its
6518 * parent's and siblings' settings, as well as the actual memory
6519 * distribution in the tree.
6521 * The following rules apply to the effective protection values:
6523 * 1. At the first level of reclaim, effective protection is equal to
6524 * the declared protection in memory.min and memory.low.
6526 * 2. To enable safe delegation of the protection configuration, at
6527 * subsequent levels the effective protection is capped to the
6528 * parent's effective protection.
6530 * 3. To make complex and dynamic subtrees easier to configure, the
6531 * user is allowed to overcommit the declared protection at a given
6532 * level. If that is the case, the parent's effective protection is
6533 * distributed to the children in proportion to how much protection
6534 * they have declared and how much of it they are utilizing.
6536 * This makes distribution proportional, but also work-conserving:
6537 * if one cgroup claims much more protection than it uses memory,
6538 * the unused remainder is available to its siblings.
6540 * 4. Conversely, when the declared protection is undercommitted at a
6541 * given level, the distribution of the larger parental protection
6542 * budget is NOT proportional. A cgroup's protection from a sibling
6543 * is capped to its own memory.min/low setting.
6545 * 5. However, to allow protecting recursive subtrees from each other
6546 * without having to declare each individual cgroup's fixed share
6547 * of the ancestor's claim to protection, any unutilized -
6548 * "floating" - protection from up the tree is distributed in
6549 * proportion to each cgroup's *usage*. This makes the protection
6550 * neutral wrt sibling cgroups and lets them compete freely over
6551 * the shared parental protection budget, but it protects the
6552 * subtree as a whole from neighboring subtrees.
6554 * Note that 4. and 5. are not in conflict: 4. is about protecting
6555 * against immediate siblings whereas 5. is about protecting against
6556 * neighboring subtrees.
6558 static unsigned long effective_protection(unsigned long usage
,
6559 unsigned long parent_usage
,
6560 unsigned long setting
,
6561 unsigned long parent_effective
,
6562 unsigned long siblings_protected
)
6564 unsigned long protected;
6567 protected = min(usage
, setting
);
6569 * If all cgroups at this level combined claim and use more
6570 * protection then what the parent affords them, distribute
6571 * shares in proportion to utilization.
6573 * We are using actual utilization rather than the statically
6574 * claimed protection in order to be work-conserving: claimed
6575 * but unused protection is available to siblings that would
6576 * otherwise get a smaller chunk than what they claimed.
6578 if (siblings_protected
> parent_effective
)
6579 return protected * parent_effective
/ siblings_protected
;
6582 * Ok, utilized protection of all children is within what the
6583 * parent affords them, so we know whatever this child claims
6584 * and utilizes is effectively protected.
6586 * If there is unprotected usage beyond this value, reclaim
6587 * will apply pressure in proportion to that amount.
6589 * If there is unutilized protection, the cgroup will be fully
6590 * shielded from reclaim, but we do return a smaller value for
6591 * protection than what the group could enjoy in theory. This
6592 * is okay. With the overcommit distribution above, effective
6593 * protection is always dependent on how memory is actually
6594 * consumed among the siblings anyway.
6599 * If the children aren't claiming (all of) the protection
6600 * afforded to them by the parent, distribute the remainder in
6601 * proportion to the (unprotected) memory of each cgroup. That
6602 * way, cgroups that aren't explicitly prioritized wrt each
6603 * other compete freely over the allowance, but they are
6604 * collectively protected from neighboring trees.
6606 * We're using unprotected memory for the weight so that if
6607 * some cgroups DO claim explicit protection, we don't protect
6608 * the same bytes twice.
6610 * Check both usage and parent_usage against the respective
6611 * protected values. One should imply the other, but they
6612 * aren't read atomically - make sure the division is sane.
6614 if (!(cgrp_dfl_root
.flags
& CGRP_ROOT_MEMORY_RECURSIVE_PROT
))
6616 if (parent_effective
> siblings_protected
&&
6617 parent_usage
> siblings_protected
&&
6618 usage
> protected) {
6619 unsigned long unclaimed
;
6621 unclaimed
= parent_effective
- siblings_protected
;
6622 unclaimed
*= usage
- protected;
6623 unclaimed
/= parent_usage
- siblings_protected
;
6632 * mem_cgroup_protected - check if memory consumption is in the normal range
6633 * @root: the top ancestor of the sub-tree being checked
6634 * @memcg: the memory cgroup to check
6636 * WARNING: This function is not stateless! It can only be used as part
6637 * of a top-down tree iteration, not for isolated queries.
6639 void mem_cgroup_calculate_protection(struct mem_cgroup
*root
,
6640 struct mem_cgroup
*memcg
)
6642 unsigned long usage
, parent_usage
;
6643 struct mem_cgroup
*parent
;
6645 if (mem_cgroup_disabled())
6649 root
= root_mem_cgroup
;
6652 * Effective values of the reclaim targets are ignored so they
6653 * can be stale. Have a look at mem_cgroup_protection for more
6655 * TODO: calculation should be more robust so that we do not need
6656 * that special casing.
6661 usage
= page_counter_read(&memcg
->memory
);
6665 parent
= parent_mem_cgroup(memcg
);
6666 /* No parent means a non-hierarchical mode on v1 memcg */
6670 if (parent
== root
) {
6671 memcg
->memory
.emin
= READ_ONCE(memcg
->memory
.min
);
6672 memcg
->memory
.elow
= READ_ONCE(memcg
->memory
.low
);
6676 parent_usage
= page_counter_read(&parent
->memory
);
6678 WRITE_ONCE(memcg
->memory
.emin
, effective_protection(usage
, parent_usage
,
6679 READ_ONCE(memcg
->memory
.min
),
6680 READ_ONCE(parent
->memory
.emin
),
6681 atomic_long_read(&parent
->memory
.children_min_usage
)));
6683 WRITE_ONCE(memcg
->memory
.elow
, effective_protection(usage
, parent_usage
,
6684 READ_ONCE(memcg
->memory
.low
),
6685 READ_ONCE(parent
->memory
.elow
),
6686 atomic_long_read(&parent
->memory
.children_low_usage
)));
6690 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6691 * @page: page to charge
6692 * @mm: mm context of the victim
6693 * @gfp_mask: reclaim mode
6695 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6696 * pages according to @gfp_mask if necessary.
6698 * Returns 0 on success. Otherwise, an error code is returned.
6700 int mem_cgroup_charge(struct page
*page
, struct mm_struct
*mm
, gfp_t gfp_mask
)
6702 unsigned int nr_pages
= thp_nr_pages(page
);
6703 struct mem_cgroup
*memcg
= NULL
;
6706 if (mem_cgroup_disabled())
6709 if (PageSwapCache(page
)) {
6710 swp_entry_t ent
= { .val
= page_private(page
), };
6714 * Every swap fault against a single page tries to charge the
6715 * page, bail as early as possible. shmem_unuse() encounters
6716 * already charged pages, too. page and memcg binding is
6717 * protected by the page lock, which serializes swap cache
6718 * removal, which in turn serializes uncharging.
6720 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
6721 if (page_memcg(compound_head(page
)))
6724 id
= lookup_swap_cgroup_id(ent
);
6726 memcg
= mem_cgroup_from_id(id
);
6727 if (memcg
&& !css_tryget_online(&memcg
->css
))
6733 memcg
= get_mem_cgroup_from_mm(mm
);
6735 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
6739 css_get(&memcg
->css
);
6740 commit_charge(page
, memcg
);
6742 local_irq_disable();
6743 mem_cgroup_charge_statistics(memcg
, page
, nr_pages
);
6744 memcg_check_events(memcg
, page
);
6748 * Cgroup1's unified memory+swap counter has been charged with the
6749 * new swapcache page, finish the transfer by uncharging the swap
6750 * slot. The swap slot would also get uncharged when it dies, but
6751 * it can stick around indefinitely and we'd count the page twice
6754 * Cgroup2 has separate resource counters for memory and swap,
6755 * so this is a non-issue here. Memory and swap charge lifetimes
6756 * correspond 1:1 to page and swap slot lifetimes: we charge the
6757 * page to memory here, and uncharge swap when the slot is freed.
6759 if (do_memsw_account() && PageSwapCache(page
)) {
6760 swp_entry_t entry
= { .val
= page_private(page
) };
6762 * The swap entry might not get freed for a long time,
6763 * let's not wait for it. The page already received a
6764 * memory+swap charge, drop the swap entry duplicate.
6766 mem_cgroup_uncharge_swap(entry
, nr_pages
);
6770 css_put(&memcg
->css
);
6775 struct uncharge_gather
{
6776 struct mem_cgroup
*memcg
;
6777 unsigned long nr_pages
;
6778 unsigned long pgpgout
;
6779 unsigned long nr_kmem
;
6780 struct page
*dummy_page
;
6783 static inline void uncharge_gather_clear(struct uncharge_gather
*ug
)
6785 memset(ug
, 0, sizeof(*ug
));
6788 static void uncharge_batch(const struct uncharge_gather
*ug
)
6790 unsigned long flags
;
6792 if (!mem_cgroup_is_root(ug
->memcg
)) {
6793 page_counter_uncharge(&ug
->memcg
->memory
, ug
->nr_pages
);
6794 if (do_memsw_account())
6795 page_counter_uncharge(&ug
->memcg
->memsw
, ug
->nr_pages
);
6796 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && ug
->nr_kmem
)
6797 page_counter_uncharge(&ug
->memcg
->kmem
, ug
->nr_kmem
);
6798 memcg_oom_recover(ug
->memcg
);
6801 local_irq_save(flags
);
6802 __count_memcg_events(ug
->memcg
, PGPGOUT
, ug
->pgpgout
);
6803 __this_cpu_add(ug
->memcg
->vmstats_percpu
->nr_page_events
, ug
->nr_pages
);
6804 memcg_check_events(ug
->memcg
, ug
->dummy_page
);
6805 local_irq_restore(flags
);
6807 /* drop reference from uncharge_page */
6808 css_put(&ug
->memcg
->css
);
6811 static void uncharge_page(struct page
*page
, struct uncharge_gather
*ug
)
6813 unsigned long nr_pages
;
6815 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6817 if (!page_memcg(page
))
6821 * Nobody should be changing or seriously looking at
6822 * page_memcg(page) at this point, we have fully
6823 * exclusive access to the page.
6826 if (ug
->memcg
!= page_memcg(page
)) {
6829 uncharge_gather_clear(ug
);
6831 ug
->memcg
= page_memcg(page
);
6833 /* pairs with css_put in uncharge_batch */
6834 css_get(&ug
->memcg
->css
);
6837 nr_pages
= compound_nr(page
);
6838 ug
->nr_pages
+= nr_pages
;
6840 if (PageMemcgKmem(page
))
6841 ug
->nr_kmem
+= nr_pages
;
6845 ug
->dummy_page
= page
;
6846 page
->memcg_data
= 0;
6847 css_put(&ug
->memcg
->css
);
6851 * mem_cgroup_uncharge - uncharge a page
6852 * @page: page to uncharge
6854 * Uncharge a page previously charged with mem_cgroup_charge().
6856 void mem_cgroup_uncharge(struct page
*page
)
6858 struct uncharge_gather ug
;
6860 if (mem_cgroup_disabled())
6863 /* Don't touch page->lru of any random page, pre-check: */
6864 if (!page_memcg(page
))
6867 uncharge_gather_clear(&ug
);
6868 uncharge_page(page
, &ug
);
6869 uncharge_batch(&ug
);
6873 * mem_cgroup_uncharge_list - uncharge a list of page
6874 * @page_list: list of pages to uncharge
6876 * Uncharge a list of pages previously charged with
6877 * mem_cgroup_charge().
6879 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
6881 struct uncharge_gather ug
;
6884 if (mem_cgroup_disabled())
6887 uncharge_gather_clear(&ug
);
6888 list_for_each_entry(page
, page_list
, lru
)
6889 uncharge_page(page
, &ug
);
6891 uncharge_batch(&ug
);
6895 * mem_cgroup_migrate - charge a page's replacement
6896 * @oldpage: currently circulating page
6897 * @newpage: replacement page
6899 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6900 * be uncharged upon free.
6902 * Both pages must be locked, @newpage->mapping must be set up.
6904 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
6906 struct mem_cgroup
*memcg
;
6907 unsigned int nr_pages
;
6908 unsigned long flags
;
6910 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
6911 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
6912 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
6913 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
6916 if (mem_cgroup_disabled())
6919 /* Page cache replacement: new page already charged? */
6920 if (page_memcg(newpage
))
6923 memcg
= page_memcg(oldpage
);
6924 VM_WARN_ON_ONCE_PAGE(!memcg
, oldpage
);
6928 /* Force-charge the new page. The old one will be freed soon */
6929 nr_pages
= thp_nr_pages(newpage
);
6931 page_counter_charge(&memcg
->memory
, nr_pages
);
6932 if (do_memsw_account())
6933 page_counter_charge(&memcg
->memsw
, nr_pages
);
6935 css_get(&memcg
->css
);
6936 commit_charge(newpage
, memcg
);
6938 local_irq_save(flags
);
6939 mem_cgroup_charge_statistics(memcg
, newpage
, nr_pages
);
6940 memcg_check_events(memcg
, newpage
);
6941 local_irq_restore(flags
);
6944 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
6945 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
6947 void mem_cgroup_sk_alloc(struct sock
*sk
)
6949 struct mem_cgroup
*memcg
;
6951 if (!mem_cgroup_sockets_enabled
)
6954 /* Do not associate the sock with unrelated interrupted task's memcg. */
6959 memcg
= mem_cgroup_from_task(current
);
6960 if (memcg
== root_mem_cgroup
)
6962 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
6964 if (css_tryget(&memcg
->css
))
6965 sk
->sk_memcg
= memcg
;
6970 void mem_cgroup_sk_free(struct sock
*sk
)
6973 css_put(&sk
->sk_memcg
->css
);
6977 * mem_cgroup_charge_skmem - charge socket memory
6978 * @memcg: memcg to charge
6979 * @nr_pages: number of pages to charge
6981 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6982 * @memcg's configured limit, %false if the charge had to be forced.
6984 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
6986 gfp_t gfp_mask
= GFP_KERNEL
;
6988 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
6989 struct page_counter
*fail
;
6991 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
6992 memcg
->tcpmem_pressure
= 0;
6995 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
6996 memcg
->tcpmem_pressure
= 1;
7000 /* Don't block in the packet receive path */
7002 gfp_mask
= GFP_NOWAIT
;
7004 mod_memcg_state(memcg
, MEMCG_SOCK
, nr_pages
);
7006 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
7009 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
7014 * mem_cgroup_uncharge_skmem - uncharge socket memory
7015 * @memcg: memcg to uncharge
7016 * @nr_pages: number of pages to uncharge
7018 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
7020 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
7021 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
7025 mod_memcg_state(memcg
, MEMCG_SOCK
, -nr_pages
);
7027 refill_stock(memcg
, nr_pages
);
7030 static int __init
cgroup_memory(char *s
)
7034 while ((token
= strsep(&s
, ",")) != NULL
) {
7037 if (!strcmp(token
, "nosocket"))
7038 cgroup_memory_nosocket
= true;
7039 if (!strcmp(token
, "nokmem"))
7040 cgroup_memory_nokmem
= true;
7044 __setup("cgroup.memory=", cgroup_memory
);
7047 * subsys_initcall() for memory controller.
7049 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7050 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7051 * basically everything that doesn't depend on a specific mem_cgroup structure
7052 * should be initialized from here.
7054 static int __init
mem_cgroup_init(void)
7059 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7060 * used for per-memcg-per-cpu caching of per-node statistics. In order
7061 * to work fine, we should make sure that the overfill threshold can't
7062 * exceed S32_MAX / PAGE_SIZE.
7064 BUILD_BUG_ON(MEMCG_CHARGE_BATCH
> S32_MAX
/ PAGE_SIZE
);
7066 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD
, "mm/memctrl:dead", NULL
,
7067 memcg_hotplug_cpu_dead
);
7069 for_each_possible_cpu(cpu
)
7070 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
7073 for_each_node(node
) {
7074 struct mem_cgroup_tree_per_node
*rtpn
;
7076 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
7077 node_online(node
) ? node
: NUMA_NO_NODE
);
7079 rtpn
->rb_root
= RB_ROOT
;
7080 rtpn
->rb_rightmost
= NULL
;
7081 spin_lock_init(&rtpn
->lock
);
7082 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
7087 subsys_initcall(mem_cgroup_init
);
7089 #ifdef CONFIG_MEMCG_SWAP
7090 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
7092 while (!refcount_inc_not_zero(&memcg
->id
.ref
)) {
7094 * The root cgroup cannot be destroyed, so it's refcount must
7097 if (WARN_ON_ONCE(memcg
== root_mem_cgroup
)) {
7101 memcg
= parent_mem_cgroup(memcg
);
7103 memcg
= root_mem_cgroup
;
7109 * mem_cgroup_swapout - transfer a memsw charge to swap
7110 * @page: page whose memsw charge to transfer
7111 * @entry: swap entry to move the charge to
7113 * Transfer the memsw charge of @page to @entry.
7115 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
7117 struct mem_cgroup
*memcg
, *swap_memcg
;
7118 unsigned int nr_entries
;
7119 unsigned short oldid
;
7121 VM_BUG_ON_PAGE(PageLRU(page
), page
);
7122 VM_BUG_ON_PAGE(page_count(page
), page
);
7124 if (mem_cgroup_disabled())
7127 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7130 memcg
= page_memcg(page
);
7132 VM_WARN_ON_ONCE_PAGE(!memcg
, page
);
7137 * In case the memcg owning these pages has been offlined and doesn't
7138 * have an ID allocated to it anymore, charge the closest online
7139 * ancestor for the swap instead and transfer the memory+swap charge.
7141 swap_memcg
= mem_cgroup_id_get_online(memcg
);
7142 nr_entries
= thp_nr_pages(page
);
7143 /* Get references for the tail pages, too */
7145 mem_cgroup_id_get_many(swap_memcg
, nr_entries
- 1);
7146 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
),
7148 VM_BUG_ON_PAGE(oldid
, page
);
7149 mod_memcg_state(swap_memcg
, MEMCG_SWAP
, nr_entries
);
7151 page
->memcg_data
= 0;
7153 if (!mem_cgroup_is_root(memcg
))
7154 page_counter_uncharge(&memcg
->memory
, nr_entries
);
7156 if (!cgroup_memory_noswap
&& memcg
!= swap_memcg
) {
7157 if (!mem_cgroup_is_root(swap_memcg
))
7158 page_counter_charge(&swap_memcg
->memsw
, nr_entries
);
7159 page_counter_uncharge(&memcg
->memsw
, nr_entries
);
7163 * Interrupts should be disabled here because the caller holds the
7164 * i_pages lock which is taken with interrupts-off. It is
7165 * important here to have the interrupts disabled because it is the
7166 * only synchronisation we have for updating the per-CPU variables.
7168 VM_BUG_ON(!irqs_disabled());
7169 mem_cgroup_charge_statistics(memcg
, page
, -nr_entries
);
7170 memcg_check_events(memcg
, page
);
7172 css_put(&memcg
->css
);
7176 * mem_cgroup_try_charge_swap - try charging swap space for a page
7177 * @page: page being added to swap
7178 * @entry: swap entry to charge
7180 * Try to charge @page's memcg for the swap space at @entry.
7182 * Returns 0 on success, -ENOMEM on failure.
7184 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
7186 unsigned int nr_pages
= thp_nr_pages(page
);
7187 struct page_counter
*counter
;
7188 struct mem_cgroup
*memcg
;
7189 unsigned short oldid
;
7191 if (mem_cgroup_disabled())
7194 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7197 memcg
= page_memcg(page
);
7199 VM_WARN_ON_ONCE_PAGE(!memcg
, page
);
7204 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7208 memcg
= mem_cgroup_id_get_online(memcg
);
7210 if (!cgroup_memory_noswap
&& !mem_cgroup_is_root(memcg
) &&
7211 !page_counter_try_charge(&memcg
->swap
, nr_pages
, &counter
)) {
7212 memcg_memory_event(memcg
, MEMCG_SWAP_MAX
);
7213 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7214 mem_cgroup_id_put(memcg
);
7218 /* Get references for the tail pages, too */
7220 mem_cgroup_id_get_many(memcg
, nr_pages
- 1);
7221 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
), nr_pages
);
7222 VM_BUG_ON_PAGE(oldid
, page
);
7223 mod_memcg_state(memcg
, MEMCG_SWAP
, nr_pages
);
7229 * mem_cgroup_uncharge_swap - uncharge swap space
7230 * @entry: swap entry to uncharge
7231 * @nr_pages: the amount of swap space to uncharge
7233 void mem_cgroup_uncharge_swap(swp_entry_t entry
, unsigned int nr_pages
)
7235 struct mem_cgroup
*memcg
;
7238 id
= swap_cgroup_record(entry
, 0, nr_pages
);
7240 memcg
= mem_cgroup_from_id(id
);
7242 if (!cgroup_memory_noswap
&& !mem_cgroup_is_root(memcg
)) {
7243 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7244 page_counter_uncharge(&memcg
->swap
, nr_pages
);
7246 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
7248 mod_memcg_state(memcg
, MEMCG_SWAP
, -nr_pages
);
7249 mem_cgroup_id_put_many(memcg
, nr_pages
);
7254 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
7256 long nr_swap_pages
= get_nr_swap_pages();
7258 if (cgroup_memory_noswap
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7259 return nr_swap_pages
;
7260 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
7261 nr_swap_pages
= min_t(long, nr_swap_pages
,
7262 READ_ONCE(memcg
->swap
.max
) -
7263 page_counter_read(&memcg
->swap
));
7264 return nr_swap_pages
;
7267 bool mem_cgroup_swap_full(struct page
*page
)
7269 struct mem_cgroup
*memcg
;
7271 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
7275 if (cgroup_memory_noswap
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7278 memcg
= page_memcg(page
);
7282 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
)) {
7283 unsigned long usage
= page_counter_read(&memcg
->swap
);
7285 if (usage
* 2 >= READ_ONCE(memcg
->swap
.high
) ||
7286 usage
* 2 >= READ_ONCE(memcg
->swap
.max
))
7293 static int __init
setup_swap_account(char *s
)
7295 if (!strcmp(s
, "1"))
7296 cgroup_memory_noswap
= false;
7297 else if (!strcmp(s
, "0"))
7298 cgroup_memory_noswap
= true;
7301 __setup("swapaccount=", setup_swap_account
);
7303 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
7306 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
7308 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
7311 static int swap_high_show(struct seq_file
*m
, void *v
)
7313 return seq_puts_memcg_tunable(m
,
7314 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.high
));
7317 static ssize_t
swap_high_write(struct kernfs_open_file
*of
,
7318 char *buf
, size_t nbytes
, loff_t off
)
7320 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7324 buf
= strstrip(buf
);
7325 err
= page_counter_memparse(buf
, "max", &high
);
7329 page_counter_set_high(&memcg
->swap
, high
);
7334 static int swap_max_show(struct seq_file
*m
, void *v
)
7336 return seq_puts_memcg_tunable(m
,
7337 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.max
));
7340 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
7341 char *buf
, size_t nbytes
, loff_t off
)
7343 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7347 buf
= strstrip(buf
);
7348 err
= page_counter_memparse(buf
, "max", &max
);
7352 xchg(&memcg
->swap
.max
, max
);
7357 static int swap_events_show(struct seq_file
*m
, void *v
)
7359 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
7361 seq_printf(m
, "high %lu\n",
7362 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_HIGH
]));
7363 seq_printf(m
, "max %lu\n",
7364 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_MAX
]));
7365 seq_printf(m
, "fail %lu\n",
7366 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_FAIL
]));
7371 static struct cftype swap_files
[] = {
7373 .name
= "swap.current",
7374 .flags
= CFTYPE_NOT_ON_ROOT
,
7375 .read_u64
= swap_current_read
,
7378 .name
= "swap.high",
7379 .flags
= CFTYPE_NOT_ON_ROOT
,
7380 .seq_show
= swap_high_show
,
7381 .write
= swap_high_write
,
7385 .flags
= CFTYPE_NOT_ON_ROOT
,
7386 .seq_show
= swap_max_show
,
7387 .write
= swap_max_write
,
7390 .name
= "swap.events",
7391 .flags
= CFTYPE_NOT_ON_ROOT
,
7392 .file_offset
= offsetof(struct mem_cgroup
, swap_events_file
),
7393 .seq_show
= swap_events_show
,
7398 static struct cftype memsw_files
[] = {
7400 .name
= "memsw.usage_in_bytes",
7401 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
7402 .read_u64
= mem_cgroup_read_u64
,
7405 .name
= "memsw.max_usage_in_bytes",
7406 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
7407 .write
= mem_cgroup_reset
,
7408 .read_u64
= mem_cgroup_read_u64
,
7411 .name
= "memsw.limit_in_bytes",
7412 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
7413 .write
= mem_cgroup_write
,
7414 .read_u64
= mem_cgroup_read_u64
,
7417 .name
= "memsw.failcnt",
7418 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
7419 .write
= mem_cgroup_reset
,
7420 .read_u64
= mem_cgroup_read_u64
,
7422 { }, /* terminate */
7426 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7427 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7428 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7429 * boot parameter. This may result in premature OOPS inside
7430 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7432 static int __init
mem_cgroup_swap_init(void)
7434 /* No memory control -> no swap control */
7435 if (mem_cgroup_disabled())
7436 cgroup_memory_noswap
= true;
7438 if (cgroup_memory_noswap
)
7441 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
, swap_files
));
7442 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
, memsw_files
));
7446 core_initcall(mem_cgroup_swap_init
);
7448 #endif /* CONFIG_MEMCG_SWAP */